From d0a7145ba99ed3a8bc3145aa785b5c86ffe65020 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 1 Jul 2024 02:09:34 +0300 Subject: [PATCH 01/38] flake.lock: Update (#8218) --- flake.lock | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/flake.lock b/flake.lock index 79bb3f63f..973ff4e56 100644 --- a/flake.lock +++ b/flake.lock @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1718895438, - "narHash": "sha256-k3JqJrkdoYwE3fHE6xGDY676AYmyh4U2Zw+0Bwe5DLU=", + "lastModified": 1719506693, + "narHash": "sha256-C8e9S7RzshSdHB7L+v9I51af1gDM5unhJ2xO1ywxNH8=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "d603719ec6e294f034936c0d0dc06f689d91b6c3", + "rev": "b2852eb9365c6de48ffb0dc2c9562591f652242a", "type": "github" }, "original": { From 197fe6c1d7bec6718ce901f0141b2725240f298c Mon Sep 17 00:00:00 2001 From: zhentaoyu Date: Mon, 1 Jul 2024 19:39:06 +0800 Subject: [PATCH 02/38] [SYCL] Update SYCL-Rope op and Refactor (#8157) * align with rope.cu and move sycl-op to a single file --- ggml/src/ggml-sycl.cpp | 305 +-------------------------------- ggml/src/ggml-sycl/backend.hpp | 1 + ggml/src/ggml-sycl/rope.cpp | 275 +++++++++++++++++++++++++++++ ggml/src/ggml-sycl/rope.hpp | 22 +++ 4 files changed, 300 insertions(+), 303 deletions(-) create mode 100644 ggml/src/ggml-sycl/rope.cpp create mode 100644 ggml/src/ggml-sycl/rope.hpp diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index 4a668a2c3..30d8a5b33 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -978,114 +978,6 @@ static void cpy_f32_q(const char * cx, char * cdst, const int ne, cpy_blck(cx + x_offset, cdst + dst_offset); } -static float rope_yarn_ramp(const float low, const float high, const int i0) { - const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low); - return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y)); -} - -struct rope_corr_dims { - float v[4]; -}; - -// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn -// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. -static void rope_yarn( - float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, - float * cos_theta, float * sin_theta -) { - // Get n-d rotational scaling corrected for extrapolation - float theta_interp = freq_scale * theta_extrap; - float theta = theta_interp; - if (ext_factor != 0.0f) { - float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; - theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; - - // Get n-d magnitude scaling corrected for interpolation - mscale *= 1.0f + 0.1f * sycl::log(1.0f / freq_scale); - } - *cos_theta = sycl::cos(theta) * mscale; - *sin_theta = sycl::sin(theta) * mscale; -} - -// rope == RoPE == rotary positional embedding -template -static void rope( - const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, - float ext_factor, float attn_factor, rope_corr_dims corr_dims -, - const sycl::nd_item<3> &item_ct1) { - const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1)); - - if (col >= ncols) { - return; - } - - const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - const int i = row*ncols + col; - const int i2 = row/p_delta_rows; - - const int p = has_pos ? pos[i2] : 0; - const float theta_base = p * dpct::pow(freq_base, -float(col) / ncols); - - float cos_theta, sin_theta; - rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta); - - const float x0 = x[i + 0]; - const float x1 = x[i + 1]; - - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + 1] = x0*sin_theta + x1*cos_theta; -} - -template -static void rope_neox( - const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, - float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, - const float * freq_factors, const sycl::nd_item<3> &item_ct1) { - const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1)); - - if (col >= ncols) { - return; - } - - const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - const int ib = col / n_dims; - const int ic = col % n_dims; - - if (ib > 0) { - const int i = row*ncols + ib*n_dims + ic; - - dst[i + 0] = x[i + 0]; - dst[i + 1] = x[i + 1]; - - return; - } - - const int i = row*ncols + ib*n_dims + ic/2; - const int i2 = row/p_delta_rows; - - float cur_rot = inv_ndims * ic - ib; - - const int p = has_pos ? pos[i2] : 0; - const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f; - - const float theta_base = - p * freq_scale * dpct::pow(theta_scale, col / 2.0f)/freq_factor; - - float cos_theta, sin_theta; - rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); - - const float x0 = x[i + 0]; - const float x1 = x[i + n_dims/2]; - - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; -} - static void k_sum_rows_f32(const float * x, float * dst, const int ncols, const sycl::nd_item<3> &item_ct1) { const int row = item_ct1.get_group(1); @@ -2241,110 +2133,6 @@ static void clamp_f32_sycl(const float *x, float *dst, const float min, }); } -template -static void rope_sycl(const T *x, T *dst, int ncols, int nrows, - const int32_t *pos, float freq_scale, int p_delta_rows, - float freq_base, float ext_factor, float attn_factor, - rope_corr_dims corr_dims, queue_ptr stream) { - GGML_ASSERT(ncols % 2 == 0); - const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); - const int num_blocks_x = (ncols + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); - const sycl::range<3> block_nums(1, num_blocks_x, nrows); - if (pos == nullptr) { - /* - DPCT1049:40: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope(x, dst, ncols, pos, freq_scale, p_delta_rows, - freq_base, ext_factor, attn_factor, corr_dims, - item_ct1); - }); - } else { - /* - DPCT1049:41: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope(x, dst, ncols, pos, freq_scale, p_delta_rows, - freq_base, ext_factor, attn_factor, corr_dims, - item_ct1); - }); - } -} - -template -static void rope_neox_sycl(const T *x, T *dst, int ncols, int n_dims, int nrows, - const int32_t *pos, float freq_scale, - int p_delta_rows, float freq_base, float ext_factor, - float attn_factor, rope_corr_dims corr_dims, - const float * freq_factors, queue_ptr stream) { - GGML_ASSERT(ncols % 2 == 0); - const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); - const int num_blocks_x = (ncols + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); - const sycl::range<3> block_nums(1, num_blocks_x, nrows); - - const float theta_scale = powf(freq_base, -2.0f/n_dims); - const float inv_ndims = -1.0f / n_dims; - - if (pos == nullptr) { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - if (freq_factors == nullptr) { - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, freq_factors, - item_ct1); - }); - } else { - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, freq_factors, - item_ct1); - }); - } - } else { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - if (freq_factors == nullptr) { - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, freq_factors, item_ct1); - }); - } else { - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, freq_factors, item_ct1); - }); - } - } -} - static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols, const int nrows, queue_ptr stream) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); @@ -3461,97 +3249,6 @@ catch (sycl::exception const &exc) { std::exit(1); } -inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - const ggml_tensor * src2 = dst->src[2]; - - GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); - GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); - GGML_ASSERT(src0->type == dst->type); - - const int64_t ne00 = src0->ne[0]; - const int64_t ne01 = src0->ne[1]; - const int64_t ne2 = dst->ne[2]; - const int64_t nrows = ggml_nrows(src0); - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - //const int n_ctx = ((int32_t *) dst->op_params)[3]; - const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; - - // RoPE alteration for extended context - float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; - memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); - memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); - memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); - memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); - - const float * freq_factors = nullptr; - const int32_t * pos = nullptr; - if ((mode & 1) == 0) { - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(src1->ne[0] == ne2); - pos = (const int32_t *) src1_dd; - } - - const bool is_neox = mode & 2; - -#pragma message("TODO: update rope NORM mode to match NEOX mode") -#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7634") - - if (is_neox) { - pos = (const int32_t *) src1_dd; - - if (src2 != nullptr) { - freq_factors = (const float *) src2->data; - } - } else { - GGML_ASSERT(src2 == nullptr && "TODO: freq_factors not implemented for !is_neox"); - } - - rope_corr_dims corr_dims; - ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v); - - // compute - if (is_neox) { - if (src0->type == GGML_TYPE_F32) { - rope_neox_sycl( - (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, freq_factors, main_stream - ); - } else if (src0->type == GGML_TYPE_F16) { - rope_neox_sycl((const sycl::half *)src0_dd, (sycl::half *)dst_dd, - ne00, n_dims, nrows, pos, freq_scale, ne01, - freq_base, ext_factor, attn_factor, corr_dims, - freq_factors, main_stream); - } else { - GGML_ASSERT(false); - } - } else { - if (src0->type == GGML_TYPE_F32) { - rope_sycl( - (const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, main_stream - ); - } else if (src0->type == GGML_TYPE_F16) { - rope_sycl((const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, - nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, main_stream); - } else { - GGML_ASSERT(false); - } - } - - (void) src1; - (void) dst; - (void) src1_dd; -} - static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const float *src0_dd, const float *src1_dd, @@ -6241,7 +5938,9 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons case GGML_OP_CONT: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + return true; case GGML_OP_ROPE: + return ggml_is_contiguous(op->src[0]); case GGML_OP_IM2COL: case GGML_OP_POOL_2D: case GGML_OP_SUM_ROWS: diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp index 2d37e271f..d5a63cd71 100644 --- a/ggml/src/ggml-sycl/backend.hpp +++ b/ggml/src/ggml-sycl/backend.hpp @@ -19,5 +19,6 @@ #include "dmmv.hpp" #include "mmq.hpp" #include "mmvq.hpp" +#include "rope.hpp" #endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp new file mode 100644 index 000000000..eabf1693e --- /dev/null +++ b/ggml/src/ggml-sycl/rope.cpp @@ -0,0 +1,275 @@ +#include "rope.hpp" + +struct rope_corr_dims { + float v[2]; +}; + +static float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low); + return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y)); +} + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static void rope_yarn( + float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * sycl::log(1.0f / freq_scale); + } + *cos_theta = sycl::cos(theta) * mscale; + *sin_theta = sycl::sin(theta) * mscale; +} + +template +static void rope_norm( + const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors, + const sycl::nd_item<3> &item_ct1) { + const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1)); + + if (i0 >= ne0) { + return; + } + + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i0 >= n_dims) { + const int i = row*ne0 + i0; + + dst[i + 0] = x[i + 0]; + dst[i + 1] = x[i + 1]; + + return; + } + + const int i = row*ne0 + i0; + const int i2 = row/p_delta_rows; + + const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f); + + const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f; + + float cos_theta; + float sin_theta; + + rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + 1]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + 1] = x0*sin_theta + x1*cos_theta; +} + +template +static void rope_neox( + const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors, + const sycl::nd_item<3> &item_ct1) { + const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1)); + + if (i0 >= ne0) { + return; + } + + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i0 >= n_dims) { + const int i = row*ne0 + i0; + + dst[i + 0] = x[i + 0]; + dst[i + 1] = x[i + 1]; + + return; + } + + const int i = row*ne0 + i0/2; + const int i2 = row/p_delta_rows; + + const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f); + + const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f; + + float cos_theta; + float sin_theta; + + rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + n_dims/2]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; +} + +template +static void rope_norm_sycl( + const T *x, T *dst, int ne0, int n_dims, int nr, const int32_t *pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, queue_ptr stream) { + GGML_ASSERT(ne0 % 2 == 0); + const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); + const int num_blocks_x = (ne0 + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); + const sycl::range<3> block_nums(1, num_blocks_x, nr); + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + if (freq_factors == nullptr) { + /* + DPCT1049:40: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_norm(x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows, + ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, + item_ct1); + }); + } else { + /* + DPCT1049:41: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_norm(x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows, + ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, + item_ct1); + }); + } +} + +template +static void rope_neox_sycl( + const T *x, T *dst, int ne0, int n_dims, int nr, const int32_t *pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, queue_ptr stream) { + GGML_ASSERT(ne0 % 2 == 0); + const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); + const int num_blocks_x = (ne0 + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); + const sycl::range<3> block_nums(1, num_blocks_x, nr); + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + if (freq_factors == nullptr) { + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ne0, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, freq_factors, + item_ct1); + }); + } else { + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ne0, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, freq_factors, + item_ct1); + }); + } +} + +void ggml_sycl_op_rope( + ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream) { + const ggml_tensor * src2 = dst->src[2]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); + GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + GGML_ASSERT(src0->type == dst->type); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t nr = ggml_nrows(src0); + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + //const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; + + // RoPE alteration for extended context + float freq_base; + float freq_scale; + float ext_factor; + float attn_factor; + float beta_fast; + float beta_slow; + + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + + const bool is_neox = mode & 2; + + const int32_t * pos = (const int32_t *) src1_dd; + + const float * freq_factors = nullptr; + if (src2 != nullptr) { + freq_factors = (const float *) src2->data; + } + + rope_corr_dims corr_dims; + ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v); + + // compute + if (is_neox) { + if (src0->type == GGML_TYPE_F32) { + rope_neox_sycl( + (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, freq_factors, main_stream + ); + } else if (src0->type == GGML_TYPE_F16) { + rope_neox_sycl( + (const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, freq_factors, main_stream + ); + } else { + GGML_ASSERT(false); + } + } else { + if (src0->type == GGML_TYPE_F32) { + rope_norm_sycl( + (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, freq_factors, main_stream + ); + } else if (src0->type == GGML_TYPE_F16) { + rope_norm_sycl( + (const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, freq_factors, main_stream + ); + } else { + GGML_ASSERT(false); + } + } + + (void) src1; + (void) dst; + (void) src1_dd; +} diff --git a/ggml/src/ggml-sycl/rope.hpp b/ggml/src/ggml-sycl/rope.hpp new file mode 100644 index 000000000..00354c313 --- /dev/null +++ b/ggml/src/ggml-sycl/rope.hpp @@ -0,0 +1,22 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_ROPE_HPP +#define GGML_SYCL_ROPE_HPP + +#include "common.hpp" + +void ggml_sycl_op_rope( + ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream); + +#endif // GGML_SYCL_ROPE_HPP From 694c59cb42d1ebd6a7d912ca65d3d7363e0f14c9 Mon Sep 17 00:00:00 2001 From: iacore <74560659+iacore@users.noreply.github.com> Date: Mon, 1 Jul 2024 11:40:58 +0000 Subject: [PATCH 03/38] Document BERT support. (#8205) * Update README.md document BERT support * Update README.md --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 99b16f6e2..153d837e3 100644 --- a/README.md +++ b/README.md @@ -108,6 +108,7 @@ Typically finetunes of the base models below are supported as well. - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon) - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2) - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne) +- [X] [BERT](https://github.com/ggerganov/llama.cpp/pull/5423) - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/) - [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft) - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila) From 257f8e41e24b5bbfc27d9e907189a3e0cdb650d4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 1 Jul 2024 14:46:18 +0300 Subject: [PATCH 04/38] nix : remove OpenCL remnants (#8235) * nix : remove OpenCL remnants * minor : remove parentheses --- .devops/nix/package.nix | 11 ++--------- 1 file changed, 2 insertions(+), 9 deletions(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 4ee0d62cb..b75d7ff9e 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -17,18 +17,15 @@ rocmPackages, vulkan-headers, vulkan-loader, - clblast, useBlas ? builtins.all (x: !x) [ useCuda useMetalKit - useOpenCL useRocm useVulkan ] && blas.meta.available, useCuda ? config.cudaSupport, - useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL, + useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin, useMpi ? false, # Increases the runtime closure size by ~700M - useOpenCL ? false, useRocm ? config.rocmSupport, useVulkan ? false, llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake @@ -56,7 +53,6 @@ let ++ lib.optionals useCuda [ "CUDA" ] ++ lib.optionals useMetalKit [ "MetalKit" ] ++ lib.optionals useMpi [ "MPI" ] - ++ lib.optionals useOpenCL [ "OpenCL" ] ++ lib.optionals useRocm [ "ROCm" ] ++ lib.optionals useVulkan [ "Vulkan" ]; @@ -198,7 +194,6 @@ effectiveStdenv.mkDerivation ( optionals effectiveStdenv.isDarwin darwinBuildInputs ++ optionals useCuda cudaBuildInputs ++ optionals useMpi [ mpi ] - ++ optionals useOpenCL [ clblast ] ++ optionals useRocm rocmBuildInputs ++ optionals useBlas [ blas ] ++ optionals useVulkan vulkanBuildInputs; @@ -210,7 +205,6 @@ effectiveStdenv.mkDerivation ( (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true) (cmakeBool "GGML_NATIVE" false) (cmakeBool "GGML_BLAS" useBlas) - (cmakeBool "GGML_CLBLAST" useOpenCL) (cmakeBool "GGML_CUDA" useCuda) (cmakeBool "GGML_HIPBLAS" useRocm) (cmakeBool "GGML_METAL" useMetalKit) @@ -254,7 +248,6 @@ effectiveStdenv.mkDerivation ( useCuda useMetalKit useMpi - useOpenCL useRocm useVulkan ; @@ -281,7 +274,7 @@ effectiveStdenv.mkDerivation ( # Configurations we don't want even the CI to evaluate. Results in the # "unsupported platform" messages. This is mostly a no-op, because # cudaPackages would've refused to evaluate anyway. - badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin; + badPlatforms = optionals useCuda lib.platforms.darwin; # Configurations that are known to result in build failures. Can be # overridden by importing Nixpkgs with `allowBroken = true`. From 3840b6f593751a0ba636bfda73b630cd6c29d7b5 Mon Sep 17 00:00:00 2001 From: Michael Francis Date: Mon, 1 Jul 2024 07:47:04 -0400 Subject: [PATCH 05/38] nix : enable curl (#8043) Co-authored-by: Georgi Gerganov --- .devops/nix/package.nix | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index b75d7ff9e..49e9b7528 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -17,6 +17,7 @@ rocmPackages, vulkan-headers, vulkan-loader, + curl, useBlas ? builtins.all (x: !x) [ useCuda useMetalKit @@ -27,6 +28,7 @@ useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin, useMpi ? false, # Increases the runtime closure size by ~700M useRocm ? config.rocmSupport, + enableCurl ? true, useVulkan ? false, llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake @@ -196,13 +198,15 @@ effectiveStdenv.mkDerivation ( ++ optionals useMpi [ mpi ] ++ optionals useRocm rocmBuildInputs ++ optionals useBlas [ blas ] - ++ optionals useVulkan vulkanBuildInputs; + ++ optionals useVulkan vulkanBuildInputs + ++ optionals enableCurl [ curl ]; cmakeFlags = [ (cmakeBool "LLAMA_BUILD_SERVER" true) (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic)) (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true) + (cmakeBool "LLAMA_CURL" enableCurl) (cmakeBool "GGML_NATIVE" false) (cmakeBool "GGML_BLAS" useBlas) (cmakeBool "GGML_CUDA" useCuda) From 0ddeff10230b88f1fa9866bbe5fe0d71ba2323a0 Mon Sep 17 00:00:00 2001 From: Roni Date: Mon, 1 Jul 2024 14:48:16 +0200 Subject: [PATCH 06/38] readme : update tool list (#8209) * Added gppm to Tool list in README * Update README.md --------- Co-authored-by: Georgi Gerganov --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 153d837e3..c136d4a5c 100644 --- a/README.md +++ b/README.md @@ -218,6 +218,7 @@ Unless otherwise noted these projects are open-source with permissive licensing: **Tools:** - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML +[crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption --- From 49122a873f54615626d1b49a2a39013ed4be98d5 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Mon, 1 Jul 2024 18:48:34 +0200 Subject: [PATCH 07/38] gemma2: add sliding window mask (#8227) * gemma2: add sliding window mask * fix data_swa uninitialized * better naming * add co-author Co-authored-by: Arlo Phoenix * replace list with single tensor * update * llama : minor styling * convert : add sanity check for query_pre_attn_scalar * fix small typo in README --------- Co-authored-by: Arlo Phoenix Co-authored-by: Georgi Gerganov --- README.md | 2 +- convert-hf-to-gguf.py | 6 +++ gguf-py/gguf/constants.py | 1 + gguf-py/gguf/gguf_writer.py | 3 ++ src/llama.cpp | 99 +++++++++++++++++++++++++------------ 5 files changed, 79 insertions(+), 32 deletions(-) diff --git a/README.md b/README.md index c136d4a5c..daba70717 100644 --- a/README.md +++ b/README.md @@ -218,7 +218,7 @@ Unless otherwise noted these projects are open-source with permissive licensing: **Tools:** - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML -[crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption +- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption --- diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 3ef2f69e7..4a7f500ff 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -2369,6 +2369,12 @@ class Gemma2Model(Model): self.gguf_writer.add_final_logit_softcapping( self.hparams["final_logit_softcapping"] ) + self.gguf_writer.add_sliding_window(self.hparams["sliding_window"]) + + # sanity check + attn_scalar = self.hparams["query_pre_attn_scalar"] + if attn_scalar != hparams["hidden_size"] / hparams["num_attention_heads"]: + raise ValueError("query_pre_attn_scalar must be equal to n_embd / n_head") def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: del bid # unusem diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 9bfa891d5..e87c58266 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -66,6 +66,7 @@ class Keys: Q_LORA_RANK = "{arch}.attention.q_lora_rank" KV_LORA_RANK = "{arch}.attention.kv_lora_rank" REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count" + SLIDING_WINDOW = "{arch}.attention.sliding_window" class Rope: DIMENSION_COUNT = "{arch}.rope.dimension_count" diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 1aeb0d9b0..75a8b2636 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -552,6 +552,9 @@ class GGUFWriter: def add_relative_attn_buckets_count(self, value: int) -> None: self.add_uint32(Keys.Attention.REL_BUCKETS_COUNT.format(arch=self.arch), value) + def add_sliding_window(self, value: int) -> None: + self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value) + def add_pooling_type(self, value: PoolingType) -> None: self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value) diff --git a/src/llama.cpp b/src/llama.cpp index 2a4d73856..eea532f6a 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -317,6 +317,7 @@ enum llm_kv { LLM_KV_ATTENTION_Q_LORA_RANK, LLM_KV_ATTENTION_KV_LORA_RANK, LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, + LLM_KV_ATTENTION_SLIDING_WINDOW, LLM_KV_ROPE_DIMENSION_COUNT, LLM_KV_ROPE_FREQ_BASE, @@ -409,6 +410,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" }, { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" }, { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" }, + { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, @@ -2085,6 +2087,7 @@ struct llama_hparams { uint32_t n_head_kv; uint32_t n_layer; uint32_t n_rot; + uint32_t n_swa = 0; // sliding window attention (SWA) uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head uint32_t n_ff; @@ -2139,6 +2142,7 @@ struct llama_hparams { if (this->n_head_kv != other.n_head_kv) return true; if (this->n_layer != other.n_layer) return true; if (this->n_rot != other.n_rot) return true; + if (this->n_swa != other.n_swa) return true; if (this->n_embd_head_k != other.n_embd_head_k) return true; if (this->n_embd_head_v != other.n_embd_head_v) return true; if (this->n_ff != other.n_ff) return true; @@ -2649,17 +2653,18 @@ struct llama_context { void * abort_callback_data = nullptr; // input tensors - struct ggml_tensor * inp_tokens; // I32 [n_batch] - struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] - struct ggml_tensor * inp_pos; // I32 [n_batch] - struct ggml_tensor * inp_out_ids; // I32 [n_outputs] - struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] - struct ggml_tensor * inp_K_shift; // I32 [kv_size] - struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch] - struct ggml_tensor * inp_cls; // I32 [n_batch] - struct ggml_tensor * inp_s_copy; // I32 [kv_size] - struct ggml_tensor * inp_s_mask; // F32 [1, n_kv] - struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch] + struct ggml_tensor * inp_tokens; // I32 [n_batch] + struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] + struct ggml_tensor * inp_pos; // I32 [n_batch] + struct ggml_tensor * inp_out_ids; // I32 [n_outputs] + struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] + struct ggml_tensor * inp_KQ_mask_swa; // F32 [kv_size, n_batch] + struct ggml_tensor * inp_K_shift; // I32 [kv_size] + struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch] + struct ggml_tensor * inp_cls; // I32 [n_batch] + struct ggml_tensor * inp_s_copy; // I32 [kv_size] + struct ggml_tensor * inp_s_mask; // F32 [1, n_kv] + struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch] // control vectors struct llama_control_vector cvec; @@ -4709,6 +4714,8 @@ static void llm_load_hparams( } break; case LLM_ARCH_GEMMA2: { + hparams.n_swa = 4096; // default value of gemma 2 + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping, false); ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false); @@ -5419,6 +5426,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv); LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); + LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa); LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k); LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v); LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa()); @@ -7775,17 +7783,18 @@ struct llm_build_context { ctx0 = ggml_init(params); - lctx.inp_tokens = nullptr; - lctx.inp_embd = nullptr; - lctx.inp_pos = nullptr; - lctx.inp_out_ids = nullptr; - lctx.inp_KQ_mask = nullptr; - lctx.inp_K_shift = nullptr; - lctx.inp_mean = nullptr; - lctx.inp_cls = nullptr; - lctx.inp_s_copy = nullptr; - lctx.inp_s_mask = nullptr; - lctx.inp_s_seq = nullptr; + lctx.inp_tokens = nullptr; + lctx.inp_embd = nullptr; + lctx.inp_pos = nullptr; + lctx.inp_out_ids = nullptr; + lctx.inp_KQ_mask = nullptr; + lctx.inp_KQ_mask_swa = nullptr; + lctx.inp_K_shift = nullptr; + lctx.inp_mean = nullptr; + lctx.inp_cls = nullptr; + lctx.inp_s_copy = nullptr; + lctx.inp_s_mask = nullptr; + lctx.inp_s_seq = nullptr; } void free() { @@ -7804,7 +7813,6 @@ struct llm_build_context { cb(lctx.inp_K_shift, "K_shift", -1); ggml_set_input(lctx.inp_K_shift); - for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * rope_factors = build_rope_factors(il); struct ggml_tensor * tmp = @@ -7939,16 +7947,27 @@ struct llm_build_context { } struct ggml_tensor * build_inp_KQ_mask(bool causal = true) { - if (causal) { - lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - } else { - lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - } + lctx.inp_KQ_mask = causal + ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)) + : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); cb(lctx.inp_KQ_mask, "KQ_mask", -1); ggml_set_input(lctx.inp_KQ_mask); + return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask, GGML_TYPE_F16) : lctx.inp_KQ_mask; } + struct ggml_tensor * build_inp_KQ_mask_swa(bool causal = true) { + GGML_ASSERT(hparams.n_swa > 0); + + lctx.inp_KQ_mask_swa = causal + ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)) + : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + cb(lctx.inp_KQ_mask_swa, "KQ_mask_swa", -1); + ggml_set_input(lctx.inp_KQ_mask_swa); + + return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask_swa, GGML_TYPE_F16) : lctx.inp_KQ_mask_swa; + } + struct ggml_tensor * build_inp_mean() { lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens); cb(lctx.inp_mean, "inp_mean", -1); @@ -11029,9 +11048,14 @@ struct llm_build_context { struct ggml_tensor * inp_pos = build_inp_pos(); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + // gemma 2 requires different mask for layers using sliding window (SWA) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(true); + struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(true); for (int il = 0; il < n_layer; ++il) { + // (il % 2) layers use SWA + struct ggml_tensor * KQ_mask_l = (il % 2 == 0) ? KQ_mask_swa : KQ_mask; + // norm cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, @@ -11067,7 +11091,7 @@ struct llm_build_context { cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask_l, n_tokens, kv_head, n_kv, 1.0f, cb, il); } cur = llm_build_norm(ctx0, cur, hparams, @@ -12670,7 +12694,12 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); - float * data = (float *) lctx.inp_KQ_mask->data; + float * data = (float *) lctx.inp_KQ_mask->data; + float * data_swa = nullptr; + + if (lctx.inp_KQ_mask_swa) { + data_swa = (float *) lctx.inp_KQ_mask_swa->data; + } // For causal attention, use only the previous KV cells // of the correct sequence for each token of the batch. @@ -12692,6 +12721,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } data[h*(n_kv*n_tokens) + j*n_kv + i] = f; + + // may need to cut off old tokens for sliding window + if (data_swa) { + if (pos - lctx.kv_self.cells[i].pos >= (int32_t)hparams.n_swa) { + f = -INFINITY; + } + data_swa[h*(n_kv*n_tokens) + j*n_kv + i] = f; + } } } From dae57a1ebc1c9bd5693ab999e19d77c5506ae559 Mon Sep 17 00:00:00 2001 From: Mateusz Charytoniuk Date: Mon, 1 Jul 2024 19:13:22 +0200 Subject: [PATCH 08/38] readme: add Paddler to the list of projects (#8239) --- README.md | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/README.md b/README.md index daba70717..3569b2bbb 100644 --- a/README.md +++ b/README.md @@ -220,6 +220,10 @@ Unless otherwise noted these projects are open-source with permissive licensing: - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption +**Infrastructure:** + +- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp + --- Here is a typical run using LLaMA v2 13B on M2 Ultra: From cb5fad4c6c2cbef92e9b8b63449e1cb7664e4846 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Mon, 1 Jul 2024 20:39:06 +0200 Subject: [PATCH 09/38] CUDA: refactor and optimize IQ MMVQ (#8215) * CUDA: refactor and optimize IQ MMVQ * uint -> uint32_t * __dp4a -> ggml_cuda_dp4a * remove MIN_CC_DP4A checks * change default * try CI fix --- ggml/src/ggml-common.h | 14 +- ggml/src/ggml-cuda.cu | 12 +- ggml/src/ggml-cuda/common.cuh | 76 +++- ggml/src/ggml-cuda/fattn-common.cuh | 50 +- ggml/src/ggml-cuda/mmvq.cu | 26 +- ggml/src/ggml-cuda/vecdotq.cuh | 682 +++++++++++++--------------- ggml/src/ggml-sycl/mmvq.cpp | 12 +- ggml/src/ggml-sycl/vecdotq.hpp | 21 - 8 files changed, 406 insertions(+), 487 deletions(-) diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index e8efceb76..c74060cc4 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -106,19 +106,19 @@ typedef sycl::half2 ggml_half2; #define QR6_K 2 #define QI2_XXS (QK_K / (4*QR2_XXS)) -#define QR2_XXS 8 +#define QR2_XXS 4 #define QI2_XS (QK_K / (4*QR2_XS)) -#define QR2_XS 8 +#define QR2_XS 4 #define QI2_S (QK_K / (4*QR2_S)) -#define QR2_S 8 +#define QR2_S 4 #define QI3_XXS (QK_K / (4*QR3_XXS)) -#define QR3_XXS 8 +#define QR3_XXS 4 #define QI3_XS (QK_K / (4*QR3_XS)) -#define QR3_XS 8 +#define QR3_XS 4 #define QI1_S (QK_K / (4*QR1_S)) #define QR1_S 8 @@ -130,10 +130,10 @@ typedef sycl::half2 ggml_half2; #define QR4_NL 2 #define QI4_XS (QK_K / (4*QR4_XS)) -#define QR4_XS 8 +#define QR4_XS 2 #define QI3_S (QK_K / (4*QR3_S)) -#define QR3_S 8 +#define QR3_S 4 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 0acfda91d..649ef5a08 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -1882,6 +1882,11 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor bool use_mul_mat_q = ggml_is_quantized(src0->type) && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32; + // if mmvq is available it's a better choice than dmmv: +#ifndef GGML_CUDA_FORCE_DMMV + use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q; +#endif // GGML_CUDA_FORCE_DMMV + bool any_gpus_with_slow_fp16 = false; if (split) { @@ -1894,22 +1899,15 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor } const int cc = ggml_cuda_info().devices[id].cc; - use_mul_mat_vec_q = use_mul_mat_vec_q && cc >= MIN_CC_DP4A; use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc); } } else { const int cc = ggml_cuda_info().devices[ctx.device].cc; - use_mul_mat_vec_q = use_mul_mat_vec_q && cc >= MIN_CC_DP4A; use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc); } - // if mmvq is available it's a better choice than dmmv: -#ifndef GGML_CUDA_FORCE_DMMV - use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q; -#endif // GGML_CUDA_FORCE_DMMV - // debug helpers //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]); //printf(" %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]); diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 8d00db6c1..472f4ace1 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -3,6 +3,7 @@ #include "ggml.h" #include "ggml-cuda.h" +#include #include #if defined(GGML_USE_HIPBLAS) @@ -268,30 +269,15 @@ static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigne return c; } -static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) { -#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) - c = __builtin_amdgcn_sdot4(a, b, c, false); -#elif defined(RDNA3) - c = __builtin_amdgcn_sudot4( true, a, true, b, c, false); -#elif defined(__gfx1010__) || defined(__gfx900__) - int tmp1; - int tmp2; - asm("\n \ - v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \ - v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \ - v_add3_u32 %0, %1, %2, %0 \n \ - v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \ - v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \ - v_add3_u32 %0, %1, %2, %0 \n \ - " - : "+v"(c), "=&v"(tmp1), "=&v"(tmp2) - : "v"(a), "v"(b) - ); -#else - const int8x4_t va = reinterpret_cast(a); - const int8x4_t vb = reinterpret_cast(b); - c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3]; -#endif +static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) { + const uint8x4_t& va = reinterpret_cast(a); + const uint8x4_t& vb = reinterpret_cast(b); + unsigned int c; + uint8x4_t& vc = reinterpret_cast(c); +#pragma unroll + for (int i = 0; i < 4; ++i) { + vc[i] = va[i] == vb[i] ? 0x00 : 0xff; + } return c; } @@ -467,8 +453,48 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half } #endif // CUDART_VERSION < 12000 +static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) { +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) + c = __builtin_amdgcn_sdot4(a, b, c, false); +#elif defined(RDNA3) + c = __builtin_amdgcn_sudot4( true, a, true, b, c, false); +#elif defined(__gfx1010__) || defined(__gfx900__) + int tmp1; + int tmp2; + asm("\n \ + v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \ + v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \ + v_add3_u32 %0, %1, %2, %0 \n \ + v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \ + v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \ + v_add3_u32 %0, %1, %2, %0 \n \ + " + : "+v"(c), "=&v"(tmp1), "=&v"(tmp2) + : "v"(a), "v"(b) + ); +#else + const int8x4_t va = reinterpret_cast(a); + const int8x4_t vb = reinterpret_cast(b); + c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3]; +#endif + return c; + +#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + +#if __CUDA_ARCH__ >= MIN_CC_DP4A + return __dp4a(a, b, c); +#else // __CUDA_ARCH__ >= MIN_CC_DP4A + const int8_t * a8 = (const int8_t *) &a; + const int8_t * b8 = (const int8_t *) &b; + return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3]; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A + +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +} + // TODO: move to ggml-common.h -static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; +static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v); diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh index bd7993595..650780fd2 100644 --- a/ggml/src/ggml-cuda/fattn-common.cuh +++ b/ggml/src/ggml-cuda/fattn-common.cuh @@ -54,12 +54,11 @@ typedef float (*vec_dot_KQ_f32_t)( template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q4_0 * K_q4_0 = (const block_q4_0 *) K_c; GGML_UNUSED(Q_v); - half sum = 0.0f; + T sum = 0.0f; #pragma unroll for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) { @@ -72,7 +71,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( const int v = (get_int_from_uint8(K_q4_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; const int u = Q_q8[k_KQ_0/WARP_SIZE]; - const int sumi = __dp4a(v, u, 0); + const int sumi = ggml_cuda_dp4a(v, u, 0); #ifdef FP16_AVAILABLE if (std::is_same::value) { @@ -90,19 +89,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q4_1 * K_q4_1 = (const block_q4_1 *) K_c; GGML_UNUSED(Q_v); @@ -120,7 +111,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( const int v = (get_int_from_uint8_aligned(K_q4_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; const int u = Q_q8[k_KQ_0/WARP_SIZE]; - const int sumi = __dp4a(v, u, 0); + const int sumi = ggml_cuda_dp4a(v, u, 0); #ifdef FP16_AVAILABLE if (std::is_same::value) { @@ -142,19 +133,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q5_0 * K_q5_0 = (const block_q5_0 *) K_c; GGML_UNUSED(Q_v); @@ -179,7 +162,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( const int u = Q_q8[k_KQ_0/WARP_SIZE]; - const int sumi = __dp4a(v, u, 0); + const int sumi = ggml_cuda_dp4a(v, u, 0); #ifdef FP16_AVAILABLE if (std::is_same::value) { @@ -197,19 +180,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q5_1 * K_q5_1 = (const block_q5_1 *) K_c; GGML_UNUSED(Q_v); @@ -234,7 +209,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( const int u = Q_q8[k_KQ_0/WARP_SIZE]; - const int sumi = __dp4a(v, u, 0); + const int sumi = ggml_cuda_dp4a(v, u, 0); #ifdef FP16_AVAILABLE if (std::is_same::value) { @@ -256,19 +231,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q8_0 * K_q8_0 = (const block_q8_0 *) K_c; GGML_UNUSED(Q_v); @@ -297,13 +264,6 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu index e8d157169..e22faf69b 100644 --- a/ggml/src/ggml-cuda/mmvq.cu +++ b/ggml/src/ggml-cuda/mmvq.cu @@ -28,16 +28,22 @@ static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) static constexpr __device__ int get_vdr_mmvq(ggml_type type) { return type == GGML_TYPE_Q4_0 ? VDR_Q4_0_Q8_1_MMVQ : - type == GGML_TYPE_Q4_1 ? VDR_Q4_1_Q8_1_MMVQ : - type == GGML_TYPE_Q5_0 ? VDR_Q5_0_Q8_1_MMVQ : - type == GGML_TYPE_Q5_1 ? VDR_Q5_1_Q8_1_MMVQ : - type == GGML_TYPE_Q8_0 ? VDR_Q8_0_Q8_1_MMVQ : - type == GGML_TYPE_Q2_K ? VDR_Q2_K_Q8_1_MMVQ : - type == GGML_TYPE_Q3_K ? VDR_Q3_K_Q8_1_MMVQ : - type == GGML_TYPE_Q4_K ? VDR_Q4_K_Q8_1_MMVQ : - type == GGML_TYPE_Q5_K ? VDR_Q5_K_Q8_1_MMVQ : - type == GGML_TYPE_Q6_K ? VDR_Q6_K_Q8_1_MMVQ : - type == GGML_TYPE_IQ4_NL ? VDR_Q4_K_Q8_1_MMVQ : + type == GGML_TYPE_Q4_1 ? VDR_Q4_1_Q8_1_MMVQ : + type == GGML_TYPE_Q5_0 ? VDR_Q5_0_Q8_1_MMVQ : + type == GGML_TYPE_Q5_1 ? VDR_Q5_1_Q8_1_MMVQ : + type == GGML_TYPE_Q8_0 ? VDR_Q8_0_Q8_1_MMVQ : + type == GGML_TYPE_Q2_K ? VDR_Q2_K_Q8_1_MMVQ : + type == GGML_TYPE_Q3_K ? VDR_Q3_K_Q8_1_MMVQ : + type == GGML_TYPE_Q4_K ? VDR_Q4_K_Q8_1_MMVQ : + type == GGML_TYPE_Q5_K ? VDR_Q5_K_Q8_1_MMVQ : + type == GGML_TYPE_Q6_K ? VDR_Q6_K_Q8_1_MMVQ : + type == GGML_TYPE_IQ2_XXS ? VDR_IQ2_XXS_Q8_1_MMVQ : + type == GGML_TYPE_IQ2_XS ? VDR_IQ2_XS_Q8_1_MMVQ : + type == GGML_TYPE_IQ2_S ? VDR_IQ2_S_Q8_1_MMVQ : + type == GGML_TYPE_IQ3_XXS ? VDR_IQ3_XXS_Q8_1_MMVQ : + type == GGML_TYPE_IQ3_S ? VDR_IQ3_S_Q8_1_MMVQ : + type == GGML_TYPE_IQ4_NL ? VDR_IQ4_NL_Q8_1_MMVQ : + type == GGML_TYPE_IQ4_XS ? VDR_IQ4_XS_Q8_1_MMVQ : 1; } diff --git a/ggml/src/ggml-cuda/vecdotq.cuh b/ggml/src/ggml-cuda/vecdotq.cuh index 3b12d6566..3f3a87c75 100644 --- a/ggml/src/ggml-cuda/vecdotq.cuh +++ b/ggml/src/ggml-cuda/vecdotq.cuh @@ -1,4 +1,5 @@ #include "common.cuh" +#include static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) { const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment @@ -28,6 +29,18 @@ static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment } +static __device__ __forceinline__ int get_int_b2(const void * x, const int & i32) { + const uint16_t * x16 = (const uint16_t *) x; + + int x32 = x16[2*i32 + 0] << 0; + x32 |= x16[2*i32 + 1] << 16; + + return x32; +} + +static __device__ __forceinline__ int get_int_b4(const void * x, const int & i32) { + return ((const int *) x)[i32]; // assume at least 4 byte alignment +} // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q @@ -38,7 +51,6 @@ static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl( const int * v, const int * u, const float & d4, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -47,17 +59,14 @@ template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_imp const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; // SIMD dot product of quantized values - sumi = __dp4a(vi0, u[2*i+0], sumi); - sumi = __dp4a(vi1, u[2*i+1], sumi); + sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); + sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); } const float2 ds8f = __half22float2(ds8); // second part effectively subtracts 8 from each quant value return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q4_1_Q8_1_MMVQ 2 @@ -66,7 +75,6 @@ template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_imp template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl( const int * v, const int * u, const half2 & dm4, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -75,8 +83,8 @@ template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; // SIMD dot product of quantized values - sumi = __dp4a(vi0, u[2*i+0], sumi); - sumi = __dp4a(vi1, u[2*i+1], sumi); + sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); + sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); } #ifdef GGML_CUDA_F16 @@ -92,9 +100,6 @@ template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1)); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q5_0_Q8_1_MMVQ 2 @@ -103,7 +108,6 @@ template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl( const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -113,23 +117,20 @@ template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_imp vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 - sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values + sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 - sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values + sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values } const float2 ds8f = __half22float2(ds8); // second part effectively subtracts 16 from each quant value return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q5_1_Q8_1_MMVQ 2 @@ -138,7 +139,6 @@ template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_imp template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl( const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -148,14 +148,14 @@ template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 - sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values + sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 - sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values + sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values } #ifdef GGML_CUDA_F16 @@ -171,10 +171,6 @@ template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it return sumi*d5d8 + m5s8 / (QI5_1 / vdr); - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q8_0_Q8_1_MMVQ 2 @@ -183,31 +179,26 @@ template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp template static __device__ __forceinline__ T vec_dot_q8_0_q8_1_impl( const int * v, const int * u, const T & d8_0, const T & d8_1) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll for (int i = 0; i < vdr; ++i) { // SIMD dot product of quantized values - sumi = __dp4a(v[i], u[i], sumi); + sumi = ggml_cuda_dp4a(v[i], u[i], sumi); } return d8_0*d8_1 * ((T) sumi); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl( const int * v, const int * u, const half2 & dm8, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll for (int i = 0; i < vdr; ++i) { // SIMD dot product of quantized values - sumi = __dp4a(v[i], u[i], sumi); + sumi = ggml_cuda_dp4a(v[i], u[i], sumi); } #ifdef GGML_CUDA_F16 @@ -223,9 +214,6 @@ template static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it return sumi*d8d8 + m8s8 / (QI8_1 / vdr); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q2_K_Q8_1_MMVQ 1 @@ -236,7 +224,6 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq( const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales, const half2 & dm2, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -246,28 +233,24 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq( const int vi = (v >> (2*i)) & 0x03030303; - sumf_d += d8[i] * (__dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product + sumf_d += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product // fill int with 4x m int m = sc >> 4; m |= m << 8; m |= m << 16; - sumf_m += d8[i] * __dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values + sumf_m += d8[i] * ggml_cuda_dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values } const float2 dm2f = __half22float2(dm2); return dm2f.x*sumf_d - dm2f.y*sumf_m; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const half2 * dm2, const float & d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -281,8 +264,8 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( #pragma unroll for (int i = i0; i < i0 + QI8_1/2; ++i) { const int vi = (vi0 >> (2*(i % (QI8_1/2)))) & 0x03030303; - sumi_d = __dp4a(vi, u[i], sumi_d); // SIMD dot product - sumi_m = __dp4a(0x01010101, u[i], sumi_m); + sumi_d = ggml_cuda_dp4a(vi, u[i], sumi_d); // SIMD dot product + sumi_m = ggml_cuda_dp4a(0x01010101, u[i], sumi_m); } sumf_d += dm2f.x * sumi_d; @@ -290,9 +273,6 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( } return d8*(sumf_d - sumf_m); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q3_K_Q8_1_MMVQ 1 @@ -303,7 +283,6 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq( const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales, const int & scale_offset, const float & d3, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf = 0.0f; #pragma unroll @@ -326,13 +305,10 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq( const int vi = __vsubss4(vil, vih); - sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product + sumf += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * sc); // SIMD dot product } return d3 * sumf; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values @@ -340,7 +316,6 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales, const float & d3, const float & d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -350,16 +325,13 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq( #pragma unroll for (int i = i0; i < i0 + QI8_1/2; ++i) { const int vi = __vsubss4((v[i/2] >> (4*(i%2))) & 0x0F0F0F0F, 0x04040404); - sumi_sc = __dp4a(vi, u[i], sumi_sc); // SIMD dot product + sumi_sc = ggml_cuda_dp4a(vi, u[i], sumi_sc); // SIMD dot product } sumi += sumi_sc * scales[i0 / (QI8_1/2)]; } return d3*d8 * sumi; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q4_K_Q8_1_MMVQ 2 @@ -370,7 +342,6 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -379,8 +350,8 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F; const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F; - const int dot1 = __dp4a(v1i, u[2*i+1], __dp4a(v0i, u[2*i+0], 0)); // SIMD dot product - const int dot2 = __dp4a(0x01010101, u[2*i+1], __dp4a(0x01010101, u[2*i+0], 0)); // sum of u + const int dot1 = ggml_cuda_dp4a(v1i, u[2*i+1], ggml_cuda_dp4a(v0i, u[2*i+0], 0)); // SIMD dot product + const int dot2 = ggml_cuda_dp4a(0x01010101, u[2*i+1], ggml_cuda_dp4a(0x01010101, u[2*i+0], 0)); // sum of u sumf_d += d8[i] * (dot1 * sc[i]); sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values @@ -389,10 +360,6 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( const float2 dm4f = __half22float2(dm4); return dm4f.x*sumf_d - dm4f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values @@ -400,7 +367,6 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -410,7 +376,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( #pragma unroll for (int j = 0; j < QI8_1; ++j) { - sumi_d = __dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product + sumi_d = ggml_cuda_dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product } const float2 ds8f = __half22float2(ds8[i]); @@ -422,10 +388,6 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( const float2 dm4f = __half22float2(dm4); return dm4f.x*sumf_d - dm4f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q5_K_Q8_1_MMVQ 2 @@ -436,7 +398,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc, const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -451,8 +412,8 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( const int v0i = vl0i | vh0i; const int v1i = vl1i | vh1i; - const int dot1 = __dp4a(v0i, u[2*i+0], __dp4a(v1i, u[2*i+1], 0)); // SIMD dot product - const int dot2 = __dp4a(0x01010101, u[2*i+0], __dp4a(0x01010101, u[2*i+1], 0)); // sum of u + const int dot1 = ggml_cuda_dp4a(v0i, u[2*i+0], ggml_cuda_dp4a(v1i, u[2*i+1], 0)); // SIMD dot product + const int dot2 = ggml_cuda_dp4a(0x01010101, u[2*i+0], ggml_cuda_dp4a(0x01010101, u[2*i+1], 0)); // sum of u sumf_d += d8[i] * (dot1 * sc[i]); sumf_m += d8[i] * (dot2 * m[i]); @@ -462,10 +423,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( const float2 dm5f = __half22float2(dm5); return dm5f.x*sumf_d - dm5f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values @@ -473,7 +430,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -483,7 +439,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( #pragma unroll for (int j = 0; j < QI8_1; ++j) { - sumi_d = __dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product + sumi_d = ggml_cuda_dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product } const float2 ds8f = __half22float2(ds8[i]); @@ -495,10 +451,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( const float2 dm4f = __half22float2(dm4); return dm4f.x*sumf_d - dm4f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q6_K_Q8_1_MMVQ 1 @@ -509,7 +461,6 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq( const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales, const float & d, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf = 0.0f; #pragma unroll @@ -522,13 +473,10 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq( const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32 - sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product + sumf += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * sc); // SIMD dot product } return d*sumf; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values @@ -536,7 +484,6 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc, const float & d6, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; #pragma unroll @@ -545,21 +492,17 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq( #pragma unroll for (int i = i0; i < i0 + 2; ++i) { - sumi_d.x = __dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product - sumi_d.x = __dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product + sumi_d.x = ggml_cuda_dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product + sumi_d.x = ggml_cuda_dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product - sumi_d.y = __dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product - sumi_d.y = __dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product + sumi_d.y = ggml_cuda_dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product + sumi_d.y = ggml_cuda_dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product } sumf_d += d8[i0/4] * (sc[i0/2+0]*sumi_d.x + sc[i0/2+1]*sumi_d.y); } return d6 * sumf_d; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } static __device__ __forceinline__ float vec_dot_q4_0_q8_1( @@ -572,9 +515,9 @@ static __device__ __forceinline__ float vec_dot_q4_0_q8_1( #pragma unroll for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_uint8(bq4_0->qs, iqs + i); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0); + v[i] = get_int_b2(bq4_0->qs, iqs + i); + u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI4_0); } return vec_dot_q4_0_q8_1_impl(v, u, bq4_0->d, bq8_1->ds); @@ -591,9 +534,9 @@ static __device__ __forceinline__ float vec_dot_q4_1_q8_1( #pragma unroll for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1); + v[i] = get_int_b4(bq4_1->qs, iqs + i); + u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI4_1); } return vec_dot_q4_1_q8_1_impl(v, u, bq4_1->dm, bq8_1->ds); @@ -610,10 +553,10 @@ static __device__ __forceinline__ float vec_dot_q5_0_q8_1( #pragma unroll for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) { - vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i); - vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i)); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0); + vl[i] = get_int_b2(bq5_0->qs, iqs + i); + vh[i] = get_int_b2(bq5_0->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI5_0); } return vec_dot_q5_0_q8_1_impl(vl, vh, u, bq5_0->d, bq8_1->ds); @@ -630,10 +573,10 @@ static __device__ __forceinline__ float vec_dot_q5_1_q8_1( #pragma unroll for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) { - vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i); - vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i)); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1); + vl[i] = get_int_b4(bq5_1->qs, iqs + i); + vh[i] = get_int_b4(bq5_1->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI5_1); } return vec_dot_q5_1_q8_1_impl(vl, vh, u, bq5_1->dm, bq8_1->ds); @@ -649,8 +592,8 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1( #pragma unroll for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_int8(bq8_0->qs, iqs + i); - u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + v[i] = get_int_b2(bq8_0->qs, iqs + i); + u[i] = get_int_b4(bq8_1->qs, iqs + i); } return vec_dot_q8_0_q8_1_impl(v, u, bq8_0->d, __low2half(bq8_1->ds)); @@ -666,13 +609,13 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1( const uint8_t * scales = bq2_K->scales + scale_offset; - const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs); + const int v = get_int_b4(bq2_K->qs, iqs); int u[QR2_K]; float d8[QR2_K]; #pragma unroll for (int i = 0; i < QR2_K; ++ i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + u[i] = get_int_b4(bq8_1[bq8_offset + i].qs, iqs % QI8_1); d8[i] = __low2float(bq8_1[bq8_offset + i].ds); } @@ -689,17 +632,17 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1( const float d = bq3_K->d; - const int vl = get_int_from_uint8(bq3_K->qs, iqs); + const int vl = get_int_b2(bq3_K->qs, iqs); // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted - const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset; + const int vh = ~get_int_b2(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset; int u[QR3_K]; float d8[QR3_K]; #pragma unroll for (int i = 0; i < QR3_K; ++i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + u[i] = get_int_b4(bq8_1[bq8_offset + i].qs, iqs % QI8_1); d8[i] = __low2float(bq8_1[bq8_offset + i].ds); } @@ -807,8 +750,8 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1( const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8); const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4)); - const int vl = get_int_from_uint8(bq6_K->ql, iqs); - const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift; + const int vl = get_int_b2(bq6_K->ql, iqs); + const int vh = get_int_b2(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift; const int8_t * scales = bq6_K->scales + scale_offset; @@ -817,335 +760,342 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1( #pragma unroll for (int i = 0; i < QR6_K; ++i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); + u[i] = get_int_b4(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds); } return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8); } +#define VDR_IQ2_XXS_Q8_1_MMVQ 2 + static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { + const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq + kbx; -#if QR2_XXS == 8 - const int ib32 = iqs; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const int8_t * q8 = bq8_1[ib32].qs; - uint32_t aux32 = q2[2] | (q2[3] << 16); + const int q2 = get_int_b2(bq2->qs, iqs); + const uint8_t * aux8 = (const uint8_t *) &q2; + const uint32_t aux32 = get_int_b2(bq2->qs, iqs + 1); + int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); - const uint8_t signs = ksigns_iq2xs[aux32 & 127]; - for (int j = 0; j < 8; ++j) { - sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - aux32 >>= 7; +#pragma unroll + for (int k0 = 0; k0 < 8; k0 += 2) { + const int * grid_pos = (const int *) (iq2xxs_grid + aux8[k0/2]); + const int signs_packed = ksigns_iq2xs[(aux32 >> (7*k0/2)) & 0x7F]; + + const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000); + const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0); + const int u0 = get_int_b4(bq8_1[iqs/2].qs, k0 + 0); + sumi = ggml_cuda_dp4a(grid0, u0, sumi); + + const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000); + const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, k0 + 1); + sumi = ggml_cuda_dp4a(grid1, u1, sumi); } - const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f; + + const int ls = aux32 >> 28; + sumi = (ls*sumi + sumi/2)/4; + const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds); return d * sumi; -#else - // iqs is 0...15 - const int ib32 = iqs/2; - const int il = iqs%2; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); - const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); - const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f; - const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; - const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; - const int8_t * q8 = bq8_1[ib32].qs + 16*il; - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1); - sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1); - } - return d * (sumi1 + sumi2); -#endif } +#define VDR_IQ2_XS_Q8_1_MMVQ 2 + static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq + kbx; - const int ib32 = iqs; - const uint16_t * q2 = bq2->qs + 4*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - const uint8_t ls1 = bq2->scales[ib32] & 0xf; - const uint8_t ls2 = bq2->scales[ib32] >> 4; + const int2 q2_packed = make_int2(get_int_b2(bq2->qs, iqs + 0), get_int_b2(bq2->qs, iqs + 1)); + const uint16_t * q2 = (const uint16_t *) &q2_packed; + const int ls0 = bq2->scales[iqs/2] & 0x0F; + const int ls1 = bq2->scales[iqs/2] >> 4; + + int sumi0 = 0; int sumi1 = 0; - for (int l = 0; l < 2; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); - const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); - const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]); - const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]); - sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1); - sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1); - q8 += 8; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l0/2] & 0x000001FF)); + const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l0/2] >> 9)); + + const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]); + const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]); + + const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1); + + if (l0 < 4) { + sumi0 = ggml_cuda_dp4a(grid_l, u0, sumi0); + sumi0 = ggml_cuda_dp4a(grid_h, u1, sumi0); + } else { + sumi1 = ggml_cuda_dp4a(grid_l, u0, sumi1); + sumi1 = ggml_cuda_dp4a(grid_h, u1, sumi1); + } } - int sumi2 = 0; - for (int l = 2; l < 4; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); - const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); - const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]); - const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]); - sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2); - sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2); - q8 += 8; - } - const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; - return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); -#else - GGML_UNUSED(ksigns64); - NO_DEVICE_CODE; -#endif + const int sumi = (sumi0*ls0 + sumi1*ls1 + (sumi0 + sumi1)/2)/4; + const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds); + return d * sumi; } -// TODO +#define VDR_IQ2_S_Q8_1_MMVQ 2 + static __device__ __forceinline__ float vec_dot_iq2_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + const block_iq2_s * bq2 = (const block_iq2_s *) vbq + kbx; - const int ib32 = iqs; - const int8_t * q8 = bq8_1[ib32].qs; - const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32; - const uint8_t ls1 = bq2->scales[ib32] & 0xf; - const uint8_t ls2 = bq2->scales[ib32] >> 4; + const int qs_packed = get_int_b2(bq2->qs, iqs/2); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bq2->qh[iqs/2]; + + const int signs_packed_32 = get_int_b2(bq2->qs, QK_K/32 + iqs/2); + const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; + + const int ls0 = bq2->scales[iqs/2] & 0x0F; + const int ls1 = bq2->scales[iqs/2] >> 4; + + int sumi0 = 0; int sumi1 = 0; - for (int l = 0; l < 2; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); - const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); - const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); - const int grid_l = __vsub4(grid[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid[1] ^ signs1, signs1); - sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1); - sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1); - q8 += 8; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int * grid_pos = (const int *)(iq2s_grid + (qs[l0/2] | ((qh << (8-l0)) & 0x300))); + + const int signs0 = __vcmpne4(((signs_packed_8[l0/2] & 0x03) << 7) | ((signs_packed_8[l0/2] & 0x0C) << 21), 0x00000000); + const int signs1 = __vcmpne4(((signs_packed_8[l0/2] & 0x30) << 3) | ((signs_packed_8[l0/2] & 0xC0) << 17), 0x00000000); + + const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0); + const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1); + + const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1); + + if (l0 < 4) { + sumi0 = ggml_cuda_dp4a(grid_l, u0, sumi0); + sumi0 = ggml_cuda_dp4a(grid_h, u1, sumi0); + } else { + sumi1 = ggml_cuda_dp4a(grid_l, u0, sumi1); + sumi1 = ggml_cuda_dp4a(grid_h, u1, sumi1); + } } - int sumi2 = 0; - for (int l = 2; l < 4; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); - const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); - const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); - const int grid_l = __vsub4(grid[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid[1] ^ signs1, signs1); - sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2); - sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2); - q8 += 8; - } - const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; - return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); -#else - GGML_UNUSED(ksigns64); - NO_DEVICE_CODE; -#endif + const int sumi = (sumi0*ls0 + sumi1*ls1 + (sumi0 + sumi1)/2)/4; + + const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds); + return d * sumi; } +#define VDR_IQ3_XXS_Q8_1_MMVQ 2 + static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq + kbx; - const int ib32 = iqs; - const uint8_t * q3 = bq2->qs + 8*ib32; - const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - uint32_t aux32 = gas[0] | (gas[1] << 16); + const block_iq3_xxs * bq3 = (const block_iq3_xxs *) vbq + kbx; + + const int2 q3_packed = make_int2(get_int_b2(bq3->qs, iqs), get_int_b2(bq3->qs, iqs+1)); + const uint8_t * q3 = (const uint8_t *) &q3_packed; + const uint32_t aux32 = get_int_b2(bq3->qs, QK_K/16 + iqs/2); + int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0]; - const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1]; - const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127)); - const int grid_l = __vsub4(grid1[0] ^ signs[0], signs[0]); - const int grid_h = __vsub4(grid2[0] ^ signs[1], signs[1]); - sumi = __dp4a(grid_l, *((int *)q8+0), sumi); - sumi = __dp4a(grid_h, *((int *)q8+1), sumi); - q8 += 8; - aux32 >>= 7; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int2 grid_pos = make_int2(iq3xxs_grid[q3[l0 + 0]], iq3xxs_grid[q3[l0 + 1]]); + + const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l0/2)) & 0x7F)); + + const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]); + const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]); + + const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1); + + sumi = ggml_cuda_dp4a(grid_l, u0, sumi); + sumi = ggml_cuda_dp4a(grid_h, u1, sumi); } - const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f; + + const int ls = aux32 >> 28; + sumi = (ls*sumi + sumi/2)/2; + const float d = __half2float(bq3->d) * __low2float(bq8_1[iqs/2].ds); return d * sumi; -#else - NO_DEVICE_CODE; -#endif } +#define VDR_IQ3_S_Q8_1_MMVQ 2 + // TODO: don't use lookup table for signs static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_iq3_s * bq2 = (const block_iq3_s *) vbq + kbx; - const int ib32 = iqs; - const uint8_t * qs = bq2->qs + 8*ib32; - const int8_t * q8 = bq8_1[ib32].qs; + const block_iq3_s * bq3 = (const block_iq3_s *) vbq + kbx; + + const int2 qs_packed = make_int2(get_int_b2(bq3->qs, iqs + 0), get_int_b2(bq3->qs, iqs + 1)); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bq3->qh[iqs/2]; + + const int signs_packed_32 = get_int_b2(bq3->signs, iqs/2); + const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; + int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)); - const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)); - uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); - uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); - const int grid_l = __vsub4(grid1[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid2[0] ^ signs1, signs1); - sumi = __dp4a(grid_l, *((int *)q8+0), sumi); - sumi = __dp4a(grid_h, *((int *)q8+1), sumi); - q8 += 8; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int2 grid_pos = make_int2( + iq3s_grid[qs[l0 + 0] | ((qh << (8 - l0)) & 0x100)], + iq3s_grid[qs[l0 + 1] | ((qh << (7 - l0)) & 0x100)]); + + const int signs0 = __vcmpne4(((signs_packed_8[l0/2] & 0x03) << 7) | ((signs_packed_8[l0/2] & 0x0C) << 21), 0x00000000); + const int signs1 = __vcmpne4(((signs_packed_8[l0/2] & 0x30) << 3) | ((signs_packed_8[l0/2] & 0xC0) << 17), 0x00000000); + + const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); + const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); + + const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1); + + sumi = ggml_cuda_dp4a(grid_l, u0, sumi); + sumi = ggml_cuda_dp4a(grid_h, u1, sumi); } - const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds); + + sumi *= 1 + 2*((bq3->scales[iqs/4] >> ((iqs << 1) & 0x04)) & 0x0F); + + const float d = __half2float(bq3->d) * __low2float(bq8_1[iqs/2].ds); return d * sumi; -#else - NO_DEVICE_CODE; -#endif } static __device__ __forceinline__ float vec_dot_iq1_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { const block_iq1_s * bq1 = (const block_iq1_s *) vbq + kbx; - const int ib32 = iqs; + const int qs_packed = get_int_b2(bq1->qs, iqs); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bq1->qh[iqs]; + int sumi = 0; -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const int * q8 = (const int *)bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); - int grid0 = grid[0] & 0x0f0f0f0f; - int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; - sumi = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi)); +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int grid = iq1s_grid_gpu[qs[l0/2] | (((qh >> 3*(l0/2)) & 0x07) << 8)]; + + const int grid0 = (grid >> 0) & 0x0F0F0F0F; + const int grid1 = (grid >> 4) & 0x0F0F0F0F; + + const int u0 = get_int_b4(bq8_1[iqs].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs].qs, l0 + 1); + + sumi = ggml_cuda_dp4a(grid0, u0, sumi); + sumi = ggml_cuda_dp4a(grid1, u1, sumi); } -#else - const int8_t * q8 = bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); - for (int j = 0; j < 4; ++j) { - sumi += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4); - } - q8 += 8; - } -#endif - const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA; - const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1); - const float d = d1q * __low2float (bq8_1[ib32].ds); - const float m = d1q * __high2float(bq8_1[ib32].ds); - return d * sumi + m * delta; + + const float d1q = __half2float(bq1->d) * (((qh >> 11) & 0x0E) + 1); + const float delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000); + const float2 ds = __half22float2(bq8_1[iqs].ds); + return d1q * (ds.x*sumi + ds.y*delta); } static __device__ __forceinline__ float vec_dot_iq1_m_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { + const block_iq1_m * bq1 = (const block_iq1_m *) vbq + kbx; - const int ib32 = iqs; - int sumi[2] = {0, 0}; - float sumf[2] = {0.f, 0.f}; -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const int * q8 = (const int *)bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 7) << 8))); - int grid0 = grid[0] & 0x0f0f0f0f; - int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; - sumi[l/2] = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi[l/2])); - const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA; - const int sumy = __dp4a(q8[2*l+1], 0x01010101, __dp4a(q8[2*l+0], 0x01010101, 0)); - sumf[l/2] += delta*sumy; - } -#else - const int8_t * q8 = bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); + const int qs_packed = get_int_b4(bq1->qs, iqs); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + int sumi[2] = {0}; + float sumf[2] = {0.0f}; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int qhl = bq1->qh[2*iqs + l0/4] >> (4 * ((l0/2) % 2)); + + const int grid = iq1s_grid_gpu[qs[l0/2] | ((qhl & 0x07) << 8)]; + + const int grid0 = (grid >> 0) & 0x0F0F0F0F; + const int grid1 = (grid >> 4) & 0x0F0F0F0F; + + const int u0 = get_int_b4(bq8_1[iqs].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs].qs, l0 + 1); + + sumi[l0/4] = ggml_cuda_dp4a(grid0, u0, sumi[l0/4]); + sumi[l0/4] = ggml_cuda_dp4a(grid1, u1, sumi[l0/4]); + + const float delta = -1.0f + IQ1M_DELTA - (qhl & 0x08) * (2.0f*IQ1M_DELTA/0x08); int sumy = 0; - for (int j = 0; j < 4; ++j) { - sumi[l/2] += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4); - sumy += q8[j] + q8[j+4]; - } - const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA; - sumf[l/2] += delta*sumy; - q8 += 8; + sumy = ggml_cuda_dp4a(u0, 0x01010101, sumy); + sumy = ggml_cuda_dp4a(u1, 0x01010101, sumy); + sumf[l0/4] += delta*sumy; } -#endif + + const uint16_t * sc = (const uint16_t *) bq1->scales; + iq1m_scale_t scale; - const uint16_t * sc = (const uint16_t *)bq1->scales; - scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); - const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds); - return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1)); + scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00F0) | ((sc[2] >> 4) & 0x0F00) | (sc[3] & 0xF000); + const float d = __half2float(scale.f16) * __low2float(bq8_1[iqs].ds); + + const int tmp = sc[iqs/2] >> (6*(iqs%2)); + const int sc0 = 2*((tmp >> 0) & 0x07) + 1; + const int sc1 = 2*((tmp >> 3) & 0x07) + 1; + return d * ((sumi[0] + sumf[0]) * sc0 + (sumi[1] + sumf[1]) * sc1); } -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -static __device__ __forceinline__ void get_int_from_table_16(const uint32_t & q4, const uint8_t * values, - int & val1, int & val2) { +static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4) { + const int q0_32 = (q4 >> 0) & 0x0F0F0F0F; + const int8_t * q0_8 = (const int8_t *) &q0_32; + const char4 val0_8 = make_char4( + kvalues_iq4nl[q0_8[0]], kvalues_iq4nl[q0_8[1]], kvalues_iq4nl[q0_8[2]], kvalues_iq4nl[q0_8[3]]); - uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32; - aux32 = q4 & 0x0f0f0f0f; - uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8); - uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8); - val1 = v1 | (v2 << 16); - aux32 = (q4 >> 4) & 0x0f0f0f0f; - v1 = values[q8[0]] | (values[q8[1]] << 8); - v2 = values[q8[2]] | (values[q8[3]] << 8); - val2 = v1 | (v2 << 16); + const int q1_32 = (q4 >> 4) & 0x0F0F0F0F; + const int8_t * q1_8 = (const int8_t *) &q1_32; + const char4 val1_8 = make_char4( + kvalues_iq4nl[q1_8[0]], kvalues_iq4nl[q1_8[1]], kvalues_iq4nl[q1_8[2]], kvalues_iq4nl[q1_8[3]]); + + return make_int2(*((const int *) &val0_8), *((const int *) &val1_8)); } -#endif + +#define VDR_IQ4_NL_Q8_1_MMVQ 2 static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { - const block_iq4_nl * bq = (const block_iq4_nl *) vbq + kbx; + const block_iq4_nl * bq4 = (const block_iq4_nl *) vbq + kbx; -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs; - const int32_t * q8 = (const int32_t *)bq8_1->qs + iqs; + const int * q8 = (const int *) bq8_1->qs + iqs; - const uint8_t * values = (const uint8_t *)kvalues_iq4nl; - - int v1, v2; - int sumi1 = 0, sumi2 = 0; + int sumi = 0; +#pragma unroll for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) { - const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16); - get_int_from_table_16(aux, values, v1, v2); - sumi1 = __dp4a(v1, q8[l+0], sumi1); - sumi2 = __dp4a(v2, q8[l+4], sumi2); + const int aux_q4 = get_int_b2(bq4->qs, iqs + l); + const int2 v = get_int_from_table_16(aux_q4); + + sumi = ggml_cuda_dp4a(v.x, q8[l + 0], sumi); + sumi = ggml_cuda_dp4a(v.y, q8[l + 4], sumi); } -#else - const uint8_t * q4 = bq->qs + 4*iqs; - const int8_t * q8 = bq8_1->qs + 4*iqs; - - int sumi1 = 0, sumi2 = 0; - for (int l = 0; l < 4*VDR_Q4_0_Q8_1_MMVQ; ++l) { - sumi1 += q8[l+ 0] * kvalues_iq4nl[q4[l] & 0xf]; - sumi2 += q8[l+16] * kvalues_iq4nl[q4[l] >> 4]; - } -#endif - const float d = (float)bq->d * __low2float(bq8_1->ds); - return d * (sumi1 + sumi2); + const float d = __half2float(bq4->d) * __low2float(bq8_1->ds); + return d * sumi; } +#define VDR_IQ4_XS_Q8_1_MMVQ 4 + static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq + kbx; - const uint8_t * values = (const uint8_t *)kvalues_iq4nl; - // iqs is 0...7 - const int ib32 = iqs; - const int32_t * q8 = (const int *)bq8_1[ib32].qs; - const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32; - const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4); - const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds); - int v1, v2; - int sumi1 = 0, sumi2 = 0; + int sumi = 0; +#pragma unroll for (int j = 0; j < 4; ++j) { - get_int_from_table_16(q4[j], values, v1, v2); - sumi1 = __dp4a(v1, q8[j+0], sumi1); - sumi2 = __dp4a(v2, q8[j+4], sumi2); + const int aux_q4 = get_int_b4(bq4->qs, iqs + j); + const int2 v = get_int_from_table_16(aux_q4); + + const int u0 = get_int_b4(bq8_1[iqs/4].qs, j + 0); + const int u1 = get_int_b4(bq8_1[iqs/4].qs, j + 4); + + sumi = ggml_cuda_dp4a(v.x, u0, sumi); + sumi = ggml_cuda_dp4a(v.y, u1, sumi); } - return d * (sumi1 + sumi2); -#else - return vec_dot_iq4_xs_q8_1(vbq, bq8_1, kbx, iqs); -#endif + + const int ls = ((bq4->scales_l[iqs/8] >> (iqs & 0x04)) & 0x0F) | (((bq4->scales_h >> (iqs/2)) & 0x03) << 4); + sumi *= ls - 32; + + const float d = __half2float(bq4->d) * __low2float(bq8_1[iqs/4].ds); + return d * sumi; } diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp index 23227649e..9b751f3c6 100644 --- a/ggml/src/ggml-sycl/mmvq.cpp +++ b/ggml/src/ggml-sycl/mmvq.cpp @@ -735,7 +735,7 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq2_xxs_q8_1( + mul_mat_vec_q_iq2_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -760,7 +760,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq2_xs_q8_1( + mul_mat_vec_q_iq2_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -785,7 +785,7 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq2_s_q8_1( + mul_mat_vec_q_iq2_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -810,7 +810,7 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq3_xxs_q8_1( + mul_mat_vec_q_iq3_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -834,7 +834,7 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq3_s_q8_1( + mul_mat_vec_q_iq3_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -924,7 +924,7 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq4_xs_q8_1( + mul_mat_vec_q_iq4_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); diff --git a/ggml/src/ggml-sycl/vecdotq.hpp b/ggml/src/ggml-sycl/vecdotq.hpp index 5e2e82546..d2dccade2 100644 --- a/ggml/src/ggml-sycl/vecdotq.hpp +++ b/ggml/src/ggml-sycl/vecdotq.hpp @@ -820,7 +820,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq, #if QK_K == 256 const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq; -#if QR2_XXS == 8 const int ib32 = iqs; const uint16_t * q2 = bq2->qs + 4*ib32; const uint8_t * aux8 = (const uint8_t *)q2; @@ -838,26 +837,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq, } const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.25f; return d * sumi; -#else - // iqs is 0...15 - const int ib32 = iqs/2; - const int il = iqs%2; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); - const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); - const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * bq8_1[ib32].ds[0] * 0.25f; - const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; - const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; - const int8_t * q8 = bq8_1[ib32].qs + 16*il; - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1); - sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1); - } - return d * (sumi1 + sumi2); -#endif #else assert(false); return 0.f; From 5fac350b9cc49d0446fc291b9c4ad53666c77591 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Tue, 2 Jul 2024 01:07:23 +0200 Subject: [PATCH 10/38] Fix gemma2 tokenizer convert (#8244) * fix gemma2 tokenizer convert * remove scores * improve code, fix new line issue --- convert-hf-to-gguf.py | 37 +++++++++++++++++++++++++++---------- 1 file changed, 27 insertions(+), 10 deletions(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 4a7f500ff..6833e9437 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -576,7 +576,19 @@ class Model: special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"]) special_vocab.add_to_gguf(self.gguf_writer) - def _set_vocab_sentencepiece(self): + def _set_vocab_sentencepiece(self, add_to_gguf=True): + tokens, scores, toktypes = self._create_vocab_sentencepiece() + + self.gguf_writer.add_tokenizer_model("llama") + self.gguf_writer.add_tokenizer_pre("default") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) + + def _create_vocab_sentencepiece(self): from sentencepiece import SentencePieceProcessor tokenizer_path = self.dir_model / 'tokenizer.model' @@ -638,14 +650,7 @@ class Model: scores.append(-1000.0) toktypes.append(SentencePieceTokenTypes.UNUSED) - self.gguf_writer.add_tokenizer_model("llama") - self.gguf_writer.add_tokenizer_pre("default") - self.gguf_writer.add_token_list(tokens) - self.gguf_writer.add_token_scores(scores) - self.gguf_writer.add_token_types(toktypes) - - special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) - special_vocab.add_to_gguf(self.gguf_writer) + return tokens, scores, toktypes def _set_vocab_llama_hf(self): vocab = gguf.LlamaHfVocab(self.dir_model) @@ -2345,7 +2350,19 @@ class Gemma2Model(Model): model_arch = gguf.MODEL_ARCH.GEMMA2 def set_vocab(self): - self._set_vocab_llama_hf() + tokens, scores, toktypes = self._create_vocab_sentencepiece() + # hack: This is required so that we can properly use start/end-of-turn for chat template + for i in range(108): + # including , , + toktypes[i] = SentencePieceTokenTypes.CONTROL + self.gguf_writer.add_tokenizer_model("llama") + self.gguf_writer.add_tokenizer_pre("default") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) self.gguf_writer.add_add_space_prefix(False) def set_gguf_parameters(self): From d08c20eddedb24515a3212e2de66bdff41a26b8c Mon Sep 17 00:00:00 2001 From: luoyu-intel Date: Tue, 2 Jul 2024 02:16:00 +0000 Subject: [PATCH 11/38] [SYCL] Fix the sub group size of Intel (#8106) * use warp_size macro for all sycl kernels * fix mask of permute_sub_group_by_xor * fix rms_norm with correct warp number * fix rms_norm_f32/group_norm_f32 * move norm to norm.cpp file * fix quantize bug * fix mmvq's batch size --- ggml/src/CMakeLists.txt | 4 +- ggml/src/ggml-sycl.cpp | 472 +++------------------------------ ggml/src/ggml-sycl/backend.hpp | 1 + ggml/src/ggml-sycl/common.hpp | 55 ++++ ggml/src/ggml-sycl/dmmv.cpp | 44 +-- ggml/src/ggml-sycl/mmvq.cpp | 113 ++++---- ggml/src/ggml-sycl/norm.cpp | 370 ++++++++++++++++++++++++++ ggml/src/ggml-sycl/norm.hpp | 35 +++ ggml/src/ggml-sycl/presets.hpp | 2 +- 9 files changed, 587 insertions(+), 509 deletions(-) create mode 100644 ggml/src/ggml-sycl/norm.cpp create mode 100644 ggml/src/ggml-sycl/norm.hpp diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index d0f4097d8..a18198f16 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -486,9 +486,11 @@ if (GGML_SYCL) add_compile_options(-I./) #include DPCT set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing") - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3") if (GGML_SYCL_TARGET STREQUAL "NVIDIA") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda") + add_compile_definitions(GGML_SYCL_WARP_SIZE=32) + else() + add_compile_definitions(GGML_SYCL_WARP_SIZE=16) endif() file(GLOB GGML_HEADERS_SYCL "ggml-sycl/*.hpp") diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index 30d8a5b33..76bad57e2 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -74,51 +74,6 @@ typedef void (*ggml_sycl_op_flatten_t)(ggml_backend_sycl_context & ctx, const gg const float *src1_dd, float *dst_dd, const queue_ptr &main_stream); -static __dpct_inline__ float warp_reduce_sum(float x, - const sycl::nd_item<3> &item_ct1) { -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - /* - DPCT1096:98: The right-most dimension of the work-group used in the SYCL - kernel that calls this function may be less than "32". The function - "dpct::permute_sub_group_by_xor" may return an unexpected result on the - CPU device. Modify the size of the work-group to ensure that the value - of the right-most dimension is a multiple of "32". - */ - x += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), x, mask); - } - return x; -} - -static __dpct_inline__ sycl::float2 -warp_reduce_sum(sycl::float2 a, const sycl::nd_item<3> &item_ct1) { -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - a.x() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.x(), - mask); - a.y() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.y(), - mask); - } - return a; -} - -static __dpct_inline__ float warp_reduce_max(float x, - const sycl::nd_item<3> &item_ct1) { -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - /* - DPCT1096:97: The right-most dimension of the work-group used in the SYCL - kernel that calls this function may be less than "32". The function - "dpct::permute_sub_group_by_xor" may return an unexpected result on the - CPU device. Modify the size of the work-group to ensure that the value - of the right-most dimension is a multiple of "32". - */ - x = sycl::fmax(x, dpct::permute_sub_group_by_xor( - item_ct1.get_sub_group(), x, mask)); - } - return x; -} - static __dpct_inline__ float op_repeat(const float a, const float b) { return b; GGML_UNUSED(a); @@ -336,47 +291,6 @@ static void sqr_f32(const float * x, float * dst, const int k, dst[i] = x[i] * x[i]; } -static void norm_f32(const float * x, float * dst, const int ncols, const float eps, - const sycl::nd_item<3> &item_ct1, sycl::float2 *s_sum, int block_size) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - const int tid = item_ct1.get_local_id(2); - - sycl::float2 mean_var = sycl::float2(0.f, 0.f); - - for (int col = tid; col < ncols; col += block_size) { - const float xi = x[row*ncols + col]; - mean_var.x() += xi; - mean_var.y() += xi * xi; - } - - // sum up partial sums - mean_var = warp_reduce_sum(mean_var, item_ct1); - if (block_size > WARP_SIZE) { - - int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = mean_var; - } - /* - DPCT1118:0: SYCL group functions and algorithms must be encountered in - converged control flow. You may need to adjust the code. - */ - item_ct1.barrier(sycl::access::fence_space::local_space); - mean_var = s_sum[lane_id]; - mean_var = warp_reduce_sum(mean_var, item_ct1); - } - - const float mean = mean_var.x() / ncols; - const float var = mean_var.y() / ncols - mean * mean; - const float inv_std = sycl::rsqrt(var + eps); - - for (int col = tid; col < ncols; col += block_size) { - dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std; - } -} - static void concat_f32(const float *x,const float *y, float *dst, const int ne0, const int ne02, const sycl::nd_item<3> &item_ct1) { int nidx = item_ct1.get_local_id(2) + @@ -444,126 +358,11 @@ static void pad_f32(const float *x, float *dst, const int ne0, const int ne00, } } -static void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps, - const sycl::nd_item<3> &item_ct1, float *s_sum, int block_size) { - int start = item_ct1.get_group(2) * group_size; - int end = start + group_size; - - start += item_ct1.get_local_id(2); - - if (end >= ne_elements) { - end = ne_elements; - } - - float tmp = 0.0f; // partial sum for thread in warp - - for (int j = start; j < end; j += block_size) { - tmp += x[j]; - } - - tmp = warp_reduce_sum(tmp, item_ct1); - if (block_size > WARP_SIZE) { - - int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = tmp; - } - /* - DPCT1118:1: SYCL group functions and algorithms must be encountered in - converged control flow. You may need to adjust the code. - */ - /* - DPCT1065:54: Consider replacing sycl::nd_item::barrier() with - sycl::nd_item::barrier(sycl::access::fence_space::local_space) for - better performance if there is no access to global memory. - */ - item_ct1.barrier(); - tmp = s_sum[lane_id]; - tmp = warp_reduce_sum(tmp, item_ct1); - } - - float mean = tmp / group_size; - tmp = 0.0f; - - for (int j = start; j < end; j += block_size) { - float xi = x[j] - mean; - dst[j] = xi; - tmp += xi * xi; - } - - tmp = warp_reduce_sum(tmp, item_ct1); - if (block_size > WARP_SIZE) { - - int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = tmp; - } - /* - DPCT1118:2: SYCL group functions and algorithms must be encountered in - converged control flow. You may need to adjust the code. - */ - /* - DPCT1065:55: Consider replacing sycl::nd_item::barrier() with - sycl::nd_item::barrier(sycl::access::fence_space::local_space) for - better performance if there is no access to global memory. - */ - item_ct1.barrier(); - tmp = s_sum[lane_id]; - tmp = warp_reduce_sum(tmp, item_ct1); - } - - float variance = tmp / group_size; - float scale = sycl::rsqrt(variance + eps); - for (int j = start; j < end; j += block_size) { - dst[j] *= scale; - } -} - -static void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps, - const sycl::nd_item<3> &item_ct1, float *s_sum, int block_size) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - const int tid = item_ct1.get_local_id(2); - - float tmp = 0.0f; // partial sum for thread in warp - - for (int col = tid; col < ncols; col += block_size) { - const float xi = x[row*ncols + col]; - tmp += xi * xi; - } - - // sum up partial sums - tmp = warp_reduce_sum(tmp, item_ct1); - if (block_size > WARP_SIZE) { - - int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = tmp; - } - /* - DPCT1118:3: SYCL group functions and algorithms must be encountered in - converged control flow. You may need to adjust the code. - */ - item_ct1.barrier(sycl::access::fence_space::local_space); - tmp = s_sum[lane_id]; - tmp = warp_reduce_sum(tmp, item_ct1); - } - - const float mean = tmp / ncols; - const float scale = sycl::rsqrt(mean + eps); - - for (int col = tid; col < ncols; col += block_size) { - dst[row*ncols + col] = scale * x[row*ncols + col]; - } -} - +template static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded, const sycl::nd_item<3> &item_ct1) { - const int ix = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); + const int ix = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2)) * QUANT_BLOCK_TILE; if (ix >= kx_padded) { return; @@ -578,23 +377,39 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int ib = i_padded / QK8_1; // block index const int iqs = i_padded % QK8_1; // quant index - - const float xi = ix < kx ? x[iy*kx + ix] : 0.0f; - float amax = sycl::fabs((float)xi); - float sum = xi; - + typedef sycl::vec TC; + typedef sycl::vec TQ; + TC zeros; + TQ qzeros; #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - amax = sycl::fmax(amax, dpct::permute_sub_group_by_xor( - item_ct1.get_sub_group(), amax, mask)); - sum += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), sum, mask); + for (int i = 0; i < QUANT_BLOCK_TILE; i++) + { + zeros[i] = 0.f; + qzeros[i] = 0; } + const TC xi = ix < kx ? *(TC *)&x[iy * kx + ix] : zeros; + float sum = xi[0]; + float amax = sycl::fabs(xi[0]); +#pragma unroll + for (int i = 1; i < QUANT_BLOCK_TILE; i++) + { + sum += xi[i]; + amax = sycl::fmax(sycl::fabs(xi[i]), amax); + } + sum = warp_reduce_sum(sum, item_ct1); + amax = warp_reduce_max(amax, item_ct1); const float d = amax / 127; - const int8_t q = amax == 0.0f ? 0 : sycl::round(xi / d); + TQ q = qzeros; + if (amax != 0.0f) + { +#pragma unroll + for (int i = 0; i < QUANT_BLOCK_TILE; i++) { + q[i] = sycl::round(xi[i] / d); + } + } - y[ib].qs[iqs] = q; + *(TQ *)&y[ib].qs[iqs] = q; if (iqs > 0) { return; @@ -728,7 +543,7 @@ static void mul_mat_p021_f16_f32( // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -781,7 +596,7 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -1643,99 +1458,6 @@ static void sqr_f32_sycl(const float *x, float *dst, const int k, }); } -static void norm_f32_sycl(const float *x, float *dst, const int ncols, - const int nrows, const float eps, - queue_ptr stream) { - GGML_ASSERT(ncols % WARP_SIZE == 0); - if (ncols < 1024) { - const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1( - sycl::range<1>(32), cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), WARP_SIZE); - }); - }); - } else { - const int work_group_size = get_work_group_size(stream->get_device()); - const sycl::range<3> block_dims(1, 1, work_group_size); - /* - DPCT1049:17: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1( - sycl::range<1>(32), cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); - }); - }); - } -} - -static void group_norm_f32_sycl(const float *x, float *dst, - const int num_groups, const int group_size, - const int ne_elements, queue_ptr stream) { - static const float eps = 1e-6f; - if (group_size < 1024) { - const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - const float eps_ct4 = eps; - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - group_norm_f32( - x, dst, group_size, ne_elements, eps_ct4, item_ct1, - s_sum_acc_ct1.get_pointer(), WARP_SIZE); - }); - }); - } else { - const int work_group_size = get_work_group_size(stream->get_device()); - const sycl::range<3> block_dims(1, 1, work_group_size); - /* - DPCT1049:18: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - const float eps_ct4 = eps; - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - group_norm_f32(x, dst, group_size, ne_elements, - eps_ct4, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); - }); - }); - } -} - static void concat_f32_sycl(const float *x, const float *y, float *dst, const int ne0, int ne1, int ne2, int ne02, queue_ptr stream) { @@ -1777,64 +1499,22 @@ static void pad_f32_sycl(const float *x, float *dst, const int ne00, }); } -static void rms_norm_f32_sycl(const float *x, float *dst, const int ncols, - const int nrows, const float eps, - queue_ptr stream) { - GGML_ASSERT(ncols % WARP_SIZE == 0); - // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); - if (ncols < 1024) { - const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - rms_norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), WARP_SIZE); - }); - }); - } else { - const int work_group_size = get_work_group_size(stream->get_device()); - const sycl::range<3> block_dims(1, 1, work_group_size); - /* - DPCT1049:19: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - rms_norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); - }); - }); - } -} - static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx, const int ky, const int kx_padded, queue_ptr stream) { const int block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE; const sycl::range<3> num_blocks(1, ky, block_num_x); - const sycl::range<3> block_size(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE); + int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE; + static_assert(QK8_1 % WARP_SIZE == 0); + const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE); { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); stream->parallel_for( sycl::nd_range<3>(num_blocks * block_size, block_size), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - quantize_q8_1(x, vy, kx, kx_padded, item_ct1); + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + quantize_q8_1(x, vy, kx, kx_padded, item_ct1); }); } } @@ -1854,7 +1534,7 @@ static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_p021_f16_f32(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y, item_ct1); }); @@ -1874,7 +1554,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl( stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_nc_f16_f32(vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y / nchannels_x, item_ct1); @@ -2139,7 +1819,7 @@ static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols, const sycl::range<3> block_nums(1, nrows, 1); stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { k_sum_rows_f32(x, dst, ncols, item_ct1); }); } @@ -2220,7 +1900,7 @@ static void soft_max_f32_submitter(const float * x, const float * mask, float * cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { soft_max_f32(x, mask, dst, ncols_par, nrows_y, scale, max_bias, m0, m1, n_head_log2, item_ct1, @@ -2400,12 +2080,6 @@ static inline int get_sycl_env(const char *env_name, int default_val) { return user_number; } -static inline int get_work_group_size(const sycl::device& device) { - dpct::device_info prop; - dpct::get_device_info(prop, device); - return prop.get_max_work_group_size(); -} - static void ggml_check_sycl() try { static bool initialized = false; @@ -2964,45 +2638,6 @@ inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor (void) src1_dd; } -inline void ggml_sycl_op_norm(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int64_t ne00 = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -inline void ggml_sycl_op_group_norm(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int num_groups = dst->op_params[0]; - int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); - group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - inline void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const float *src0_dd, const float *src1_dd, @@ -3066,28 +2701,6 @@ inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, const ggml_tensor (void) src1_dd; } -inline void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int64_t ne00 = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - static int64_t get_row_rounding(ggml_type type, const std::array & tensor_split) { int64_t min_compute_capability = INT_MAX; int64_t max_compute_capability = INT_MIN; @@ -4273,7 +3886,6 @@ bool ggml_sycl_supports_dmmv(enum ggml_type type) { static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer); - int64_t min_compute_capability = INT_MAX; if (split) { diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp index d5a63cd71..3afa33919 100644 --- a/ggml/src/ggml-sycl/backend.hpp +++ b/ggml/src/ggml-sycl/backend.hpp @@ -20,5 +20,6 @@ #include "mmq.hpp" #include "mmvq.hpp" #include "rope.hpp" +#include "norm.hpp" #endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index e01f91633..dfd4a7c2c 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -295,5 +295,60 @@ struct ggml_backend_sycl_context { } }; +// common host functions + +static inline int get_work_group_size(const sycl::device& device) { + dpct::device_info prop; + dpct::get_device_info(prop, device); + return prop.get_max_work_group_size(); +} + + +// common device functions + +static __dpct_inline__ float warp_reduce_sum(float x, + const sycl::nd_item<3>& item_ct1) { +#pragma unroll + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + /* + DPCT1096:98: The right-most dimension of the work-group used in the SYCL + kernel that calls this function may be less than "32". The function + "dpct::permute_sub_group_by_xor" may return an unexpected result on the + CPU device. Modify the size of the work-group to ensure that the value + of the right-most dimension is a multiple of "32". + */ + x += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), x, mask); + } + return x; +} + +static __dpct_inline__ sycl::float2 +warp_reduce_sum(sycl::float2 a, const sycl::nd_item<3>& item_ct1) { +#pragma unroll + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + a.x() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.x(), + mask); + a.y() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.y(), + mask); + } + return a; +} + +static __dpct_inline__ float warp_reduce_max(float x, + const sycl::nd_item<3>& item_ct1) { +#pragma unroll + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + /* + DPCT1096:97: The right-most dimension of the work-group used in the SYCL + kernel that calls this function may be less than "32". The function + "dpct::permute_sub_group_by_xor" may return an unexpected result on the + CPU device. Modify the size of the work-group to ensure that the value + of the right-most dimension is a multiple of "32". + */ + x = sycl::fmax(x, dpct::permute_sub_group_by_xor( + item_ct1.get_sub_group(), x, mask)); + } + return x; +} #endif // GGML_SYCL_COMMON_HPP diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp index 3a87d3ef8..927819281 100644 --- a/ggml/src/ggml-sycl/dmmv.cpp +++ b/ggml/src/ggml-sycl/dmmv.cpp @@ -76,7 +76,7 @@ static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -104,7 +104,7 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, nrows, item_ct1); }); @@ -227,7 +227,7 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -346,7 +346,7 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -499,7 +499,7 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -633,7 +633,7 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -748,7 +748,7 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -774,7 +774,7 @@ static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -795,7 +795,7 @@ static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -816,7 +816,7 @@ static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -837,7 +837,7 @@ static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -858,7 +858,7 @@ static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -873,10 +873,10 @@ static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y, const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2 const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, 32); + const sycl::range<3> block_dims(1, ny, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -889,10 +889,10 @@ static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y, const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, 32); + const sycl::range<3> block_dims(1, ny, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -905,10 +905,10 @@ static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y, const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, 32); + const sycl::range<3> block_dims(1, ny, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -918,10 +918,10 @@ static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y, const int nrows, dpct::queue_ptr stream) { GGML_ASSERT(ncols % QK_K == 0); - const sycl::range<3> block_dims(1, 1, 32); + const sycl::range<3> block_dims(1, 1, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1); }); } @@ -934,10 +934,10 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y, const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, 32); + const sycl::range<3> block_dims(1, ny, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); }); } diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp index 9b751f3c6..3fbc4dd60 100644 --- a/ggml/src/ggml-sycl/mmvq.cpp +++ b/ggml/src/ggml-sycl/mmvq.cpp @@ -37,7 +37,7 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_ // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -85,7 +85,7 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -133,7 +133,7 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -181,7 +181,7 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -229,7 +229,7 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -277,7 +277,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -325,7 +325,7 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -373,7 +373,7 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -421,7 +421,7 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -470,7 +470,7 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -495,7 +495,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -519,7 +519,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -543,7 +543,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -567,7 +567,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -591,7 +591,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -615,7 +615,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -639,7 +639,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -663,7 +663,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -687,7 +687,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -711,7 +711,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -734,7 +734,7 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -759,7 +759,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -784,7 +784,7 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -809,7 +809,7 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq3_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -833,7 +833,7 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq3_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -858,7 +858,7 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq1_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -879,7 +879,7 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq1_m_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -901,7 +901,7 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq4_nl_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -923,7 +923,7 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq4_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -936,7 +936,7 @@ void ggml_sycl_op_mul_mat_vec_q( const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, float *dst_dd_i, const int64_t row_low, const int64_t row_high, - const int64_t src1_ncols, const int64_t src1_padded_row_size, + const int64_t src1_ncols, const int64_t src1_padded_col_size, const dpct::queue_ptr &stream) { const int64_t ne10 = src1->ne[0]; @@ -948,77 +948,80 @@ void ggml_sycl_op_mul_mat_vec_q( int id; SYCL_CHECK( CHECK_TRY_ERROR(id = get_current_device_id())); - + const size_t q8_1_ts = sizeof(block_q8_1); + const size_t q8_1_bs = QK8_1; // the main device has a larger memory buffer to hold the results from all GPUs // nrows_dst == nrows of the matrix that the kernel writes into const int64_t nrows_dst = id == ctx.device ? ne00 : row_diff; - - switch (src0->type) { + for (int i = 0; i < src1_ncols; i++) + { + const size_t src1_ddq_i_offset = i * src1_padded_col_size * q8_1_ts / q8_1_bs; + const char* src1_ddq_i_bs = src1_ddq_i + src1_ddq_i_offset; + float* dst_dd_i_bs = dst_dd_i + i * dst->ne[0]; + switch (src0->type) { case GGML_TYPE_Q4_0: - mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q4_1: - mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q5_0: - mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q5_1: - mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q8_0: - mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q2_K: - mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q3_K: - mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q4_K: - mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q5_K: - mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q6_K: - mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ1_S: - mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ1_M: - mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ2_XXS: - mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ2_XS: - mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ2_S: - mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ3_XXS: - mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ3_S: - mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ4_NL: - mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ4_XS: - mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; default: GGML_ASSERT(false); break; + } } - (void) src1; (void) dst; (void) src1_ddf_i; - (void) src1_ncols; - (void) src1_padded_row_size; } diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp new file mode 100644 index 000000000..a77f7852c --- /dev/null +++ b/ggml/src/ggml-sycl/norm.cpp @@ -0,0 +1,370 @@ +#include "norm.hpp" + +static void norm_f32(const float* x, float* dst, const int ncols, const float eps, + const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + const int tid = item_ct1.get_local_id(2); + + const int nthreads = item_ct1.get_local_range(2); + const int nwarps = nthreads / WARP_SIZE; + assert(nwarps % WARP_SIZE == 0); + sycl::float2 mean_var = sycl::float2(0.f, 0.f); + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[row * ncols + col]; + mean_var.x() += xi; + mean_var.y() += xi * xi; + } + + // sum up partial sums + mean_var = warp_reduce_sum(mean_var, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = mean_var; + } + /* + DPCT1118:0: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + item_ct1.barrier(sycl::access::fence_space::local_space); + mean_var = 0.f; + int nreduce = nwarps / WARP_SIZE; + for (size_t i = 0; i < nreduce; i += 1) + { + mean_var += s_sum[lane_id + i * WARP_SIZE]; + } + mean_var = warp_reduce_sum(mean_var, item_ct1); + } + + const float mean = mean_var.x() / ncols; + const float var = mean_var.y() / ncols - mean * mean; + const float inv_std = sycl::rsqrt(var + eps); + + for (int col = tid; col < ncols; col += block_size) { + dst[row * ncols + col] = (x[row * ncols + col] - mean) * inv_std; + } +} + +static void group_norm_f32(const float* x, float* dst, const int group_size, const int ne_elements, const float eps, + const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) { + int start = item_ct1.get_group(2) * group_size; + int end = start + group_size; + const int nthreads = item_ct1.get_local_range(2); + const int nwarps = nthreads / WARP_SIZE; + assert(nwarps % WARP_SIZE == 0); + start += item_ct1.get_local_id(2); + + if (end >= ne_elements) { + end = ne_elements; + } + + float tmp = 0.0f; // partial sum for thread in warp + + for (int j = start; j < end; j += block_size) { + tmp += x[j]; + } + + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:1: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:54: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + tmp = 0.f; + int nreduce = nwarps / WARP_SIZE; + for (size_t i = 0; i < nreduce; i += 1) + { + tmp += s_sum[lane_id + i * WARP_SIZE]; + } + tmp = warp_reduce_sum(tmp, item_ct1); + } + + float mean = tmp / group_size; + tmp = 0.0f; + + for (int j = start; j < end; j += block_size) { + float xi = x[j] - mean; + dst[j] = xi; + tmp += xi * xi; + } + + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:2: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:55: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + tmp = s_sum[lane_id]; + tmp = warp_reduce_sum(tmp, item_ct1); + } + + float variance = tmp / group_size; + float scale = sycl::rsqrt(variance + eps); + for (int j = start; j < end; j += block_size) { + dst[j] *= scale; + } +} + +static void rms_norm_f32(const float* x, float* dst, const int ncols, const float eps, + const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + const int tid = item_ct1.get_local_id(2); + const int nthreads = item_ct1.get_local_range(2); + const int nwarps = nthreads / WARP_SIZE; + assert(nwarps % WARP_SIZE == 0); + float tmp = 0.0f; // partial sum for thread in warp + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[row * ncols + col]; + tmp += xi * xi; + } + + // sum up partial sums + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:3: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + item_ct1.barrier(sycl::access::fence_space::local_space); + int nreduce = nwarps / WARP_SIZE; + tmp = 0.f; + for (size_t i = 0; i < nreduce; i += 1) + { + tmp += s_sum[lane_id + i * WARP_SIZE]; + } + tmp = warp_reduce_sum(tmp, item_ct1); + } + + const float mean = tmp / ncols; + const float scale = sycl::rsqrt(mean + eps); + + for (int col = tid; col < ncols; col += block_size) { + dst[row * ncols + col] = scale * x[row * ncols + col]; + } +} + +static void norm_f32_sycl(const float* x, float* dst, const int ncols, + const int nrows, const float eps, + queue_ptr stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + if (ncols < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler& cgh) { + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + norm_f32(x, dst, ncols, eps, item_ct1, + nullptr, WARP_SIZE); + }); + }); + } + else { + const int work_group_size = get_work_group_size(stream->get_device()); + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:17: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor s_sum_acc_ct1( + sycl::range<1>(work_group_size / WARP_SIZE), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); + }); + }); + } +} + +static void group_norm_f32_sycl(const float* x, float* dst, + const int num_groups, const int group_size, + const int ne_elements, queue_ptr stream) { + static const float eps = 1e-6f; + if (group_size < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler& cgh) { + const float eps_ct4 = eps; + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + group_norm_f32( + x, dst, group_size, ne_elements, eps_ct4, item_ct1, + nullptr, WARP_SIZE); + }); + }); + } + else { + const int work_group_size = get_work_group_size(stream->get_device()); + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:18: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), + cgh); + + const float eps_ct4 = eps; + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + group_norm_f32(x, dst, group_size, ne_elements, + eps_ct4, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); + }); + }); + } +} + +static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, + const int nrows, const float eps, + queue_ptr stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); + if (ncols < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler& cgh) { + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + rms_norm_f32(x, dst, ncols, eps, item_ct1, + nullptr, WARP_SIZE); + }); + }); + } + else { + const int work_group_size = get_work_group_size(stream->get_device()); + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:19: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), + cgh); + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + rms_norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); + }); + }); + } +} + +void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1, + ggml_tensor* dst, const float* src0_dd, + const float* src1_dd, float* dst_dd, + const queue_ptr& main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); + + (void)src1; + (void)dst; + (void)src1_dd; +} + +void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, + const ggml_tensor* src1, ggml_tensor* dst, + const float* src0_dd, const float* src1_dd, + float* dst_dd, + const queue_ptr& main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + int num_groups = dst->op_params[0]; + int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); + group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream); + + (void)src1; + (void)dst; + (void)src1_dd; +} + +void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, + const ggml_tensor* src1, ggml_tensor* dst, + const float* src0_dd, const float* src1_dd, + float* dst_dd, + const queue_ptr& main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); + + (void)src1; + (void)dst; + (void)src1_dd; +} diff --git a/ggml/src/ggml-sycl/norm.hpp b/ggml/src/ggml-sycl/norm.hpp new file mode 100644 index 000000000..a9ad9156f --- /dev/null +++ b/ggml/src/ggml-sycl/norm.hpp @@ -0,0 +1,35 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_NORM_HPP +#define GGML_SYCL_NORM_HPP + +#include "common.hpp" + +void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1, + ggml_tensor* dst, const float* src0_dd, + const float* src1_dd, float* dst_dd, + const queue_ptr& main_stream); + +void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, + const ggml_tensor* src1, ggml_tensor* dst, + const float* src0_dd, const float* src1_dd, + float* dst_dd, + const queue_ptr& main_stream); + +void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, + const ggml_tensor* src1, ggml_tensor* dst, + const float* src0_dd, const float* src1_dd, + float* dst_dd, + const queue_ptr& main_stream); + +#endif // GGML_SYCL_NORM_HPP diff --git a/ggml/src/ggml-sycl/presets.hpp b/ggml/src/ggml-sycl/presets.hpp index fe9d41770..c09c75dc7 100644 --- a/ggml/src/ggml-sycl/presets.hpp +++ b/ggml/src/ggml-sycl/presets.hpp @@ -16,7 +16,7 @@ #define GGML_SYCL_MAX_STREAMS 8 #define GGML_SYCL_MAX_BUFFERS 256 -#define WARP_SIZE 32 +#define WARP_SIZE GGML_SYCL_WARP_SIZE #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses #define SYCL_GELU_BLOCK_SIZE 256 From a9f3b102157ba992cfe058909b7f6e1906d2d647 Mon Sep 17 00:00:00 2001 From: luoyu-intel Date: Tue, 2 Jul 2024 04:50:07 +0000 Subject: [PATCH 12/38] [SYCL] Fix win build conflict of math library (#8230) * fix win build conflict of math library * fix the condition: !(win32 & SYCL) * revert warp_size=16 --- CMakePresets.json | 1 + ggml/src/CMakeLists.txt | 6 ++++-- 2 files changed, 5 insertions(+), 2 deletions(-) diff --git a/CMakePresets.json b/CMakePresets.json index d69bc0344..bdad38952 100644 --- a/CMakePresets.json +++ b/CMakePresets.json @@ -19,6 +19,7 @@ "cacheVariables": { "CMAKE_EXPORT_COMPILE_COMMANDS": "ON", "CMAKE_CXX_COMPILER": "icx", + "CMAKE_C_COMPILER": "cl", "GGML_SYCL": "ON", "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.." } diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index a18198f16..08b71d410 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -490,7 +490,7 @@ if (GGML_SYCL) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda") add_compile_definitions(GGML_SYCL_WARP_SIZE=32) else() - add_compile_definitions(GGML_SYCL_WARP_SIZE=16) + add_compile_definitions(GGML_SYCL_WARP_SIZE=32) endif() file(GLOB GGML_HEADERS_SYCL "ggml-sycl/*.hpp") @@ -1168,7 +1168,9 @@ target_link_libraries(ggml PRIVATE Threads::Threads ${GGML_EXTRA_LIBS}) find_library(MATH_LIBRARY m) if (MATH_LIBRARY) - target_link_libraries(ggml PRIVATE ${MATH_LIBRARY}) + if (NOT WIN32 OR NOT GGML_SYCL) + target_link_libraries(ggml PRIVATE ${MATH_LIBRARY}) + endif() endif() if (BUILD_SHARED_LIBS) From 0e0590adab9f367b15ae2bf090a6d24f9df47ff1 Mon Sep 17 00:00:00 2001 From: slaren Date: Tue, 2 Jul 2024 08:39:38 +0200 Subject: [PATCH 13/38] cuda : update supports_op for matrix multiplication (#8245) --- ggml/src/ggml-cuda.cu | 47 ++++++++++++++++++++++++-------------- tests/test-backend-ops.cpp | 1 + 2 files changed, 31 insertions(+), 17 deletions(-) diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 649ef5a08..1c9ccc8a1 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -2711,27 +2711,40 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: { - struct ggml_tensor * a; - struct ggml_tensor * b; + struct ggml_tensor * a = op->src[0]; if (op->op == GGML_OP_MUL_MAT) { - a = op->src[0]; - b = op->src[1]; - } else { - a = op->src[2]; - b = op->src[1]; - } - if (a->ne[3] != b->ne[3]) { - return false; - } - ggml_type a_type = a->type; - if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS || - a_type == GGML_TYPE_IQ1_S || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S || - a_type == GGML_TYPE_IQ1_M || a_type == GGML_TYPE_IQ2_S || a_type == GGML_TYPE_IQ4_XS) { - if (b->ne[1] == 1 && ggml_nrows(b) > 1) { + struct ggml_tensor * b = op->src[1]; + if (a->ne[3] != b->ne[3]) { return false; } } - return true; + switch (a->type) { + case GGML_TYPE_F32: + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_Q8_K: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + return true; + default: + return false; + } } break; case GGML_OP_GET_ROWS: { diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index f74c0db47..2bb71ac03 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -2052,6 +2052,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M, GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS, + GGML_TYPE_BF16, }; // unary ops From 023b8807e10bc3ade24a255f01c1ad2a01bb4228 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Tue, 2 Jul 2024 08:40:49 +0200 Subject: [PATCH 14/38] convert-hf : print output file name when completed (#8181) * convert-hf : print output file name when completed This commit adds the output file name to the log message when the conversion is completed. The motivation for this change is that when `--outfile` option is not specified it migth not be obvious where the output file is written. With this change the output of running the script will be something like the following: ```console INFO:hf-to-gguf:Model successfully exported to models/gemma-2-9b-it.gguf. ``` Signed-off-by: Daniel Bevenius * squash! convert-hf : print output file name when completed Updates the output of to support printing the directory if the output is split into multiple files. Also the output file name is now retrieved from the model_instance object. Signed-off-by: Daniel Bevenius * squash! convert-hf : print output file name when completed Use parent attribute of Path object and string interpolation. Signed-off-by: Daniel Bevenius * squash! convert-hf : print output file name when completed Use os.sep instead of hardcoding the path separator. Signed-off-by: Daniel Bevenius --------- Signed-off-by: Daniel Bevenius --- convert-hf-to-gguf.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 6833e9437..05fd70171 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -3120,7 +3120,8 @@ def main() -> None: "auto": gguf.LlamaFileType.GUESSED, } - if args.use_temp_file and (args.split_max_tensors > 0 or args.split_max_size != "0"): + is_split = args.split_max_tensors > 0 or args.split_max_size != "0" + if args.use_temp_file and is_split: logger.error("Error: Cannot use temp file when splitting") sys.exit(1) @@ -3157,11 +3158,12 @@ def main() -> None: if args.vocab_only: logger.info("Exporting model vocab...") model_instance.write_vocab() - logger.info("Model vocab successfully exported.") + logger.info(f"Model vocab successfully exported to {model_instance.fname_out}") else: logger.info("Exporting model...") model_instance.write() - logger.info("Model successfully exported.") + out_path = f"{model_instance.fname_out.parent}{os.sep}" if is_split else model_instance.fname_out + logger.info(f"Model successfully exported to {out_path}") if __name__ == '__main__': From 968967376dc2c018d29f897c4883d335bbf384fb Mon Sep 17 00:00:00 2001 From: Faisal Zaghloul Date: Tue, 2 Jul 2024 10:36:00 -0400 Subject: [PATCH 15/38] Add `JAIS` model(s) (#8118) * Add `JAIS` model(s) * cleanup * address review comments * remove hack * un-hardcode max-alibi-bias * minor tweaks --------- Co-authored-by: fmz --- convert-hf-to-gguf-update.py | 1 + convert-hf-to-gguf.py | 93 ++++++++++++++++++ gguf-py/gguf/constants.py | 14 +++ gguf-py/gguf/tensor_mapping.py | 19 ++-- include/llama.h | 1 + src/llama.cpp | 169 +++++++++++++++++++++++++++++++++ 6 files changed, 288 insertions(+), 9 deletions(-) diff --git a/convert-hf-to-gguf-update.py b/convert-hf-to-gguf-update.py index 2758214fa..944e9d15a 100755 --- a/convert-hf-to-gguf-update.py +++ b/convert-hf-to-gguf-update.py @@ -86,6 +86,7 @@ models = [ {"name": "poro-chat", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", }, {"name": "jina-v2-code", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", }, {"name": "viking", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B + {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, ] diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 05fd70171..6add27cbb 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -490,6 +490,9 @@ class Model: if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee": # ref: https://huggingface.co/LumiOpen/Viking-7B res = "viking" + if chkhsh == "b53802fb28e26d645c3a310b34bfe07da813026ec7c7716883404d5e0f8b1901": + # ref: https://huggingface.co/core42/jais-13b + res = "jais" if res is None: logger.warning("\n") @@ -2965,6 +2968,96 @@ class T5Model(Model): return [(self.map_tensor_name(name), data_torch)] +@Model.register("JAISLMHeadModel") +class JaisModel(Model): + model_arch = gguf.MODEL_ARCH.JAIS + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # SwigLU activation + assert self.hparams["activation_function"] == "swiglu" + # ALiBi position embedding + assert self.hparams["position_embedding_type"] == "alibi" + + # Embeddings scale + self.embeddings_scale = 1.0 + # note: For some JAIS flavors, output is tied to (same as) wte in original model + self.output_is_wte = False + if 'mup_embeddings_scale' in self.hparams: + self.output_is_wte = True # Hack (?) + self.embeddings_scale = self.hparams['mup_embeddings_scale'] + elif 'embeddings_scale' in self.hparams: + self.embeddings_scale = self.hparams['embeddings_scale'] + else: + assert False + + self.width_scale = 1.0 + if 'mup_output_alpha' in self.hparams: + assert 'mup_width_scale' in self.hparams + self.width_scale = self.hparams['mup_output_alpha'] * self.hparams['mup_width_scale'] + elif 'width_scale' in self.hparams: + self.width_scale = self.hparams['width_scale'] + else: + assert False + + self.max_alibi_bias = 8.0 + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_block_count(self.hparams["n_layer"]) + self.gguf_writer.add_context_length(self.hparams["n_positions"]) + self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) + self.gguf_writer.add_feed_forward_length(self.hparams["n_inner"]) + self.gguf_writer.add_head_count(self.hparams["n_head"]) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + + tensors: list[tuple[str, Tensor]] = [] + + # we don't need these + if name.endswith((".attn.bias")): + return tensors + + if name.endswith(("relative_pe.slopes")): + # Calculate max ALiBi bias (this is the inverse of the ALiBi calculation) + # Some other models has max_alibi_bias spelled out explicitly in the hyperparams, + # but Jais's PyTorch model simply precalculates the slope values and places them + # in relative_pes.slopes + n_head_closest_log2 = 2 ** math.floor(math.log2(self.hparams["n_head"])) + first_val = float(data_torch._data[0]) + self.max_alibi_bias = -round(math.log2(first_val) * n_head_closest_log2) + + return tensors + + if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_fc2.weight")): + data_torch = data_torch.transpose(1, 0) + + new_name = self.map_tensor_name(name) + + if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD): + tensors.append((new_name, data_torch * self.embeddings_scale)) + if self.output_is_wte: + tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch * self.width_scale)) + elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT): + assert not self.output_is_wte + tensors.append((new_name, data_torch * self.width_scale)) + else: + tensors.append((new_name, data_torch)) + + return tensors + + def write_tensors(self): + super().write_tensors() + self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias) + + ###### CONVERSION LOGIC ###### diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index e87c58266..419f10cee 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -164,6 +164,7 @@ class MODEL_ARCH(IntEnum): DEEPSEEK2 = auto() BITNET = auto() T5 = auto() + JAIS = auto() class MODEL_TENSOR(IntEnum): @@ -288,6 +289,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.DEEPSEEK2: "deepseek2", MODEL_ARCH.BITNET: "bitnet", MODEL_ARCH.T5: "t5", + MODEL_ARCH.JAIS: "jais", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -954,6 +956,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ENC_FFN_UP, MODEL_TENSOR.ENC_OUTPUT_NORM, ], + MODEL_ARCH.JAIS: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_UP, + ], # TODO } diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 0bed43939..20e28423b 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -10,7 +10,7 @@ class TensorNameMap: # Token embeddings MODEL_TENSOR.TOKEN_EMBD: ( "gpt_neox.embed_in", # gptneox - "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx + "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais "transformer.word_embeddings", # falcon "word_embeddings", # bloom "model.embed_tokens", # llama-hf @@ -49,7 +49,7 @@ class TensorNameMap: # Output MODEL_TENSOR.OUTPUT: ( "embed_out", # gptneox - "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx + "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais "output", # llama-pth bloom internlm2 "word_embeddings_for_head", # persimmon "lm_head.linear", # phi2 @@ -58,7 +58,7 @@ class TensorNameMap: # Output norm MODEL_TENSOR.OUTPUT_NORM: ( "gpt_neox.final_layer_norm", # gptneox - "transformer.ln_f", # gpt2 gpt-j falcon + "transformer.ln_f", # gpt2 gpt-j falcon jais "model.norm", # llama-hf baichuan internlm2 "norm", # llama-pth "transformer.norm_f", # mpt dbrx @@ -81,7 +81,7 @@ class TensorNameMap: # Attention norm MODEL_TENSOR.ATTN_NORM: ( "gpt_neox.layers.{bid}.input_layernorm", # gptneox - "transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen + "transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen jais "transformer.blocks.{bid}.norm_1", # mpt "transformer.h.{bid}.input_layernorm", # falcon7b "h.{bid}.input_layernorm", # bloom @@ -109,7 +109,7 @@ class TensorNameMap: # Attention query-key-value MODEL_TENSOR.ATTN_QKV: ( "gpt_neox.layers.{bid}.attention.query_key_value", # gptneox - "transformer.h.{bid}.attn.c_attn", # gpt2 qwen + "transformer.h.{bid}.attn.c_attn", # gpt2 qwen jais "transformer.blocks.{bid}.attn.Wqkv", # mpt "transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv", # dbrx "transformer.h.{bid}.self_attention.query_key_value", # falcon @@ -160,7 +160,7 @@ class TensorNameMap: # Attention output MODEL_TENSOR.ATTN_OUT: ( "gpt_neox.layers.{bid}.attention.dense", # gptneox - "transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen + "transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen jais "transformer.blocks.{bid}.attn.out_proj", # mpt "transformer.h.{bid}.self_attention.dense", # falcon "h.{bid}.self_attention.dense", # bloom @@ -202,7 +202,7 @@ class TensorNameMap: # Feed-forward norm MODEL_TENSOR.FFN_NORM: ( "gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox - "transformer.h.{bid}.ln_2", # gpt2 refact qwen + "transformer.h.{bid}.ln_2", # gpt2 refact qwen jais "h.{bid}.post_attention_layernorm", # bloom "transformer.blocks.{bid}.norm_2", # mpt "model.layers.{bid}.post_attention_layernorm", # llama-hf @@ -239,7 +239,7 @@ class TensorNameMap: # Feed-forward up MODEL_TENSOR.FFN_UP: ( "gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox - "transformer.h.{bid}.mlp.c_fc", # gpt2 + "transformer.h.{bid}.mlp.c_fc", # gpt2 jais "transformer.blocks.{bid}.ffn.up_proj", # mpt "transformer.h.{bid}.mlp.dense_h_to_4h", # falcon "h.{bid}.mlp.dense_h_to_4h", # bloom @@ -285,6 +285,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.gate_proj", # llama-hf refact "layers.{bid}.feed_forward.w1", # llama-pth "transformer.h.{bid}.mlp.w2", # qwen + "transformer.h.{bid}.mlp.c_fc2", # jais "model.layers.layers.{bid}.mlp.gate_proj", # plamo "model.layers.{bid}.feed_forward.w1", # internlm2 "encoder.layers.{bid}.mlp.fc12", # nomic-bert @@ -308,7 +309,7 @@ class TensorNameMap: # Feed-forward down MODEL_TENSOR.FFN_DOWN: ( "gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox - "transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen + "transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen jais "transformer.blocks.{bid}.ffn.down_proj", # mpt "transformer.h.{bid}.mlp.dense_4h_to_h", # falcon "h.{bid}.mlp.dense_4h_to_h", # bloom diff --git a/include/llama.h b/include/llama.h index cafeafb85..c5b618292 100644 --- a/include/llama.h +++ b/include/llama.h @@ -89,6 +89,7 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, LLAMA_VOCAB_PRE_TYPE_PORO = 15, LLAMA_VOCAB_PRE_TYPE_VIKING = 16, + LLAMA_VOCAB_PRE_TYPE_JAIS = 17, }; // note: these values should be synchronized with ggml_rope diff --git a/src/llama.cpp b/src/llama.cpp index eea532f6a..73f52435a 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -228,6 +228,7 @@ enum llm_arch { LLM_ARCH_DEEPSEEK2, LLM_ARCH_BITNET, LLM_ARCH_T5, + LLM_ARCH_JAIS, LLM_ARCH_UNKNOWN, }; @@ -269,6 +270,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_DEEPSEEK2, "deepseek2" }, { LLM_ARCH_BITNET, "bitnet" }, { LLM_ARCH_T5, "t5" }, + { LLM_ARCH_JAIS, "jais" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -1236,6 +1238,21 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_JAIS, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -2035,6 +2052,7 @@ enum e_model { MODEL_410M, MODEL_0_5B, MODEL_1B, + MODEL_1_3B, MODEL_1_4B, MODEL_2B, MODEL_2_8B, @@ -4276,6 +4294,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_410M: return "410M"; case MODEL_0_5B: return "0.5B"; case MODEL_1B: return "1B"; + case MODEL_1_3B: return "1.3B"; case MODEL_1_4B: return "1.4B"; case MODEL_2B: return "2B"; case MODEL_2_8B: return "2.8B"; @@ -4898,6 +4917,18 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_JAIS: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias); + + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1_3B; break; + case 40: model.type = e_model::MODEL_13B; break; + /* TODO: add variants */ + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -5129,6 +5160,9 @@ static void llm_load_vocab( } else if ( tokenizer_pre == "viking") { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_VIKING; + } else if ( + tokenizer_pre == "jais") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } @@ -6962,6 +6996,44 @@ static bool llm_load_tensors( layer.ffn_up_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "scale", i), {1}); } } break; + case LLM_ARCH_JAIS: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // Output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); + layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}); + + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_gate_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}); + + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -12354,6 +12426,97 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_jais() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/float(n_embd_head), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = llm_build_norm(ctx0, inpL, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } }; static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { @@ -12585,6 +12748,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_bitnet(); } break; + case LLM_ARCH_JAIS: + { + result = llm.build_jais(); + } break; default: GGML_ASSERT(false); } @@ -13947,6 +14114,7 @@ struct llm_tokenizer_bpe { break; case LLAMA_VOCAB_PRE_TYPE_GPT2: case LLAMA_VOCAB_PRE_TYPE_OLMO: + case LLAMA_VOCAB_PRE_TYPE_JAIS: regex_exprs = { "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", }; @@ -17826,6 +17994,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_MAMBA: case LLM_ARCH_JINA_BERT_V2: case LLM_ARCH_T5: + case LLM_ARCH_JAIS: return LLAMA_ROPE_TYPE_NONE; // use what we call a normal RoPE, operating on pairs of consecutive head values From 07a3fc0608a68c0c93a5fbfa9c58f4c9ec64cb81 Mon Sep 17 00:00:00 2001 From: Clint Herron Date: Tue, 2 Jul 2024 12:18:10 -0400 Subject: [PATCH 16/38] Removes multiple newlines at the end of files that is breaking the editorconfig step of CI. (#8258) --- .github/ISSUE_TEMPLATE/config.yml | 2 -- common/common.h | 1 - examples/embedding/README.md | 1 - examples/infill/infill.cpp | 1 - examples/lookup/README.md | 1 - examples/main-cmake-pkg/.gitignore | 1 - examples/main-cmake-pkg/CMakeLists.txt | 1 - examples/server-embd.py | 1 - examples/server/tests/features/passkey.feature | 1 - examples/server/themes/buttons-top/index.html | 1 - examples/server/themes/wild/index.html | 1 - examples/sycl/run-llama2.sh | 1 - examples/sycl/win-build-sycl.bat | 1 - examples/sycl/win-run-llama2.bat | 2 -- ggml/include/ggml-metal.h | 1 - ggml/src/ggml-cuda/cpy.cu | 1 - ggml/src/ggml-metal.metal | 1 - ggml/src/ggml-quants.h | 1 - ggml/src/ggml-vulkan-shaders.hpp | 1 - scripts/pod-llama.sh | 1 - src/unicode-data.cpp | 1 - tests/test-rope.cpp | 1 - 22 files changed, 24 deletions(-) diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml index c88134dbb..eb8c4b472 100644 --- a/.github/ISSUE_TEMPLATE/config.yml +++ b/.github/ISSUE_TEMPLATE/config.yml @@ -9,5 +9,3 @@ contact_links: - name: Want to contribute? url: https://github.com/ggerganov/llama.cpp/wiki/contribute about: Head to the contribution guide page of the wiki for areas you can help with - - diff --git a/common/common.h b/common/common.h index 627b7ed85..65c0ef81a 100644 --- a/common/common.h +++ b/common/common.h @@ -459,4 +459,3 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha void yaml_dump_non_result_info( FILE * stream, const gpt_params & params, const llama_context * lctx, const std::string & timestamp, const std::vector & prompt_tokens, const char * model_desc); - diff --git a/examples/embedding/README.md b/examples/embedding/README.md index 86df18958..e3705b454 100644 --- a/examples/embedding/README.md +++ b/examples/embedding/README.md @@ -58,4 +58,3 @@ The above command will output space-separated float values. ```powershell embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null ``` - diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index ca71dd687..0e682154d 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -659,4 +659,3 @@ int main(int argc, char ** argv) { return 0; } - diff --git a/examples/lookup/README.md b/examples/lookup/README.md index 5bfb0de93..71c345c03 100644 --- a/examples/lookup/README.md +++ b/examples/lookup/README.md @@ -10,4 +10,3 @@ More info: https://github.com/ggerganov/llama.cpp/pull/4484 https://github.com/ggerganov/llama.cpp/issues/4226 - diff --git a/examples/main-cmake-pkg/.gitignore b/examples/main-cmake-pkg/.gitignore index e32c11c7f..67c01d64c 100644 --- a/examples/main-cmake-pkg/.gitignore +++ b/examples/main-cmake-pkg/.gitignore @@ -48,4 +48,3 @@ build*/ out/ tmp/ - diff --git a/examples/main-cmake-pkg/CMakeLists.txt b/examples/main-cmake-pkg/CMakeLists.txt index a97ded365..3b38db292 100644 --- a/examples/main-cmake-pkg/CMakeLists.txt +++ b/examples/main-cmake-pkg/CMakeLists.txt @@ -30,4 +30,3 @@ target_include_directories(${TARGET} PRIVATE ${_common_path}) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) - diff --git a/examples/server-embd.py b/examples/server-embd.py index 118e04271..a9a36a44c 100644 --- a/examples/server-embd.py +++ b/examples/server-embd.py @@ -31,4 +31,3 @@ for i in range(n-1): embedding2 = np.array(result[j]) similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2)) print(f"Similarity between {i} and {j}: {similarity:.2f}") - diff --git a/examples/server/tests/features/passkey.feature b/examples/server/tests/features/passkey.feature index 1bde7aab8..6a5a84e6a 100644 --- a/examples/server/tests/features/passkey.feature +++ b/examples/server/tests/features/passkey.feature @@ -52,4 +52,3 @@ Feature: Passkey / Self-extend with context shift #| TheBloke/Llama-2-7B-GGUF | llama-2-7b.Q2_K.gguf | 4096 | 3 | 16384 | 512 | 4 | 512 | 500 | 300 | 1234 | 5 | 1234 | #| TheBloke/Mixtral-8x7B-v0.1-GGUF | mixtral-8x7b-v0.1.Q2_K.gguf | 32768 | 2 | 16384 | 512 | 4 | 512 | 500 | 100 | 0987 | 5 | 0 # 987 | - diff --git a/examples/server/themes/buttons-top/index.html b/examples/server/themes/buttons-top/index.html index 6af30d307..8334bcde5 100644 --- a/examples/server/themes/buttons-top/index.html +++ b/examples/server/themes/buttons-top/index.html @@ -1054,4 +1054,3 @@ - diff --git a/examples/server/themes/wild/index.html b/examples/server/themes/wild/index.html index 772e716cd..8361c5774 100644 --- a/examples/server/themes/wild/index.html +++ b/examples/server/themes/wild/index.html @@ -1058,4 +1058,3 @@ - diff --git a/examples/sycl/run-llama2.sh b/examples/sycl/run-llama2.sh index da0e4aaba..111366fb0 100755 --- a/examples/sycl/run-llama2.sh +++ b/examples/sycl/run-llama2.sh @@ -34,4 +34,3 @@ fi #use multiple GPUs with same max compute units #ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 - diff --git a/examples/sycl/win-build-sycl.bat b/examples/sycl/win-build-sycl.bat index cdae5a528..17dd1ff5c 100644 --- a/examples/sycl/win-build-sycl.bat +++ b/examples/sycl/win-build-sycl.bat @@ -31,4 +31,3 @@ exit /B 0 :ERROR echo comomand error: %errorlevel% exit /B %errorlevel% - diff --git a/examples/sycl/win-run-llama2.bat b/examples/sycl/win-run-llama2.bat index 1d4d7d2cd..f0385cdf0 100644 --- a/examples/sycl/win-run-llama2.bat +++ b/examples/sycl/win-run-llama2.bat @@ -7,5 +7,3 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:" .\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0 - - diff --git a/ggml/include/ggml-metal.h b/ggml/include/ggml-metal.h index e7543ae79..6c3226c37 100644 --- a/ggml/include/ggml-metal.h +++ b/ggml/include/ggml-metal.h @@ -63,4 +63,3 @@ GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend); #ifdef __cplusplus } #endif - diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu index 12d741f01..3db57034b 100644 --- a/ggml/src/ggml-cuda/cpy.cu +++ b/ggml/src/ggml-cuda/cpy.cu @@ -487,4 +487,3 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) { GGML_ASSERT(false); } } - diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal index e2796fd60..c3503479b 100644 --- a/ggml/src/ggml-metal.metal +++ b/ggml/src/ggml-metal.metal @@ -6537,4 +6537,3 @@ template [[host_name("kernel_mul_mv_id_iq3_s_f32")]] kernel kernel_mul_mv_id_t template [[host_name("kernel_mul_mv_id_iq2_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; - diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h index 4d436a8f0..30983b872 100644 --- a/ggml/src/ggml-quants.h +++ b/ggml/src/ggml-quants.h @@ -130,4 +130,3 @@ void iq3xs_free_impl(int grid_size); #ifdef __cplusplus } #endif - diff --git a/ggml/src/ggml-vulkan-shaders.hpp b/ggml/src/ggml-vulkan-shaders.hpp index 01ff66f71..f0c4c6baf 100644 --- a/ggml/src/ggml-vulkan-shaders.hpp +++ b/ggml/src/ggml-vulkan-shaders.hpp @@ -144954,4 +144954,3 @@ unsigned char sum_rows_f32_data[] = { }; const uint64_t sum_rows_f32_len = 2112; - diff --git a/scripts/pod-llama.sh b/scripts/pod-llama.sh index 586d6ea18..0d6d4032d 100644 --- a/scripts/pod-llama.sh +++ b/scripts/pod-llama.sh @@ -210,4 +210,3 @@ fi # more benches #GGML_CUDA=1 make -j && ./llama-batched-bench ./models/codellama-7b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1 #GGML_CUDA=1 make -j && ./llama-batched-bench ./models/codellama-13b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1 - diff --git a/src/unicode-data.cpp b/src/unicode-data.cpp index 4a939898b..02bdf7823 100644 --- a/src/unicode-data.cpp +++ b/src/unicode-data.cpp @@ -7030,4 +7030,3 @@ const std::vector unicode_ranges_nfd = { // start, last, nfd {0x02FA1C, 0x02FA1C, 0x009F3B}, {0x02FA1D, 0x02FA1D, 0x02A600}, }; - diff --git a/tests/test-rope.cpp b/tests/test-rope.cpp index f0895ffaa..8159e276a 100644 --- a/tests/test-rope.cpp +++ b/tests/test-rope.cpp @@ -218,4 +218,3 @@ int main(int /*argc*/, const char ** /*argv*/) { return 0; } - From 3e2618bc7bf9e9fbf58c32cc3c8dd7d5df1de27e Mon Sep 17 00:00:00 2001 From: Clint Herron Date: Tue, 2 Jul 2024 13:19:56 -0400 Subject: [PATCH 17/38] Adding step to `clean` target to remove legacy binary names to reduce upgrade / migration confusion arising from #7809. (#8257) --- Makefile | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/Makefile b/Makefile index 8ae4f1dc4..2730d8b60 100644 --- a/Makefile +++ b/Makefile @@ -62,6 +62,11 @@ TEST_TARGETS = \ tests/test-tokenizer-1-bpe \ tests/test-tokenizer-1-spm +# Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned +LEGACY_TARGETS = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ + simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \ + retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm + # Deprecation aliases ifdef LLAMA_CUBLAS $(error LLAMA_CUBLAS is removed. Use GGML_CUDA instead.) @@ -1086,6 +1091,7 @@ clean: rm -vrf ggml/src/ggml-cuda/template-instances/*.o rm -rvf $(BUILD_TARGETS) rm -rvf $(TEST_TARGETS) + rm -rvf $(LEGACY_TARGETS) find examples pocs -type f -name "*.o" -delete # From a27152b602b369e76f85b7cb7b872a321b7218f7 Mon Sep 17 00:00:00 2001 From: MistApproach <98988043+MistApproach@users.noreply.github.com> Date: Tue, 2 Jul 2024 22:56:46 +0200 Subject: [PATCH 18/38] fix: add missing short command line argument -mli for multiline-input (#8261) --- common/common.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/common/common.cpp b/common/common.cpp index 5a0d0ee03..2c05a4d4a 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -757,7 +757,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.cache_type_v = argv[++i]; return true; } - if (arg == "--multiline-input") { + if (arg == "-mli" || arg == "--multiline-input") { params.multiline_input = true; return true; } From fadde6713506d9e6c124f5680ab8c7abebe31837 Mon Sep 17 00:00:00 2001 From: AidanBeltonS Date: Wed, 3 Jul 2024 02:55:34 +0100 Subject: [PATCH 19/38] Dequant improvements rebase (#8255) * Single load for half2 * Store scales in local mem * Vec load quantized values --- ggml/src/ggml-sycl/common.hpp | 6 ++++++ ggml/src/ggml-sycl/convert.cpp | 7 +++++-- ggml/src/ggml-sycl/dequantize.hpp | 30 +++++++++++++++++++----------- 3 files changed, 30 insertions(+), 13 deletions(-) diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index dfd4a7c2c..476d847ca 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -351,4 +351,10 @@ static __dpct_inline__ float warp_reduce_max(float x, return x; } +// Helper for vec loading aligned data +template +inline sycl::vec vec_aligned_load(const Tp* aligned_ptr) { + return *reinterpret_cast*>(aligned_ptr); +} + #endif // GGML_SYCL_COMMON_HPP diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp index ce9de2b42..a15271b51 100644 --- a/ggml/src/ggml-sycl/convert.cpp +++ b/ggml/src/ggml-sycl/convert.cpp @@ -152,12 +152,15 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor scale_local_acc(sycl::range<1>(12), cgh); + cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_K(vx, y, item_ct1); + dequantize_block_q4_K(vx, y, scale_local_acc.get_pointer(), item_ct1); }); + }); } } diff --git a/ggml/src/ggml-sycl/dequantize.hpp b/ggml/src/ggml-sycl/dequantize.hpp index b6080d83a..ed8ad098b 100644 --- a/ggml/src/ggml-sycl/dequantize.hpp +++ b/ggml/src/ggml-sycl/dequantize.hpp @@ -293,7 +293,8 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri #if QK_K == 256 static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { if (j < 4) { - d = q[j] & 63; m = q[j + 4] & 63; + d = q[j] & 63; + m = q[j + 4] & 63; } else { d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4); m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); @@ -303,7 +304,7 @@ static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8 template static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1) { + uint8_t* scales_local, const sycl::nd_item<3> &item_ct1) { const block_q4_K * x = (const block_q4_K *) vx; const int i = item_ct1.get_group(2); @@ -318,19 +319,26 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri dst_t * y = yy + i*QK_K + 64*il + n*ir; - const float dall = x[i].dm[0]; - const float dmin = x[i].dm[1]; + const sycl::half2 dm = x[i].dm; + const float dall = dm[0]; + const float dmin = dm[1]; - const uint8_t * q = x[i].qs + 32*il + n*ir; + if (tid < 12) + scales_local[tid] = x[i].scales[tid]; + item_ct1.barrier(sycl::access::fence_space::local_space); uint8_t sc, m; - get_scale_min_k4(is + 0, x[i].scales, sc, m); - const float d1 = dall * sc; const float m1 = dmin * m; - get_scale_min_k4(is + 1, x[i].scales, sc, m); - const float d2 = dall * sc; const float m2 = dmin * m; + get_scale_min_k4(is + 0, scales_local, sc, m); + const float d1 = dall * sc; + const float m1 = dmin * m; + get_scale_min_k4(is + 1, scales_local, sc, m); + const float d2 = dall * sc; + const float m2 = dmin * m; + + sycl::vec q_vec = vec_aligned_load(x[i].qs + 32*il + n*ir); for (int l = 0; l < n; ++l) { - y[l + 0] = d1 * (q[l] & 0xF) - m1; - y[l +32] = d2 * (q[l] >> 4) - m2; + y[l + 0] = d1 * (q_vec[l] & 0xF) - m1; + y[l +32] = d2 * (q_vec[l] >> 4) - m2; } #else const int tid = item_ct1.get_local_id(2); From f8d6a23804f3798ff2869da68c1223b618df09ec Mon Sep 17 00:00:00 2001 From: Judd Date: Wed, 3 Jul 2024 20:40:16 +0800 Subject: [PATCH 20/38] fix typo (#8267) Co-authored-by: Judd --- ggml/src/ggml.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index f5502afbe..bc91ac3a7 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -5312,7 +5312,7 @@ void ggml_mul_mat_set_prec( as -> [cols, rows, n_expert] ids -> [n_experts_used, n_tokens] (i32) b -> [cols, n_expert_used, n_tokens] - c -> [cols, n_expert_used, n_tokens] + c -> [rows, n_expert_used, n_tokens] in b, n_experts_used can be broadcasted to match the n_expert_used of ids From 916248af1f3c16abd7408de848e025da095c621c Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Wed, 3 Jul 2024 16:01:54 +0200 Subject: [PATCH 21/38] fix phi 3 conversion (#8262) --- convert-hf-to-gguf.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 6add27cbb..d01aed224 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -1942,7 +1942,7 @@ class Phi3MiniModel(Model): if len(rope_scaling_type) == 0: raise KeyError('Missing the required key rope_scaling.type') - if rope_scaling_type == 'su': + if rope_scaling_type == 'su' or rope_scaling_type == 'longrope': attn_factor = math.sqrt(1 + math.log(scale) / math.log(orig_max_pos_embds)) if scale > 1.0 else 1.0 elif rope_scaling_type == 'yarn': attn_factor = 0.1 * math.log(scale) + 1.0 if scale > 1.0 else 1.0 From 5f2d4e60e202aabee10051e6615bb821e51787be Mon Sep 17 00:00:00 2001 From: slaren Date: Wed, 3 Jul 2024 19:33:31 +0200 Subject: [PATCH 22/38] ppl : fix n_seq_max for perplexity (#8277) * ppl : fix n_seq_max for perplexity * use 1 seq for kl_divergence --- examples/perplexity/perplexity.cpp | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index efde8dfdf..dbe445391 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1991,6 +1991,12 @@ int main(int argc, char ** argv) { params.n_batch = std::min(params.n_batch, n_kv); } else { params.n_batch = std::min(params.n_batch, params.n_ctx); + if (params.kl_divergence) { + params.n_parallel = 1; + } else { + // ensure there's at least enough seq_ids for HellaSwag + params.n_parallel = std::max(4, params.n_parallel); + } } if (params.ppl_stride > 0) { @@ -2015,9 +2021,6 @@ int main(int argc, char ** argv) { llama_model * model; llama_context * ctx; - // ensure there's at least enough seq_ids for HellaSwag - params.n_parallel = std::max(4, params.n_parallel); - // load the model and apply lora adapter, if any std::tie(model, ctx) = llama_init_from_gpt_params(params); if (model == NULL) { From d23287f122c34ebef368742116d53a0ccb2041ee Mon Sep 17 00:00:00 2001 From: Daniele <57776841+daniandtheweb@users.noreply.github.com> Date: Wed, 3 Jul 2024 23:02:58 +0000 Subject: [PATCH 23/38] Define and optimize RDNA1 (#8085) --- ggml/src/ggml-cuda/common.cuh | 4 ++++ ggml/src/ggml-cuda/mmq.cuh | 10 +++++++--- 2 files changed, 11 insertions(+), 3 deletions(-) diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 472f4ace1..4ff06b871 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -227,6 +227,10 @@ typedef float2 dfloat2; #define RDNA2 #endif +#if defined(__gfx1010__) || defined(__gfx1012__) +#define RDNA1 +#endif + #ifndef __has_builtin #define __has_builtin(x) 0 #endif diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index 1396e7a75..deaed066f 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -60,12 +60,16 @@ static constexpr __device__ int get_mmq_x_max_device() { } static constexpr int get_mmq_y_host(const int cc) { - return int8_mma_available(cc) || cc >= CC_VOLTA ? 128 : 64; + return cc >= CC_OFFSET_AMD ? (cc == CC_RDNA1 ? 64 : 128) : (cc >= CC_VOLTA ? 128 : 64); } static constexpr __device__ int get_mmq_y_device() { #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA1) + return 64; +#else return 128; +#endif // defined RDNA1 #else #if __CUDA_ARCH__ >= CC_VOLTA return 128; @@ -2259,9 +2263,9 @@ static __device__ void mul_mat_q_process_tile( template #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) +#if defined(RDNA3) || defined(RDNA2) || defined(RDNA1) __launch_bounds__(WARP_SIZE*nwarps, 2) -#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(RDNA3) || defined(RDNA2) || defined(RDNA1) #else #if __CUDA_ARCH__ >= CC_VOLTA __launch_bounds__(WARP_SIZE*nwarps, 1) From f619024764e72261f14d7c31d892b8fb976603b4 Mon Sep 17 00:00:00 2001 From: AidanBeltonS Date: Thu, 4 Jul 2024 02:07:19 +0100 Subject: [PATCH 24/38] [SYCL] Remove unneeded semicolons (#8280) --- ggml/src/ggml-sycl/dpct/helper.hpp | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp index 1ff297218..5e98660dc 100644 --- a/ggml/src/ggml-sycl/dpct/helper.hpp +++ b/ggml/src/ggml-sycl/dpct/helper.hpp @@ -255,7 +255,7 @@ namespace dpct void set_pitch(size_t pitch) { _pitch = pitch; } size_t get_x() { return _x; } - void set_x(size_t x) { _x = x; }; + void set_x(size_t x) { _x = x; } size_t get_y() { return _y; } void set_y(size_t y) { _y = y; } @@ -1056,7 +1056,7 @@ namespace dpct #error "Only support Windows and Linux." #endif next_free = mapped_address_space; - }; + } public: using buffer_id_t = int; @@ -1077,7 +1077,7 @@ namespace dpct #else #error "Only support Windows and Linux." #endif - }; + } mem_mgr(const mem_mgr &) = delete; mem_mgr &operator=(const mem_mgr &) = delete; From 20fc3804bfb727074bc270b6eacb60af8d0bf7d4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 4 Jul 2024 10:41:03 +0300 Subject: [PATCH 25/38] convert : fix gemma v1 tokenizer convert (#8248) ggml-ci --- convert-hf-to-gguf-update.py | 5 ++++- convert-hf-to-gguf.py | 3 +++ models/ggml-vocab-bert-bge.gguf.inp | 4 ++++ models/ggml-vocab-bert-bge.gguf.out | 2 ++ models/ggml-vocab-command-r.gguf.inp | 4 ++++ models/ggml-vocab-command-r.gguf.out | 2 ++ models/ggml-vocab-deepseek-coder.gguf.inp | 4 ++++ models/ggml-vocab-deepseek-coder.gguf.out | 2 ++ models/ggml-vocab-deepseek-llm.gguf.inp | 4 ++++ models/ggml-vocab-deepseek-llm.gguf.out | 2 ++ models/ggml-vocab-falcon.gguf.inp | 4 ++++ models/ggml-vocab-falcon.gguf.out | 2 ++ models/ggml-vocab-gpt-2.gguf.inp | 4 ++++ models/ggml-vocab-gpt-2.gguf.out | 2 ++ models/ggml-vocab-llama-bpe.gguf.inp | 6 ++++-- models/ggml-vocab-llama-bpe.gguf.out | 3 ++- models/ggml-vocab-llama-spm.gguf.inp | 4 ++++ models/ggml-vocab-llama-spm.gguf.out | 2 ++ models/ggml-vocab-mpt.gguf.inp | 4 ++++ models/ggml-vocab-mpt.gguf.out | 2 ++ models/ggml-vocab-phi-3.gguf.inp | 4 ++++ models/ggml-vocab-phi-3.gguf.out | 2 ++ models/ggml-vocab-qwen2.gguf.inp | 4 ++++ models/ggml-vocab-qwen2.gguf.out | 2 ++ models/ggml-vocab-refact.gguf.inp | 4 ++++ models/ggml-vocab-refact.gguf.out | 2 ++ models/ggml-vocab-starcoder.gguf.inp | 4 ++++ models/ggml-vocab-starcoder.gguf.out | 2 ++ 28 files changed, 85 insertions(+), 4 deletions(-) diff --git a/convert-hf-to-gguf-update.py b/convert-hf-to-gguf-update.py index 944e9d15a..9eb406cb8 100755 --- a/convert-hf-to-gguf-update.py +++ b/convert-hf-to-gguf-update.py @@ -86,6 +86,8 @@ models = [ {"name": "poro-chat", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", }, {"name": "jina-v2-code", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", }, {"name": "viking", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B + {"name": "gemma", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", }, + {"name": "gemma-2", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", }, {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, ] @@ -273,7 +275,8 @@ tests = [ "3333333", "33333333", "333333333", - # "Cửa Việt", # llama-bpe fails on this + "Cửa Việt", # llama-bpe fails on this + " discards", chktxt, ] diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index d01aed224..bae145589 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -2316,6 +2316,8 @@ class GemmaModel(Model): special_vocab._set_special_token("eot", 107) special_vocab.add_to_gguf(self.gguf_writer) + self.gguf_writer.add_add_space_prefix(False) + def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] @@ -2366,6 +2368,7 @@ class Gemma2Model(Model): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) + self.gguf_writer.add_add_space_prefix(False) def set_gguf_parameters(self): diff --git a/models/ggml-vocab-bert-bge.gguf.inp b/models/ggml-vocab-bert-bge.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-bert-bge.gguf.inp +++ b/models/ggml-vocab-bert-bge.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-bert-bge.gguf.out b/models/ggml-vocab-bert-bge.gguf.out index e4a76cdb0..82d4ed1c1 100644 --- a/models/ggml-vocab-bert-bge.gguf.out +++ b/models/ggml-vocab-bert-bge.gguf.out @@ -40,4 +40,6 @@ 21211 22394 22394 21211 22394 22394 2509 21211 22394 22394 22394 + 12731 2050 19710 + 5860 18117 100 1006 3671 1007 100 1006 3674 7861 29147 2483 9530 16280 23854 1007 100 100 1017 3943 21211 21211 2509 21211 22394 21211 22394 2509 21211 22394 22394 21211 22394 22394 2509 1017 1012 1017 1017 1012 1012 1017 1017 1012 1012 1012 1017 100 1029 1855 100 100 6207 100 100 14677 23632 22203 1811 1995 1011 1011 1011 1011 1011 1011 1027 1027 1027 1027 1027 1027 1027 1192 15290 29754 14150 1192 10260 1181 29755 29436 29741 10260 16856 29747 23925 10325 1005 1005 1005 1005 1005 1005 1036 1036 1036 1036 1036 1036 1036 1000 1000 1000 1000 1012 1012 1012 1012 1012 1012 999 999 999 999 999 999 1029 1029 1029 1029 1029 1029 1045 1005 2310 2042 1005 2409 2002 1005 1055 2045 1010 1005 2128 2017 2469 1029 1005 1049 2025 2469 1045 1005 2222 2191 2009 1010 1005 1040 2017 2066 2070 5572 1029 2057 1005 2310 1037 1005 2222 diff --git a/models/ggml-vocab-command-r.gguf.inp b/models/ggml-vocab-command-r.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-command-r.gguf.inp +++ b/models/ggml-vocab-command-r.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-command-r.gguf.out b/models/ggml-vocab-command-r.gguf.out index cc4277daa..939b9dc30 100644 --- a/models/ggml-vocab-command-r.gguf.out +++ b/models/ggml-vocab-command-r.gguf.out @@ -40,4 +40,6 @@ 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 + 42 30719 12584 + 3642 4388 127731 51628 205 57788 18494 97469 126134 206 2226 256 230 1737 18258 16 80503 122 35927 2226 242 112 57462 1737 54457 223165 106230 2096 16 48389 11254 107 255 2226 107 255 228 26 228 26 26 228 26 26 26 228 26 26 26 26 228 26 26 26 26 26 228 26 26 26 26 26 26 228 26 26 26 26 26 26 26 228 26 26 26 26 26 26 26 26 228 26 21 26 228 26 2271 26 228 26 3834 26 182018 230 174833 38111 249 86325 241 38111 245 86325 232 38111 252 38111 123 38111 261 165 24629 38111 261 38111 103 174833 38111 235 188568 231 5691 12081 13336 2648 29325 14315 24 26 24 27 24 28 24 5123 18372 8391 158343 3512 40071 2196 3236 8750 1764 37097 41168 29721 32797 25646 3802 4975 4975 116167 57178 10251 154048 27292 1767 5125 2632 2155 91 2378 1919 1914 2782 19 2155 3354 1933 5470 38 2155 52 2068 5470 1767 4961 3059 1894 19 2155 43 1933 3026 2725 23186 38 2930 14 20676 1671 14 83 51 diff --git a/models/ggml-vocab-deepseek-coder.gguf.inp b/models/ggml-vocab-deepseek-coder.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-deepseek-coder.gguf.inp +++ b/models/ggml-vocab-deepseek-coder.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-deepseek-coder.gguf.out b/models/ggml-vocab-deepseek-coder.gguf.out index 9ccc560d6..a43e3f0f1 100644 --- a/models/ggml-vocab-deepseek-coder.gguf.out +++ b/models/ggml-vocab-deepseek-coder.gguf.out @@ -40,4 +40,6 @@ 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 + 34 155 119 242 64 24297 155 119 216 83 + 1607 2539 185 207 185 185 207 185 185 185 207 12405 459 22758 185 243 185 315 185 251 185 730 185 10047 235 209 334 8760 8 12394 233 114 350 222 10047 221 104 169 116 224 334 4684 3909 992 24330 262 29651 612 8 207 156 237 214 12394 99 234 10047 99 234 207 18 207 18 18 207 18 18 18 207 18 18 18 18 207 18 18 18 18 18 207 18 18 18 18 18 18 207 18 18 18 18 18 18 18 207 18 18 18 18 18 18 18 18 207 18 13 18 207 18 524 18 207 18 1202 18 207 155 239 209 155 239 114 155 239 228 155 240 220 155 239 224 155 240 211 155 239 231 155 239 115 155 239 240 155 240 210 155 239 240 155 239 95 155 239 114 155 239 214 10047 233 210 3015 19100 608 9413 2668 16 18 16 19 16 20 16 1393 169 121 239 18155 374 17194 28 2861 6478 616 2251 14994 31269 4191 6 4686 4686 10252 3358 3358 3409 524 15330 3023 15031 5668 303 6 312 798 651 83 839 362 6 82 741 11 651 1369 340 2037 30 651 44 441 2037 303 6 642 1098 359 11 651 35 340 833 738 10860 30 998 6 10709 245 6 75 43 diff --git a/models/ggml-vocab-deepseek-llm.gguf.inp b/models/ggml-vocab-deepseek-llm.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-deepseek-llm.gguf.inp +++ b/models/ggml-vocab-deepseek-llm.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-deepseek-llm.gguf.out b/models/ggml-vocab-deepseek-llm.gguf.out index fd94b896d..d31ac1cc6 100644 --- a/models/ggml-vocab-deepseek-llm.gguf.out +++ b/models/ggml-vocab-deepseek-llm.gguf.out @@ -40,4 +40,6 @@ 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 + 34 32555 242 64 23708 32555 216 83 + 1763 2550 185 207 185 185 207 185 185 185 207 11969 486 22504 185 243 185 300 185 251 185 663 185 10044 95300 334 8754 8 33701 114 350 222 10044 221 104 46713 334 34732 996 24250 262 80923 8 207 37103 214 12356 99 234 10044 99 234 207 18 207 18 18 207 18 18 18 207 18 18 18 18 207 18 18 18 18 18 207 18 18 18 18 18 18 207 18 18 18 18 18 18 18 207 18 18 18 18 18 18 18 18 207 18 13 18 207 18 526 18 207 18 1204 18 207 71374 209 71374 114 71374 228 155 240 220 71374 224 155 240 211 71374 231 71374 115 71374 240 155 240 210 71374 240 71374 95 71374 114 71374 214 71899 210 3025 19017 612 9407 2681 16 18 16 19 16 20 16 1398 68940 239 78827 55170 76659 620 91754 31116 36804 4885 4885 10897 4390 4390 41047 15278 3033 14986 5675 304 6 313 803 655 33326 362 6 82 745 11 655 1374 340 2049 30 655 44 441 2049 304 6 647 1099 359 11 655 35 340 837 742 10842 30 1003 6 10699 245 6 75 43 diff --git a/models/ggml-vocab-falcon.gguf.inp b/models/ggml-vocab-falcon.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-falcon.gguf.inp +++ b/models/ggml-vocab-falcon.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-falcon.gguf.out b/models/ggml-vocab-falcon.gguf.out index 209b04cda..1ab70fe70 100644 --- a/models/ggml-vocab-falcon.gguf.out +++ b/models/ggml-vocab-falcon.gguf.out @@ -40,4 +40,6 @@ 22287 22287 30 22287 22287 3138 22287 22287 22287 + 46 19768 239 76 9634 19768 213 95 + 1080 1502 1212 4824 1001 1212 192 204 663 49453 2069 742 561 1501 193 2571 232 206 204 19 11003 20 8196 126 283 219 48778 116 13392 204 19 51831 732 63209 1741 7955 522 20 22438 211 3346 111 231 2571 111 231 204 30 204 3138 204 22287 204 22287 30 204 22287 3138 204 22287 22287 204 22287 22287 30 204 22287 22287 3138 204 30 25 30 204 30 513 30 204 30 951 30 27171 236 206 38154 126 38154 225 167 237 217 38154 221 167 237 208 38154 228 38154 127 38154 237 167 237 207 38154 237 38154 107 38154 126 38154 211 20589 207 204 42 50087 123 2727 20300 32022 133 234 17419 30137 28 7858 181 133 236 204 37057 2228 10666 5052 133 6207 151 215 150 134 5052 133 6279 5052 223 151 216 49679 123 53110 47043 7795 204 7544 7544 7544 8543 8543 17593 3513 3513 12844 51520 17664 4247 295 18 298 650 204 18 95 693 332 18 94 629 23 204 18 1553 299 1310 42 204 18 56 416 1310 295 18 567 717 334 23 204 18 47 299 606 596 6696 42 703 18 16139 241 18 87 55 diff --git a/models/ggml-vocab-gpt-2.gguf.inp b/models/ggml-vocab-gpt-2.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-gpt-2.gguf.inp +++ b/models/ggml-vocab-gpt-2.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-gpt-2.gguf.out b/models/ggml-vocab-gpt-2.gguf.out index 78430f0d3..88217d3fa 100644 --- a/models/ggml-vocab-gpt-2.gguf.out +++ b/models/ggml-vocab-gpt-2.gguf.out @@ -40,4 +40,6 @@ 24840 20370 24840 24840 24840 2091 20370 + 34 157 119 255 64 16049 157 119 229 83 + 1221 1371 198 220 628 220 628 198 220 197 220 197 197 220 197 198 220 220 198 220 220 220 198 220 220 220 220 198 220 220 220 220 220 198 8582 248 222 357 11265 8 30325 114 447 235 8582 234 104 37929 357 48101 795 13210 271 1673 36686 515 8 14519 227 12520 99 247 8582 99 247 513 4747 23460 513 20370 23460 2091 23460 20370 23460 24840 23460 2091 20370 513 13 18 513 492 18 513 986 18 28053 252 222 157 252 114 157 252 241 157 253 233 157 252 237 157 253 224 157 252 244 157 252 115 157 252 253 157 253 223 157 252 253 157 252 95 157 252 114 157 252 227 47249 223 5633 22755 239 46349 111 28839 101 18040 32432 98 43291 1485 1415 24309 25465 171 121 252 40103 1421 18604 12466 121 16843 141 231 15166 12466 121 16142 12466 239 141 232 30143 140 111 16142 21169 21727 31583 18849 705 39115 6 33153 15506 63 15931 15931 16317 13896 3228 9805 3548 314 1053 587 705 44040 339 338 612 11 705 2200 345 1654 30 705 44 407 1654 314 1183 787 340 11 705 35 345 588 617 8887 30 775 6 26979 257 6 75 43 diff --git a/models/ggml-vocab-llama-bpe.gguf.inp b/models/ggml-vocab-llama-bpe.gguf.inp index 9380bf355..5b4aeb31a 100644 --- a/models/ggml-vocab-llama-bpe.gguf.inp +++ b/models/ggml-vocab-llama-bpe.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ @@ -104,5 +108,3 @@ __ggml_vocab_test__ 🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL __ggml_vocab_test__ - Việt -__ggml_vocab_test__ diff --git a/models/ggml-vocab-llama-bpe.gguf.out b/models/ggml-vocab-llama-bpe.gguf.out index 1f3607fb6..bb1fe229c 100644 --- a/models/ggml-vocab-llama-bpe.gguf.out +++ b/models/ggml-vocab-llama-bpe.gguf.out @@ -40,5 +40,6 @@ 8765 8765 18 8765 8765 1644 8765 8765 8765 + 34 91163 101798 + 2624 2402 198 4815 15073 66597 8004 1602 2355 79772 11187 9468 248 222 320 8416 8 27623 114 102470 9468 234 104 31643 320 36773 100166 98634 8 26602 227 11410 99 247 9468 99 247 220 18 220 1644 220 8765 220 8765 18 220 8765 1644 220 8765 8765 220 8765 8765 18 220 8765 8765 1644 220 18 13 18 220 18 497 18 220 18 1131 18 220 21549 222 98629 241 45358 233 21549 237 45358 224 21549 244 21549 115 21549 253 45358 223 21549 253 21549 95 98629 227 76460 223 949 37046 101067 19000 23182 102301 9263 18136 16 36827 21909 56560 54337 19175 102118 13373 64571 34694 3114 112203 80112 3436 106451 14196 14196 74694 3089 3089 29249 17523 3001 27708 7801 358 3077 1027 364 83 820 568 596 1070 11 364 793 499 2771 30 364 44 539 2771 358 3358 1304 433 11 364 35 499 1093 1063 15600 30 1226 6 43712 264 64966 43 - 101798 diff --git a/models/ggml-vocab-llama-spm.gguf.inp b/models/ggml-vocab-llama-spm.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-llama-spm.gguf.inp +++ b/models/ggml-vocab-llama-spm.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-llama-spm.gguf.out b/models/ggml-vocab-llama-spm.gguf.out index 9c3327cb5..1c3b0a2c9 100644 --- a/models/ggml-vocab-llama-spm.gguf.out +++ b/models/ggml-vocab-llama-spm.gguf.out @@ -40,4 +40,6 @@ 29871 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29941 + 315 228 190 176 29874 10630 30529 29873 + 29871 2313 3163 29871 13 29871 13 13 29871 13 13 13 29871 12 29871 12 12 29871 12 13 259 13 1678 13 268 13 418 13 243 162 157 131 313 8945 29897 29871 243 162 155 185 30722 243 162 143 174 30598 313 20787 953 3848 275 16125 630 29897 29871 31681 29871 243 162 169 156 243 162 169 156 29871 29941 29871 29941 29941 29871 29941 29941 29941 29871 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29871 29941 29889 29941 29871 29941 636 29941 29871 29941 856 29941 29871 31849 31324 31934 228 162 142 228 161 146 228 162 133 228 161 153 228 161 186 31708 228 162 132 31708 228 161 165 31324 228 161 136 243 162 155 132 1577 30672 31522 30505 11548 31041 30732 29896 29941 29896 29946 29896 29945 29896 30408 30739 448 23648 2751 25512 1538 4851 665 1386 29713 1305 14550 4907 11120 16159 16159 16159 15945 15945 3045 636 6824 6824 6824 8773 8773 8773 306 29915 345 1063 525 29873 1025 540 29915 29879 727 29892 525 1525 366 1854 29973 525 29924 451 1854 306 29915 645 1207 372 29892 525 29928 366 763 777 23429 29973 1334 29915 29963 29872 263 29915 29880 29931 diff --git a/models/ggml-vocab-mpt.gguf.inp b/models/ggml-vocab-mpt.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-mpt.gguf.inp +++ b/models/ggml-vocab-mpt.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-mpt.gguf.out b/models/ggml-vocab-mpt.gguf.out index d8d0fe909..d4a877d1c 100644 --- a/models/ggml-vocab-mpt.gguf.out +++ b/models/ggml-vocab-mpt.gguf.out @@ -40,4 +40,6 @@ 26409 20084 26409 26409 26409 1610 20084 + 36 6829 244 66 17721 35177 85 + 1262 2196 586 1744 33525 186 209 623 28910 187 50276 187 50275 187 50274 187 50273 187 14931 237 211 313 6320 10 49042 116 325 224 14931 223 106 171 118 226 313 34263 802 13511 261 32147 456 10 3384 239 216 22692 101 236 14931 101 236 495 5922 30057 495 20084 495 26409 30057 20084 495 26409 1610 495 26409 20084 495 15 20 495 537 20 495 1051 20 209 18081 211 18081 116 18081 230 39936 222 18081 226 39936 213 18081 233 18081 117 18081 242 39936 212 18081 242 18081 97 18081 116 18081 216 14931 235 212 3736 15367 41197 13610 19934 41869 21275 1012 1047 18795 40120 20422 241 16081 6877 12880 11514 1068 8713 38177 13396 3415 9925 12559 10453 1389 42011 35033 34842 11202 9739 9739 33021 18963 4672 25561 8220 309 1849 644 686 42618 344 434 627 13 686 1848 368 2119 32 686 46 417 2119 309 1833 1056 352 13 686 37 368 751 690 10331 32 844 8 31516 247 8 77 45 diff --git a/models/ggml-vocab-phi-3.gguf.inp b/models/ggml-vocab-phi-3.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-phi-3.gguf.inp +++ b/models/ggml-vocab-phi-3.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-phi-3.gguf.out b/models/ggml-vocab-phi-3.gguf.out index 9c3327cb5..1c3b0a2c9 100644 --- a/models/ggml-vocab-phi-3.gguf.out +++ b/models/ggml-vocab-phi-3.gguf.out @@ -40,4 +40,6 @@ 29871 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29941 + 315 228 190 176 29874 10630 30529 29873 + 29871 2313 3163 29871 13 29871 13 13 29871 13 13 13 29871 12 29871 12 12 29871 12 13 259 13 1678 13 268 13 418 13 243 162 157 131 313 8945 29897 29871 243 162 155 185 30722 243 162 143 174 30598 313 20787 953 3848 275 16125 630 29897 29871 31681 29871 243 162 169 156 243 162 169 156 29871 29941 29871 29941 29941 29871 29941 29941 29941 29871 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29871 29941 29889 29941 29871 29941 636 29941 29871 29941 856 29941 29871 31849 31324 31934 228 162 142 228 161 146 228 162 133 228 161 153 228 161 186 31708 228 162 132 31708 228 161 165 31324 228 161 136 243 162 155 132 1577 30672 31522 30505 11548 31041 30732 29896 29941 29896 29946 29896 29945 29896 30408 30739 448 23648 2751 25512 1538 4851 665 1386 29713 1305 14550 4907 11120 16159 16159 16159 15945 15945 3045 636 6824 6824 6824 8773 8773 8773 306 29915 345 1063 525 29873 1025 540 29915 29879 727 29892 525 1525 366 1854 29973 525 29924 451 1854 306 29915 645 1207 372 29892 525 29928 366 763 777 23429 29973 1334 29915 29963 29872 263 29915 29880 29931 diff --git a/models/ggml-vocab-qwen2.gguf.inp b/models/ggml-vocab-qwen2.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-qwen2.gguf.inp +++ b/models/ggml-vocab-qwen2.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-qwen2.gguf.out b/models/ggml-vocab-qwen2.gguf.out index 401a510e8..4ab275396 100644 --- a/models/ggml-vocab-qwen2.gguf.out +++ b/models/ggml-vocab-qwen2.gguf.out @@ -40,4 +40,6 @@ 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 + 34 90063 128324 + 2560 2347 198 4710 14731 65497 7847 1572 2303 78672 10947 145836 320 8252 8 26525 114 378 235 149921 30543 320 35673 99066 97534 8 25521 227 11162 99 247 149955 220 18 220 18 18 220 18 18 18 220 18 18 18 18 220 18 18 18 18 18 220 18 18 18 18 18 18 220 18 18 18 18 18 18 18 220 18 18 18 18 18 18 18 18 220 18 13 18 220 18 496 18 220 18 1112 18 220 146394 97529 241 44258 233 146568 44258 224 147603 20879 115 146280 44258 223 146280 147272 97529 227 144534 937 104100 18493 22377 99257 16 18 16 19 16 20 16 35727 21216 55460 53237 18658 14144 1456 13073 63471 33594 3038 133178 79012 3355 4605 4605 13874 13874 73594 3014 3014 28149 17085 2928 26610 7646 358 3003 1012 364 83 813 566 594 1052 11 364 787 498 2704 30 364 44 537 2704 358 3278 1281 432 11 364 35 498 1075 1045 15243 30 1205 6 42612 264 63866 43 diff --git a/models/ggml-vocab-refact.gguf.inp b/models/ggml-vocab-refact.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-refact.gguf.inp +++ b/models/ggml-vocab-refact.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-refact.gguf.out b/models/ggml-vocab-refact.gguf.out index 06b15c090..46d8b4aec 100644 --- a/models/ggml-vocab-refact.gguf.out +++ b/models/ggml-vocab-refact.gguf.out @@ -40,4 +40,6 @@ 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 + 53 33934 83 33217 17102 102 + 1214 12258 334 719 8878 202 10885 4222 16104 28570 203 3807 253 227 308 4382 27 18458 133 46113 44967 123 13868 308 12565 19775 33071 40824 733 27 41889 5945 118 252 3807 118 252 225 37 225 37 37 225 37 37 37 225 37 37 37 37 225 37 37 37 37 37 225 37 37 37 37 37 37 225 37 37 37 37 37 37 37 225 37 37 37 37 37 37 37 37 225 37 32 37 225 37 497 37 225 37 1179 37 225 14574 227 14574 133 14574 246 30457 238 14574 242 30457 229 14574 249 14574 134 14574 258 30457 228 14574 258 14574 114 14574 133 14574 232 36628 228 1018 4982 13368 2909 9513 17827 35 37 35 38 35 39 35 11873 47838 20921 16623 13028 8372 1039 9446 40242 13852 2053 8949 12531 1520 10700 5881 9592 13299 914 31753 31359 9163 3202 35472 10397 439 4763 2583 330 102 1455 938 1182 2017 30 330 613 844 3654 49 330 63 646 3654 439 4621 1930 561 30 330 54 844 2124 1629 35993 49 2688 25 7709 312 25 94 62 diff --git a/models/ggml-vocab-starcoder.gguf.inp b/models/ggml-vocab-starcoder.gguf.inp index 0a89107c6..5b4aeb31a 100644 --- a/models/ggml-vocab-starcoder.gguf.inp +++ b/models/ggml-vocab-starcoder.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-starcoder.gguf.out b/models/ggml-vocab-starcoder.gguf.out index ccb55c7fe..9ce2698a9 100644 --- a/models/ggml-vocab-starcoder.gguf.out +++ b/models/ggml-vocab-starcoder.gguf.out @@ -40,4 +40,6 @@ 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 + 72 34269 102 33245 17234 121 + 1236 12266 353 736 8886 221 10883 4238 16101 28540 222 3822 272 246 327 4434 46 18445 152 46030 45022 142 13878 327 12585 19884 33773 40920 751 46 41839 5954 137 271 3822 137 271 244 56 244 56 56 244 56 56 56 244 56 56 56 56 244 56 56 56 56 56 244 56 56 56 56 56 56 244 56 56 56 56 56 56 56 244 56 56 56 56 56 56 56 56 244 56 51 56 244 56 516 56 244 56 1198 56 244 14566 246 14566 152 14566 265 30428 257 14566 261 30428 248 14566 268 14566 153 14566 277 30428 247 14566 277 14566 133 14566 152 14566 251 36570 247 1037 4995 13379 2924 9515 17823 54 56 54 57 54 58 54 11904 47892 20895 16625 13047 8389 1059 9504 40216 13858 2073 8983 12571 1539 10721 5918 9643 13298 932 31723 31330 9221 3226 35426 10400 457 4783 2602 349 121 1477 957 1200 2038 49 349 632 863 3673 68 349 82 666 3673 457 4650 1949 580 49 349 73 863 2144 1649 35941 68 2726 44 7728 331 44 113 81 From 402d6feffa0572d1c7a957901b6d1702bd188484 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Thu, 4 Jul 2024 12:50:57 +0200 Subject: [PATCH 26/38] llama : suppress unref var in Windows MSVC (#8150) * llama : suppress unref var in Windows MSVC This commit suppresses two warnings that are currently generated for src/llama.cpp when building on Windows MSVC ```console C:\llama.cpp\src\llama.cpp(14349,45): warning C4101: 'ex': unreferenced local variable [C:\llama.cpp\build\src\llama.vcxproj] C:\llama.cpp\src\llama.cpp(19285,44): warning C4101: 'e': unreferenced local variable [C:\llama.cpp\build\src\llama.vcxproj] ``` * Update src/llama.cpp --------- Co-authored-by: Georgi Gerganov --- src/llama.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/llama.cpp b/src/llama.cpp index 73f52435a..3d131b325 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -14781,7 +14781,7 @@ private: size_t prefix_offset = input_offset; unicode_cpt_from_utf8(input, prefix_offset); return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset }; - } catch(std::invalid_argument & ex) { + } catch (std::invalid_argument & /*ex*/) { // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER return { "\xEF\xBF\xBD", 3, 1 }; } @@ -19725,7 +19725,7 @@ static std::string llama_decode_text(const std::string & text) { const auto utf8 = unicode_cpt_to_utf8(cpt); try { decoded_text += unicode_utf8_to_byte(utf8); - } catch (const std::out_of_range & e) { + } catch (const std::out_of_range & /*e*/) { decoded_text += "[UNK_BYTE_0x"; for (const auto c : utf8) { decoded_text += format("%02x", (uint8_t) c); From f8c4c0738d72d2162736edd72dd5db8b269adca1 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Thu, 4 Jul 2024 12:53:42 +0200 Subject: [PATCH 27/38] tests : add _CRT_SECURE_NO_WARNINGS for WIN32 (#8231) This commit adds the compile definition `_CRT_SECURE_NO_WARNINGS` to the root cmake subproject. The motivation for this is that currently the following warnings are displayed when compiling the tests and common cmake subprojects: ```console test-llama-grammar.cpp C:\llama.cpp\src\.\llama.cpp(1406,77): warning C4996: 'strerror': This function or variable may be unsafe. Consider using strerror_s instead. To disable deprecation, use _CRT_SECURE_NO_WARNINGS. See online help for details. [C:\llama.cpp\build\tests\test-llama-grammar.vcxproj] ... ``` This compile definition is currently set for the `src` subproject and this change moves into the root cmake project so that it is applied to all cmake subprojects. --- CMakeLists.txt | 4 ++++ src/CMakeLists.txt | 2 -- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index e3a0cc369..d95414d71 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -42,6 +42,10 @@ endif() option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT}) +if (WIN32) + add_compile_definitions(_CRT_SECURE_NO_WARNINGS) +endif() + # # option list # diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index ccb607e56..c2049df79 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -1,7 +1,5 @@ # TODO: should not use this if (WIN32) - add_compile_definitions(_CRT_SECURE_NO_WARNINGS) - if (BUILD_SHARED_LIBS) set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON) endif() From 807b0c49ff7071094f97ebc3a0a8e2b9e274f503 Mon Sep 17 00:00:00 2001 From: fairydreaming <166155368+fairydreaming@users.noreply.github.com> Date: Thu, 4 Jul 2024 15:46:11 +0200 Subject: [PATCH 28/38] Inference support for T5 and FLAN-T5 model families (#5763) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * llama : add inference support and model types for T5 and FLAN-T5 model families * llama : add new API functions to support encoder-decoder models: llama_encode(), llama_model_has_encoder(), llama_model_decoder_start_token() * common, llama-cli, llama-batched : add support for encoder-decoder models * convert-hf : handle shared token embeddings tensors in T5Model * convert-hf : add support for SentencePiece BPE tokenizer in T5Model (for Pile-T5 models) * convert-hf : add MT5ForConditionalGeneration and UMT5ForConditionalGeneration to architectures supported by T5Model * convert : add t5 tokenizer tests, use "slow" HF tokenizer for t5 --------- Co-authored-by: Stanisław Szymczyk Co-authored-by: Georgi Gerganov --- common/common.cpp | 19 +- convert-hf-to-gguf-update.py | 19 +- convert-hf-to-gguf.py | 46 +- examples/batched/batched.cpp | 34 +- examples/main/main.cpp | 22 +- include/llama.h | 15 + models/ggml-vocab-bert-bge.gguf.inp | 2 + models/ggml-vocab-bert-bge.gguf.out | 1 + models/ggml-vocab-command-r.gguf.inp | 2 + models/ggml-vocab-command-r.gguf.out | 1 + models/ggml-vocab-deepseek-coder.gguf.inp | 2 + models/ggml-vocab-deepseek-coder.gguf.out | 1 + models/ggml-vocab-deepseek-llm.gguf.inp | 2 + models/ggml-vocab-deepseek-llm.gguf.out | 1 + models/ggml-vocab-falcon.gguf.inp | 2 + models/ggml-vocab-falcon.gguf.out | 1 + models/ggml-vocab-gpt-2.gguf.inp | 2 + models/ggml-vocab-gpt-2.gguf.out | 1 + models/ggml-vocab-llama-bpe.gguf.inp | 2 + models/ggml-vocab-llama-bpe.gguf.out | 1 + models/ggml-vocab-llama-spm.gguf.inp | 2 + models/ggml-vocab-llama-spm.gguf.out | 1 + models/ggml-vocab-mpt.gguf.inp | 2 + models/ggml-vocab-mpt.gguf.out | 1 + models/ggml-vocab-phi-3.gguf.inp | 2 + models/ggml-vocab-phi-3.gguf.out | 1 + models/ggml-vocab-qwen2.gguf.inp | 2 + models/ggml-vocab-qwen2.gguf.out | 1 + models/ggml-vocab-refact.gguf.inp | 2 + models/ggml-vocab-refact.gguf.out | 1 + models/ggml-vocab-starcoder.gguf.inp | 2 + models/ggml-vocab-starcoder.gguf.out | 1 + src/llama.cpp | 783 +++++++++++++++++++++- 33 files changed, 946 insertions(+), 31 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 2c05a4d4a..4138dc3b6 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -2070,7 +2070,24 @@ std::tuple llama_init_from_gpt_par if (params.warmup) { LOG("warming up the model with an empty run\n"); - std::vector tmp = { llama_token_bos(model), llama_token_eos(model), }; + std::vector tmp; + llama_token bos = llama_token_bos(model); + llama_token eos = llama_token_eos(model); + // some models (e.g. T5) don't have a BOS token + if (bos != -1) { + tmp.push_back(bos); + } + tmp.push_back(eos); + + if (llama_model_has_encoder(model)) { + llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0)); + llama_token decoder_start_token_id = llama_model_decoder_start_token(model); + if (decoder_start_token_id == -1) { + decoder_start_token_id = bos; + } + tmp.clear(); + tmp.push_back(decoder_start_token_id); + } llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0)); llama_kv_cache_clear(lctx); llama_synchronize(lctx); diff --git a/convert-hf-to-gguf-update.py b/convert-hf-to-gguf-update.py index 9eb406cb8..21a306255 100755 --- a/convert-hf-to-gguf-update.py +++ b/convert-hf-to-gguf-update.py @@ -45,6 +45,7 @@ class TOKENIZER_TYPE(IntEnum): SPM = auto() BPE = auto() WPM = auto() + UGM = auto() # TODO: this string has to exercise as much pre-tokenizer functionality as possible @@ -89,6 +90,7 @@ models = [ {"name": "gemma", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", }, {"name": "gemma-2", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", }, {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, + {"name": "t5", "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", }, ] @@ -110,9 +112,13 @@ def download_model(model): os.makedirs(f"models/tokenizers/{name}", exist_ok=True) files = ["config.json", "tokenizer.json", "tokenizer_config.json"] + if tokt == TOKENIZER_TYPE.SPM: files.append("tokenizer.model") + if tokt == TOKENIZER_TYPE.UGM: + files.append("spiece.model") + for file in files: save_path = f"models/tokenizers/{name}/{file}" if os.path.isfile(save_path): @@ -135,7 +141,7 @@ for model in models: name = model["name"] tokt = model["tokt"] - if tokt == TOKENIZER_TYPE.SPM: + if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM: continue # Skip if the tokenizer folder does not exist or there are other download issues previously @@ -145,7 +151,10 @@ for model in models: # create the tokenizer try: - tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") + if name == "t5": + tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False) + else: + tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") except OSError as e: logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}") continue # Skip to the next model if the tokenizer can't be loaded @@ -266,6 +275,7 @@ tests = [ "\n =", "' era", "Hello, y'all! How are you 😁 ?我想在apple工作1314151天～", + "!!!!!!", "3", "33", "333", @@ -304,7 +314,10 @@ for model in models: # create the tokenizer try: - tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") + if name == "t5": + tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False) + else: + tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") except OSError as e: logger.error(f"Failed to load tokenizer for model {name}. Error: {e}") continue # Skip this model and continue with the next one in the loop diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index bae145589..a1d165023 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -2853,11 +2853,17 @@ class DeepseekV2Model(Model): raise ValueError(f"Unprocessed experts: {experts}") -@Model.register("T5ForConditionalGeneration") @Model.register("T5WithLMHeadModel") +@Model.register("T5ForConditionalGeneration") +@Model.register("MT5ForConditionalGeneration") +@Model.register("UMT5ForConditionalGeneration") class T5Model(Model): model_arch = gguf.MODEL_ARCH.T5 + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.shared_token_embeddings_found = False + def set_vocab(self): # to avoid TypeError: Descriptors cannot be created directly # exception when importing sentencepiece_model_pb2 @@ -2865,17 +2871,29 @@ class T5Model(Model): from sentencepiece import SentencePieceProcessor from sentencepiece import sentencepiece_model_pb2 as model - tokenizer_path = self.dir_model / 'spiece.model' + tokenizer_path = self.dir_model / 'tokenizer.model' + + # many older models use spiece.model tokenizer model filename + if not tokenizer_path.is_file(): + tokenizer_path = self.dir_model / 'spiece.model' if not tokenizer_path.is_file(): raise FileNotFoundError(f"File not found: {tokenizer_path}") sentencepiece_model = model.ModelProto() sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read()) + + # some models like Pile-T5 family use BPE tokenizer instead of Unigram + if sentencepiece_model.trainer_spec.model_type == 2: # BPE + # assure the tokenizer model file name is correct + assert tokenizer_path.name == 'tokenizer.model' + return self._set_vocab_sentencepiece() + else: + assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM + add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap - assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM tokenizer = SentencePieceProcessor() tokenizer.LoadFromFile(str(tokenizer_path)) @@ -2945,7 +2963,10 @@ class T5Model(Model): def set_gguf_parameters(self): self.gguf_writer.add_name("T5") - self.gguf_writer.add_context_length(self.hparams["n_positions"]) + if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None: + logger.warning("Couldn't find context length in config.json, assuming default value of 512") + n_ctx = 512 + self.gguf_writer.add_context_length(n_ctx) self.gguf_writer.add_embedding_length(self.hparams["d_model"]) self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"]) self.gguf_writer.add_block_count(self.hparams["num_layers"]) @@ -2961,12 +2982,17 @@ class T5Model(Model): def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: del bid # unused - # Sometimes T5 and Flan-T5 based models contain "encoder.embed_tokens.weight" tensor or - # "decoder.embed_tokens.weight" tensors that are duplicates of "shared.weight" tensor - # To prevent errors caused by an unnecessary unmapped tensor, skip both of them and use only "shared.weight". - if name == "decoder.embed_tokens.weight" or name == "encoder.embed_tokens.weight": - logger.debug(f"Skipping tensor {name!r} in safetensors so that convert can end normally.") - return [] + # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight", + # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored + # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder + # and decoder and ignore the remaining ones. + if name in ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "shared.weight"]: + if not self.shared_token_embeddings_found: + name = "shared.weight" + self.shared_token_embeddings_found = True + else: + logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.") + return [] return [(self.map_tensor_name(name), data_torch)] diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index 62d9b144d..2442e954d 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -93,14 +93,34 @@ int main(int argc, char ** argv) { // create a llama_batch // we use this object to submit token data for decoding - llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0, 1); + llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel); + + std::vector seq_ids(n_parallel, 0); + for (int32_t i = 0; i < n_parallel; ++i) { + seq_ids[i] = i; + } // evaluate the initial prompt for (size_t i = 0; i < tokens_list.size(); ++i) { - llama_batch_add(batch, tokens_list[i], i, { 0 }, false); + llama_batch_add(batch, tokens_list[i], i, seq_ids, false); } GGML_ASSERT(batch.n_tokens == (int) tokens_list.size()); + if (llama_model_has_encoder(model)) { + if (llama_encode(ctx, batch)) { + LOG_TEE("%s : failed to eval\n", __func__); + return 1; + } + + llama_token decoder_start_token_id = llama_model_decoder_start_token(model); + if (decoder_start_token_id == -1) { + decoder_start_token_id = llama_token_bos(model); + } + + llama_batch_clear(batch); + llama_batch_add(batch, decoder_start_token_id, 0, seq_ids, false); + } + // llama_decode will output logits only for the last token of the prompt batch.logits[batch.n_tokens - 1] = true; @@ -109,11 +129,11 @@ int main(int argc, char ** argv) { return 1; } - // assign the system KV cache to all parallel sequences - // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them - for (int32_t i = 1; i < n_parallel; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); - } + //// assign the system KV cache to all parallel sequences + //// this way, the parallel sequences will "reuse" the prompt tokens without having to copy them + //for (int32_t i = 1; i < n_parallel; ++i) { + // llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); + //} if (n_parallel > 1) { LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel); diff --git a/examples/main/main.cpp b/examples/main/main.cpp index d512953b9..22bb37889 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -255,7 +255,9 @@ int main(int argc, char ** argv) { } const bool add_bos = llama_should_add_bos_token(model); - GGML_ASSERT(llama_add_eos_token(model) != 1); + if (!llama_model_has_encoder(model)) { + GGML_ASSERT(llama_add_eos_token(model) != 1); + } LOG("add_bos: %d\n", add_bos); std::vector embd_inp; @@ -517,6 +519,24 @@ int main(int argc, char ** argv) { exit(1); } + if (llama_model_has_encoder(model)) { + int enc_input_size = embd_inp.size(); + llama_token * enc_input_buf = embd_inp.data(); + + if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) { + LOG_TEE("%s : failed to eval\n", __func__); + return 1; + } + + llama_token decoder_start_token_id = llama_model_decoder_start_token(model); + if (decoder_start_token_id == -1) { + decoder_start_token_id = llama_token_bos(model); + } + + embd_inp.clear(); + embd_inp.push_back(decoder_start_token_id); + } + while ((n_remain != 0 && !is_antiprompt) || params.interactive) { // predict if (!embd.empty()) { diff --git a/include/llama.h b/include/llama.h index c5b618292..aca79d2b8 100644 --- a/include/llama.h +++ b/include/llama.h @@ -485,6 +485,13 @@ extern "C" { // Get a llama model tensor LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name); + // Returns true if the model contains an encoder that requires llama_encode() call + LLAMA_API bool llama_model_has_encoder(const struct llama_model * model); + + // For encoder-decoder models, this function returns id of the token that must be provided + // to the decoder to start generating output sequence. For other models, it returns -1. + LLAMA_API llama_token llama_model_decoder_start_token(const struct llama_model * model); + // Returns 0 on success LLAMA_API uint32_t llama_model_quantize( const char * fname_inp, @@ -770,6 +777,14 @@ extern "C" { // Frees a batch of tokens allocated with llama_batch_init() LLAMA_API void llama_batch_free(struct llama_batch batch); + // Processes a batch of tokens with the ecoder part of the encoder-decoder model. + // Stores the encoder output internally for later use by the decoder cross-attention layers. + // 0 - success + // < 0 - error + LLAMA_API int32_t llama_encode( + struct llama_context * ctx, + struct llama_batch batch); + // Positive return values does not mean a fatal error, but rather a warning. // 0 - success // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) diff --git a/models/ggml-vocab-bert-bge.gguf.inp b/models/ggml-vocab-bert-bge.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-bert-bge.gguf.inp +++ b/models/ggml-vocab-bert-bge.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-bert-bge.gguf.out b/models/ggml-vocab-bert-bge.gguf.out index 82d4ed1c1..a62566ce7 100644 --- a/models/ggml-vocab-bert-bge.gguf.out +++ b/models/ggml-vocab-bert-bge.gguf.out @@ -31,6 +31,7 @@ 1027 1005 3690 7592 1010 1061 1005 2035 999 2129 2024 2017 100 1029 1855 100 100 6207 100 100 14677 23632 22203 1811 1995 + 999 999 999 999 999 999 1017 3943 21211 diff --git a/models/ggml-vocab-command-r.gguf.inp b/models/ggml-vocab-command-r.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-command-r.gguf.inp +++ b/models/ggml-vocab-command-r.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-command-r.gguf.out b/models/ggml-vocab-command-r.gguf.out index 939b9dc30..3f6b41888 100644 --- a/models/ggml-vocab-command-r.gguf.out +++ b/models/ggml-vocab-command-r.gguf.out @@ -31,6 +31,7 @@ 206 1857 14 4515 28339 19 1770 14 1954 8 4070 1955 1933 80503 231 5691 12081 13336 2648 29325 14315 24 26 24 27 24 28 24 5123 18372 + 57178 10251 26 26 26 26 26 26 diff --git a/models/ggml-vocab-deepseek-coder.gguf.inp b/models/ggml-vocab-deepseek-coder.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-deepseek-coder.gguf.inp +++ b/models/ggml-vocab-deepseek-coder.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-deepseek-coder.gguf.out b/models/ggml-vocab-deepseek-coder.gguf.out index a43e3f0f1..52c4111a1 100644 --- a/models/ggml-vocab-deepseek-coder.gguf.out +++ b/models/ggml-vocab-deepseek-coder.gguf.out @@ -31,6 +31,7 @@ 185 405 6 2895 17535 11 320 6 435 0 1717 417 340 12394 233 210 3015 19100 608 9413 2668 16 18 16 19 16 20 16 1393 169 121 239 + 15330 3023 18 18 18 18 18 18 diff --git a/models/ggml-vocab-deepseek-llm.gguf.inp b/models/ggml-vocab-deepseek-llm.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-deepseek-llm.gguf.inp +++ b/models/ggml-vocab-deepseek-llm.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-deepseek-llm.gguf.out b/models/ggml-vocab-deepseek-llm.gguf.out index d31ac1cc6..0191b7a11 100644 --- a/models/ggml-vocab-deepseek-llm.gguf.out +++ b/models/ggml-vocab-deepseek-llm.gguf.out @@ -31,6 +31,7 @@ 185 403 6 2906 17464 11 320 6 436 0 1724 418 340 33701 210 3025 19017 612 9407 2681 16 18 16 19 16 20 16 1398 68940 239 + 15278 3033 18 18 18 18 18 18 diff --git a/models/ggml-vocab-falcon.gguf.inp b/models/ggml-vocab-falcon.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-falcon.gguf.inp +++ b/models/ggml-vocab-falcon.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-falcon.gguf.out b/models/ggml-vocab-falcon.gguf.out index 1ab70fe70..64a48d97f 100644 --- a/models/ggml-vocab-falcon.gguf.out +++ b/models/ggml-vocab-falcon.gguf.out @@ -31,6 +31,7 @@ 1212 40 18 4932 9856 23 291 18 436 12 1265 362 299 8196 207 204 42 50087 123 2727 20300 32022 133 234 17419 30137 28 7858 181 133 236 + 51520 30 3138 22287 diff --git a/models/ggml-vocab-gpt-2.gguf.inp b/models/ggml-vocab-gpt-2.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-gpt-2.gguf.inp +++ b/models/ggml-vocab-gpt-2.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-gpt-2.gguf.out b/models/ggml-vocab-gpt-2.gguf.out index 88217d3fa..17a13bdfc 100644 --- a/models/ggml-vocab-gpt-2.gguf.out +++ b/models/ggml-vocab-gpt-2.gguf.out @@ -31,6 +31,7 @@ 198 796 6 6980 15496 11 331 6 439 0 1374 389 345 30325 223 5633 22755 239 46349 111 28839 101 18040 32432 98 43291 1485 1415 24309 25465 171 121 252 + 13896 3228 18 2091 20370 diff --git a/models/ggml-vocab-llama-bpe.gguf.inp b/models/ggml-vocab-llama-bpe.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-llama-bpe.gguf.inp +++ b/models/ggml-vocab-llama-bpe.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-llama-bpe.gguf.out b/models/ggml-vocab-llama-bpe.gguf.out index bb1fe229c..4b35cf93f 100644 --- a/models/ggml-vocab-llama-bpe.gguf.out +++ b/models/ggml-vocab-llama-bpe.gguf.out @@ -31,6 +31,7 @@ 198 284 6 11639 9906 11 379 65948 0 2650 527 499 27623 223 949 37046 101067 19000 23182 102301 9263 18136 16 36827 21909 + 17523 3001 18 1644 8765 diff --git a/models/ggml-vocab-llama-spm.gguf.inp b/models/ggml-vocab-llama-spm.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-llama-spm.gguf.inp +++ b/models/ggml-vocab-llama-spm.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-llama-spm.gguf.out b/models/ggml-vocab-llama-spm.gguf.out index 1c3b0a2c9..93aacf8ba 100644 --- a/models/ggml-vocab-llama-spm.gguf.out +++ b/models/ggml-vocab-llama-spm.gguf.out @@ -31,6 +31,7 @@ 29871 13 353 525 3152 15043 29892 343 29915 497 29991 1128 526 366 29871 243 162 155 132 1577 30672 31522 30505 11548 31041 30732 29896 29941 29896 29946 29896 29945 29896 30408 30739 + 1738 6824 21004 29871 29941 29871 29941 29941 29871 29941 29941 29941 diff --git a/models/ggml-vocab-mpt.gguf.inp b/models/ggml-vocab-mpt.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-mpt.gguf.inp +++ b/models/ggml-vocab-mpt.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-mpt.gguf.out b/models/ggml-vocab-mpt.gguf.out index d4a877d1c..372c751bf 100644 --- a/models/ggml-vocab-mpt.gguf.out +++ b/models/ggml-vocab-mpt.gguf.out @@ -31,6 +31,7 @@ 187 426 8 8685 12092 13 340 8 455 2 1359 403 368 49042 212 3736 15367 41197 13610 19934 41869 21275 1012 1047 18795 40120 20422 241 + 18963 4672 20 1610 20084 diff --git a/models/ggml-vocab-phi-3.gguf.inp b/models/ggml-vocab-phi-3.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-phi-3.gguf.inp +++ b/models/ggml-vocab-phi-3.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-phi-3.gguf.out b/models/ggml-vocab-phi-3.gguf.out index 1c3b0a2c9..93aacf8ba 100644 --- a/models/ggml-vocab-phi-3.gguf.out +++ b/models/ggml-vocab-phi-3.gguf.out @@ -31,6 +31,7 @@ 29871 13 353 525 3152 15043 29892 343 29915 497 29991 1128 526 366 29871 243 162 155 132 1577 30672 31522 30505 11548 31041 30732 29896 29941 29896 29946 29896 29945 29896 30408 30739 + 1738 6824 21004 29871 29941 29871 29941 29941 29871 29941 29941 29941 diff --git a/models/ggml-vocab-qwen2.gguf.inp b/models/ggml-vocab-qwen2.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-qwen2.gguf.inp +++ b/models/ggml-vocab-qwen2.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-qwen2.gguf.out b/models/ggml-vocab-qwen2.gguf.out index 4ab275396..18b4b45cd 100644 --- a/models/ggml-vocab-qwen2.gguf.out +++ b/models/ggml-vocab-qwen2.gguf.out @@ -31,6 +31,7 @@ 198 284 6 11385 9707 11 379 64848 0 2585 525 498 26525 223 937 104100 18493 22377 99257 16 18 16 19 16 20 16 35727 21216 + 17085 2928 18 18 18 18 18 18 diff --git a/models/ggml-vocab-refact.gguf.inp b/models/ggml-vocab-refact.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-refact.gguf.inp +++ b/models/ggml-vocab-refact.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-refact.gguf.out b/models/ggml-vocab-refact.gguf.out index 46d8b4aec..63d8305c3 100644 --- a/models/ggml-vocab-refact.gguf.out +++ b/models/ggml-vocab-refact.gguf.out @@ -31,6 +31,7 @@ 203 280 25 34666 8279 30 533 25 464 19 4971 884 844 18458 228 1018 4982 13368 2909 9513 17827 35 37 35 38 35 39 35 11873 47838 + 9163 3202 37 37 37 37 37 37 diff --git a/models/ggml-vocab-starcoder.gguf.inp b/models/ggml-vocab-starcoder.gguf.inp index 5b4aeb31a..9baf7d77a 100644 --- a/models/ggml-vocab-starcoder.gguf.inp +++ b/models/ggml-vocab-starcoder.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 diff --git a/models/ggml-vocab-starcoder.gguf.out b/models/ggml-vocab-starcoder.gguf.out index 9ce2698a9..87e2465d3 100644 --- a/models/ggml-vocab-starcoder.gguf.out +++ b/models/ggml-vocab-starcoder.gguf.out @@ -31,6 +31,7 @@ 222 299 44 34719 8302 49 553 44 483 38 4998 904 863 18445 247 1037 4995 13379 2924 9515 17823 54 56 54 57 54 58 54 11904 47892 + 9221 3226 56 56 56 56 56 56 diff --git a/src/llama.cpp b/src/llama.cpp index 3d131b325..1fe2b9f79 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -2044,12 +2044,18 @@ enum e_model { MODEL_17M, MODEL_22M, MODEL_33M, + MODEL_60M, MODEL_70M, + MODEL_80M, MODEL_109M, MODEL_137M, MODEL_160M, + MODEL_220M, + MODEL_250M, MODEL_335M, MODEL_410M, + MODEL_770M, + MODEL_780M, MODEL_0_5B, MODEL_1B, MODEL_1_3B, @@ -2061,6 +2067,7 @@ enum e_model { MODEL_6_9B, MODEL_7B, MODEL_8B, + MODEL_11B, MODEL_12B, MODEL_13B, MODEL_14B, @@ -2112,6 +2119,7 @@ struct llama_hparams { uint32_t n_expert = 0; uint32_t n_expert_used = 0; uint32_t n_vocab_type = 0; // for BERT-style token types + uint32_t n_rel_attn_bkts = 0; uint32_t n_layer_dense_lead = 0; uint32_t n_lora_q = 0; @@ -2147,6 +2155,10 @@ struct llama_hparams { bool use_alibi = false; bool attn_soft_cap = false; + // needed by encoder-decoder models (e.g. T5, FLAN-T5) + // ref: https://github.com/ggerganov/llama.cpp/pull/8141 + llama_token dec_start_token_id = -1; + enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE; enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE; @@ -2167,6 +2179,7 @@ struct llama_hparams { if (this->n_expert != other.n_expert) return true; if (this->n_expert_used != other.n_expert_used) return true; + if (this->n_rel_attn_bkts != other.n_rel_attn_bkts) return true; if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true; if (this->n_lora_q != other.n_lora_q) return true; if (this->n_lora_kv != other.n_lora_kv) return true; @@ -2182,6 +2195,8 @@ struct llama_hparams { if (this->ssm_d_state != other.ssm_d_state) return true; if (this->ssm_dt_rank != other.ssm_dt_rank) return true; + if (this->dec_start_token_id != other.dec_start_token_id) return true; + const float EPSILON = 1e-9f; if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true; @@ -2254,6 +2269,7 @@ struct llama_cparams { void * cb_eval_user_data; }; +// TODO: separate into "llama_layer_enc" and "llama_layer_dec" struct llama_layer { // normalization struct ggml_tensor * attn_norm; @@ -2271,6 +2287,8 @@ struct llama_layer { struct ggml_tensor * attn_sub_norm; struct ggml_tensor * attn_post_norm; struct ggml_tensor * ffn_sub_norm; + struct ggml_tensor * attn_norm_cross; + struct ggml_tensor * attn_norm_enc; // attention struct ggml_tensor * wq; @@ -2282,6 +2300,14 @@ struct llama_layer { struct ggml_tensor * wq_b; struct ggml_tensor * wkv_a_mqa; struct ggml_tensor * wkv_b; + struct ggml_tensor * wq_cross; + struct ggml_tensor * wk_cross; + struct ggml_tensor * wv_cross; + struct ggml_tensor * wo_cross; + struct ggml_tensor * wq_enc; + struct ggml_tensor * wk_enc; + struct ggml_tensor * wv_enc; + struct ggml_tensor * wo_enc; // attention bias struct ggml_tensor * bq; @@ -2290,6 +2316,11 @@ struct llama_layer { struct ggml_tensor * bo; struct ggml_tensor * bqkv; + // relative position bias + struct ggml_tensor * attn_rel_b; + struct ggml_tensor * attn_rel_b_enc; + struct ggml_tensor * attn_rel_b_cross; + // normalization struct ggml_tensor * ffn_norm; struct ggml_tensor * ffn_norm_b; @@ -2297,11 +2328,15 @@ struct llama_layer { struct ggml_tensor * layer_out_norm; struct ggml_tensor * layer_out_norm_b; struct ggml_tensor * ffn_norm_exps; + struct ggml_tensor * ffn_norm_enc; // ff struct ggml_tensor * ffn_gate; // w1 struct ggml_tensor * ffn_down; // w2 struct ggml_tensor * ffn_up; // w3 + struct ggml_tensor * ffn_gate_enc; + struct ggml_tensor * ffn_down_enc; + struct ggml_tensor * ffn_up_enc; // ff MoE struct ggml_tensor * ffn_gate_inp; @@ -2535,6 +2570,7 @@ struct llama_model { struct ggml_tensor * output_norm_b; struct ggml_tensor * output; struct ggml_tensor * output_b; + struct ggml_tensor * output_norm_enc; std::vector layers; @@ -2663,6 +2699,13 @@ struct llama_context { // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE std::map> embd_seq; + // whether we are computing encoder output or decoder output + bool is_encoding = false; + + // output of the encoder part of the encoder-decoder models + std::vector embd_enc; + std::vector> seq_ids_enc; + // memory buffers used to evaluate the model std::vector buf_compute_meta; ggml_backend_sched_t sched = nullptr; @@ -2683,6 +2726,9 @@ struct llama_context { struct ggml_tensor * inp_s_copy; // I32 [kv_size] struct ggml_tensor * inp_s_mask; // F32 [1, n_kv] struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch] + struct ggml_tensor * inp_pos_bucket; // I32 [n_batch|n_kv, n_batch] + struct ggml_tensor * inp_embd_enc; // F32 [n_embd, n_outputs_enc] + struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch] // control vectors struct llama_control_vector cvec; @@ -4286,12 +4332,18 @@ static const char * llama_model_type_name(e_model type) { case MODEL_17M: return "17M"; case MODEL_22M: return "22M"; case MODEL_33M: return "33M"; + case MODEL_60M: return "60M"; case MODEL_70M: return "70M"; + case MODEL_80M: return "80M"; case MODEL_109M: return "109M"; case MODEL_137M: return "137M"; case MODEL_160M: return "160M"; + case MODEL_220M: return "220M"; + case MODEL_250M: return "250M"; case MODEL_335M: return "335M"; case MODEL_410M: return "410M"; + case MODEL_770M: return "770M"; + case MODEL_780M: return "780M"; case MODEL_0_5B: return "0.5B"; case MODEL_1B: return "1B"; case MODEL_1_3B: return "1.3B"; @@ -4303,6 +4355,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_6_9B: return "6.9B"; case MODEL_7B: return "7B"; case MODEL_8B: return "8B"; + case MODEL_11B: return "11B"; case MODEL_12B: return "12B"; case MODEL_13B: return "13B"; case MODEL_14B: return "14B"; @@ -4917,6 +4970,38 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_T5: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, hparams.n_rel_attn_bkts); + + uint32_t dec_start_token_id; + if (ml.get_key(LLM_KV_DECODER_START_TOKEN_ID, dec_start_token_id, false)) { + hparams.dec_start_token_id = dec_start_token_id; + } + + switch (hparams.n_layer) { + case 6: model.type = e_model::MODEL_60M; break; // t5-small + case 8: model.type = e_model::MODEL_80M; break; // flan-t5-small + case 12: + switch (hparams.n_ff) { + case 3072: model.type = e_model::MODEL_220M; break; // t5-base + case 2048: model.type = e_model::MODEL_250M; break; // flan-t5-base + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 24: + switch (hparams.n_ff) { + case 4096: model.type = e_model::MODEL_770M; break; // t5-large + case 2816: model.type = e_model::MODEL_780M; break; // flan-t5-large + case 16384: model.type = e_model::MODEL_3B; break; // t5-3b + case 5120: model.type = e_model::MODEL_3B; break; // flan-t5-xl + case 65536: model.type = e_model::MODEL_11B; break; // t5-11b + case 10240: model.type = e_model::MODEL_11B; break; // flan-t5-xxl + default: model.type = e_model::MODEL_UNKNOWN; + } break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_JAIS: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -6996,6 +7081,64 @@ static bool llm_load_tensors( layer.ffn_up_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "scale", i), {1}); } } break; + case LLM_ARCH_T5: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm_enc = ml.create_tensor(ctx_output, tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_DEC_OUTPUT_NORM, "weight"), {n_embd}); + + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); + } + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_rel_b_enc = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ENC_ATTN_REL_B, "weight", i), {hparams.n_head, hparams.n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + + layer.wq_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wk_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wv_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); + layer.wo_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}); + + layer.ffn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_gate_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_GATE, "weight", i), {n_embd, n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_down_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_UP, "weight", i), {n_embd, n_ff}); + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_rel_b = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_ATTN_REL_B, "weight", i), {hparams.n_head, hparams.n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}); + + layer.attn_norm_cross = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_CROSS_ATTN_NORM, "weight", i), {n_embd}); + // this tensor seems to be unused in HF transformers implementation + layer.attn_rel_b_cross = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "weight", i), {hparams.n_head, hparams.n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + + layer.wq_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wk_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wv_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); + layer.wo_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_gate = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_FFN_GATE, "weight", i), {n_embd, n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; case LLM_ARCH_JAIS: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -7787,6 +7930,7 @@ struct llm_build_context { const int32_t n_tokens; const int32_t n_kv; // size of KV cache to consider (n_kv <= kv_self.size) const int32_t n_outputs; + const int32_t n_outputs_enc; const int32_t kv_head; // index of where we store new KV data in the cache const int32_t n_ctx_orig; @@ -7836,6 +7980,7 @@ struct llm_build_context { n_tokens (batch.n_tokens), n_kv (worst_case ? kv_self.size : kv_self.n), n_outputs (worst_case ? n_tokens : lctx.n_outputs), + n_outputs_enc (worst_case ? n_tokens : lctx.embd_enc.size() / hparams.n_embd), kv_head (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head), n_ctx_orig (cparams.n_ctx_orig_yarn), flash_attn (cparams.flash_attn), @@ -7867,6 +8012,9 @@ struct llm_build_context { lctx.inp_s_copy = nullptr; lctx.inp_s_mask = nullptr; lctx.inp_s_seq = nullptr; + lctx.inp_pos_bucket = nullptr; + lctx.inp_embd_enc = nullptr; + lctx.inp_KQ_mask_cross = nullptr; } void free() { @@ -8119,6 +8267,53 @@ struct llm_build_context { return gf; } + struct ggml_tensor * llm_build_pos_bucket(bool causal) { + if (causal) { + lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv, n_tokens); + } else { + lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_tokens); + } + + ggml_set_input(lctx.inp_pos_bucket); + cb(lctx.inp_pos_bucket, "pos_bucket", -1); + + return lctx.inp_pos_bucket; + } + + struct ggml_tensor * llm_build_pos_bias(struct ggml_tensor * pos_bucket, struct ggml_tensor * attn_rel_b) { + struct ggml_tensor * pos_bucket_1d = ggml_view_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1], 0); + cb(pos_bucket_1d, "pos_bucket_1d", -1); + + struct ggml_tensor * pos_bias = ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d); + cb(pos_bias, "pos_bias", -1); + + pos_bias = ggml_view_3d(ctx0, pos_bias, pos_bias->ne[0], lctx.inp_pos_bucket->ne[0], lctx.inp_pos_bucket->ne[1], ggml_element_size(pos_bias) * pos_bias->ne[0], ggml_element_size(pos_bias) * pos_bias->ne[0] * lctx.inp_pos_bucket->ne[0], 0); + cb(pos_bias, "pos_bias", -1); + + pos_bias = ggml_permute(ctx0, pos_bias, 2, 0, 1, 3); + cb(pos_bias, "pos_bias", -1); + + pos_bias = ggml_cont(ctx0, pos_bias); + cb(pos_bias, "pos_bias", -1); + + return pos_bias; + } + + struct ggml_tensor * llm_build_inp_embd_enc() { + const int64_t n_embd = hparams.n_embd; + lctx.inp_embd_enc = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_outputs_enc); + ggml_set_input(lctx.inp_embd_enc); + cb(lctx.inp_embd_enc, "embd_enc", -1); + return lctx.inp_embd_enc; + } + + struct ggml_tensor * llm_build_inp_KQ_mask_cross() { + lctx.inp_KQ_mask_cross = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_outputs_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + ggml_set_input(lctx.inp_KQ_mask_cross); + cb(lctx.inp_KQ_mask_cross, "KQ_mask_cross", -1); + return lctx.inp_KQ_mask_cross; + } + struct ggml_cgraph * build_llama() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -12426,6 +12621,321 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_t5() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + if (lctx.is_encoding) { + struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm_enc, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_enc, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_enc, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_enc, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); + + struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; + struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b); + struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); + cb(kq_b, "kq_b", il); + + kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + + struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens))); + cb(v, "v", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); + + ggml_build_forward_expand(gf, cur); + + cur = ggml_mul_mat(ctx0, model.layers[il].wo_enc, cur); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm_enc, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up_enc, NULL, NULL, + model.layers[il].ffn_gate_enc, NULL, NULL, + model.layers[il].ffn_down_enc, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cb(cur, "result_embd", -1); + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm_enc, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + } else { + struct ggml_tensor * embd_enc = llm_build_inp_embd_enc(); + struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true); + + struct ggml_tensor * KQ_mask_dec = build_inp_KQ_mask(); + struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross(); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); + + struct ggml_tensor * k = + ggml_view_3d(ctx0, kv_self.k_l[il], + n_embd_head_k, n_kv, n_head_kv, + ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k), + 0); + cb(k, "k", il); + + struct ggml_tensor * v = + ggml_view_3d(ctx0, kv_self.v_l[il], + n_kv, n_embd_head_v, n_head_kv, + ggml_element_size(kv_self.v_l[il])*n_ctx, + ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v, + 0); + cb(v, "v", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); + + struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b; + struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b); + struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); + cb(kq_b, "kq_b", il); + + kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); + + ggml_build_forward_expand(gf, cur); + + cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); + cb(cur, "kqv_out", il); + } + + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "cross_inp", il); + + struct ggml_tensor * inpCA = cur; + + // norm + cur = llm_build_norm(ctx0, cur, hparams, + model.layers[il].attn_norm_cross, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm_cross", il); + + // cross-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_cross, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_cross, embd_enc); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_cross, embd_enc); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc); + + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); + + kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + + struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc))); + cb(v, "v", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); + + ggml_build_forward_expand(gf, cur); + + cur = ggml_mul_mat(ctx0, model.layers[il].wo_cross, cur); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cb(cur, "result_embd", -1); + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + } + + ggml_build_forward_expand(gf, cur); + + return gf; + } + struct ggml_cgraph * build_jais() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -12748,6 +13258,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_bitnet(); } break; + case LLM_ARCH_T5: + { + result = llm.build_t5(); + } break; case LLM_ARCH_JAIS: { result = llm.build_jais(); @@ -12790,6 +13304,30 @@ static void llama_set_s_copy(llama_context & lctx) { } } +static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) { + // TODO move to hparams if a T5 variant appears that uses a different value + const int64_t max_distance = 128; + + if (bidirectional) { + n_buckets >>= 1; + } + + const int64_t max_exact = n_buckets >> 1; + + int32_t relative_position = x - y; + int32_t relative_bucket = 0; + if (bidirectional) { + relative_bucket += (relative_position > 0) * n_buckets; + relative_position = abs(relative_position); + } else { + relative_position = -std::min(relative_position, 0); + } + int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact)); + relative_position_if_large = std::min(relative_position_if_large, n_buckets - 1); + relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large); + return relative_bucket; +} + static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { // // set input data @@ -12855,7 +13393,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { if (lctx.inp_KQ_mask) { // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache. - if (cparams.causal_attn) { + if (cparams.causal_attn && !lctx.is_encoding) { const int64_t n_kv = kv_self.n; const int64_t n_tokens = batch.n_tokens; @@ -12908,7 +13446,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } else { // when using kv cache, the mask needs to match the kv cache size const int64_t n_tokens = batch.n_tokens; - const int64_t n_stride = hparams.causal_attn ? kv_self.n : n_tokens; + const int64_t n_stride = hparams.causal_attn && !lctx.is_encoding ? kv_self.n : n_tokens; GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); @@ -13072,6 +13610,70 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } } + + if (lctx.inp_pos_bucket) { + const int64_t n_tokens = batch.n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_pos_bucket->buffer)); + + int32_t * data = (int32_t *) lctx.inp_pos_bucket->data; + + if (!lctx.is_encoding) { + const int64_t n_kv = kv_self.n; + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_kv; ++i) { + data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding); + } + } + } + } else { + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_tokens; ++i) { + data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(batch.pos[i], batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding); + } + } + } + } + } + + if (!lctx.is_encoding && lctx.inp_embd_enc) { + assert(lctx.inp_embd_enc->type == GGML_TYPE_F32); + assert((size_t) ggml_nelements(lctx.inp_embd_enc) == lctx.embd_enc.size()); + + ggml_backend_tensor_set(lctx.inp_embd_enc, lctx.embd_enc.data(), 0, ggml_nbytes(lctx.inp_embd_enc)); + } + + if (!lctx.is_encoding && lctx.inp_KQ_mask_cross) { + const int64_t n_output_enc = lctx.embd_enc.size() / hparams.n_embd; + const int64_t n_tokens = batch.n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_cross->buffer)); + + float * data = (float *) lctx.inp_KQ_mask_cross->data; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_output_enc; ++i) { + float f = -INFINITY; + for (int s = 0; s < batch.n_seq_id[j]; ++s) { + const llama_seq_id seq_id = batch.seq_id[j][s]; + if (lctx.seq_ids_enc[i].find(seq_id) != lctx.seq_ids_enc[i].end()) { + f = 0.0f; + } + } + data[h*(n_output_enc*n_tokens) + j*n_output_enc + i] = f; + } + } + + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_output_enc; ++j) { + data[h*(n_output_enc*n_tokens) + i*n_output_enc + j] = -INFINITY; + } + } + } + } } // Make sure enough space is available for outputs. @@ -13088,7 +13690,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { // TODO: use a per-batch flag for logits presence instead const bool has_logits = !cparams.embeddings; - const bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE); + const bool has_embd = lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE)); const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0; const size_t embd_size = has_embd ? n_embd*n_outputs_max : 0; @@ -13180,6 +13782,7 @@ static int llama_decode_internal( llama_context & lctx, llama_batch batch_all) { // TODO: rename back to batch + lctx.is_encoding = false; const uint32_t n_tokens_all = batch_all.n_tokens; if (n_tokens_all == 0) { @@ -13212,6 +13815,7 @@ static int llama_decode_internal( const auto n_ubatch = cparams.n_ubatch; + // TODO: simplify or deprecate std::vector pos; std::vector n_seq_id; std::vector seq_id_arr; @@ -13475,6 +14079,138 @@ static int llama_decode_internal( return 0; } +// encode a batch of tokens by evaluating the encoder part of the transformer +// +// - lctx: llama context +// - batch: batch to evaluate +// +// return 0 on success +// return positive int on warning +// return negative int on error +// +static int llama_encode_internal( + llama_context & lctx, + llama_batch batch) { + + lctx.is_encoding = true; + + const uint32_t n_tokens = batch.n_tokens; + + if (n_tokens == 0) { + LLAMA_LOG_ERROR("%s: n_tokens == 0", __func__); + return -1; + } + + const auto & model = lctx.model; + const auto & hparams = model.hparams; + const auto & cparams = lctx.cparams; + + GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + + // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot + GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens"); + + if (lctx.t_compute_start_us == 0) { + lctx.t_compute_start_us = ggml_time_us(); + } + + lctx.n_queued_tokens += n_tokens; + + const int64_t n_embd = hparams.n_embd; + + // TODO: simplify or deprecate + std::vector pos; + std::vector n_seq_id; + std::vector seq_id_arr; + std::vector> seq_id; + + // reserve output buffer + if (llama_output_reserve(lctx, n_tokens) < n_tokens) { + LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens); + return -2; + }; + + for (uint32_t i = 0; i < n_tokens; ++i) { + lctx.output_ids[i] = i; + } + + lctx.inp_embd_enc = NULL; + lctx.n_outputs = n_tokens; + + const int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; + GGML_ASSERT(n_threads > 0); + + // helpers for smoother batch API transition + // after deprecating the llama_eval calls, these will be removed + if (batch.pos == nullptr) { + pos.resize(n_tokens); + for (uint32_t i = 0; i < n_tokens; i++) { + pos[i] = batch.all_pos_0 + i*batch.all_pos_1; + } + + batch.pos = pos.data(); + } + + if (batch.seq_id == nullptr) { + n_seq_id.resize(n_tokens); + seq_id.resize(n_tokens); + seq_id_arr.resize(n_tokens); + for (uint32_t i = 0; i < n_tokens; i++) { + n_seq_id[i] = 1; + seq_id[i].resize(1); + seq_id[i][0] = batch.all_seq_id; + seq_id_arr[i] = seq_id[i].data(); + } + + batch.n_seq_id = n_seq_id.data(); + batch.seq_id = seq_id_arr.data(); + } + + ggml_backend_sched_reset(lctx.sched); + ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); + + ggml_cgraph * gf = llama_build_graph(lctx, batch, false); + + // the output embeddings after the final encoder normalization + struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 1]; + + GGML_ASSERT(strcmp(embd->name, "result_norm") == 0); + + ggml_backend_sched_alloc_graph(lctx.sched, gf); + + llama_set_inputs(lctx, batch); + + llama_graph_compute(lctx, gf, n_threads); + + // extract embeddings + if (embd) { + ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd); + GGML_ASSERT(backend_embd != nullptr); + + // extract token embeddings + GGML_ASSERT(lctx.embd != nullptr); + + lctx.embd_enc.resize(n_tokens*n_embd); + float * embd_out = lctx.embd_enc.data(); + + ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float)); + + // remember the sequence ids used during the encoding - needed for cross attention later + lctx.seq_ids_enc.resize(n_tokens); + for (uint32_t i = 0; i < n_tokens; i++) { + for (int s = 0; s < batch.n_seq_id[i]; s++) { + llama_seq_id seq_id = batch.seq_id[i][s]; + lctx.seq_ids_enc[i].insert(seq_id); + } + } + } + + // Reset state for the next token before backend sync, to allow the CPU activities in the reset to + // overlap with device computation. + ggml_backend_sched_reset(lctx.sched); + + return 0; +} // find holes from the beginning of the KV cache and fill them by moving data from the end of the cache static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { @@ -14547,11 +15283,14 @@ struct llm_tokenizer_ugm { std::string normalized; normalize(text, &normalized); size_t input_len = normalized.size(); + if (input_len == 0) { + return; + } // initialize score_sum to -FLT_MAX so it will be always lower than sums of token scores - std::vector tokenization_results(input_len + 1, {0, 0, -FLT_MAX}); + std::vector tokenization_results(input_len + 1, {vocab.special_unk_id, 0, -FLT_MAX}); // at the beginning tokenization score is zero - tokenization_results[0] = { 0, 0, 0 }; + tokenization_results[0] = { vocab.special_unk_id, 0, 0 }; for (size_t input_offset = 0; input_offset < input_len;) { size_t prefix_offset = input_offset; @@ -14571,7 +15310,7 @@ struct llm_tokenizer_ugm { single_codepoint_token_found = true; } llama_token token_id = node->value; - const auto &token_data = vocab.id_to_token[token_id]; + const auto & token_data = vocab.id_to_token[token_id]; // we set the user-defined token scores to 0 to make them more likely to be selected // (normal token scores are log probabilities, so they are negative) @@ -16854,10 +17593,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // sanity checks // - // - qs.n_attention_wv == 0 for Mamba models - // - qs.n_attention_wv == model.hparams.n_layer for Transformer models + // - qs.n_attention_wv == 0 for Mamba models + // - qs.n_attention_wv == model.hparams.n_layer for Transformer models + // - qs.n_attention_wv == 3 * model.hparams.n_layer for Encoder-Decoder models // - GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer) && "n_attention_wv is unexpected"); + GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer || qs.n_attention_wv == 3 * (int)model.hparams.n_layer) && "n_attention_wv is unexpected"); size_t total_size_org = 0; size_t total_size_new = 0; @@ -16982,6 +17722,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s quantize &= name.find("ssm_x.weight") == std::string::npos; quantize &= name.find("ssm_dt.weight") == std::string::npos; + // do not quantize relative position bias (T5) + quantize &= name.find("attn_rel_b.weight") == std::string::npos; + enum ggml_type new_type; void * new_data; size_t new_size; @@ -18138,6 +18881,17 @@ struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const ch return it->second; } +bool llama_model_has_encoder(const struct llama_model * model) { + switch (model->arch) { + case LLM_ARCH_T5: return true; + default: return false; + } +} + +llama_token llama_model_decoder_start_token(const struct llama_model * model) { + return model->hparams.dec_start_token_id; +} + uint32_t llama_model_quantize( const char * fname_inp, const char * fname_out, @@ -19484,6 +20238,17 @@ void llama_batch_free(struct llama_batch batch) { if (batch.logits) free(batch.logits); } +int32_t llama_encode( + struct llama_context * ctx, + struct llama_batch batch) { + const int ret = llama_encode_internal(*ctx, batch); + if (ret < 0) { + LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret); + } + + return ret; +} + int32_t llama_decode( struct llama_context * ctx, struct llama_batch batch) { From b0a46993dfbf8b8127598f319d4dcfdd83824ba8 Mon Sep 17 00:00:00 2001 From: ditsuke Date: Tue, 27 Feb 2024 12:01:02 +0530 Subject: [PATCH 29/38] build(python): Package scripts with pip-0517 compliance --- .gitignore | 11 +- __init__.py | 0 ...ert-hf-to-gguf.py => convert_hf_to_gguf.py | 0 ...-update.py => convert_hf_to_gguf_update.py | 20 +- ...o-gguf.py => convert_llama_ggml_to_gguf.py | 0 convert_lora_to_ggml.py | 149 ++ convert_persimmon_to_gguf.py | 137 ++ poetry.lock | 1197 +++++++++++++++++ pyproject.toml | 44 + 9 files changed, 1546 insertions(+), 12 deletions(-) create mode 100644 __init__.py rename convert-hf-to-gguf.py => convert_hf_to_gguf.py (100%) rename convert-hf-to-gguf-update.py => convert_hf_to_gguf_update.py (94%) mode change 100755 => 100644 rename convert-llama-ggml-to-gguf.py => convert_llama_ggml_to_gguf.py (100%) create mode 100755 convert_lora_to_ggml.py create mode 100755 convert_persimmon_to_gguf.py create mode 100644 poetry.lock create mode 100644 pyproject.toml diff --git a/.gitignore b/.gitignore index 177e6a8db..7a63cda19 100644 --- a/.gitignore +++ b/.gitignore @@ -98,13 +98,14 @@ examples/server/*.mjs.hpp # Python -__pycache__ -.venv -/Pipfile -dist -poetry.lock +/.venv +/__pycache__/ +*/poetry.lock poetry.toml +# Nix +/result + # Test binaries /tests/test-backend-ops /tests/test-double-float diff --git a/__init__.py b/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/convert-hf-to-gguf.py b/convert_hf_to_gguf.py similarity index 100% rename from convert-hf-to-gguf.py rename to convert_hf_to_gguf.py diff --git a/convert-hf-to-gguf-update.py b/convert_hf_to_gguf_update.py old mode 100755 new mode 100644 similarity index 94% rename from convert-hf-to-gguf-update.py rename to convert_hf_to_gguf_update.py index 21a306255..ca337af2c --- a/convert-hf-to-gguf-update.py +++ b/convert_hf_to_gguf_update.py @@ -50,7 +50,7 @@ class TOKENIZER_TYPE(IntEnum): # TODO: this string has to exercise as much pre-tokenizer functionality as possible # will be updated with time - contributions welcome -chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL' +chktxt = "\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````\"\"\"\"......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL" if len(sys.argv) == 2: token = sys.argv[1] @@ -99,7 +99,7 @@ def download_file_with_auth(url, token, save_path): response = sess.get(url, headers=headers) response.raise_for_status() os.makedirs(os.path.dirname(save_path), exist_ok=True) - with open(save_path, 'wb') as f: + with open(save_path, "wb") as f: f.write(response.content) logger.info(f"File {save_path} downloaded successfully") @@ -156,7 +156,9 @@ for model in models: else: tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") except OSError as e: - logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}") + logger.error( + f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}" + ) continue # Skip to the next model if the tokenizer can't be loaded chktok = tokenizer.encode(chktxt) @@ -176,13 +178,15 @@ for model in models: pre_tokenizer = cfg["pre_tokenizer"] logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4)) if "ignore_merges" in cfg["model"]: - logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4)) + logger.info( + "ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4) + ) logger.info("") - src_ifs += f" if chkhsh == \"{chkhsh}\":\n" + src_ifs += f' if chkhsh == "{chkhsh}":\n' src_ifs += f" # ref: {model['repo']}\n" - src_ifs += f" res = \"{name}\"\n" + src_ifs += f' res = "{name}"\n' src_func = f""" def get_vocab_base_pre(self, tokenizer) -> str: @@ -343,6 +347,8 @@ logger.info("\nRun the following commands to generate the vocab files for testin for model in models: name = model["name"] - print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100 + print( + f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only" + ) # noqa: NP100 logger.info("\n") diff --git a/convert-llama-ggml-to-gguf.py b/convert_llama_ggml_to_gguf.py similarity index 100% rename from convert-llama-ggml-to-gguf.py rename to convert_llama_ggml_to_gguf.py diff --git a/convert_lora_to_ggml.py b/convert_lora_to_ggml.py new file mode 100755 index 000000000..276d0d63a --- /dev/null +++ b/convert_lora_to_ggml.py @@ -0,0 +1,149 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import json +import os +import struct +import sys +from pathlib import Path +from typing import Any, BinaryIO, Sequence + +import numpy as np +import torch + +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) +import gguf + +NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1} + + +def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None: + fout.write(b"ggla"[::-1]) # magic (ggml lora) + fout.write(struct.pack("i", 1)) # file version + fout.write(struct.pack("i", params["r"])) + # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int + # but some models ship a float value instead + # let's convert to int, but fail if lossless conversion is not possible + assert ( + int(params["lora_alpha"]) == params["lora_alpha"] + ), "cannot convert float to int losslessly" + fout.write(struct.pack("i", int(params["lora_alpha"]))) + + +def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None: + sname = name.encode("utf-8") + fout.write( + struct.pack( + "iii", + len(shape), + len(sname), + NUMPY_TYPE_TO_FTYPE[data_type.name], + ) + ) + fout.write(struct.pack("i" * len(shape), *shape[::-1])) + fout.write(sname) + fout.seek((fout.tell() + 31) & -32) + + +if __name__ == '__main__': + if len(sys.argv) < 2: + print(f"Usage: python {sys.argv[0]} [arch]") + print( + "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'" + ) + print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)") + sys.exit(1) + + input_json = os.path.join(sys.argv[1], "adapter_config.json") + input_model = os.path.join(sys.argv[1], "adapter_model.bin") + output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin") + + if os.path.exists(input_model): + model = torch.load(input_model, map_location="cpu") + else: + input_model = os.path.join(sys.argv[1], "adapter_model.safetensors") + # lazy import load_file only if lora is in safetensors format. + from safetensors.torch import load_file + model = load_file(input_model, device="cpu") + + arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama" + + if arch_name not in gguf.MODEL_ARCH_NAMES.values(): + print(f"Error: unsupported architecture {arch_name}") + sys.exit(1) + + arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)] + name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone + + with open(input_json, "r") as f: + params = json.load(f) + + if params["peft_type"] != "LORA": + print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA") + sys.exit(1) + + if params["fan_in_fan_out"] is True: + print("Error: param fan_in_fan_out is not supported") + sys.exit(1) + + if params["bias"] is not None and params["bias"] != "none": + print("Error: param bias is not supported") + sys.exit(1) + + # TODO: these seem to be layers that have been trained but without lora. + # doesn't seem widely used but eventually should be supported + if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0: + print("Error: param modules_to_save is not supported") + sys.exit(1) + + with open(output_path, "wb") as fout: + fout.truncate() + + write_file_header(fout, params) + for k, v in model.items(): + orig_k = k + if k.endswith(".default.weight"): + k = k.replace(".default.weight", ".weight") + if k in ["llama_proj.weight", "llama_proj.bias"]: + continue + if k.endswith("lora_A.weight"): + if v.dtype != torch.float16 and v.dtype != torch.float32: + v = v.float() + v = v.T + else: + v = v.float() + + t = v.detach().numpy() + + prefix = "base_model.model." + if k.startswith(prefix): + k = k[len(prefix) :] + + lora_suffixes = (".lora_A.weight", ".lora_B.weight") + if k.endswith(lora_suffixes): + suffix = k[-len(lora_suffixes[0]):] + k = k[: -len(lora_suffixes[0])] + else: + print(f"Error: unrecognized tensor name {orig_k}") + sys.exit(1) + + tname = name_map.get_name(k) + if tname is None: + print(f"Error: could not map tensor name {orig_k}") + print(" Note: the arch parameter must be specified if the model is not llama") + sys.exit(1) + + if suffix == ".lora_A.weight": + tname += ".weight.loraA" + elif suffix == ".lora_B.weight": + tname += ".weight.loraB" + else: + assert False + + print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB") + write_tensor_header(fout, tname, t.shape, t.dtype) + t.tofile(fout) + + print(f"Converted {input_json} and {input_model} to {output_path}") + diff --git a/convert_persimmon_to_gguf.py b/convert_persimmon_to_gguf.py new file mode 100755 index 000000000..e1fe3c521 --- /dev/null +++ b/convert_persimmon_to_gguf.py @@ -0,0 +1,137 @@ +#!/usr/bin/env python3 +import argparse +import os +import sys +from pathlib import Path +from pprint import pprint + +import torch +from sentencepiece import SentencePieceProcessor + +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) +import gguf + + +def _flatten_dict(dct, tensors, prefix=None): + assert isinstance(dct, dict) + for key in dct.keys(): + new_prefix = prefix + '.' + key if prefix is not None else key + if isinstance(dct[key], torch.Tensor): + tensors[new_prefix] = dct[key] + elif isinstance(dct[key], dict): + _flatten_dict(dct[key], tensors, new_prefix) + else: + raise ValueError(type(dct[key])) + return None + + +def _get_sentencepiece_tokenizer_info(dir_model: Path): + tokenizer_path = dir_model / 'adept_vocab.model' + print('gguf: getting sentencepiece tokenizer from', tokenizer_path) + tokenizer = SentencePieceProcessor(str(tokenizer_path)) + print('gguf: adding tokens') + tokens: list[bytes] = [] + scores: list[float] = [] + toktypes: list[int] = [] + + for i in range(tokenizer.vocab_size()): + text: bytes + score: float + + piece = tokenizer.id_to_piece(i) + text = piece.encode("utf-8") + score = tokenizer.get_score(i) + + toktype = 1 + if tokenizer.is_unknown(i): + toktype = 2 + if tokenizer.is_control(i): + toktype = 3 + if tokenizer.is_unused(i): + toktype = 5 + if tokenizer.is_byte(i): + toktype = 6 + + tokens.append(text) + scores.append(score) + toktypes.append(toktype) + pass + return tokens, scores, toktypes + + +def main(): + parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("--ckpt-path", type=Path, help="path to persimmon checkpoint .pt file") + parser.add_argument("--model-dir", type=Path, help="directory containing model e.g. 8b_chat_model_release") + parser.add_argument("--adept-inference-dir", type=str, help="path to adept-inference code directory") + args = parser.parse_args() + sys.path.append(str(args.adept_inference_dir)) + persimmon_model = torch.load(args.ckpt_path) + hparams = persimmon_model['args'] + pprint(hparams) + tensors: dict[str, torch.Tensor] = {} + _flatten_dict(persimmon_model['model'], tensors, None) + + arch = gguf.MODEL_ARCH.PERSIMMON + gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch]) + + block_count = hparams.num_layers + head_count = hparams.num_attention_heads + head_count_kv = head_count + ctx_length = hparams.seq_length + hidden_size = hparams.hidden_size + + gguf_writer.add_name('persimmon-8b-chat') + gguf_writer.add_context_length(ctx_length) + gguf_writer.add_embedding_length(hidden_size) + gguf_writer.add_block_count(block_count) + gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size) + # ref: https://github.com/ggerganov/llama.cpp/pull/4889/commits/eea19039fc52ea2dbd1aab45b59ab4e3e29a3443 + gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2) + gguf_writer.add_head_count(head_count) + gguf_writer.add_head_count_kv(head_count_kv) + gguf_writer.add_rope_freq_base(hparams.rotary_emb_base) + gguf_writer.add_layer_norm_eps(hparams.layernorm_epsilon) + + tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir) + gguf_writer.add_tokenizer_model('llama') + gguf_writer.add_token_list(tokens) + gguf_writer.add_token_scores(scores) + gguf_writer.add_token_types(toktypes) + gguf_writer.add_bos_token_id(71013) + gguf_writer.add_eos_token_id(71013) + + tensor_map = gguf.get_tensor_name_map(arch, block_count) + print(tensor_map) + for name in tensors.keys(): + data = tensors[name] + if name.endswith(".self_attention.rotary_emb.inv_freq"): + continue + old_dtype = data.dtype + # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?) + data = data.to(torch.float32).squeeze().numpy() + new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) + if new_name is None: + print("Can not map tensor '" + name + "'") + sys.exit() + n_dims = len(data.shape) + print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + gguf_writer.add_tensor(new_name, data) + print("gguf: write header") + gguf_writer.write_header_to_file() + print("gguf: write metadata") + gguf_writer.write_kv_data_to_file() + print("gguf: write tensors") + gguf_writer.write_tensors_to_file() + + gguf_writer.close() + + print(f"gguf: model successfully exported to '{args.outfile}'") + print("") + + +if __name__ == '__main__': + main() + diff --git a/poetry.lock b/poetry.lock new file mode 100644 index 000000000..eb6baa6c7 --- /dev/null +++ b/poetry.lock @@ -0,0 +1,1197 @@ +# This file is automatically @generated by Poetry 1.7.1 and should not be changed by hand. + +[[package]] +name = "atomicwrites" +version = "1.4.1" +description = "Atomic file writes." +optional = false +python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*" +files = [ + {file = "atomicwrites-1.4.1.tar.gz", hash = "sha256:81b2c9071a49367a7f770170e5eec8cb66567cfbbc8c73d20ce5ca4a8d71cf11"}, +] + +[[package]] +name = "attrs" +version = "23.2.0" +description = "Classes Without Boilerplate" +optional = false +python-versions = ">=3.7" +files = [ + {file = "attrs-23.2.0-py3-none-any.whl", hash = "sha256:99b87a485a5820b23b879f04c2305b44b951b502fd64be915879d77a7e8fc6f1"}, + {file = "attrs-23.2.0.tar.gz", hash = "sha256:935dc3b529c262f6cf76e50877d35a4bd3c1de194fd41f47a2b7ae8f19971f30"}, +] + +[package.extras] +cov = ["attrs[tests]", "coverage[toml] (>=5.3)"] +dev = ["attrs[tests]", "pre-commit"] +docs = ["furo", "myst-parser", "sphinx", "sphinx-notfound-page", "sphinxcontrib-towncrier", "towncrier", "zope-interface"] +tests = ["attrs[tests-no-zope]", "zope-interface"] +tests-mypy = ["mypy (>=1.6)", "pytest-mypy-plugins"] +tests-no-zope = ["attrs[tests-mypy]", "cloudpickle", "hypothesis", "pympler", "pytest (>=4.3.0)", "pytest-xdist[psutil]"] + +[[package]] +name = "certifi" +version = "2024.2.2" +description = "Python package for providing Mozilla's CA Bundle." +optional = false +python-versions = ">=3.6" +files = [ + {file = "certifi-2024.2.2-py3-none-any.whl", hash = "sha256:dc383c07b76109f368f6106eee2b593b04a011ea4d55f652c6ca24a754d1cdd1"}, + {file = "certifi-2024.2.2.tar.gz", hash = "sha256:0569859f95fc761b18b45ef421b1290a0f65f147e92a1e5eb3e635f9a5e4e66f"}, +] + +[[package]] +name = "charset-normalizer" +version = "3.3.2" +description = "The Real First Universal Charset Detector. Open, modern and actively maintained alternative to Chardet." +optional = false +python-versions = ">=3.7.0" +files = [ + {file = "charset-normalizer-3.3.2.tar.gz", hash = "sha256:f30c3cb33b24454a82faecaf01b19c18562b1e89558fb6c56de4d9118a032fd5"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:25baf083bf6f6b341f4121c2f3c548875ee6f5339300e08be3f2b2ba1721cdd3"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:06435b539f889b1f6f4ac1758871aae42dc3a8c0e24ac9e60c2384973ad73027"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:9063e24fdb1e498ab71cb7419e24622516c4a04476b17a2dab57e8baa30d6e03"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6897af51655e3691ff853668779c7bad41579facacf5fd7253b0133308cf000d"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1d3193f4a680c64b4b6a9115943538edb896edc190f0b222e73761716519268e"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:cd70574b12bb8a4d2aaa0094515df2463cb429d8536cfb6c7ce983246983e5a6"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8465322196c8b4d7ab6d1e049e4c5cb460d0394da4a27d23cc242fbf0034b6b5"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:a9a8e9031d613fd2009c182b69c7b2c1ef8239a0efb1df3f7c8da66d5dd3d537"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:beb58fe5cdb101e3a055192ac291b7a21e3b7ef4f67fa1d74e331a7f2124341c"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:e06ed3eb3218bc64786f7db41917d4e686cc4856944f53d5bdf83a6884432e12"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:2e81c7b9c8979ce92ed306c249d46894776a909505d8f5a4ba55b14206e3222f"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_s390x.whl", hash = "sha256:572c3763a264ba47b3cf708a44ce965d98555f618ca42c926a9c1616d8f34269"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:fd1abc0d89e30cc4e02e4064dc67fcc51bd941eb395c502aac3ec19fab46b519"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-win32.whl", hash = "sha256:3d47fa203a7bd9c5b6cee4736ee84ca03b8ef23193c0d1ca99b5089f72645c73"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-win_amd64.whl", hash = "sha256:10955842570876604d404661fbccbc9c7e684caf432c09c715ec38fbae45ae09"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:802fe99cca7457642125a8a88a084cef28ff0cf9407060f7b93dca5aa25480db"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:573f6eac48f4769d667c4442081b1794f52919e7edada77495aaed9236d13a96"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:549a3a73da901d5bc3ce8d24e0600d1fa85524c10287f6004fbab87672bf3e1e"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f27273b60488abe721a075bcca6d7f3964f9f6f067c8c4c605743023d7d3944f"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1ceae2f17a9c33cb48e3263960dc5fc8005351ee19db217e9b1bb15d28c02574"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:65f6f63034100ead094b8744b3b97965785388f308a64cf8d7c34f2f2e5be0c4"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:753f10e867343b4511128c6ed8c82f7bec3bd026875576dfd88483c5c73b2fd8"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4a78b2b446bd7c934f5dcedc588903fb2f5eec172f3d29e52a9096a43722adfc"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:e537484df0d8f426ce2afb2d0f8e1c3d0b114b83f8850e5f2fbea0e797bd82ae"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:eb6904c354526e758fda7167b33005998fb68c46fbc10e013ca97f21ca5c8887"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:deb6be0ac38ece9ba87dea880e438f25ca3eddfac8b002a2ec3d9183a454e8ae"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_s390x.whl", hash = "sha256:4ab2fe47fae9e0f9dee8c04187ce5d09f48eabe611be8259444906793ab7cbce"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:80402cd6ee291dcb72644d6eac93785fe2c8b9cb30893c1af5b8fdd753b9d40f"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-win32.whl", hash = "sha256:7cd13a2e3ddeed6913a65e66e94b51d80a041145a026c27e6bb76c31a853c6ab"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-win_amd64.whl", hash = "sha256:663946639d296df6a2bb2aa51b60a2454ca1cb29835324c640dafb5ff2131a77"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:0b2b64d2bb6d3fb9112bafa732def486049e63de9618b5843bcdd081d8144cd8"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:ddbb2551d7e0102e7252db79ba445cdab71b26640817ab1e3e3648dad515003b"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:55086ee1064215781fff39a1af09518bc9255b50d6333f2e4c74ca09fac6a8f6"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8f4a014bc36d3c57402e2977dada34f9c12300af536839dc38c0beab8878f38a"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a10af20b82360ab00827f916a6058451b723b4e65030c5a18577c8b2de5b3389"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:8d756e44e94489e49571086ef83b2bb8ce311e730092d2c34ca8f7d925cb20aa"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:90d558489962fd4918143277a773316e56c72da56ec7aa3dc3dbbe20fdfed15b"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6ac7ffc7ad6d040517be39eb591cac5ff87416c2537df6ba3cba3bae290c0fed"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:7ed9e526742851e8d5cc9e6cf41427dfc6068d4f5a3bb03659444b4cabf6bc26"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:8bdb58ff7ba23002a4c5808d608e4e6c687175724f54a5dade5fa8c67b604e4d"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_ppc64le.whl", hash = "sha256:6b3251890fff30ee142c44144871185dbe13b11bab478a88887a639655be1068"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_s390x.whl", hash = "sha256:b4a23f61ce87adf89be746c8a8974fe1c823c891d8f86eb218bb957c924bb143"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:efcb3f6676480691518c177e3b465bcddf57cea040302f9f4e6e191af91174d4"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-win32.whl", hash = "sha256:d965bba47ddeec8cd560687584e88cf699fd28f192ceb452d1d7ee807c5597b7"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-win_amd64.whl", hash = "sha256:96b02a3dc4381e5494fad39be677abcb5e6634bf7b4fa83a6dd3112607547001"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:95f2a5796329323b8f0512e09dbb7a1860c46a39da62ecb2324f116fa8fdc85c"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c002b4ffc0be611f0d9da932eb0f704fe2602a9a949d1f738e4c34c75b0863d5"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a981a536974bbc7a512cf44ed14938cf01030a99e9b3a06dd59578882f06f985"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3287761bc4ee9e33561a7e058c72ac0938c4f57fe49a09eae428fd88aafe7bb6"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:42cb296636fcc8b0644486d15c12376cb9fa75443e00fb25de0b8602e64c1714"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:0a55554a2fa0d408816b3b5cedf0045f4b8e1a6065aec45849de2d6f3f8e9786"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:c083af607d2515612056a31f0a8d9e0fcb5876b7bfc0abad3ecd275bc4ebc2d5"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:87d1351268731db79e0f8e745d92493ee2841c974128ef629dc518b937d9194c"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:bd8f7df7d12c2db9fab40bdd87a7c09b1530128315d047a086fa3ae3435cb3a8"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_s390x.whl", hash = "sha256:c180f51afb394e165eafe4ac2936a14bee3eb10debc9d9e4db8958fe36afe711"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:8c622a5fe39a48f78944a87d4fb8a53ee07344641b0562c540d840748571b811"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-win32.whl", hash = "sha256:db364eca23f876da6f9e16c9da0df51aa4f104a972735574842618b8c6d999d4"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-win_amd64.whl", hash = "sha256:86216b5cee4b06df986d214f664305142d9c76df9b6512be2738aa72a2048f99"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:6463effa3186ea09411d50efc7d85360b38d5f09b870c48e4600f63af490e56a"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:6c4caeef8fa63d06bd437cd4bdcf3ffefe6738fb1b25951440d80dc7df8c03ac"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:37e55c8e51c236f95b033f6fb391d7d7970ba5fe7ff453dad675e88cf303377a"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:fb69256e180cb6c8a894fee62b3afebae785babc1ee98b81cdf68bbca1987f33"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ae5f4161f18c61806f411a13b0310bea87f987c7d2ecdbdaad0e94eb2e404238"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:b2b0a0c0517616b6869869f8c581d4eb2dd83a4d79e0ebcb7d373ef9956aeb0a"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:45485e01ff4d3630ec0d9617310448a8702f70e9c01906b0d0118bdf9d124cf2"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:eb00ed941194665c332bf8e078baf037d6c35d7c4f3102ea2d4f16ca94a26dc8"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:2127566c664442652f024c837091890cb1942c30937add288223dc895793f898"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:a50aebfa173e157099939b17f18600f72f84eed3049e743b68ad15bd69b6bf99"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:4d0d1650369165a14e14e1e47b372cfcb31d6ab44e6e33cb2d4e57265290044d"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_s390x.whl", hash = "sha256:923c0c831b7cfcb071580d3f46c4baf50f174be571576556269530f4bbd79d04"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:06a81e93cd441c56a9b65d8e1d043daeb97a3d0856d177d5c90ba85acb3db087"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-win32.whl", hash = "sha256:6ef1d82a3af9d3eecdba2321dc1b3c238245d890843e040e41e470ffa64c3e25"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-win_amd64.whl", hash = "sha256:eb8821e09e916165e160797a6c17edda0679379a4be5c716c260e836e122f54b"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:c235ebd9baae02f1b77bcea61bce332cb4331dc3617d254df3323aa01ab47bd4"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:5b4c145409bef602a690e7cfad0a15a55c13320ff7a3ad7ca59c13bb8ba4d45d"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:68d1f8a9e9e37c1223b656399be5d6b448dea850bed7d0f87a8311f1ff3dabb0"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:22afcb9f253dac0696b5a4be4a1c0f8762f8239e21b99680099abd9b2b1b2269"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e27ad930a842b4c5eb8ac0016b0a54f5aebbe679340c26101df33424142c143c"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:1f79682fbe303db92bc2b1136016a38a42e835d932bab5b3b1bfcfbf0640e519"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b261ccdec7821281dade748d088bb6e9b69e6d15b30652b74cbbac25e280b796"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:122c7fa62b130ed55f8f285bfd56d5f4b4a5b503609d181f9ad85e55c89f4185"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:d0eccceffcb53201b5bfebb52600a5fb483a20b61da9dbc885f8b103cbe7598c"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:9f96df6923e21816da7e0ad3fd47dd8f94b2a5ce594e00677c0013018b813458"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:7f04c839ed0b6b98b1a7501a002144b76c18fb1c1850c8b98d458ac269e26ed2"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_s390x.whl", hash = "sha256:34d1c8da1e78d2e001f363791c98a272bb734000fcef47a491c1e3b0505657a8"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:ff8fa367d09b717b2a17a052544193ad76cd49979c805768879cb63d9ca50561"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-win32.whl", hash = "sha256:aed38f6e4fb3f5d6bf81bfa990a07806be9d83cf7bacef998ab1a9bd660a581f"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-win_amd64.whl", hash = "sha256:b01b88d45a6fcb69667cd6d2f7a9aeb4bf53760d7fc536bf679ec94fe9f3ff3d"}, + {file = "charset_normalizer-3.3.2-py3-none-any.whl", hash = "sha256:3e4d1f6587322d2788836a99c69062fbb091331ec940e02d12d179c1d53e25fc"}, +] + +[[package]] +name = "colorama" +version = "0.4.6" +description = "Cross-platform colored terminal text." +optional = false +python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7" +files = [ + {file = "colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6"}, + {file = "colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44"}, +] + +[[package]] +name = "filelock" +version = "3.13.1" +description = "A platform independent file lock." +optional = false +python-versions = ">=3.8" +files = [ + {file = "filelock-3.13.1-py3-none-any.whl", hash = "sha256:57dbda9b35157b05fb3e58ee91448612eb674172fab98ee235ccb0b5bee19a1c"}, + {file = "filelock-3.13.1.tar.gz", hash = "sha256:521f5f56c50f8426f5e03ad3b281b490a87ef15bc6c526f168290f0c7148d44e"}, +] + +[package.extras] +docs = ["furo (>=2023.9.10)", "sphinx (>=7.2.6)", "sphinx-autodoc-typehints (>=1.24)"] +testing = ["covdefaults (>=2.3)", "coverage (>=7.3.2)", "diff-cover (>=8)", "pytest (>=7.4.3)", "pytest-cov (>=4.1)", "pytest-mock (>=3.12)", "pytest-timeout (>=2.2)"] +typing = ["typing-extensions (>=4.8)"] + +[[package]] +name = "fsspec" +version = "2024.2.0" +description = "File-system specification" +optional = false +python-versions = ">=3.8" +files = [ + {file = "fsspec-2024.2.0-py3-none-any.whl", hash = "sha256:817f969556fa5916bc682e02ca2045f96ff7f586d45110fcb76022063ad2c7d8"}, + {file = "fsspec-2024.2.0.tar.gz", hash = "sha256:b6ad1a679f760dda52b1168c859d01b7b80648ea6f7f7c7f5a8a91dc3f3ecb84"}, +] + +[package.extras] +abfs = ["adlfs"] +adl = ["adlfs"] +arrow = ["pyarrow (>=1)"] +dask = ["dask", "distributed"] +devel = ["pytest", "pytest-cov"] +dropbox = ["dropbox", "dropboxdrivefs", "requests"] +full = ["adlfs", "aiohttp (!=4.0.0a0,!=4.0.0a1)", "dask", "distributed", "dropbox", "dropboxdrivefs", "fusepy", "gcsfs", "libarchive-c", "ocifs", "panel", "paramiko", "pyarrow (>=1)", "pygit2", "requests", "s3fs", "smbprotocol", "tqdm"] +fuse = ["fusepy"] +gcs = ["gcsfs"] +git = ["pygit2"] +github = ["requests"] +gs = ["gcsfs"] +gui = ["panel"] +hdfs = ["pyarrow (>=1)"] +http = ["aiohttp (!=4.0.0a0,!=4.0.0a1)"] +libarchive = ["libarchive-c"] +oci = ["ocifs"] +s3 = ["s3fs"] +sftp = ["paramiko"] +smb = ["smbprotocol"] +ssh = ["paramiko"] +tqdm = ["tqdm"] + +[[package]] +name = "gguf" +version = "0.7.0" +description = "Read and write ML models in GGUF for GGML" +optional = false +python-versions = ">=3.8" +files = [] +develop = false + +[package.dependencies] +numpy = ">=1.17" + +[package.source] +type = "directory" +url = "gguf-py" + +[[package]] +name = "huggingface-hub" +version = "0.20.3" +description = "Client library to download and publish models, datasets and other repos on the huggingface.co hub" +optional = false +python-versions = ">=3.8.0" +files = [ + {file = "huggingface_hub-0.20.3-py3-none-any.whl", hash = "sha256:d988ae4f00d3e307b0c80c6a05ca6dbb7edba8bba3079f74cda7d9c2e562a7b6"}, + {file = "huggingface_hub-0.20.3.tar.gz", hash = "sha256:94e7f8e074475fbc67d6a71957b678e1b4a74ff1b64a644fd6cbb83da962d05d"}, +] + +[package.dependencies] +filelock = "*" +fsspec = ">=2023.5.0" +packaging = ">=20.9" +pyyaml = ">=5.1" +requests = "*" +tqdm = ">=4.42.1" +typing-extensions = ">=3.7.4.3" + +[package.extras] +all = ["InquirerPy (==0.3.4)", "Jinja2", "Pillow", "aiohttp", "gradio", "jedi", "mypy (==1.5.1)", "numpy", "pydantic (>1.1,<2.0)", "pydantic (>1.1,<3.0)", "pytest", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-rerunfailures", "pytest-vcr", "pytest-xdist", "ruff (>=0.1.3)", "soundfile", "types-PyYAML", "types-requests", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)", "urllib3 (<2.0)"] +cli = ["InquirerPy (==0.3.4)"] +dev = ["InquirerPy (==0.3.4)", "Jinja2", "Pillow", "aiohttp", "gradio", "jedi", "mypy (==1.5.1)", "numpy", "pydantic (>1.1,<2.0)", "pydantic (>1.1,<3.0)", "pytest", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-rerunfailures", "pytest-vcr", "pytest-xdist", "ruff (>=0.1.3)", "soundfile", "types-PyYAML", "types-requests", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)", "urllib3 (<2.0)"] +fastai = ["fastai (>=2.4)", "fastcore (>=1.3.27)", "toml"] +inference = ["aiohttp", "pydantic (>1.1,<2.0)", "pydantic (>1.1,<3.0)"] +quality = ["mypy (==1.5.1)", "ruff (>=0.1.3)"] +tensorflow = ["graphviz", "pydot", "tensorflow"] +testing = ["InquirerPy (==0.3.4)", "Jinja2", "Pillow", "aiohttp", "gradio", "jedi", "numpy", "pydantic (>1.1,<2.0)", "pydantic (>1.1,<3.0)", "pytest", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-rerunfailures", "pytest-vcr", "pytest-xdist", "soundfile", "urllib3 (<2.0)"] +torch = ["torch"] +typing = ["types-PyYAML", "types-requests", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)"] + +[[package]] +name = "idna" +version = "3.6" +description = "Internationalized Domain Names in Applications (IDNA)" +optional = false +python-versions = ">=3.5" +files = [ + {file = "idna-3.6-py3-none-any.whl", hash = "sha256:c05567e9c24a6b9faaa835c4821bad0590fbb9d5779e7caa6e1cc4978e7eb24f"}, + {file = "idna-3.6.tar.gz", hash = "sha256:9ecdbbd083b06798ae1e86adcbfe8ab1479cf864e4ee30fe4e46a003d12491ca"}, +] + +[[package]] +name = "jinja2" +version = "3.1.3" +description = "A very fast and expressive template engine." +optional = false +python-versions = ">=3.7" +files = [ + {file = "Jinja2-3.1.3-py3-none-any.whl", hash = "sha256:7d6d50dd97d52cbc355597bd845fabfbac3f551e1f99619e39a35ce8c370b5fa"}, + {file = "Jinja2-3.1.3.tar.gz", hash = "sha256:ac8bd6544d4bb2c9792bf3a159e80bba8fda7f07e81bc3aed565432d5925ba90"}, +] + +[package.dependencies] +MarkupSafe = ">=2.0" + +[package.extras] +i18n = ["Babel (>=2.7)"] + +[[package]] +name = "markupsafe" +version = "2.1.5" +description = "Safely add untrusted strings to HTML/XML markup." +optional = false +python-versions = ">=3.7" +files = [ + {file = "MarkupSafe-2.1.5-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:a17a92de5231666cfbe003f0e4b9b3a7ae3afb1ec2845aadc2bacc93ff85febc"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:72b6be590cc35924b02c78ef34b467da4ba07e4e0f0454a2c5907f473fc50ce5"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e61659ba32cf2cf1481e575d0462554625196a1f2fc06a1c777d3f48e8865d46"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2174c595a0d73a3080ca3257b40096db99799265e1c27cc5a610743acd86d62f"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ae2ad8ae6ebee9d2d94b17fb62763125f3f374c25618198f40cbb8b525411900"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:075202fa5b72c86ad32dc7d0b56024ebdbcf2048c0ba09f1cde31bfdd57bcfff"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:598e3276b64aff0e7b3451b72e94fa3c238d452e7ddcd893c3ab324717456bad"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:fce659a462a1be54d2ffcacea5e3ba2d74daa74f30f5f143fe0c58636e355fdd"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-win32.whl", hash = "sha256:d9fad5155d72433c921b782e58892377c44bd6252b5af2f67f16b194987338a4"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-win_amd64.whl", hash = "sha256:bf50cd79a75d181c9181df03572cdce0fbb75cc353bc350712073108cba98de5"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:629ddd2ca402ae6dbedfceeba9c46d5f7b2a61d9749597d4307f943ef198fc1f"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:5b7b716f97b52c5a14bffdf688f971b2d5ef4029127f1ad7a513973cfd818df2"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6ec585f69cec0aa07d945b20805be741395e28ac1627333b1c5b0105962ffced"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b91c037585eba9095565a3556f611e3cbfaa42ca1e865f7b8015fe5c7336d5a5"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:7502934a33b54030eaf1194c21c692a534196063db72176b0c4028e140f8f32c"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:0e397ac966fdf721b2c528cf028494e86172b4feba51d65f81ffd65c63798f3f"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:c061bb86a71b42465156a3ee7bd58c8c2ceacdbeb95d05a99893e08b8467359a"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:3a57fdd7ce31c7ff06cdfbf31dafa96cc533c21e443d57f5b1ecc6cdc668ec7f"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-win32.whl", hash = "sha256:397081c1a0bfb5124355710fe79478cdbeb39626492b15d399526ae53422b906"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-win_amd64.whl", hash = "sha256:2b7c57a4dfc4f16f7142221afe5ba4e093e09e728ca65c51f5620c9aaeb9a617"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:8dec4936e9c3100156f8a2dc89c4b88d5c435175ff03413b443469c7c8c5f4d1"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:3c6b973f22eb18a789b1460b4b91bf04ae3f0c4234a0a6aa6b0a92f6f7b951d4"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ac07bad82163452a6884fe8fa0963fb98c2346ba78d779ec06bd7a6262132aee"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f5dfb42c4604dddc8e4305050aa6deb084540643ed5804d7455b5df8fe16f5e5"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ea3d8a3d18833cf4304cd2fc9cbb1efe188ca9b5efef2bdac7adc20594a0e46b"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:d050b3361367a06d752db6ead6e7edeb0009be66bc3bae0ee9d97fb326badc2a"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:bec0a414d016ac1a18862a519e54b2fd0fc8bbfd6890376898a6c0891dd82e9f"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:58c98fee265677f63a4385256a6d7683ab1832f3ddd1e66fe948d5880c21a169"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-win32.whl", hash = "sha256:8590b4ae07a35970728874632fed7bd57b26b0102df2d2b233b6d9d82f6c62ad"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-win_amd64.whl", hash = "sha256:823b65d8706e32ad2df51ed89496147a42a2a6e01c13cfb6ffb8b1e92bc910bb"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:c8b29db45f8fe46ad280a7294f5c3ec36dbac9491f2d1c17345be8e69cc5928f"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ec6a563cff360b50eed26f13adc43e61bc0c04d94b8be985e6fb24b81f6dcfdf"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a549b9c31bec33820e885335b451286e2969a2d9e24879f83fe904a5ce59d70a"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4f11aa001c540f62c6166c7726f71f7573b52c68c31f014c25cc7901deea0b52"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:7b2e5a267c855eea6b4283940daa6e88a285f5f2a67f2220203786dfa59b37e9"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:2d2d793e36e230fd32babe143b04cec8a8b3eb8a3122d2aceb4a371e6b09b8df"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:ce409136744f6521e39fd8e2a24c53fa18ad67aa5bc7c2cf83645cce5b5c4e50"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-win32.whl", hash = "sha256:4096e9de5c6fdf43fb4f04c26fb114f61ef0bf2e5604b6ee3019d51b69e8c371"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-win_amd64.whl", hash = "sha256:4275d846e41ecefa46e2015117a9f491e57a71ddd59bbead77e904dc02b1bed2"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:656f7526c69fac7f600bd1f400991cc282b417d17539a1b228617081106feb4a"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:97cafb1f3cbcd3fd2b6fbfb99ae11cdb14deea0736fc2b0952ee177f2b813a46"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1f3fbcb7ef1f16e48246f704ab79d79da8a46891e2da03f8783a5b6fa41a9532"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fa9db3f79de01457b03d4f01b34cf91bc0048eb2c3846ff26f66687c2f6d16ab"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ffee1f21e5ef0d712f9033568f8344d5da8cc2869dbd08d87c84656e6a2d2f68"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:5dedb4db619ba5a2787a94d877bc8ffc0566f92a01c0ef214865e54ecc9ee5e0"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:30b600cf0a7ac9234b2638fbc0fb6158ba5bdcdf46aeb631ead21248b9affbc4"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:8dd717634f5a044f860435c1d8c16a270ddf0ef8588d4887037c5028b859b0c3"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-win32.whl", hash = "sha256:daa4ee5a243f0f20d528d939d06670a298dd39b1ad5f8a72a4275124a7819eff"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-win_amd64.whl", hash = "sha256:619bc166c4f2de5caa5a633b8b7326fbe98e0ccbfacabd87268a2b15ff73a029"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:7a68b554d356a91cce1236aa7682dc01df0edba8d043fd1ce607c49dd3c1edcf"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:db0b55e0f3cc0be60c1f19efdde9a637c32740486004f20d1cff53c3c0ece4d2"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3e53af139f8579a6d5f7b76549125f0d94d7e630761a2111bc431fd820e163b8"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:17b950fccb810b3293638215058e432159d2b71005c74371d784862b7e4683f3"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4c31f53cdae6ecfa91a77820e8b151dba54ab528ba65dfd235c80b086d68a465"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:bff1b4290a66b490a2f4719358c0cdcd9bafb6b8f061e45c7a2460866bf50c2e"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:bc1667f8b83f48511b94671e0e441401371dfd0f0a795c7daa4a3cd1dde55bea"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:5049256f536511ee3f7e1b3f87d1d1209d327e818e6ae1365e8653d7e3abb6a6"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-win32.whl", hash = "sha256:00e046b6dd71aa03a41079792f8473dc494d564611a8f89bbbd7cb93295ebdcf"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-win_amd64.whl", hash = "sha256:fa173ec60341d6bb97a89f5ea19c85c5643c1e7dedebc22f5181eb73573142c5"}, + {file = "MarkupSafe-2.1.5.tar.gz", hash = "sha256:d283d37a890ba4c1ae73ffadf8046435c76e7bc2247bbb63c00bd1a709c6544b"}, +] + +[[package]] +name = "more-itertools" +version = "10.2.0" +description = "More routines for operating on iterables, beyond itertools" +optional = false +python-versions = ">=3.8" +files = [ + {file = "more-itertools-10.2.0.tar.gz", hash = "sha256:8fccb480c43d3e99a00087634c06dd02b0d50fbf088b380de5a41a015ec239e1"}, + {file = "more_itertools-10.2.0-py3-none-any.whl", hash = "sha256:686b06abe565edfab151cb8fd385a05651e1fdf8f0a14191e4439283421f8684"}, +] + +[[package]] +name = "mpmath" +version = "1.3.0" +description = "Python library for arbitrary-precision floating-point arithmetic" +optional = false +python-versions = "*" +files = [ + {file = "mpmath-1.3.0-py3-none-any.whl", hash = "sha256:a0b2b9fe80bbcd81a6647ff13108738cfb482d481d826cc0e02f5b35e5c88d2c"}, + {file = "mpmath-1.3.0.tar.gz", hash = "sha256:7a28eb2a9774d00c7bc92411c19a89209d5da7c4c9a9e227be8330a23a25b91f"}, +] + +[package.extras] +develop = ["codecov", "pycodestyle", "pytest (>=4.6)", "pytest-cov", "wheel"] +docs = ["sphinx"] +gmpy = ["gmpy2 (>=2.1.0a4)"] +tests = ["pytest (>=4.6)"] + +[[package]] +name = "networkx" +version = "3.2.1" +description = "Python package for creating and manipulating graphs and networks" +optional = false +python-versions = ">=3.9" +files = [ + {file = "networkx-3.2.1-py3-none-any.whl", hash = "sha256:f18c69adc97877c42332c170849c96cefa91881c99a7cb3e95b7c659ebdc1ec2"}, + {file = "networkx-3.2.1.tar.gz", hash = "sha256:9f1bb5cf3409bf324e0a722c20bdb4c20ee39bf1c30ce8ae499c8502b0b5e0c6"}, +] + +[package.extras] +default = ["matplotlib (>=3.5)", "numpy (>=1.22)", "pandas (>=1.4)", "scipy (>=1.9,!=1.11.0,!=1.11.1)"] +developer = ["changelist (==0.4)", "mypy (>=1.1)", "pre-commit (>=3.2)", "rtoml"] +doc = ["nb2plots (>=0.7)", "nbconvert (<7.9)", "numpydoc (>=1.6)", "pillow (>=9.4)", "pydata-sphinx-theme (>=0.14)", "sphinx (>=7)", "sphinx-gallery (>=0.14)", "texext (>=0.6.7)"] +extra = ["lxml (>=4.6)", "pydot (>=1.4.2)", "pygraphviz (>=1.11)", "sympy (>=1.10)"] +test = ["pytest (>=7.2)", "pytest-cov (>=4.0)"] + +[[package]] +name = "numpy" +version = "1.26.4" +description = "Fundamental package for array computing in Python" +optional = false +python-versions = ">=3.9" +files = [ + {file = "numpy-1.26.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:9ff0f4f29c51e2803569d7a51c2304de5554655a60c5d776e35b4a41413830d0"}, + {file = "numpy-1.26.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2e4ee3380d6de9c9ec04745830fd9e2eccb3e6cf790d39d7b98ffd19b0dd754a"}, + {file = "numpy-1.26.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d209d8969599b27ad20994c8e41936ee0964e6da07478d6c35016bc386b66ad4"}, + {file = "numpy-1.26.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ffa75af20b44f8dba823498024771d5ac50620e6915abac414251bd971b4529f"}, + {file = "numpy-1.26.4-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:62b8e4b1e28009ef2846b4c7852046736bab361f7aeadeb6a5b89ebec3c7055a"}, + {file = "numpy-1.26.4-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:a4abb4f9001ad2858e7ac189089c42178fcce737e4169dc61321660f1a96c7d2"}, + {file = "numpy-1.26.4-cp310-cp310-win32.whl", hash = "sha256:bfe25acf8b437eb2a8b2d49d443800a5f18508cd811fea3181723922a8a82b07"}, + {file = "numpy-1.26.4-cp310-cp310-win_amd64.whl", hash = "sha256:b97fe8060236edf3662adfc2c633f56a08ae30560c56310562cb4f95500022d5"}, + {file = "numpy-1.26.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:4c66707fabe114439db9068ee468c26bbdf909cac0fb58686a42a24de1760c71"}, + {file = "numpy-1.26.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:edd8b5fe47dab091176d21bb6de568acdd906d1887a4584a15a9a96a1dca06ef"}, + {file = "numpy-1.26.4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7ab55401287bfec946ced39700c053796e7cc0e3acbef09993a9ad2adba6ca6e"}, + {file = "numpy-1.26.4-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:666dbfb6ec68962c033a450943ded891bed2d54e6755e35e5835d63f4f6931d5"}, + {file = "numpy-1.26.4-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:96ff0b2ad353d8f990b63294c8986f1ec3cb19d749234014f4e7eb0112ceba5a"}, + {file = "numpy-1.26.4-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:60dedbb91afcbfdc9bc0b1f3f402804070deed7392c23eb7a7f07fa857868e8a"}, + {file = "numpy-1.26.4-cp311-cp311-win32.whl", hash = "sha256:1af303d6b2210eb850fcf03064d364652b7120803a0b872f5211f5234b399f20"}, + {file = "numpy-1.26.4-cp311-cp311-win_amd64.whl", hash = "sha256:cd25bcecc4974d09257ffcd1f098ee778f7834c3ad767fe5db785be9a4aa9cb2"}, + {file = "numpy-1.26.4-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:b3ce300f3644fb06443ee2222c2201dd3a89ea6040541412b8fa189341847218"}, + {file = "numpy-1.26.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:03a8c78d01d9781b28a6989f6fa1bb2c4f2d51201cf99d3dd875df6fbd96b23b"}, + {file = "numpy-1.26.4-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9fad7dcb1aac3c7f0584a5a8133e3a43eeb2fe127f47e3632d43d677c66c102b"}, + {file = "numpy-1.26.4-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:675d61ffbfa78604709862923189bad94014bef562cc35cf61d3a07bba02a7ed"}, + {file = "numpy-1.26.4-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:ab47dbe5cc8210f55aa58e4805fe224dac469cde56b9f731a4c098b91917159a"}, + {file = "numpy-1.26.4-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:1dda2e7b4ec9dd512f84935c5f126c8bd8b9f2fc001e9f54af255e8c5f16b0e0"}, + {file = "numpy-1.26.4-cp312-cp312-win32.whl", hash = "sha256:50193e430acfc1346175fcbdaa28ffec49947a06918b7b92130744e81e640110"}, + {file = "numpy-1.26.4-cp312-cp312-win_amd64.whl", hash = "sha256:08beddf13648eb95f8d867350f6a018a4be2e5ad54c8d8caed89ebca558b2818"}, + {file = "numpy-1.26.4-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7349ab0fa0c429c82442a27a9673fc802ffdb7c7775fad780226cb234965e53c"}, + {file = "numpy-1.26.4-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:52b8b60467cd7dd1e9ed082188b4e6bb35aa5cdd01777621a1658910745b90be"}, + {file = "numpy-1.26.4-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d5241e0a80d808d70546c697135da2c613f30e28251ff8307eb72ba696945764"}, + {file = "numpy-1.26.4-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f870204a840a60da0b12273ef34f7051e98c3b5961b61b0c2c1be6dfd64fbcd3"}, + {file = "numpy-1.26.4-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:679b0076f67ecc0138fd2ede3a8fd196dddc2ad3254069bcb9faf9a79b1cebcd"}, + {file = "numpy-1.26.4-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:47711010ad8555514b434df65f7d7b076bb8261df1ca9bb78f53d3b2db02e95c"}, + {file = "numpy-1.26.4-cp39-cp39-win32.whl", hash = "sha256:a354325ee03388678242a4d7ebcd08b5c727033fcff3b2f536aea978e15ee9e6"}, + {file = "numpy-1.26.4-cp39-cp39-win_amd64.whl", hash = "sha256:3373d5d70a5fe74a2c1bb6d2cfd9609ecf686d47a2d7b1d37a8f3b6bf6003aea"}, + {file = "numpy-1.26.4-pp39-pypy39_pp73-macosx_10_9_x86_64.whl", hash = "sha256:afedb719a9dcfc7eaf2287b839d8198e06dcd4cb5d276a3df279231138e83d30"}, + {file = "numpy-1.26.4-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:95a7476c59002f2f6c590b9b7b998306fba6a5aa646b1e22ddfeaf8f78c3a29c"}, + {file = "numpy-1.26.4-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:7e50d0a0cc3189f9cb0aeb3a6a6af18c16f59f004b866cd2be1c14b36134a4a0"}, + {file = "numpy-1.26.4.tar.gz", hash = "sha256:2a02aba9ed12e4ac4eb3ea9421c420301a0c6460d9830d74a9df87efa4912010"}, +] + +[[package]] +name = "packaging" +version = "23.2" +description = "Core utilities for Python packages" +optional = false +python-versions = ">=3.7" +files = [ + {file = "packaging-23.2-py3-none-any.whl", hash = "sha256:8c491190033a9af7e1d931d0b5dacc2ef47509b34dd0de67ed209b5203fc88c7"}, + {file = "packaging-23.2.tar.gz", hash = "sha256:048fb0e9405036518eaaf48a55953c750c11e1a1b68e0dd1a9d62ed0c092cfc5"}, +] + +[[package]] +name = "pluggy" +version = "0.13.1" +description = "plugin and hook calling mechanisms for python" +optional = false +python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*" +files = [ + {file = "pluggy-0.13.1-py2.py3-none-any.whl", hash = "sha256:966c145cd83c96502c3c3868f50408687b38434af77734af1e9ca461a4081d2d"}, + {file = "pluggy-0.13.1.tar.gz", hash = "sha256:15b2acde666561e1298d71b523007ed7364de07029219b604cf808bfa1c765b0"}, +] + +[package.extras] +dev = ["pre-commit", "tox"] + +[[package]] +name = "protobuf" +version = "4.25.3" +description = "" +optional = false +python-versions = ">=3.8" +files = [ + {file = "protobuf-4.25.3-cp310-abi3-win32.whl", hash = "sha256:d4198877797a83cbfe9bffa3803602bbe1625dc30d8a097365dbc762e5790faa"}, + {file = "protobuf-4.25.3-cp310-abi3-win_amd64.whl", hash = "sha256:209ba4cc916bab46f64e56b85b090607a676f66b473e6b762e6f1d9d591eb2e8"}, + {file = "protobuf-4.25.3-cp37-abi3-macosx_10_9_universal2.whl", hash = "sha256:f1279ab38ecbfae7e456a108c5c0681e4956d5b1090027c1de0f934dfdb4b35c"}, + {file = "protobuf-4.25.3-cp37-abi3-manylinux2014_aarch64.whl", hash = "sha256:e7cb0ae90dd83727f0c0718634ed56837bfeeee29a5f82a7514c03ee1364c019"}, + {file = "protobuf-4.25.3-cp37-abi3-manylinux2014_x86_64.whl", hash = "sha256:7c8daa26095f82482307bc717364e7c13f4f1c99659be82890dcfc215194554d"}, + {file = "protobuf-4.25.3-cp38-cp38-win32.whl", hash = "sha256:f4f118245c4a087776e0a8408be33cf09f6c547442c00395fbfb116fac2f8ac2"}, + {file = "protobuf-4.25.3-cp38-cp38-win_amd64.whl", hash = "sha256:c053062984e61144385022e53678fbded7aea14ebb3e0305ae3592fb219ccfa4"}, + {file = "protobuf-4.25.3-cp39-cp39-win32.whl", hash = "sha256:19b270aeaa0099f16d3ca02628546b8baefe2955bbe23224aaf856134eccf1e4"}, + {file = "protobuf-4.25.3-cp39-cp39-win_amd64.whl", hash = "sha256:e3c97a1555fd6388f857770ff8b9703083de6bf1f9274a002a332d65fbb56c8c"}, + {file = "protobuf-4.25.3-py3-none-any.whl", hash = "sha256:f0700d54bcf45424477e46a9f0944155b46fb0639d69728739c0e47bab83f2b9"}, + {file = "protobuf-4.25.3.tar.gz", hash = "sha256:25b5d0b42fd000320bd7830b349e3b696435f3b329810427a6bcce6a5492cc5c"}, +] + +[[package]] +name = "py" +version = "1.11.0" +description = "library with cross-python path, ini-parsing, io, code, log facilities" +optional = false +python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*, !=3.4.*" +files = [ + {file = "py-1.11.0-py2.py3-none-any.whl", hash = "sha256:607c53218732647dff4acdfcd50cb62615cedf612e72d1724fb1a0cc6405b378"}, + {file = "py-1.11.0.tar.gz", hash = "sha256:51c75c4126074b472f746a24399ad32f6053d1b34b68d2fa41e558e6f4a98719"}, +] + +[[package]] +name = "pytest" +version = "5.4.3" +description = "pytest: simple powerful testing with Python" +optional = false +python-versions = ">=3.5" +files = [ + {file = "pytest-5.4.3-py3-none-any.whl", hash = "sha256:5c0db86b698e8f170ba4582a492248919255fcd4c79b1ee64ace34301fb589a1"}, + {file = "pytest-5.4.3.tar.gz", hash = "sha256:7979331bfcba207414f5e1263b5a0f8f521d0f457318836a7355531ed1a4c7d8"}, +] + +[package.dependencies] +atomicwrites = {version = ">=1.0", markers = "sys_platform == \"win32\""} +attrs = ">=17.4.0" +colorama = {version = "*", markers = "sys_platform == \"win32\""} +more-itertools = ">=4.0.0" +packaging = "*" +pluggy = ">=0.12,<1.0" +py = ">=1.5.0" +wcwidth = "*" + +[package.extras] +checkqa-mypy = ["mypy (==v0.761)"] +testing = ["argcomplete", "hypothesis (>=3.56)", "mock", "nose", "requests", "xmlschema"] + +[[package]] +name = "pyyaml" +version = "6.0.1" +description = "YAML parser and emitter for Python" +optional = false +python-versions = ">=3.6" +files = [ + {file = "PyYAML-6.0.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:d858aa552c999bc8a8d57426ed01e40bef403cd8ccdd0fc5f6f04a00414cac2a"}, + {file = "PyYAML-6.0.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:fd66fc5d0da6d9815ba2cebeb4205f95818ff4b79c3ebe268e75d961704af52f"}, + {file = "PyYAML-6.0.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:69b023b2b4daa7548bcfbd4aa3da05b3a74b772db9e23b982788168117739938"}, + {file = "PyYAML-6.0.1-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:81e0b275a9ecc9c0c0c07b4b90ba548307583c125f54d5b6946cfee6360c733d"}, + {file = "PyYAML-6.0.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ba336e390cd8e4d1739f42dfe9bb83a3cc2e80f567d8805e11b46f4a943f5515"}, + {file = "PyYAML-6.0.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:326c013efe8048858a6d312ddd31d56e468118ad4cdeda36c719bf5bb6192290"}, + {file = "PyYAML-6.0.1-cp310-cp310-win32.whl", hash = "sha256:bd4af7373a854424dabd882decdc5579653d7868b8fb26dc7d0e99f823aa5924"}, + {file = "PyYAML-6.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:fd1592b3fdf65fff2ad0004b5e363300ef59ced41c2e6b3a99d4089fa8c5435d"}, + {file = "PyYAML-6.0.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:6965a7bc3cf88e5a1c3bd2e0b5c22f8d677dc88a455344035f03399034eb3007"}, + {file = "PyYAML-6.0.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:f003ed9ad21d6a4713f0a9b5a7a0a79e08dd0f221aff4525a2be4c346ee60aab"}, + {file = "PyYAML-6.0.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:42f8152b8dbc4fe7d96729ec2b99c7097d656dc1213a3229ca5383f973a5ed6d"}, + {file = "PyYAML-6.0.1-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:062582fca9fabdd2c8b54a3ef1c978d786e0f6b3a1510e0ac93ef59e0ddae2bc"}, + {file = "PyYAML-6.0.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d2b04aac4d386b172d5b9692e2d2da8de7bfb6c387fa4f801fbf6fb2e6ba4673"}, + {file = "PyYAML-6.0.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:e7d73685e87afe9f3b36c799222440d6cf362062f78be1013661b00c5c6f678b"}, + {file = "PyYAML-6.0.1-cp311-cp311-win32.whl", hash = "sha256:1635fd110e8d85d55237ab316b5b011de701ea0f29d07611174a1b42f1444741"}, + {file = "PyYAML-6.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:bf07ee2fef7014951eeb99f56f39c9bb4af143d8aa3c21b1677805985307da34"}, + {file = "PyYAML-6.0.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:855fb52b0dc35af121542a76b9a84f8d1cd886ea97c84703eaa6d88e37a2ad28"}, + {file = "PyYAML-6.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40df9b996c2b73138957fe23a16a4f0ba614f4c0efce1e9406a184b6d07fa3a9"}, + {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a08c6f0fe150303c1c6b71ebcd7213c2858041a7e01975da3a99aed1e7a378ef"}, + {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6c22bec3fbe2524cde73d7ada88f6566758a8f7227bfbf93a408a9d86bcc12a0"}, + {file = "PyYAML-6.0.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8d4e9c88387b0f5c7d5f281e55304de64cf7f9c0021a3525bd3b1c542da3b0e4"}, + {file = "PyYAML-6.0.1-cp312-cp312-win32.whl", hash = "sha256:d483d2cdf104e7c9fa60c544d92981f12ad66a457afae824d146093b8c294c54"}, + {file = "PyYAML-6.0.1-cp312-cp312-win_amd64.whl", hash = "sha256:0d3304d8c0adc42be59c5f8a4d9e3d7379e6955ad754aa9d6ab7a398b59dd1df"}, + {file = "PyYAML-6.0.1-cp36-cp36m-macosx_10_9_x86_64.whl", hash = "sha256:50550eb667afee136e9a77d6dc71ae76a44df8b3e51e41b77f6de2932bfe0f47"}, + {file = "PyYAML-6.0.1-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1fe35611261b29bd1de0070f0b2f47cb6ff71fa6595c077e42bd0c419fa27b98"}, + {file = "PyYAML-6.0.1-cp36-cp36m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:704219a11b772aea0d8ecd7058d0082713c3562b4e271b849ad7dc4a5c90c13c"}, + {file = "PyYAML-6.0.1-cp36-cp36m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:afd7e57eddb1a54f0f1a974bc4391af8bcce0b444685d936840f125cf046d5bd"}, + {file = "PyYAML-6.0.1-cp36-cp36m-win32.whl", hash = "sha256:fca0e3a251908a499833aa292323f32437106001d436eca0e6e7833256674585"}, + {file = "PyYAML-6.0.1-cp36-cp36m-win_amd64.whl", hash = "sha256:f22ac1c3cac4dbc50079e965eba2c1058622631e526bd9afd45fedd49ba781fa"}, + {file = "PyYAML-6.0.1-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:b1275ad35a5d18c62a7220633c913e1b42d44b46ee12554e5fd39c70a243d6a3"}, + {file = "PyYAML-6.0.1-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:18aeb1bf9a78867dc38b259769503436b7c72f7a1f1f4c93ff9a17de54319b27"}, + {file = "PyYAML-6.0.1-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:596106435fa6ad000c2991a98fa58eeb8656ef2325d7e158344fb33864ed87e3"}, + {file = "PyYAML-6.0.1-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:baa90d3f661d43131ca170712d903e6295d1f7a0f595074f151c0aed377c9b9c"}, + {file = "PyYAML-6.0.1-cp37-cp37m-win32.whl", hash = "sha256:9046c58c4395dff28dd494285c82ba00b546adfc7ef001486fbf0324bc174fba"}, + {file = "PyYAML-6.0.1-cp37-cp37m-win_amd64.whl", hash = "sha256:4fb147e7a67ef577a588a0e2c17b6db51dda102c71de36f8549b6816a96e1867"}, + {file = "PyYAML-6.0.1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:1d4c7e777c441b20e32f52bd377e0c409713e8bb1386e1099c2415f26e479595"}, + {file = "PyYAML-6.0.1-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a0cd17c15d3bb3fa06978b4e8958dcdc6e0174ccea823003a106c7d4d7899ac5"}, + {file = "PyYAML-6.0.1-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:28c119d996beec18c05208a8bd78cbe4007878c6dd15091efb73a30e90539696"}, + {file = "PyYAML-6.0.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7e07cbde391ba96ab58e532ff4803f79c4129397514e1413a7dc761ccd755735"}, + {file = "PyYAML-6.0.1-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:49a183be227561de579b4a36efbb21b3eab9651dd81b1858589f796549873dd6"}, + {file = "PyYAML-6.0.1-cp38-cp38-win32.whl", hash = "sha256:184c5108a2aca3c5b3d3bf9395d50893a7ab82a38004c8f61c258d4428e80206"}, + {file = "PyYAML-6.0.1-cp38-cp38-win_amd64.whl", hash = "sha256:1e2722cc9fbb45d9b87631ac70924c11d3a401b2d7f410cc0e3bbf249f2dca62"}, + {file = "PyYAML-6.0.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:9eb6caa9a297fc2c2fb8862bc5370d0303ddba53ba97e71f08023b6cd73d16a8"}, + {file = "PyYAML-6.0.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:c8098ddcc2a85b61647b2590f825f3db38891662cfc2fc776415143f599bb859"}, + {file = "PyYAML-6.0.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5773183b6446b2c99bb77e77595dd486303b4faab2b086e7b17bc6bef28865f6"}, + {file = "PyYAML-6.0.1-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:b786eecbdf8499b9ca1d697215862083bd6d2a99965554781d0d8d1ad31e13a0"}, + {file = "PyYAML-6.0.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bc1bf2925a1ecd43da378f4db9e4f799775d6367bdb94671027b73b393a7c42c"}, + {file = "PyYAML-6.0.1-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:04ac92ad1925b2cff1db0cfebffb6ffc43457495c9b3c39d3fcae417d7125dc5"}, + {file = "PyYAML-6.0.1-cp39-cp39-win32.whl", hash = "sha256:faca3bdcf85b2fc05d06ff3fbc1f83e1391b3e724afa3feba7d13eeab355484c"}, + {file = "PyYAML-6.0.1-cp39-cp39-win_amd64.whl", hash = "sha256:510c9deebc5c0225e8c96813043e62b680ba2f9c50a08d3724c7f28a747d1486"}, + {file = "PyYAML-6.0.1.tar.gz", hash = "sha256:bfdf460b1736c775f2ba9f6a92bca30bc2095067b8a9d77876d1fad6cc3b4a43"}, +] + +[[package]] +name = "regex" +version = "2023.12.25" +description = "Alternative regular expression module, to replace re." +optional = false +python-versions = ">=3.7" +files = [ + {file = "regex-2023.12.25-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:0694219a1d54336fd0445ea382d49d36882415c0134ee1e8332afd1529f0baa5"}, + {file = "regex-2023.12.25-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:b014333bd0217ad3d54c143de9d4b9a3ca1c5a29a6d0d554952ea071cff0f1f8"}, + {file = "regex-2023.12.25-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:d865984b3f71f6d0af64d0d88f5733521698f6c16f445bb09ce746c92c97c586"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1e0eabac536b4cc7f57a5f3d095bfa557860ab912f25965e08fe1545e2ed8b4c"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:c25a8ad70e716f96e13a637802813f65d8a6760ef48672aa3502f4c24ea8b400"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:a9b6d73353f777630626f403b0652055ebfe8ff142a44ec2cf18ae470395766e"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a9cc99d6946d750eb75827cb53c4371b8b0fe89c733a94b1573c9dd16ea6c9e4"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:88d1f7bef20c721359d8675f7d9f8e414ec5003d8f642fdfd8087777ff7f94b5"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:cb3fe77aec8f1995611f966d0c656fdce398317f850d0e6e7aebdfe61f40e1cd"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:7aa47c2e9ea33a4a2a05f40fcd3ea36d73853a2aae7b4feab6fc85f8bf2c9704"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:df26481f0c7a3f8739fecb3e81bc9da3fcfae34d6c094563b9d4670b047312e1"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:c40281f7d70baf6e0db0c2f7472b31609f5bc2748fe7275ea65a0b4601d9b392"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_s390x.whl", hash = "sha256:d94a1db462d5690ebf6ae86d11c5e420042b9898af5dcf278bd97d6bda065423"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:ba1b30765a55acf15dce3f364e4928b80858fa8f979ad41f862358939bdd1f2f"}, + {file = "regex-2023.12.25-cp310-cp310-win32.whl", hash = "sha256:150c39f5b964e4d7dba46a7962a088fbc91f06e606f023ce57bb347a3b2d4630"}, + {file = "regex-2023.12.25-cp310-cp310-win_amd64.whl", hash = "sha256:09da66917262d9481c719599116c7dc0c321ffcec4b1f510c4f8a066f8768105"}, + {file = "regex-2023.12.25-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:1b9d811f72210fa9306aeb88385b8f8bcef0dfbf3873410413c00aa94c56c2b6"}, + {file = "regex-2023.12.25-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:d902a43085a308cef32c0d3aea962524b725403fd9373dea18110904003bac97"}, + {file = "regex-2023.12.25-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:d166eafc19f4718df38887b2bbe1467a4f74a9830e8605089ea7a30dd4da8887"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c7ad32824b7f02bb3c9f80306d405a1d9b7bb89362d68b3c5a9be53836caebdb"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:636ba0a77de609d6510235b7f0e77ec494d2657108f777e8765efc060094c98c"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:0fda75704357805eb953a3ee15a2b240694a9a514548cd49b3c5124b4e2ad01b"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f72cbae7f6b01591f90814250e636065850c5926751af02bb48da94dfced7baa"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:db2a0b1857f18b11e3b0e54ddfefc96af46b0896fb678c85f63fb8c37518b3e7"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:7502534e55c7c36c0978c91ba6f61703faf7ce733715ca48f499d3dbbd7657e0"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:e8c7e08bb566de4faaf11984af13f6bcf6a08f327b13631d41d62592681d24fe"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:283fc8eed679758de38fe493b7d7d84a198b558942b03f017b1f94dda8efae80"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_s390x.whl", hash = "sha256:f44dd4d68697559d007462b0a3a1d9acd61d97072b71f6d1968daef26bc744bd"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:67d3ccfc590e5e7197750fcb3a2915b416a53e2de847a728cfa60141054123d4"}, + {file = "regex-2023.12.25-cp311-cp311-win32.whl", hash = "sha256:68191f80a9bad283432385961d9efe09d783bcd36ed35a60fb1ff3f1ec2efe87"}, + {file = "regex-2023.12.25-cp311-cp311-win_amd64.whl", hash = "sha256:7d2af3f6b8419661a0c421584cfe8aaec1c0e435ce7e47ee2a97e344b98f794f"}, + {file = "regex-2023.12.25-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:8a0ccf52bb37d1a700375a6b395bff5dd15c50acb745f7db30415bae3c2b0715"}, + {file = "regex-2023.12.25-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:c3c4a78615b7762740531c27cf46e2f388d8d727d0c0c739e72048beb26c8a9d"}, + {file = "regex-2023.12.25-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:ad83e7545b4ab69216cef4cc47e344d19622e28aabec61574b20257c65466d6a"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b7a635871143661feccce3979e1727c4e094f2bdfd3ec4b90dfd4f16f571a87a"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d498eea3f581fbe1b34b59c697512a8baef88212f92e4c7830fcc1499f5b45a5"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:43f7cd5754d02a56ae4ebb91b33461dc67be8e3e0153f593c509e21d219c5060"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:51f4b32f793812714fd5307222a7f77e739b9bc566dc94a18126aba3b92b98a3"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ba99d8077424501b9616b43a2d208095746fb1284fc5ba490139651f971d39d9"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:4bfc2b16e3ba8850e0e262467275dd4d62f0d045e0e9eda2bc65078c0110a11f"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:8c2c19dae8a3eb0ea45a8448356ed561be843b13cbc34b840922ddf565498c1c"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_ppc64le.whl", hash = "sha256:60080bb3d8617d96f0fb7e19796384cc2467447ef1c491694850ebd3670bc457"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_s390x.whl", hash = "sha256:b77e27b79448e34c2c51c09836033056a0547aa360c45eeeb67803da7b0eedaf"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:518440c991f514331f4850a63560321f833979d145d7d81186dbe2f19e27ae3d"}, + {file = "regex-2023.12.25-cp312-cp312-win32.whl", hash = "sha256:e2610e9406d3b0073636a3a2e80db05a02f0c3169b5632022b4e81c0364bcda5"}, + {file = "regex-2023.12.25-cp312-cp312-win_amd64.whl", hash = "sha256:cc37b9aeebab425f11f27e5e9e6cf580be7206c6582a64467a14dda211abc232"}, + {file = "regex-2023.12.25-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:da695d75ac97cb1cd725adac136d25ca687da4536154cdc2815f576e4da11c69"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d126361607b33c4eb7b36debc173bf25d7805847346dd4d99b5499e1fef52bc7"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4719bb05094d7d8563a450cf8738d2e1061420f79cfcc1fa7f0a44744c4d8f73"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:5dd58946bce44b53b06d94aa95560d0b243eb2fe64227cba50017a8d8b3cd3e2"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:22a86d9fff2009302c440b9d799ef2fe322416d2d58fc124b926aa89365ec482"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:2aae8101919e8aa05ecfe6322b278f41ce2994c4a430303c4cd163fef746e04f"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:e692296c4cc2873967771345a876bcfc1c547e8dd695c6b89342488b0ea55cd8"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:263ef5cc10979837f243950637fffb06e8daed7f1ac1e39d5910fd29929e489a"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:d6f7e255e5fa94642a0724e35406e6cb7001c09d476ab5fce002f652b36d0c39"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:88ad44e220e22b63b0f8f81f007e8abbb92874d8ced66f32571ef8beb0643b2b"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_s390x.whl", hash = "sha256:3a17d3ede18f9cedcbe23d2daa8a2cd6f59fe2bf082c567e43083bba3fb00347"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:d15b274f9e15b1a0b7a45d2ac86d1f634d983ca40d6b886721626c47a400bf39"}, + {file = "regex-2023.12.25-cp37-cp37m-win32.whl", hash = "sha256:ed19b3a05ae0c97dd8f75a5d8f21f7723a8c33bbc555da6bbe1f96c470139d3c"}, + {file = "regex-2023.12.25-cp37-cp37m-win_amd64.whl", hash = "sha256:a6d1047952c0b8104a1d371f88f4ab62e6275567d4458c1e26e9627ad489b445"}, + {file = "regex-2023.12.25-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:b43523d7bc2abd757119dbfb38af91b5735eea45537ec6ec3a5ec3f9562a1c53"}, + {file = "regex-2023.12.25-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:efb2d82f33b2212898f1659fb1c2e9ac30493ac41e4d53123da374c3b5541e64"}, + {file = "regex-2023.12.25-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:b7fca9205b59c1a3d5031f7e64ed627a1074730a51c2a80e97653e3e9fa0d415"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:086dd15e9435b393ae06f96ab69ab2d333f5d65cbe65ca5a3ef0ec9564dfe770"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e81469f7d01efed9b53740aedd26085f20d49da65f9c1f41e822a33992cb1590"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:34e4af5b27232f68042aa40a91c3b9bb4da0eeb31b7632e0091afc4310afe6cb"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9852b76ab558e45b20bf1893b59af64a28bd3820b0c2efc80e0a70a4a3ea51c1"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ff100b203092af77d1a5a7abe085b3506b7eaaf9abf65b73b7d6905b6cb76988"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:cc038b2d8b1470364b1888a98fd22d616fba2b6309c5b5f181ad4483e0017861"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:094ba386bb5c01e54e14434d4caabf6583334090865b23ef58e0424a6286d3dc"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:5cd05d0f57846d8ba4b71d9c00f6f37d6b97d5e5ef8b3c3840426a475c8f70f4"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:9aa1a67bbf0f957bbe096375887b2505f5d8ae16bf04488e8b0f334c36e31360"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_s390x.whl", hash = "sha256:98a2636994f943b871786c9e82bfe7883ecdaba2ef5df54e1450fa9869d1f756"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:37f8e93a81fc5e5bd8db7e10e62dc64261bcd88f8d7e6640aaebe9bc180d9ce2"}, + {file = "regex-2023.12.25-cp38-cp38-win32.whl", hash = "sha256:d78bd484930c1da2b9679290a41cdb25cc127d783768a0369d6b449e72f88beb"}, + {file = "regex-2023.12.25-cp38-cp38-win_amd64.whl", hash = "sha256:b521dcecebc5b978b447f0f69b5b7f3840eac454862270406a39837ffae4e697"}, + {file = "regex-2023.12.25-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:f7bc09bc9c29ebead055bcba136a67378f03d66bf359e87d0f7c759d6d4ffa31"}, + {file = "regex-2023.12.25-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e14b73607d6231f3cc4622809c196b540a6a44e903bcfad940779c80dffa7be7"}, + {file = "regex-2023.12.25-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:9eda5f7a50141291beda3edd00abc2d4a5b16c29c92daf8d5bd76934150f3edc"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cc6bb9aa69aacf0f6032c307da718f61a40cf970849e471254e0e91c56ffca95"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:298dc6354d414bc921581be85695d18912bea163a8b23cac9a2562bbcd5088b1"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:2f4e475a80ecbd15896a976aa0b386c5525d0ed34d5c600b6d3ebac0a67c7ddf"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:531ac6cf22b53e0696f8e1d56ce2396311254eb806111ddd3922c9d937151dae"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:22f3470f7524b6da61e2020672df2f3063676aff444db1daa283c2ea4ed259d6"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:89723d2112697feaa320c9d351e5f5e7b841e83f8b143dba8e2d2b5f04e10923"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:0ecf44ddf9171cd7566ef1768047f6e66975788258b1c6c6ca78098b95cf9a3d"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:905466ad1702ed4acfd67a902af50b8db1feeb9781436372261808df7a2a7bca"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:4558410b7a5607a645e9804a3e9dd509af12fb72b9825b13791a37cd417d73a5"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_s390x.whl", hash = "sha256:7e316026cc1095f2a3e8cc012822c99f413b702eaa2ca5408a513609488cb62f"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:3b1de218d5375cd6ac4b5493e0b9f3df2be331e86520f23382f216c137913d20"}, + {file = "regex-2023.12.25-cp39-cp39-win32.whl", hash = "sha256:11a963f8e25ab5c61348d090bf1b07f1953929c13bd2309a0662e9ff680763c9"}, + {file = "regex-2023.12.25-cp39-cp39-win_amd64.whl", hash = "sha256:e693e233ac92ba83a87024e1d32b5f9ab15ca55ddd916d878146f4e3406b5c91"}, + {file = "regex-2023.12.25.tar.gz", hash = "sha256:29171aa128da69afdf4bde412d5bedc335f2ca8fcfe4489038577d05f16181e5"}, +] + +[[package]] +name = "requests" +version = "2.31.0" +description = "Python HTTP for Humans." +optional = false +python-versions = ">=3.7" +files = [ + {file = "requests-2.31.0-py3-none-any.whl", hash = "sha256:58cd2187c01e70e6e26505bca751777aa9f2ee0b7f4300988b709f44e013003f"}, + {file = "requests-2.31.0.tar.gz", hash = "sha256:942c5a758f98d790eaed1a29cb6eefc7ffb0d1cf7af05c3d2791656dbd6ad1e1"}, +] + +[package.dependencies] +certifi = ">=2017.4.17" +charset-normalizer = ">=2,<4" +idna = ">=2.5,<4" +urllib3 = ">=1.21.1,<3" + +[package.extras] +socks = ["PySocks (>=1.5.6,!=1.5.7)"] +use-chardet-on-py3 = ["chardet (>=3.0.2,<6)"] + +[[package]] +name = "safetensors" +version = "0.4.2" +description = "" +optional = false +python-versions = ">=3.7" +files = [ + {file = "safetensors-0.4.2-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:69d8bb8384dc2cb5b72c36c4d6980771b293d1a1377b378763f5e37b6bb8d133"}, + {file = "safetensors-0.4.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:3d420e19fcef96d0067f4de4699682b4bbd85fc8fea0bd45fcd961fdf3e8c82c"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9ca54742122fa3c4821754adb67318e1cd25c3a22bbf0c5520d5176e77a099ac"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:8b47aa643afdfd66cf7ce4c184092ae734e15d10aba2c2948f24270211801c3c"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d88a16bbc330f27e7f2d4caaf6fb061ad0b8a756ecc4033260b0378e128ce8a2"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e9223b8ac21085db614a510eb3445e7083cae915a9202357555fa939695d4f57"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ce6cb86133dc8930a7ab5e7438545a7f205f7a1cdd5aaf108c1d0da6bdcfbc2b"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b8a628e0ae2bbc334b62952c384aa5f41621d01850f8d67b04a96b9c39dd7326"}, + {file = "safetensors-0.4.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:88d6beb7f811a081e0e5f1d9669fdac816c45340c04b1eaf7ebfda0ce93ea403"}, + {file = "safetensors-0.4.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:b57fc5b1b54cb12d8690a58a4cf4b7144730d4bde9d98aa0e1dab6295a1cd579"}, + {file = "safetensors-0.4.2-cp310-none-win32.whl", hash = "sha256:9d87a1c98803c16cf113b9ba03f07b2dce5e8eabfd1811a7f7323fcaa2a1bf47"}, + {file = "safetensors-0.4.2-cp310-none-win_amd64.whl", hash = "sha256:18930ec1d1ecb526d3d9835abc2489b8f1530877518f0c541e77ef0b7abcbd99"}, + {file = "safetensors-0.4.2-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:c5dd2ed788730ed56b415d1a11c62026b8cc8c573f55a2092afb3ab383e94fff"}, + {file = "safetensors-0.4.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:cc41791b33efb9c83a59b731619f3d15f543dfe71f3a793cb8fbf9bd5d0d5d71"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4c888bf71d5ca12a720f1ed87d407c4918afa022fb247a6546d8fac15b1f112b"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e6b2feb4b47226a16a792e6fac3f49442714884a3d4c1008569d5068a3941be9"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f41cc0ee4b838ae8f4d8364a1b162067693d11a3893f0863be8c228d40e4d0ee"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:51b7228e46c0a483c40ba4b9470dea00fb1ff8685026bb4766799000f6328ac2"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:02697f8f2be8ca3c37a4958702dbdb1864447ef765e18b5328a1617022dcf164"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:27fd8f65cf7c80e4280cae1ee6bcd85c483882f6580821abe71ee1a0d3dcfca7"}, + {file = "safetensors-0.4.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:c487b5f113b0924c9534a07dc034830fb4ef05ce9bb6d78cfe016a7dedfe281f"}, + {file = "safetensors-0.4.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:da7f6483f3fe67ff39b3a55552552c67930ea10a36e9f2539d36fc205273d767"}, + {file = "safetensors-0.4.2-cp311-none-win32.whl", hash = "sha256:52a7012f6cb9cb4a132760b6308daede18a9f5f8952ce08adc7c67a7d865c2d8"}, + {file = "safetensors-0.4.2-cp311-none-win_amd64.whl", hash = "sha256:4d1361a097ac430b310ce9eed8ed4746edee33ddafdfbb965debc8966fc34dc2"}, + {file = "safetensors-0.4.2-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:77af8aa0edcc2863760fd6febbfdb82e88fd75d0e60c1ce4ba57208ba5e4a89b"}, + {file = "safetensors-0.4.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:846666c1c5a8c8888d2dfda8d3921cb9cb8e2c5f78365be756c11021e75a0a2a"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4f4bfc7ea19b446bfad41510d4b4c76101698c00caaa8a332c8edd8090a412ef"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:233436fd30f27ffeb3c3780d0b84f496518868445c7a8db003639a649cc98453"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7a09237a795d11cd11f9dae505d170a29b5616151db1e10c14f892b11caadc7d"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:de01c9a3a3b7b69627d624ff69d9f11d28ce9908eea2fb6245adafa4b1d43df6"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8c1f25c5069ee42a5bcffdc66c300a407941edd73f3239e9fdefd26216407391"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:7a73b3649456d09ca8506140d44484b63154a7378434cc1e8719f8056550b224"}, + {file = "safetensors-0.4.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:e1625a8d07d046e968bd5c4961810aba1225984e4fb9243626f9d04a06ed3fee"}, + {file = "safetensors-0.4.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8f74c86b25615cb24ad4cff765a2eefc09d71bf0fed97588cf585aad9c38fbb4"}, + {file = "safetensors-0.4.2-cp312-none-win32.whl", hash = "sha256:8523b9c5777d771bcde5c2389c03f1cdf7ebe8797432a1bd5e345efe25c55987"}, + {file = "safetensors-0.4.2-cp312-none-win_amd64.whl", hash = "sha256:dcff0243e1737a21f83d664c63fed89d1f532c23fc6830d0427279fabd789ccb"}, + {file = "safetensors-0.4.2-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:96ad3d7d472612e26cbe413922b4fb13933310f0511d346ea5cc9a1e856e52eb"}, + {file = "safetensors-0.4.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:88250922401b5ae4e37de929178caf46be47ed16c817b2237b81679bec07c120"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d40443554142fc0ab30652d5cc8554c4b7a613513bde00373e18afd5de8cbe4b"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:27f53f70106224d32d874aacecbeb4a6e4c5b16a1d2006d0e876d97229086d71"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:cc068afe23734dfb26ce19db0a7877499ddf73b1d55ceb762417e8da4a1b05fb"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9be1918eb8d43a11a6f8806759fccfa0eeb0542b12924caba66af8a7800ad01a"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:41911087d20a7bbd78cb4ad4f98aab0c431533107584df6635d8b54b99945573"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:50771c662aab909f31e94d048e76861fd027d66076ea773eef2e66c717766e24"}, + {file = "safetensors-0.4.2-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:13f2e57be007b7ea9329133d2399e6bdfcf1910f655440a4da17df3a45afcd30"}, + {file = "safetensors-0.4.2-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:c772147e6395bc829842e0a98e1b30c67fe25d816299c28196488511d5a5e951"}, + {file = "safetensors-0.4.2-cp37-cp37m-macosx_10_12_x86_64.whl", hash = "sha256:36239a0060b537a3e8c473df78cffee14c3ec4f51d5f1a853af99371a2fb2a35"}, + {file = "safetensors-0.4.2-cp37-cp37m-macosx_11_0_arm64.whl", hash = "sha256:d0cbb7664fad2c307f95195f951b7059e95dc23e0e1822e5978c8b500098543c"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2b3e55adb6bd9dc1c2a341e72f48f075953fa35d173dd8e29a95b3b02d0d1462"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:42f743b3cca863fba53ca57a193f510e5ec359b97f38c282437716b6768e4a25"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:04e6af4a6dbeb06c4e6e7d46cf9c716cbc4cc5ef62584fd8a7c0fe558562df45"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:a492ba21b5c8f14ee5ec9b20f42ba969e53ca1f909a4d04aad736b66a341dcc2"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b25b8233a1a85dc67e39838951cfb01595d792f3b7b644add63edb652992e030"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:fd27e063fbdafe776f7b1714da59110e88f270e86db00788a8fd65f4eacfeba7"}, + {file = "safetensors-0.4.2-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:1b6fa399f251bbeb52029bf5a0ac2878d7705dd3612a2f8895b48e9c11f0367d"}, + {file = "safetensors-0.4.2-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:de642d46b459e4afd5c2020b26c0d6d869a171ea00411897d5776c127cac74f0"}, + {file = "safetensors-0.4.2-cp37-none-win32.whl", hash = "sha256:77b72d17754c93bb68f3598182f14d78776e0b9b31682ca5bb2c7c5bd9a75267"}, + {file = "safetensors-0.4.2-cp37-none-win_amd64.whl", hash = "sha256:d36ee3244d461cd655aeef493792c3bccf4875282f8407fd9af99e9a41cf2530"}, + {file = "safetensors-0.4.2-cp38-cp38-macosx_10_12_x86_64.whl", hash = "sha256:16b6b3884f7876c6b3b23a742428223a7170a5a9dac819d8c12a1569422c4b5a"}, + {file = "safetensors-0.4.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:ee25d311493fbbe0be9d395faee46e9d79e8948f461e388ff39e59875ed9a350"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:eed8097968585cd752a1171f86fce9aa1d89a29033e5cd8bec5a502e29f6b7af"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:880e6865cf72cb67f9ab8d04a3c4b49dd95ae92fb1583929ce65aed94e1f685f"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:91290f83daf80ce6d1a7f629b244443c200060a80f908b29d879021409e5ea94"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3517d568486ab3508a7acc360b82d7a4a3e26b86efdf210a9ecd9d233c40708a"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e1f43a77eb38540f782999e5dc5645164fe9027d3f0194f6c9a5126168017efa"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b684d9818aa5d63fddc65f7d0151968037d255d91adf74eba82125b41c680aaa"}, + {file = "safetensors-0.4.2-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:ab1f5d84185f9fefaf21413efb764e4908057b8a9a0b987ede890c353490fd70"}, + {file = "safetensors-0.4.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:2bd979642e6c3a517ef4b84ff36c2fee4015664fea05a61154fc565978347553"}, + {file = "safetensors-0.4.2-cp38-none-win32.whl", hash = "sha256:11be6e7afed29e5a5628f0aa6214e34bc194da73f558dc69fc7d56e07037422a"}, + {file = "safetensors-0.4.2-cp38-none-win_amd64.whl", hash = "sha256:2f7a6e5d29bd2cc340cffaa391fa437b1be9d21a2bd8b8724d2875d13a6ef2a9"}, + {file = "safetensors-0.4.2-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:a5a921b4fe6925f9942adff3ebae8c16e0487908c54586a5a42f35b59fd69794"}, + {file = "safetensors-0.4.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:b691727228c28f2d82d8a92b2bc26e7a1f129ee40b2f2a3185b5974e038ed47c"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:91ca1056decc4e981248786e87b2a202d4841ee5f99d433f1adf3d44d4bcfa0e"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:55969fd2e6fdb38dc221b0ab380668c21b0efa12a7562db9924759faa3c51757"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6ae429bfaecc10ab5fe78c93009b3d1656c1581da560041e700eadb497dbe7a4"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:4ff88f194fe4ac50b463a4a6f0c03af9ad72eb5d24ec6d6730af59522e37fedb"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a80cb48d0a447f8dd18e61813efa7d3f8f8d52edf0f05806abc0c59b83431f57"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b286fb7adfee70a4189898ac2342b8a67d5f493e6b21b0af89ca8eac1b967cbf"}, + {file = "safetensors-0.4.2-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:0ceeff9ddbab4f78738489eb6682867ae946178776f33699737b2129b5394dc1"}, + {file = "safetensors-0.4.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:a26fae748a7488cb3aac381eddfa818c42052c87b5e689fb4c6e82ed58cec209"}, + {file = "safetensors-0.4.2-cp39-none-win32.whl", hash = "sha256:039a42ab33c9d68b39706fd38f1922ace26866eff246bf20271edb619f5f848b"}, + {file = "safetensors-0.4.2-cp39-none-win_amd64.whl", hash = "sha256:b3a3e1f5b85859e398773f064943b62a4059f225008a2a8ee6add1edcf77cacf"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:4e70d442ad17e8b153ef9095bf48ea64f15a66bf26dc2b6ca94660c154edbc24"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:b90f1d9809caf4ff395951b4703295a68d12907f6945bbc3129e934ff8ae46f6"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8c7ac9ad3728838006598e296b3ae9f27d80b489effd4685b92d97b3fc4c98f6"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:de5730d77e6ff7f4c7039e20913661ad0ea2f86c09e71c039e73dfdd1f394f08"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:44feb8cb156d6803dcd19fc6b81b27235f29b877660605a6ac35e1da7d64f0e4"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:523a241c33e7c827ab9a3a23760d75c7d062f43dfe55b6b019409f89b0fb52d1"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:fb18300e8eb74291225214f26c9a8ae2110fd61a6c9b5a2ff4c4e0eb1bb9a998"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-macosx_10_12_x86_64.whl", hash = "sha256:fe5437ff9fb116e44f2ab558981249ae63f978392b4576e62fcfe167d353edbc"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d9304a0934ced5a5d272f39de36291dc141dfc152d277f03fb4d65f2fb2ffa7c"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:160ba1b1e11cf874602c233ab80a14f588571d09556cbc3586900121d622b5ed"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:04fcd6fcf7d9c13c7e5dc7e08de5e492ee4daa8f4ad74b4d8299d3eb0224292f"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:906d14c4a677d35834fb0f3a5455ef8305e1bba10a5e0f2e0f357b3d1ad989f2"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:df3fcdec0cd543084610d1f09c65cdb10fb3079f79bceddc092b0d187c6a265b"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-macosx_10_12_x86_64.whl", hash = "sha256:5ca76f13fb1cef242ea3ad2cb37388e7d005994f42af8b44bee56ba48b2d45ce"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-macosx_11_0_arm64.whl", hash = "sha256:278a1a3414c020785decdcd741c578725721274d2f9f787fcc930882e83b89cc"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:05b5a461cc68ecd42d9d546e5e1268a39d8ede7934a68d1ce17c3c659cb829d6"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c2341411412a41671d25e26bed59ec121e46bf4fadb8132895e610411c4b9681"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:3497ac3895acf17c5f98197f1fa4769f09c5e7ede07fcb102f1c201e663e052c"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:01b5e71d3754d2201294f1eb7a6d59cce3a5702ff96d83d226571b2ca2183837"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:3627dbd1ea488dd8046a0491de5087f3c0d641e7acc80c0189a33c69398f1cd1"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-macosx_10_12_x86_64.whl", hash = "sha256:9d56f0ef53afad26ec54ceede78a43e9a23a076dadbbda7b44d304c591abf4c1"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-macosx_11_0_arm64.whl", hash = "sha256:b259ca73d42daf658a1bda463f1f83885ae4d93a60869be80d7f7dfcc9d8bbb5"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1ebc3cd401e4eb54e7c0a70346be565e81942d9a41fafd5f4bf7ab3a55d10378"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5bc384a0309b706aa0425c93abb0390508a61bf029ce99c7d9df4220f25871a5"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:af2d8f7235d8a08fbccfb8394387890e7fa38942b349a94e6eff13c52ac98087"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:0911315bbcc5289087d063c2c2c7ccd711ea97a7e557a7bce005ac2cf80146aa"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:1efe31673be91832d73439a2af426743e1395fc9ef7b081914e9e1d567bd7b5f"}, + {file = "safetensors-0.4.2.tar.gz", hash = "sha256:acc85dcb09ec5e8aa787f588d7ad4d55c103f31e4ff060e17d92cc0e8b8cac73"}, +] + +[package.extras] +all = ["safetensors[jax]", "safetensors[numpy]", "safetensors[paddlepaddle]", "safetensors[pinned-tf]", "safetensors[quality]", "safetensors[testing]", "safetensors[torch]"] +dev = ["safetensors[all]"] +jax = ["flax (>=0.6.3)", "jax (>=0.3.25)", "jaxlib (>=0.3.25)", "safetensors[numpy]"] +mlx = ["mlx (>=0.0.9)"] +numpy = ["numpy (>=1.21.6)"] +paddlepaddle = ["paddlepaddle (>=2.4.1)", "safetensors[numpy]"] +pinned-tf = ["safetensors[numpy]", "tensorflow (==2.11.0)"] +quality = ["black (==22.3)", "click (==8.0.4)", "flake8 (>=3.8.3)", "isort (>=5.5.4)"] +tensorflow = ["safetensors[numpy]", "tensorflow (>=2.11.0)"] +testing = ["h5py (>=3.7.0)", "huggingface_hub (>=0.12.1)", "hypothesis (>=6.70.2)", "pytest (>=7.2.0)", "pytest-benchmark (>=4.0.0)", "safetensors[numpy]", "setuptools_rust (>=1.5.2)"] +torch = ["safetensors[numpy]", "torch (>=1.10)"] + +[[package]] +name = "sentencepiece" +version = "0.1.99" +description = "SentencePiece python wrapper" +optional = false +python-versions = "*" +files = [ + {file = "sentencepiece-0.1.99-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:0eb528e70571b7c02723e5804322469b82fe7ea418c96051d0286c0fa028db73"}, + {file = "sentencepiece-0.1.99-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:77d7fafb2c4e4659cbdf303929503f37a26eabc4ff31d3a79bf1c5a1b338caa7"}, + {file = "sentencepiece-0.1.99-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:be9cf5b9e404c245aeb3d3723c737ba7a8f5d4ba262ef233a431fa6c45f732a0"}, + {file = "sentencepiece-0.1.99-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:baed1a26464998f9710d20e52607c29ffd4293e7c71c6a1f83f51ad0911ec12c"}, + {file = "sentencepiece-0.1.99-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:9832f08bb372d4c8b567612f8eab9e36e268dff645f1c28f9f8e851be705f6d1"}, + {file = "sentencepiece-0.1.99-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:019e7535108e309dae2b253a75834fc3128240aa87c00eb80732078cdc182588"}, + {file = "sentencepiece-0.1.99-cp310-cp310-win32.whl", hash = "sha256:fa16a830416bb823fa2a52cbdd474d1f7f3bba527fd2304fb4b140dad31bb9bc"}, + {file = "sentencepiece-0.1.99-cp310-cp310-win_amd64.whl", hash = "sha256:14b0eccb7b641d4591c3e12ae44cab537d68352e4d3b6424944f0c447d2348d5"}, + {file = "sentencepiece-0.1.99-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:6d3c56f24183a1e8bd61043ff2c58dfecdc68a5dd8955dc13bab83afd5f76b81"}, + {file = "sentencepiece-0.1.99-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ed6ea1819fd612c989999e44a51bf556d0ef6abfb553080b9be3d347e18bcfb7"}, + {file = "sentencepiece-0.1.99-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:a2a0260cd1fb7bd8b4d4f39dc2444a8d5fd4e0a0c4d5c899810ef1abf99b2d45"}, + {file = "sentencepiece-0.1.99-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8a1abff4d1ff81c77cac3cc6fefa34fa4b8b371e5ee51cb7e8d1ebc996d05983"}, + {file = "sentencepiece-0.1.99-cp311-cp311-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:004e6a621d4bc88978eecb6ea7959264239a17b70f2cbc348033d8195c9808ec"}, + {file = "sentencepiece-0.1.99-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:db361e03342c41680afae5807590bc88aa0e17cfd1a42696a160e4005fcda03b"}, + {file = "sentencepiece-0.1.99-cp311-cp311-win32.whl", hash = "sha256:2d95e19168875b70df62916eb55428a0cbcb834ac51d5a7e664eda74def9e1e0"}, + {file = "sentencepiece-0.1.99-cp311-cp311-win_amd64.whl", hash = "sha256:f90d73a6f81248a909f55d8e6ef56fec32d559e1e9af045f0b0322637cb8e5c7"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-macosx_10_9_x86_64.whl", hash = "sha256:62e24c81e74bd87a6e0d63c51beb6527e4c0add67e1a17bac18bcd2076afcfeb"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:57efcc2d51caff20d9573567d9fd3f854d9efe613ed58a439c78c9f93101384a"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6a904c46197993bd1e95b93a6e373dca2f170379d64441041e2e628ad4afb16f"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d89adf59854741c0d465f0e1525b388c0d174f611cc04af54153c5c4f36088c4"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-win32.whl", hash = "sha256:47c378146928690d1bc106fdf0da768cebd03b65dd8405aa3dd88f9c81e35dba"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-win_amd64.whl", hash = "sha256:9ba142e7a90dd6d823c44f9870abdad45e6c63958eb60fe44cca6828d3b69da2"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:b7b1a9ae4d7c6f1f867e63370cca25cc17b6f4886729595b885ee07a58d3cec3"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d0f644c9d4d35c096a538507b2163e6191512460035bf51358794a78515b74f7"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:c8843d23a0f686d85e569bd6dcd0dd0e0cbc03731e63497ca6d5bacd18df8b85"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:33e6f690a1caebb4867a2e367afa1918ad35be257ecdb3455d2bbd787936f155"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-win32.whl", hash = "sha256:8a321866c2f85da7beac74a824b4ad6ddc2a4c9bccd9382529506d48f744a12c"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-win_amd64.whl", hash = "sha256:c42f753bcfb7661c122a15b20be7f684b61fc8592c89c870adf52382ea72262d"}, + {file = "sentencepiece-0.1.99-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:85b476406da69c70586f0bb682fcca4c9b40e5059814f2db92303ea4585c650c"}, + {file = "sentencepiece-0.1.99-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:cfbcfe13c69d3f87b7fcd5da168df7290a6d006329be71f90ba4f56bc77f8561"}, + {file = "sentencepiece-0.1.99-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:445b0ec381af1cd4eef95243e7180c63d9c384443c16c4c47a28196bd1cda937"}, + {file = "sentencepiece-0.1.99-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c6890ea0f2b4703f62d0bf27932e35808b1f679bdb05c7eeb3812b935ba02001"}, + {file = "sentencepiece-0.1.99-cp38-cp38-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:fb71af492b0eefbf9f2501bec97bcd043b6812ab000d119eaf4bd33f9e283d03"}, + {file = "sentencepiece-0.1.99-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:27b866b5bd3ddd54166bbcbf5c8d7dd2e0b397fac8537991c7f544220b1f67bc"}, + {file = "sentencepiece-0.1.99-cp38-cp38-win32.whl", hash = "sha256:b133e8a499eac49c581c3c76e9bdd08c338cc1939e441fee6f92c0ccb5f1f8be"}, + {file = "sentencepiece-0.1.99-cp38-cp38-win_amd64.whl", hash = "sha256:0eaf3591dd0690a87f44f4df129cf8d05d8a4029b5b6709b489b8e27f9a9bcff"}, + {file = "sentencepiece-0.1.99-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:38efeda9bbfb55052d482a009c6a37e52f42ebffcea9d3a98a61de7aee356a28"}, + {file = "sentencepiece-0.1.99-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:6c030b081dc1e1bcc9fadc314b19b740715d3d566ad73a482da20d7d46fd444c"}, + {file = "sentencepiece-0.1.99-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:84dbe53e02e4f8a2e45d2ac3e430d5c83182142658e25edd76539b7648928727"}, + {file = "sentencepiece-0.1.99-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0b0f55d0a0ee1719b4b04221fe0c9f0c3461dc3dabd77a035fa2f4788eb3ef9a"}, + {file = "sentencepiece-0.1.99-cp39-cp39-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:18e800f206cd235dc27dc749299e05853a4e4332e8d3dfd81bf13d0e5b9007d9"}, + {file = "sentencepiece-0.1.99-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2ae1c40cda8f9d5b0423cfa98542735c0235e7597d79caf318855cdf971b2280"}, + {file = "sentencepiece-0.1.99-cp39-cp39-win32.whl", hash = "sha256:c84ce33af12ca222d14a1cdd37bd76a69401e32bc68fe61c67ef6b59402f4ab8"}, + {file = "sentencepiece-0.1.99-cp39-cp39-win_amd64.whl", hash = "sha256:350e5c74d739973f1c9643edb80f7cc904dc948578bcb1d43c6f2b173e5d18dd"}, + {file = "sentencepiece-0.1.99.tar.gz", hash = "sha256:189c48f5cb2949288f97ccdb97f0473098d9c3dcf5a3d99d4eabe719ec27297f"}, +] + +[[package]] +name = "sympy" +version = "1.12" +description = "Computer algebra system (CAS) in Python" +optional = false +python-versions = ">=3.8" +files = [ + {file = "sympy-1.12-py3-none-any.whl", hash = "sha256:c3588cd4295d0c0f603d0f2ae780587e64e2efeedb3521e46b9bb1d08d184fa5"}, + {file = "sympy-1.12.tar.gz", hash = "sha256:ebf595c8dac3e0fdc4152c51878b498396ec7f30e7a914d6071e674d49420fb8"}, +] + +[package.dependencies] +mpmath = ">=0.19" + +[[package]] +name = "tokenizers" +version = "0.15.2" +description = "" +optional = false +python-versions = ">=3.7" +files = [ + {file = "tokenizers-0.15.2-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:52f6130c9cbf70544287575a985bf44ae1bda2da7e8c24e97716080593638012"}, + {file = "tokenizers-0.15.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:054c1cc9c6d68f7ffa4e810b3d5131e0ba511b6e4be34157aa08ee54c2f8d9ee"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:a9b9b070fdad06e347563b88c278995735292ded1132f8657084989a4c84a6d5"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ea621a7eef4b70e1f7a4e84dd989ae3f0eeb50fc8690254eacc08acb623e82f1"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:cf7fd9a5141634fa3aa8d6b7be362e6ae1b4cda60da81388fa533e0b552c98fd"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:44f2a832cd0825295f7179eaf173381dc45230f9227ec4b44378322d900447c9"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:8b9ec69247a23747669ec4b0ca10f8e3dfb3545d550258129bd62291aabe8605"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:40b6a4c78da863ff26dbd5ad9a8ecc33d8a8d97b535172601cf00aee9d7ce9ce"}, + {file = "tokenizers-0.15.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:5ab2a4d21dcf76af60e05af8063138849eb1d6553a0d059f6534357bce8ba364"}, + {file = "tokenizers-0.15.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:a47acfac7e511f6bbfcf2d3fb8c26979c780a91e06fb5b9a43831b2c0153d024"}, + {file = "tokenizers-0.15.2-cp310-none-win32.whl", hash = "sha256:064ff87bb6acdbd693666de9a4b692add41308a2c0ec0770d6385737117215f2"}, + {file = "tokenizers-0.15.2-cp310-none-win_amd64.whl", hash = "sha256:3b919afe4df7eb6ac7cafd2bd14fb507d3f408db7a68c43117f579c984a73843"}, + {file = "tokenizers-0.15.2-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:89cd1cb93e4b12ff39bb2d626ad77e35209de9309a71e4d3d4672667b4b256e7"}, + {file = "tokenizers-0.15.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:cfed5c64e5be23d7ee0f0e98081a25c2a46b0b77ce99a4f0605b1ec43dd481fa"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:a907d76dcfda37023ba203ab4ceeb21bc5683436ebefbd895a0841fd52f6f6f2"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:20ea60479de6fc7b8ae756b4b097572372d7e4032e2521c1bbf3d90c90a99ff0"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:48e2b9335be2bc0171df9281385c2ed06a15f5cf121c44094338306ab7b33f2c"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:112a1dd436d2cc06e6ffdc0b06d55ac019a35a63afd26475205cb4b1bf0bfbff"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:4620cca5c2817177ee8706f860364cc3a8845bc1e291aaf661fb899e5d1c45b0"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ccd73a82751c523b3fc31ff8194702e4af4db21dc20e55b30ecc2079c5d43cb7"}, + {file = "tokenizers-0.15.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:107089f135b4ae7817affe6264f8c7a5c5b4fd9a90f9439ed495f54fcea56fb4"}, + {file = "tokenizers-0.15.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:0ff110ecc57b7aa4a594396525a3451ad70988e517237fe91c540997c4e50e29"}, + {file = "tokenizers-0.15.2-cp311-none-win32.whl", hash = "sha256:6d76f00f5c32da36c61f41c58346a4fa7f0a61be02f4301fd30ad59834977cc3"}, + {file = "tokenizers-0.15.2-cp311-none-win_amd64.whl", hash = "sha256:cc90102ed17271cf0a1262babe5939e0134b3890345d11a19c3145184b706055"}, + {file = "tokenizers-0.15.2-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:f86593c18d2e6248e72fb91c77d413a815153b8ea4e31f7cd443bdf28e467670"}, + {file = "tokenizers-0.15.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0774bccc6608eca23eb9d620196687c8b2360624619623cf4ba9dc9bd53e8b51"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:d0222c5b7c9b26c0b4822a82f6a7011de0a9d3060e1da176f66274b70f846b98"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3835738be1de66624fff2f4f6f6684775da4e9c00bde053be7564cbf3545cc66"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:0143e7d9dcd811855c1ce1ab9bf5d96d29bf5e528fd6c7824d0465741e8c10fd"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:db35825f6d54215f6b6009a7ff3eedee0848c99a6271c870d2826fbbedf31a38"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3f5e64b0389a2be47091d8cc53c87859783b837ea1a06edd9d8e04004df55a5c"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9e0480c452217edd35eca56fafe2029fb4d368b7c0475f8dfa3c5c9c400a7456"}, + {file = "tokenizers-0.15.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:a33ab881c8fe70474980577e033d0bc9a27b7ab8272896e500708b212995d834"}, + {file = "tokenizers-0.15.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:a308a607ca9de2c64c1b9ba79ec9a403969715a1b8ba5f998a676826f1a7039d"}, + {file = "tokenizers-0.15.2-cp312-none-win32.whl", hash = "sha256:b8fcfa81bcb9447df582c5bc96a031e6df4da2a774b8080d4f02c0c16b42be0b"}, + {file = "tokenizers-0.15.2-cp312-none-win_amd64.whl", hash = "sha256:38d7ab43c6825abfc0b661d95f39c7f8af2449364f01d331f3b51c94dcff7221"}, + {file = "tokenizers-0.15.2-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:38bfb0204ff3246ca4d5e726e8cc8403bfc931090151e6eede54d0e0cf162ef0"}, + {file = "tokenizers-0.15.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:9c861d35e8286a53e06e9e28d030b5a05bcbf5ac9d7229e561e53c352a85b1fc"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:936bf3842db5b2048eaa53dade907b1160f318e7c90c74bfab86f1e47720bdd6"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:620beacc3373277700d0e27718aa8b25f7b383eb8001fba94ee00aeea1459d89"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:2735ecbbf37e52db4ea970e539fd2d450d213517b77745114f92867f3fc246eb"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:473c83c5e2359bb81b0b6fde870b41b2764fcdd36d997485e07e72cc3a62264a"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:968fa1fb3c27398b28a4eca1cbd1e19355c4d3a6007f7398d48826bbe3a0f728"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:865c60ae6eaebdde7da66191ee9b7db52e542ed8ee9d2c653b6d190a9351b980"}, + {file = "tokenizers-0.15.2-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:7c0d8b52664ab2d4a8d6686eb5effc68b78608a9008f086a122a7b2996befbab"}, + {file = "tokenizers-0.15.2-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:f33dfbdec3784093a9aebb3680d1f91336c56d86cc70ddf88708251da1fe9064"}, + {file = "tokenizers-0.15.2-cp37-cp37m-macosx_10_12_x86_64.whl", hash = "sha256:d44ba80988ff9424e33e0a49445072ac7029d8c0e1601ad25a0ca5f41ed0c1d6"}, + {file = "tokenizers-0.15.2-cp37-cp37m-macosx_11_0_arm64.whl", hash = "sha256:dce74266919b892f82b1b86025a613956ea0ea62a4843d4c4237be2c5498ed3a"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:0ef06b9707baeb98b316577acb04f4852239d856b93e9ec3a299622f6084e4be"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c73e2e74bbb07910da0d37c326869f34113137b23eadad3fc00856e6b3d9930c"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:4eeb12daf02a59e29f578a865f55d87cd103ce62bd8a3a5874f8fdeaa82e336b"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9ba9f6895af58487ca4f54e8a664a322f16c26bbb442effd01087eba391a719e"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ccec77aa7150e38eec6878a493bf8c263ff1fa8a62404e16c6203c64c1f16a26"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f3f40604f5042ff210ba82743dda2b6aa3e55aa12df4e9f2378ee01a17e2855e"}, + {file = "tokenizers-0.15.2-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:5645938a42d78c4885086767c70923abad047163d809c16da75d6b290cb30bbe"}, + {file = "tokenizers-0.15.2-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:05a77cbfebe28a61ab5c3891f9939cc24798b63fa236d84e5f29f3a85a200c00"}, + {file = "tokenizers-0.15.2-cp37-none-win32.whl", hash = "sha256:361abdc068e8afe9c5b818769a48624687fb6aaed49636ee39bec4e95e1a215b"}, + {file = "tokenizers-0.15.2-cp37-none-win_amd64.whl", hash = "sha256:7ef789f83eb0f9baeb4d09a86cd639c0a5518528f9992f38b28e819df397eb06"}, + {file = "tokenizers-0.15.2-cp38-cp38-macosx_10_12_x86_64.whl", hash = "sha256:4fe1f74a902bee74a3b25aff180fbfbf4f8b444ab37c4d496af7afd13a784ed2"}, + {file = "tokenizers-0.15.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:4c4b89038a684f40a6b15d6b09f49650ac64d951ad0f2a3ea9169687bbf2a8ba"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:d05a1b06f986d41aed5f2de464c003004b2df8aaf66f2b7628254bcbfb72a438"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:508711a108684111ec8af89d3a9e9e08755247eda27d0ba5e3c50e9da1600f6d"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:daa348f02d15160cb35439098ac96e3a53bacf35885072611cd9e5be7d333daa"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:494fdbe5932d3416de2a85fc2470b797e6f3226c12845cadf054dd906afd0442"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:c2d60f5246f4da9373f75ff18d64c69cbf60c3bca597290cea01059c336d2470"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:93268e788825f52de4c7bdcb6ebc1fcd4a5442c02e730faa9b6b08f23ead0e24"}, + {file = "tokenizers-0.15.2-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:6fc7083ab404019fc9acafe78662c192673c1e696bd598d16dc005bd663a5cf9"}, + {file = "tokenizers-0.15.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:41e39b41e5531d6b2122a77532dbea60e171ef87a3820b5a3888daa847df4153"}, + {file = "tokenizers-0.15.2-cp38-none-win32.whl", hash = "sha256:06cd0487b1cbfabefb2cc52fbd6b1f8d4c37799bd6c6e1641281adaa6b2504a7"}, + {file = "tokenizers-0.15.2-cp38-none-win_amd64.whl", hash = "sha256:5179c271aa5de9c71712e31cb5a79e436ecd0d7532a408fa42a8dbfa4bc23fd9"}, + {file = "tokenizers-0.15.2-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:82f8652a74cc107052328b87ea8b34291c0f55b96d8fb261b3880216a9f9e48e"}, + {file = "tokenizers-0.15.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:02458bee6f5f3139f1ebbb6d042b283af712c0981f5bc50edf771d6b762d5e4f"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:c9a09cd26cca2e1c349f91aa665309ddb48d71636370749414fbf67bc83c5343"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:158be8ea8554e5ed69acc1ce3fbb23a06060bd4bbb09029431ad6b9a466a7121"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:1ddba9a2b0c8c81633eca0bb2e1aa5b3a15362b1277f1ae64176d0f6eba78ab1"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:3ef5dd1d39797044642dbe53eb2bc56435308432e9c7907728da74c69ee2adca"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:454c203164e07a860dbeb3b1f4a733be52b0edbb4dd2e5bd75023ffa8b49403a"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0cf6b7f1d4dc59af960e6ffdc4faffe6460bbfa8dce27a58bf75755ffdb2526d"}, + {file = "tokenizers-0.15.2-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:2ef09bbc16519f6c25d0c7fc0c6a33a6f62923e263c9d7cca4e58b8c61572afb"}, + {file = "tokenizers-0.15.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:c9a2ebdd2ad4ec7a68e7615086e633857c85e2f18025bd05d2a4399e6c5f7169"}, + {file = "tokenizers-0.15.2-cp39-none-win32.whl", hash = "sha256:918fbb0eab96fe08e72a8c2b5461e9cce95585d82a58688e7f01c2bd546c79d0"}, + {file = "tokenizers-0.15.2-cp39-none-win_amd64.whl", hash = "sha256:524e60da0135e106b254bd71f0659be9f89d83f006ea9093ce4d1fab498c6d0d"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:6a9b648a58281c4672212fab04e60648fde574877d0139cd4b4f93fe28ca8944"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:7c7d18b733be6bbca8a55084027f7be428c947ddf871c500ee603e375013ffba"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:13ca3611de8d9ddfbc4dc39ef54ab1d2d4aaa114ac8727dfdc6a6ec4be017378"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:237d1bf3361cf2e6463e6c140628e6406766e8b27274f5fcc62c747ae3c6f094"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:67a0fe1e49e60c664915e9fb6b0cb19bac082ab1f309188230e4b2920230edb3"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:4e022fe65e99230b8fd89ebdfea138c24421f91c1a4f4781a8f5016fd5cdfb4d"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:d857be2df69763362ac699f8b251a8cd3fac9d21893de129bc788f8baaef2693"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-macosx_10_12_x86_64.whl", hash = "sha256:708bb3e4283177236309e698da5fcd0879ce8fd37457d7c266d16b550bcbbd18"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:64c35e09e9899b72a76e762f9854e8750213f67567787d45f37ce06daf57ca78"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c1257f4394be0d3b00de8c9e840ca5601d0a4a8438361ce9c2b05c7d25f6057b"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:02272fe48280e0293a04245ca5d919b2c94a48b408b55e858feae9618138aeda"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:dc3ad9ebc76eabe8b1d7c04d38be884b8f9d60c0cdc09b0aa4e3bcf746de0388"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:32e16bdeffa7c4f46bf2152172ca511808b952701d13e7c18833c0b73cb5c23f"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-macosx_10_12_x86_64.whl", hash = "sha256:fb16ba563d59003028b678d2361a27f7e4ae0ab29c7a80690efa20d829c81fdb"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-macosx_11_0_arm64.whl", hash = "sha256:2277c36d2d6cdb7876c274547921a42425b6810d38354327dd65a8009acf870c"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:1cf75d32e8d250781940d07f7eece253f2fe9ecdb1dc7ba6e3833fa17b82fcbc"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f1b3b31884dc8e9b21508bb76da80ebf7308fdb947a17affce815665d5c4d028"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b10122d8d8e30afb43bb1fe21a3619f62c3e2574bff2699cf8af8b0b6c5dc4a3"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:d88b96ff0fe8e91f6ef01ba50b0d71db5017fa4e3b1d99681cec89a85faf7bf7"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:37aaec5a52e959892870a7c47cef80c53797c0db9149d458460f4f31e2fb250e"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-macosx_10_12_x86_64.whl", hash = "sha256:e2ea752f2b0fe96eb6e2f3adbbf4d72aaa1272079b0dfa1145507bd6a5d537e6"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-macosx_11_0_arm64.whl", hash = "sha256:4b19a808d8799fda23504a5cd31d2f58e6f52f140380082b352f877017d6342b"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:64c86e5e068ac8b19204419ed8ca90f9d25db20578f5881e337d203b314f4104"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:de19c4dc503c612847edf833c82e9f73cd79926a384af9d801dcf93f110cea4e"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ea09acd2fe3324174063d61ad620dec3bcf042b495515f27f638270a7d466e8b"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:cf27fd43472e07b57cf420eee1e814549203d56de00b5af8659cb99885472f1f"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:7ca22bd897537a0080521445d91a58886c8c04084a6a19e6c78c586e0cfa92a5"}, + {file = "tokenizers-0.15.2.tar.gz", hash = "sha256:e6e9c6e019dd5484be5beafc775ae6c925f4c69a3487040ed09b45e13df2cb91"}, +] + +[package.dependencies] +huggingface_hub = ">=0.16.4,<1.0" + +[package.extras] +dev = ["tokenizers[testing]"] +docs = ["setuptools_rust", "sphinx", "sphinx_rtd_theme"] +testing = ["black (==22.3)", "datasets", "numpy", "pytest", "requests"] + +[[package]] +name = "torch" +version = "2.2.1+cpu" +description = "Tensors and Dynamic neural networks in Python with strong GPU acceleration" +optional = false +python-versions = ">=3.8.0" +files = [ + {file = "torch-2.2.1+cpu-cp310-cp310-linux_x86_64.whl", hash = "sha256:5d82422cf04797f1b2a8574b64a916070ec83eef58ad4900615ee0218d7b8b8e"}, + {file = "torch-2.2.1+cpu-cp310-cp310-win_amd64.whl", hash = "sha256:f8914dd0f5f0e5c66fdecd9559403eea9feac82d1ea639b672fde0073c6addbd"}, + {file = "torch-2.2.1+cpu-cp311-cp311-linux_x86_64.whl", hash = "sha256:6bc973d5632374b92b4b293817b4d2ff8c8ce1c784c748b471dba1fffcd9c333"}, + {file = "torch-2.2.1+cpu-cp311-cp311-win_amd64.whl", hash = "sha256:abdec34b0ade8fca0520055e72c3094425ae0ef210718e9c0278121cd3608c32"}, + {file = "torch-2.2.1+cpu-cp312-cp312-linux_x86_64.whl", hash = "sha256:d7339580135da4105c1244a8621faa076990409afeab5a7b642c3c1ee70a5622"}, + {file = "torch-2.2.1+cpu-cp312-cp312-win_amd64.whl", hash = "sha256:039128fcb5548122465b15f679b8831c47d14f0d6c28c1f1b631f8019c104720"}, + {file = "torch-2.2.1+cpu-cp38-cp38-linux_x86_64.whl", hash = "sha256:2b447f7bb50b393b4544b4036d587e39ab524d4353e77c197f6a2727f22b0d47"}, + {file = "torch-2.2.1+cpu-cp38-cp38-win_amd64.whl", hash = "sha256:2ccdf3e5f71e6426ea9e34d21c3cc333b29d4f48299b981d28aeb5112b5495e1"}, + {file = "torch-2.2.1+cpu-cp39-cp39-linux_x86_64.whl", hash = "sha256:2fb340b289760040a16a77a6d70b8a48961abba1822e6f58705c97c80befa03e"}, + {file = "torch-2.2.1+cpu-cp39-cp39-win_amd64.whl", hash = "sha256:e03dc4654ecceeb5b03f0a6f60b342c0e0d267b3ebc61e4f672cace1df8cd930"}, +] + +[package.dependencies] +filelock = "*" +fsspec = "*" +jinja2 = "*" +networkx = "*" +sympy = "*" +typing-extensions = ">=4.8.0" + +[package.extras] +opt-einsum = ["opt-einsum (>=3.3)"] +optree = ["optree (>=0.9.1)"] + +[package.source] +type = "legacy" +url = "https://download.pytorch.org/whl/cpu" +reference = "pytorch" + +[[package]] +name = "tqdm" +version = "4.66.2" +description = "Fast, Extensible Progress Meter" +optional = false +python-versions = ">=3.7" +files = [ + {file = "tqdm-4.66.2-py3-none-any.whl", hash = "sha256:1ee4f8a893eb9bef51c6e35730cebf234d5d0b6bd112b0271e10ed7c24a02bd9"}, + {file = "tqdm-4.66.2.tar.gz", hash = "sha256:6cd52cdf0fef0e0f543299cfc96fec90d7b8a7e88745f411ec33eb44d5ed3531"}, +] + +[package.dependencies] +colorama = {version = "*", markers = "platform_system == \"Windows\""} + +[package.extras] +dev = ["pytest (>=6)", "pytest-cov", "pytest-timeout", "pytest-xdist"] +notebook = ["ipywidgets (>=6)"] +slack = ["slack-sdk"] +telegram = ["requests"] + +[[package]] +name = "transformers" +version = "4.38.1" +description = "State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow" +optional = false +python-versions = ">=3.8.0" +files = [ + {file = "transformers-4.38.1-py3-none-any.whl", hash = "sha256:a7a9265fb060183e9d975cbbadc4d531b10281589c43f6d07563f86322728973"}, + {file = "transformers-4.38.1.tar.gz", hash = "sha256:86dc84ccbe36123647e84cbd50fc31618c109a41e6be92514b064ab55bf1304c"}, +] + +[package.dependencies] +filelock = "*" +huggingface-hub = ">=0.19.3,<1.0" +numpy = ">=1.17" +packaging = ">=20.0" +pyyaml = ">=5.1" +regex = "!=2019.12.17" +requests = "*" +safetensors = ">=0.4.1" +tokenizers = ">=0.14,<0.19" +tqdm = ">=4.27" + +[package.extras] +accelerate = ["accelerate (>=0.21.0)"] +agents = ["Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "datasets (!=2.5.0)", "diffusers", "opencv-python", "sentencepiece (>=0.1.91,!=0.1.92)", "torch"] +all = ["Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "av (==9.2.0)", "codecarbon (==1.2.0)", "decord (==0.6.0)", "flax (>=0.4.1,<=0.7.0)", "jax (>=0.4.1,<=0.4.13)", "jaxlib (>=0.4.1,<=0.4.13)", "kenlm", "keras-nlp (>=0.3.1)", "librosa", "onnxconverter-common", "optax (>=0.0.8,<=0.1.4)", "optuna", "phonemizer", "protobuf", "pyctcdecode (>=0.4.0)", "ray[tune] (>=2.7.0)", "sentencepiece (>=0.1.91,!=0.1.92)", "sigopt", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx", "timm", "tokenizers (>=0.14,<0.19)", "torch", "torchaudio", "torchvision"] +audio = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)"] +codecarbon = ["codecarbon (==1.2.0)"] +deepspeed = ["accelerate (>=0.21.0)", "deepspeed (>=0.9.3)"] +deepspeed-testing = ["GitPython (<3.1.19)", "accelerate (>=0.21.0)", "beautifulsoup4", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "deepspeed (>=0.9.3)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "hf-doc-builder (>=0.3.0)", "nltk", "optuna", "parameterized", "protobuf", "psutil", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "sentencepiece (>=0.1.91,!=0.1.92)", "tensorboard", "timeout-decorator"] +dev = ["GitPython (<3.1.19)", "Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "av (==9.2.0)", "beautifulsoup4", "codecarbon (==1.2.0)", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "decord (==0.6.0)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "flax (>=0.4.1,<=0.7.0)", "fugashi (>=1.0)", "hf-doc-builder", "hf-doc-builder (>=0.3.0)", "ipadic (>=1.0.0,<2.0)", "isort (>=5.5.4)", "jax (>=0.4.1,<=0.4.13)", "jaxlib (>=0.4.1,<=0.4.13)", "kenlm", "keras-nlp (>=0.3.1)", "librosa", "nltk", "onnxconverter-common", "optax (>=0.0.8,<=0.1.4)", "optuna", "parameterized", "phonemizer", "protobuf", "psutil", "pyctcdecode (>=0.4.0)", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "ray[tune] (>=2.7.0)", "rhoknp (>=1.1.0,<1.3.1)", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "scikit-learn", "sentencepiece (>=0.1.91,!=0.1.92)", "sigopt", "sudachidict-core (>=20220729)", "sudachipy (>=0.6.6)", "tensorboard", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx", "timeout-decorator", "timm", "tokenizers (>=0.14,<0.19)", "torch", "torchaudio", "torchvision", "unidic (>=1.0.2)", "unidic-lite (>=1.0.7)", "urllib3 (<2.0.0)"] +dev-tensorflow = ["GitPython (<3.1.19)", "Pillow (>=10.0.1,<=15.0)", "beautifulsoup4", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "hf-doc-builder", "hf-doc-builder (>=0.3.0)", "isort (>=5.5.4)", "kenlm", "keras-nlp (>=0.3.1)", "librosa", "nltk", "onnxconverter-common", "onnxruntime (>=1.4.0)", "onnxruntime-tools (>=1.4.2)", "parameterized", "phonemizer", "protobuf", "psutil", "pyctcdecode (>=0.4.0)", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "scikit-learn", "sentencepiece (>=0.1.91,!=0.1.92)", "tensorboard", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx", "timeout-decorator", "tokenizers (>=0.14,<0.19)", "urllib3 (<2.0.0)"] +dev-torch = ["GitPython (<3.1.19)", "Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "beautifulsoup4", "codecarbon (==1.2.0)", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "fugashi (>=1.0)", "hf-doc-builder", "hf-doc-builder (>=0.3.0)", "ipadic (>=1.0.0,<2.0)", "isort (>=5.5.4)", "kenlm", "librosa", "nltk", "onnxruntime (>=1.4.0)", "onnxruntime-tools (>=1.4.2)", "optuna", "parameterized", "phonemizer", "protobuf", "psutil", "pyctcdecode (>=0.4.0)", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "ray[tune] (>=2.7.0)", "rhoknp (>=1.1.0,<1.3.1)", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "scikit-learn", "sentencepiece (>=0.1.91,!=0.1.92)", "sigopt", "sudachidict-core (>=20220729)", "sudachipy (>=0.6.6)", "tensorboard", "timeout-decorator", "timm", "tokenizers (>=0.14,<0.19)", "torch", "torchaudio", "torchvision", "unidic (>=1.0.2)", "unidic-lite (>=1.0.7)", "urllib3 (<2.0.0)"] +docs = ["Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "av (==9.2.0)", "codecarbon (==1.2.0)", "decord (==0.6.0)", "flax (>=0.4.1,<=0.7.0)", "hf-doc-builder", "jax (>=0.4.1,<=0.4.13)", "jaxlib (>=0.4.1,<=0.4.13)", "kenlm", "keras-nlp (>=0.3.1)", "librosa", "onnxconverter-common", "optax (>=0.0.8,<=0.1.4)", "optuna", "phonemizer", "protobuf", "pyctcdecode (>=0.4.0)", "ray[tune] (>=2.7.0)", "sentencepiece (>=0.1.91,!=0.1.92)", "sigopt", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx", "timm", "tokenizers (>=0.14,<0.19)", "torch", "torchaudio", "torchvision"] +docs-specific = ["hf-doc-builder"] +flax = ["flax (>=0.4.1,<=0.7.0)", "jax (>=0.4.1,<=0.4.13)", "jaxlib (>=0.4.1,<=0.4.13)", "optax (>=0.0.8,<=0.1.4)"] +flax-speech = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)"] +ftfy = ["ftfy"] +integrations = ["optuna", "ray[tune] (>=2.7.0)", "sigopt"] +ja = ["fugashi (>=1.0)", "ipadic (>=1.0.0,<2.0)", "rhoknp (>=1.1.0,<1.3.1)", "sudachidict-core (>=20220729)", "sudachipy (>=0.6.6)", "unidic (>=1.0.2)", "unidic-lite (>=1.0.7)"] +modelcreation = ["cookiecutter (==1.7.3)"] +natten = ["natten (>=0.14.6,<0.15.0)"] +onnx = ["onnxconverter-common", "onnxruntime (>=1.4.0)", "onnxruntime-tools (>=1.4.2)", "tf2onnx"] +onnxruntime = ["onnxruntime (>=1.4.0)", "onnxruntime-tools (>=1.4.2)"] +optuna = ["optuna"] +quality = ["GitPython (<3.1.19)", "datasets (!=2.5.0)", "hf-doc-builder (>=0.3.0)", "isort (>=5.5.4)", "ruff (==0.1.5)", "urllib3 (<2.0.0)"] +ray = ["ray[tune] (>=2.7.0)"] +retrieval = ["datasets (!=2.5.0)", "faiss-cpu"] +sagemaker = ["sagemaker (>=2.31.0)"] +sentencepiece = ["protobuf", "sentencepiece (>=0.1.91,!=0.1.92)"] +serving = ["fastapi", "pydantic", "starlette", "uvicorn"] +sigopt = ["sigopt"] +sklearn = ["scikit-learn"] +speech = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)", "torchaudio"] +testing = ["GitPython (<3.1.19)", "beautifulsoup4", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "hf-doc-builder (>=0.3.0)", "nltk", "parameterized", "protobuf", "psutil", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "tensorboard", "timeout-decorator"] +tf = ["keras-nlp (>=0.3.1)", "onnxconverter-common", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx"] +tf-cpu = ["keras-nlp (>=0.3.1)", "onnxconverter-common", "tensorflow-cpu (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx"] +tf-speech = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)"] +timm = ["timm"] +tokenizers = ["tokenizers (>=0.14,<0.19)"] +torch = ["accelerate (>=0.21.0)", "torch"] +torch-speech = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)", "torchaudio"] +torch-vision = ["Pillow (>=10.0.1,<=15.0)", "torchvision"] +torchhub = ["filelock", "huggingface-hub (>=0.19.3,<1.0)", "importlib-metadata", "numpy (>=1.17)", "packaging (>=20.0)", "protobuf", "regex (!=2019.12.17)", "requests", "sentencepiece (>=0.1.91,!=0.1.92)", "tokenizers (>=0.14,<0.19)", "torch", "tqdm (>=4.27)"] +video = ["av (==9.2.0)", "decord (==0.6.0)"] +vision = ["Pillow (>=10.0.1,<=15.0)"] + +[[package]] +name = "typing-extensions" +version = "4.9.0" +description = "Backported and Experimental Type Hints for Python 3.8+" +optional = false +python-versions = ">=3.8" +files = [ + {file = "typing_extensions-4.9.0-py3-none-any.whl", hash = "sha256:af72aea155e91adfc61c3ae9e0e342dbc0cba726d6cba4b6c72c1f34e47291cd"}, + {file = "typing_extensions-4.9.0.tar.gz", hash = "sha256:23478f88c37f27d76ac8aee6c905017a143b0b1b886c3c9f66bc2fd94f9f5783"}, +] + +[[package]] +name = "urllib3" +version = "2.2.1" +description = "HTTP library with thread-safe connection pooling, file post, and more." +optional = false +python-versions = ">=3.8" +files = [ + {file = "urllib3-2.2.1-py3-none-any.whl", hash = "sha256:450b20ec296a467077128bff42b73080516e71b56ff59a60a02bef2232c4fa9d"}, + {file = "urllib3-2.2.1.tar.gz", hash = "sha256:d0570876c61ab9e520d776c38acbbb5b05a776d3f9ff98a5c8fd5162a444cf19"}, +] + +[package.extras] +brotli = ["brotli (>=1.0.9)", "brotlicffi (>=0.8.0)"] +h2 = ["h2 (>=4,<5)"] +socks = ["pysocks (>=1.5.6,!=1.5.7,<2.0)"] +zstd = ["zstandard (>=0.18.0)"] + +[[package]] +name = "wcwidth" +version = "0.2.13" +description = "Measures the displayed width of unicode strings in a terminal" +optional = false +python-versions = "*" +files = [ + {file = "wcwidth-0.2.13-py2.py3-none-any.whl", hash = "sha256:3da69048e4540d84af32131829ff948f1e022c1c6bdb8d6102117aac784f6859"}, + {file = "wcwidth-0.2.13.tar.gz", hash = "sha256:72ea0c06399eb286d978fdedb6923a9eb47e1c486ce63e9b4e64fc18303972b5"}, +] + +[metadata] +lock-version = "2.0" +python-versions = ">=3.9" +content-hash = "c8c4cc87637266a7b85debcbafa8887c5ad81cc8ef40e98a3f52c7c50af05c03" diff --git a/pyproject.toml b/pyproject.toml new file mode 100644 index 000000000..8759d99e1 --- /dev/null +++ b/pyproject.toml @@ -0,0 +1,44 @@ +[tool.poetry] +name = "llama-cpp-scripts" +version = "0.0.0" +description = "Scripts that ship with llama.cpp" +authors = ["GGML "] +readme = "README.md" +homepage = "https://ggml.ai" +repository = "https://github.com/ggerganov/llama.cpp" +keywords = ["ggml", "gguf", "llama.cpp"] +packages = [{ include = "__init__.py", from = "." }] +classifiers = [ + "Programming Language :: Python :: 3", + "License :: OSI Approved :: MIT License", + "Operating System :: OS Independent", +] + +[tool.poetry.dependencies] +python = ">=3.9" +numpy = "^1.25.0" +sentencepiece = ">=0.1.98,<0.2.0" +transformers = ">=4.35.2,<5.0.0" +protobuf = ">=4.21.0,<5.0.0" +gguf = { path = "./gguf-py" } +torch = {version = "^2.2.0", source = "pytorch"} + +[tool.poetry.dev-dependencies] +pytest = "^5.2" + +[[tool.poetry.source]] +name = "pytorch" +url = "https://download.pytorch.org/whl/cpu" +priority = "explicit" + +[build-system] +requires = ["poetry-core>=1.0.0"] +build-backend = "poetry.core.masonry.api" + +[tool.poetry.scripts] +llama-convert-hf-to-gguf = "convert_hf_to_gguf:main" +llama-convert-llama-ggml-to-gguf = "convert_llama_ggml_to_gguf:main" +llama-convert-lora-to-ggml = "convert_lora_to_ggml:main" +llama-convert-persimmon-to-gguf = "convert_persimmon_to_gguf:main" +llama-convert = "convert:main" +llama-ggml_vk_generate_shaders = "ggml_vk_generate_shaders:main" From b1c3f26e5e882fa46d7a6704d893b31a025e9000 Mon Sep 17 00:00:00 2001 From: ditsuke Date: Thu, 29 Feb 2024 01:47:15 +0530 Subject: [PATCH 30/38] fix: Actually include scripts in build Not namespaced though :( --- __init__.py | 0 pyproject.toml | 4 ++-- 2 files changed, 2 insertions(+), 2 deletions(-) delete mode 100644 __init__.py diff --git a/__init__.py b/__init__.py deleted file mode 100644 index e69de29bb..000000000 diff --git a/pyproject.toml b/pyproject.toml index 8759d99e1..2c6e855e3 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -7,7 +7,7 @@ readme = "README.md" homepage = "https://ggml.ai" repository = "https://github.com/ggerganov/llama.cpp" keywords = ["ggml", "gguf", "llama.cpp"] -packages = [{ include = "__init__.py", from = "." }] +packages = [{ include = "*.py", from = "." }] classifiers = [ "Programming Language :: Python :: 3", "License :: OSI Approved :: MIT License", @@ -21,7 +21,7 @@ sentencepiece = ">=0.1.98,<0.2.0" transformers = ">=4.35.2,<5.0.0" protobuf = ">=4.21.0,<5.0.0" gguf = { path = "./gguf-py" } -torch = {version = "^2.2.0", source = "pytorch"} +torch = { version = "^2.2.0", source = "pytorch" } [tool.poetry.dev-dependencies] pytest = "^5.2" From 821922916f95cc3853fc7f58ea2f1e15a0ffa669 Mon Sep 17 00:00:00 2001 From: ditsuke Date: Sun, 10 Mar 2024 23:21:46 +0530 Subject: [PATCH 31/38] fix: Update script paths in CI scripts --- ci/run.sh | 2 +- requirements.txt | 6 +++--- ...-gguf-update.txt => requirements-convert_hf_to_gguf.txt} | 0 ...-gguf.txt => requirements-convert_hf_to_gguf_update.txt} | 0 ...gguf.txt => requirements-convert_llama_ggml_to_gguf.txt} | 0 requirements/requirements-convert_lora_to_ggml.txt | 3 +++ requirements/requirements-convert_persimmon_to_gguf.txt | 3 +++ scripts/check-requirements.sh | 4 ++-- 8 files changed, 12 insertions(+), 6 deletions(-) rename requirements/{requirements-convert-hf-to-gguf-update.txt => requirements-convert_hf_to_gguf.txt} (100%) rename requirements/{requirements-convert-hf-to-gguf.txt => requirements-convert_hf_to_gguf_update.txt} (100%) rename requirements/{requirements-convert-llama-ggml-to-gguf.txt => requirements-convert_llama_ggml_to_gguf.txt} (100%) create mode 100644 requirements/requirements-convert_lora_to_ggml.txt create mode 100644 requirements/requirements-convert_persimmon_to_gguf.txt diff --git a/ci/run.sh b/ci/run.sh index e0cedb24f..067ac405b 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -688,7 +688,7 @@ function gg_run_embd_bge_small { (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log - python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf + python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf model_f16="${path_models}/ggml-model-f16.gguf" model_q8_0="${path_models}/ggml-model-q8_0.gguf" diff --git a/requirements.txt b/requirements.txt index e5cfbf10b..1eca1a13f 100644 --- a/requirements.txt +++ b/requirements.txt @@ -6,6 +6,6 @@ -r ./requirements/requirements-convert-legacy-llama.txt --r ./requirements/requirements-convert-hf-to-gguf.txt --r ./requirements/requirements-convert-hf-to-gguf-update.txt --r ./requirements/requirements-convert-llama-ggml-to-gguf.txt +-r ./requirements/requirements-convert_hf_to_gguf.txt +-r ./requirements/requirements-convert_hf_to_gguf_update.txt +-r ./requirements/requirements-convert_llama_ggml_to_gguf.txt diff --git a/requirements/requirements-convert-hf-to-gguf-update.txt b/requirements/requirements-convert_hf_to_gguf.txt similarity index 100% rename from requirements/requirements-convert-hf-to-gguf-update.txt rename to requirements/requirements-convert_hf_to_gguf.txt diff --git a/requirements/requirements-convert-hf-to-gguf.txt b/requirements/requirements-convert_hf_to_gguf_update.txt similarity index 100% rename from requirements/requirements-convert-hf-to-gguf.txt rename to requirements/requirements-convert_hf_to_gguf_update.txt diff --git a/requirements/requirements-convert-llama-ggml-to-gguf.txt b/requirements/requirements-convert_llama_ggml_to_gguf.txt similarity index 100% rename from requirements/requirements-convert-llama-ggml-to-gguf.txt rename to requirements/requirements-convert_llama_ggml_to_gguf.txt diff --git a/requirements/requirements-convert_lora_to_ggml.txt b/requirements/requirements-convert_lora_to_ggml.txt new file mode 100644 index 000000000..fe5d13dd8 --- /dev/null +++ b/requirements/requirements-convert_lora_to_ggml.txt @@ -0,0 +1,3 @@ +-r ./requirements-convert.txt +torch~=2.1.1 + diff --git a/requirements/requirements-convert_persimmon_to_gguf.txt b/requirements/requirements-convert_persimmon_to_gguf.txt new file mode 100644 index 000000000..fe5d13dd8 --- /dev/null +++ b/requirements/requirements-convert_persimmon_to_gguf.txt @@ -0,0 +1,3 @@ +-r ./requirements-convert.txt +torch~=2.1.1 + diff --git a/scripts/check-requirements.sh b/scripts/check-requirements.sh index 0c6afdd59..69a08c841 100755 --- a/scripts/check-requirements.sh +++ b/scripts/check-requirements.sh @@ -167,11 +167,11 @@ if (( do_cleanup )); then fi check_convert_script examples/convert-legacy-llama.py -for py in convert-*.py; do +for py in convert_*.py; do # skip convert-hf-to-gguf-update.py # TODO: the check is failing for some reason: # https://github.com/ggerganov/llama.cpp/actions/runs/8875330981/job/24364557177?pr=6920 - [[ $py == convert-hf-to-gguf-update.py ]] && continue + [[ $py == convert_hf_to_gguf_update.py ]] && continue check_convert_script "$py" done From de14e2ea2b542386dcb800226eb360be76f433fd Mon Sep 17 00:00:00 2001 From: ditsuke Date: Tue, 2 Jul 2024 15:18:13 +0530 Subject: [PATCH 32/38] chore: ignore all __pychache__ --- .gitignore | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.gitignore b/.gitignore index 7a63cda19..6125e350c 100644 --- a/.gitignore +++ b/.gitignore @@ -99,7 +99,7 @@ examples/server/*.mjs.hpp # Python /.venv -/__pycache__/ +__pycache__/ */poetry.lock poetry.toml From 07786a61a2097306ee496f565f78796f1aa205e6 Mon Sep 17 00:00:00 2001 From: ditsuke Date: Tue, 2 Jul 2024 15:35:43 +0530 Subject: [PATCH 33/38] chore: Fixup requirements and build --- CMakeLists.txt | 4 +++- examples/llava/requirements.txt | 2 +- pyproject.toml | 2 +- requirements.txt | 2 ++ requirements/requirements-convert_lora_to_ggml.txt | 4 ++-- requirements/requirements-convert_persimmon_to_gguf.txt | 4 ++-- 6 files changed, 11 insertions(+), 7 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index d95414d71..f7b6601b3 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -156,7 +156,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama) install( - FILES convert-hf-to-gguf.py + FILES convert_hf_to_gguf.py PERMISSIONS OWNER_READ OWNER_WRITE @@ -165,6 +165,8 @@ install( GROUP_EXECUTE WORLD_READ WORLD_EXECUTE + RENAME + convert-hf-to-gguf.py DESTINATION ${CMAKE_INSTALL_BINDIR}) configure_file(cmake/llama.pc.in diff --git a/examples/llava/requirements.txt b/examples/llava/requirements.txt index 17cb4d5e5..21149d6fe 100644 --- a/examples/llava/requirements.txt +++ b/examples/llava/requirements.txt @@ -1,3 +1,3 @@ -r ../../requirements/requirements-convert-legacy-llama.txt pillow~=10.2.0 -torch~=2.1.1 +torch~=2.2.1 diff --git a/pyproject.toml b/pyproject.toml index 2c6e855e3..e94cd1ba6 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -41,4 +41,4 @@ llama-convert-llama-ggml-to-gguf = "convert_llama_ggml_to_gguf:main" llama-convert-lora-to-ggml = "convert_lora_to_ggml:main" llama-convert-persimmon-to-gguf = "convert_persimmon_to_gguf:main" llama-convert = "convert:main" -llama-ggml_vk_generate_shaders = "ggml_vk_generate_shaders:main" +llama-ggml-vk-generate-shaders = "ggml_vk_generate_shaders:main" diff --git a/requirements.txt b/requirements.txt index 1eca1a13f..4f9bcc9ca 100644 --- a/requirements.txt +++ b/requirements.txt @@ -9,3 +9,5 @@ -r ./requirements/requirements-convert_hf_to_gguf.txt -r ./requirements/requirements-convert_hf_to_gguf_update.txt -r ./requirements/requirements-convert_llama_ggml_to_gguf.txt +-r ./requirements/requirements-convert_lora_to_ggml.txt +-r ./requirements/requirements-convert_persimmon_to_gguf.txt diff --git a/requirements/requirements-convert_lora_to_ggml.txt b/requirements/requirements-convert_lora_to_ggml.txt index fe5d13dd8..4135cd655 100644 --- a/requirements/requirements-convert_lora_to_ggml.txt +++ b/requirements/requirements-convert_lora_to_ggml.txt @@ -1,3 +1,3 @@ --r ./requirements-convert.txt -torch~=2.1.1 +-r ./requirements-convert-legacy-llama.txt +torch~=2.2.1 diff --git a/requirements/requirements-convert_persimmon_to_gguf.txt b/requirements/requirements-convert_persimmon_to_gguf.txt index fe5d13dd8..4135cd655 100644 --- a/requirements/requirements-convert_persimmon_to_gguf.txt +++ b/requirements/requirements-convert_persimmon_to_gguf.txt @@ -1,3 +1,3 @@ --r ./requirements-convert.txt -torch~=2.1.1 +-r ./requirements-convert-legacy-llama.txt +torch~=2.2.1 From 01a5f06550a866bd63717635e5e7e1ff4203b873 Mon Sep 17 00:00:00 2001 From: ditsuke Date: Tue, 2 Jul 2024 15:48:13 +0530 Subject: [PATCH 34/38] chore: Remove rebase artifacts --- convert_hf_to_gguf_update.py | 20 +-- convert_lora_to_ggml.py | 149 ------------------ convert_persimmon_to_gguf.py | 137 ---------------- pyproject.toml | 3 - requirements.txt | 2 - .../requirements-convert_lora_to_ggml.txt | 3 - ...requirements-convert_persimmon_to_gguf.txt | 3 - 7 files changed, 7 insertions(+), 310 deletions(-) mode change 100644 => 100755 convert_hf_to_gguf_update.py delete mode 100755 convert_lora_to_ggml.py delete mode 100755 convert_persimmon_to_gguf.py delete mode 100644 requirements/requirements-convert_lora_to_ggml.txt delete mode 100644 requirements/requirements-convert_persimmon_to_gguf.txt diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py old mode 100644 new mode 100755 index ca337af2c..21a306255 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -50,7 +50,7 @@ class TOKENIZER_TYPE(IntEnum): # TODO: this string has to exercise as much pre-tokenizer functionality as possible # will be updated with time - contributions welcome -chktxt = "\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````\"\"\"\"......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL" +chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL' if len(sys.argv) == 2: token = sys.argv[1] @@ -99,7 +99,7 @@ def download_file_with_auth(url, token, save_path): response = sess.get(url, headers=headers) response.raise_for_status() os.makedirs(os.path.dirname(save_path), exist_ok=True) - with open(save_path, "wb") as f: + with open(save_path, 'wb') as f: f.write(response.content) logger.info(f"File {save_path} downloaded successfully") @@ -156,9 +156,7 @@ for model in models: else: tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") except OSError as e: - logger.error( - f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}" - ) + logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}") continue # Skip to the next model if the tokenizer can't be loaded chktok = tokenizer.encode(chktxt) @@ -178,15 +176,13 @@ for model in models: pre_tokenizer = cfg["pre_tokenizer"] logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4)) if "ignore_merges" in cfg["model"]: - logger.info( - "ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4) - ) + logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4)) logger.info("") - src_ifs += f' if chkhsh == "{chkhsh}":\n' + src_ifs += f" if chkhsh == \"{chkhsh}\":\n" src_ifs += f" # ref: {model['repo']}\n" - src_ifs += f' res = "{name}"\n' + src_ifs += f" res = \"{name}\"\n" src_func = f""" def get_vocab_base_pre(self, tokenizer) -> str: @@ -347,8 +343,6 @@ logger.info("\nRun the following commands to generate the vocab files for testin for model in models: name = model["name"] - print( - f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only" - ) # noqa: NP100 + print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100 logger.info("\n") diff --git a/convert_lora_to_ggml.py b/convert_lora_to_ggml.py deleted file mode 100755 index 276d0d63a..000000000 --- a/convert_lora_to_ggml.py +++ /dev/null @@ -1,149 +0,0 @@ -#!/usr/bin/env python3 -from __future__ import annotations - -import json -import os -import struct -import sys -from pathlib import Path -from typing import Any, BinaryIO, Sequence - -import numpy as np -import torch - -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) -import gguf - -NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1} - - -def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None: - fout.write(b"ggla"[::-1]) # magic (ggml lora) - fout.write(struct.pack("i", 1)) # file version - fout.write(struct.pack("i", params["r"])) - # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int - # but some models ship a float value instead - # let's convert to int, but fail if lossless conversion is not possible - assert ( - int(params["lora_alpha"]) == params["lora_alpha"] - ), "cannot convert float to int losslessly" - fout.write(struct.pack("i", int(params["lora_alpha"]))) - - -def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None: - sname = name.encode("utf-8") - fout.write( - struct.pack( - "iii", - len(shape), - len(sname), - NUMPY_TYPE_TO_FTYPE[data_type.name], - ) - ) - fout.write(struct.pack("i" * len(shape), *shape[::-1])) - fout.write(sname) - fout.seek((fout.tell() + 31) & -32) - - -if __name__ == '__main__': - if len(sys.argv) < 2: - print(f"Usage: python {sys.argv[0]} [arch]") - print( - "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'" - ) - print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)") - sys.exit(1) - - input_json = os.path.join(sys.argv[1], "adapter_config.json") - input_model = os.path.join(sys.argv[1], "adapter_model.bin") - output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin") - - if os.path.exists(input_model): - model = torch.load(input_model, map_location="cpu") - else: - input_model = os.path.join(sys.argv[1], "adapter_model.safetensors") - # lazy import load_file only if lora is in safetensors format. - from safetensors.torch import load_file - model = load_file(input_model, device="cpu") - - arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama" - - if arch_name not in gguf.MODEL_ARCH_NAMES.values(): - print(f"Error: unsupported architecture {arch_name}") - sys.exit(1) - - arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)] - name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone - - with open(input_json, "r") as f: - params = json.load(f) - - if params["peft_type"] != "LORA": - print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA") - sys.exit(1) - - if params["fan_in_fan_out"] is True: - print("Error: param fan_in_fan_out is not supported") - sys.exit(1) - - if params["bias"] is not None and params["bias"] != "none": - print("Error: param bias is not supported") - sys.exit(1) - - # TODO: these seem to be layers that have been trained but without lora. - # doesn't seem widely used but eventually should be supported - if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0: - print("Error: param modules_to_save is not supported") - sys.exit(1) - - with open(output_path, "wb") as fout: - fout.truncate() - - write_file_header(fout, params) - for k, v in model.items(): - orig_k = k - if k.endswith(".default.weight"): - k = k.replace(".default.weight", ".weight") - if k in ["llama_proj.weight", "llama_proj.bias"]: - continue - if k.endswith("lora_A.weight"): - if v.dtype != torch.float16 and v.dtype != torch.float32: - v = v.float() - v = v.T - else: - v = v.float() - - t = v.detach().numpy() - - prefix = "base_model.model." - if k.startswith(prefix): - k = k[len(prefix) :] - - lora_suffixes = (".lora_A.weight", ".lora_B.weight") - if k.endswith(lora_suffixes): - suffix = k[-len(lora_suffixes[0]):] - k = k[: -len(lora_suffixes[0])] - else: - print(f"Error: unrecognized tensor name {orig_k}") - sys.exit(1) - - tname = name_map.get_name(k) - if tname is None: - print(f"Error: could not map tensor name {orig_k}") - print(" Note: the arch parameter must be specified if the model is not llama") - sys.exit(1) - - if suffix == ".lora_A.weight": - tname += ".weight.loraA" - elif suffix == ".lora_B.weight": - tname += ".weight.loraB" - else: - assert False - - print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB") - write_tensor_header(fout, tname, t.shape, t.dtype) - t.tofile(fout) - - print(f"Converted {input_json} and {input_model} to {output_path}") - diff --git a/convert_persimmon_to_gguf.py b/convert_persimmon_to_gguf.py deleted file mode 100755 index e1fe3c521..000000000 --- a/convert_persimmon_to_gguf.py +++ /dev/null @@ -1,137 +0,0 @@ -#!/usr/bin/env python3 -import argparse -import os -import sys -from pathlib import Path -from pprint import pprint - -import torch -from sentencepiece import SentencePieceProcessor - -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) -import gguf - - -def _flatten_dict(dct, tensors, prefix=None): - assert isinstance(dct, dict) - for key in dct.keys(): - new_prefix = prefix + '.' + key if prefix is not None else key - if isinstance(dct[key], torch.Tensor): - tensors[new_prefix] = dct[key] - elif isinstance(dct[key], dict): - _flatten_dict(dct[key], tensors, new_prefix) - else: - raise ValueError(type(dct[key])) - return None - - -def _get_sentencepiece_tokenizer_info(dir_model: Path): - tokenizer_path = dir_model / 'adept_vocab.model' - print('gguf: getting sentencepiece tokenizer from', tokenizer_path) - tokenizer = SentencePieceProcessor(str(tokenizer_path)) - print('gguf: adding tokens') - tokens: list[bytes] = [] - scores: list[float] = [] - toktypes: list[int] = [] - - for i in range(tokenizer.vocab_size()): - text: bytes - score: float - - piece = tokenizer.id_to_piece(i) - text = piece.encode("utf-8") - score = tokenizer.get_score(i) - - toktype = 1 - if tokenizer.is_unknown(i): - toktype = 2 - if tokenizer.is_control(i): - toktype = 3 - if tokenizer.is_unused(i): - toktype = 5 - if tokenizer.is_byte(i): - toktype = 6 - - tokens.append(text) - scores.append(score) - toktypes.append(toktype) - pass - return tokens, scores, toktypes - - -def main(): - parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file") - parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("--ckpt-path", type=Path, help="path to persimmon checkpoint .pt file") - parser.add_argument("--model-dir", type=Path, help="directory containing model e.g. 8b_chat_model_release") - parser.add_argument("--adept-inference-dir", type=str, help="path to adept-inference code directory") - args = parser.parse_args() - sys.path.append(str(args.adept_inference_dir)) - persimmon_model = torch.load(args.ckpt_path) - hparams = persimmon_model['args'] - pprint(hparams) - tensors: dict[str, torch.Tensor] = {} - _flatten_dict(persimmon_model['model'], tensors, None) - - arch = gguf.MODEL_ARCH.PERSIMMON - gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch]) - - block_count = hparams.num_layers - head_count = hparams.num_attention_heads - head_count_kv = head_count - ctx_length = hparams.seq_length - hidden_size = hparams.hidden_size - - gguf_writer.add_name('persimmon-8b-chat') - gguf_writer.add_context_length(ctx_length) - gguf_writer.add_embedding_length(hidden_size) - gguf_writer.add_block_count(block_count) - gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size) - # ref: https://github.com/ggerganov/llama.cpp/pull/4889/commits/eea19039fc52ea2dbd1aab45b59ab4e3e29a3443 - gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2) - gguf_writer.add_head_count(head_count) - gguf_writer.add_head_count_kv(head_count_kv) - gguf_writer.add_rope_freq_base(hparams.rotary_emb_base) - gguf_writer.add_layer_norm_eps(hparams.layernorm_epsilon) - - tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir) - gguf_writer.add_tokenizer_model('llama') - gguf_writer.add_token_list(tokens) - gguf_writer.add_token_scores(scores) - gguf_writer.add_token_types(toktypes) - gguf_writer.add_bos_token_id(71013) - gguf_writer.add_eos_token_id(71013) - - tensor_map = gguf.get_tensor_name_map(arch, block_count) - print(tensor_map) - for name in tensors.keys(): - data = tensors[name] - if name.endswith(".self_attention.rotary_emb.inv_freq"): - continue - old_dtype = data.dtype - # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?) - data = data.to(torch.float32).squeeze().numpy() - new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) - if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() - n_dims = len(data.shape) - print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - gguf_writer.add_tensor(new_name, data) - print("gguf: write header") - gguf_writer.write_header_to_file() - print("gguf: write metadata") - gguf_writer.write_kv_data_to_file() - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - - gguf_writer.close() - - print(f"gguf: model successfully exported to '{args.outfile}'") - print("") - - -if __name__ == '__main__': - main() - diff --git a/pyproject.toml b/pyproject.toml index e94cd1ba6..198cb43dd 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -38,7 +38,4 @@ build-backend = "poetry.core.masonry.api" [tool.poetry.scripts] llama-convert-hf-to-gguf = "convert_hf_to_gguf:main" llama-convert-llama-ggml-to-gguf = "convert_llama_ggml_to_gguf:main" -llama-convert-lora-to-ggml = "convert_lora_to_ggml:main" -llama-convert-persimmon-to-gguf = "convert_persimmon_to_gguf:main" -llama-convert = "convert:main" llama-ggml-vk-generate-shaders = "ggml_vk_generate_shaders:main" diff --git a/requirements.txt b/requirements.txt index 4f9bcc9ca..1eca1a13f 100644 --- a/requirements.txt +++ b/requirements.txt @@ -9,5 +9,3 @@ -r ./requirements/requirements-convert_hf_to_gguf.txt -r ./requirements/requirements-convert_hf_to_gguf_update.txt -r ./requirements/requirements-convert_llama_ggml_to_gguf.txt --r ./requirements/requirements-convert_lora_to_ggml.txt --r ./requirements/requirements-convert_persimmon_to_gguf.txt diff --git a/requirements/requirements-convert_lora_to_ggml.txt b/requirements/requirements-convert_lora_to_ggml.txt deleted file mode 100644 index 4135cd655..000000000 --- a/requirements/requirements-convert_lora_to_ggml.txt +++ /dev/null @@ -1,3 +0,0 @@ --r ./requirements-convert-legacy-llama.txt -torch~=2.2.1 - diff --git a/requirements/requirements-convert_persimmon_to_gguf.txt b/requirements/requirements-convert_persimmon_to_gguf.txt deleted file mode 100644 index 4135cd655..000000000 --- a/requirements/requirements-convert_persimmon_to_gguf.txt +++ /dev/null @@ -1,3 +0,0 @@ --r ./requirements-convert-legacy-llama.txt -torch~=2.2.1 - From 1e920018d38aee270a044cc48eaefe7c64cb8750 Mon Sep 17 00:00:00 2001 From: ditsuke Date: Wed, 3 Jul 2024 01:02:56 +0530 Subject: [PATCH 35/38] doc: Add context for why we add an explicit pytorch source --- pyproject.toml | 3 +++ 1 file changed, 3 insertions(+) diff --git a/pyproject.toml b/pyproject.toml index 198cb43dd..25e2e20b2 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -26,6 +26,9 @@ torch = { version = "^2.2.0", source = "pytorch" } [tool.poetry.dev-dependencies] pytest = "^5.2" + +# Force wheel + cpu +# For discussion and context see https://github.com/python-poetry/poetry#6409 [[tool.poetry.source]] name = "pytorch" url = "https://download.pytorch.org/whl/cpu" From 51d2ebadbbf365b894f3888361df42dbacb12b7a Mon Sep 17 00:00:00 2001 From: ditsuke Date: Thu, 4 Jul 2024 20:54:35 +0530 Subject: [PATCH 36/38] build: Export hf-to-gguf as snakecase --- CMakeLists.txt | 2 -- 1 file changed, 2 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index f7b6601b3..4f6cd6872 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -165,8 +165,6 @@ install( GROUP_EXECUTE WORLD_READ WORLD_EXECUTE - RENAME - convert-hf-to-gguf.py DESTINATION ${CMAKE_INSTALL_BINDIR}) configure_file(cmake/llama.pc.in From 6f63d646c1a06a6e09f721009a2676864ae04e31 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Thu, 4 Jul 2024 18:38:58 +0200 Subject: [PATCH 37/38] tokenize : add --show-count (token) option (#8299) This commit adds a new option to the tokenize example, --show-count. When this is set the total number of tokens are printed to stdout. This was added as an option as I was concerned that there might be scripts that use the output from this program and it might be better to not print this information by default. The motivation for this is that can be useful to find out how many tokens a file contains, for example when trying to determine prompt input file sizes for testing. Signed-off-by: Daniel Bevenius --- examples/tokenize/tokenize.cpp | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/examples/tokenize/tokenize.cpp b/examples/tokenize/tokenize.cpp index 54c9834af..0180c87d8 100644 --- a/examples/tokenize/tokenize.cpp +++ b/examples/tokenize/tokenize.cpp @@ -30,6 +30,7 @@ static void print_usage_information(const char * argv0, FILE * stream) { fprintf(stream, " --stdin read prompt from standard input.\n"); fprintf(stream, " --no-bos do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n"); fprintf(stream, " --log-disable disable logs. Makes stderr quiet when loading the model.\n"); + fprintf(stream, " --show-count print the total number of tokens.\n"); } static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) { @@ -195,6 +196,7 @@ int main(int raw_argc, char ** raw_argv) { bool printing_ids = false; bool no_bos = false; bool disable_logging = false; + bool show_token_count = false; const char * model_path = NULL; const char * prompt_path = NULL; const char * prompt_arg = NULL; @@ -249,6 +251,9 @@ int main(int raw_argc, char ** raw_argv) { else if (arg == "--log-disable") { disable_logging = true; } + else if (arg == "--show-count") { + show_token_count = true; + } else { fprintf(stderr, "Error: unknown option '%s'\n", argv[iarg].c_str()); return 1; @@ -384,6 +389,9 @@ int main(int raw_argc, char ** raw_argv) { printf("]\n"); } + if (show_token_count) { + printf("Total number of tokens: %ld\n", tokens.size()); + } // silence valgrind llama_free(ctx); llama_free_model(model); From d7fd29fff16456ce9c3a23fd2d09a66256b05aff Mon Sep 17 00:00:00 2001 From: Icecream95 Date: Fri, 5 Jul 2024 05:14:21 +1200 Subject: [PATCH 38/38] llama : add OpenELM support (#7359) * Initial OpenELM support (270M only so far) * Fill out missing entries in llama_model_type_name * fixup! Initial OpenELM support (270M only so far) Fix formatting * llama : support all OpenELM models * llama : add variable GQA and variable FFN sizes Some metadata keys can now also be arrays to support setting their value per-layer for models like OpenELM. * llama : minor spacing changes Co-authored-by: Georgi Gerganov * llama : use std::array for per-layer hparams * llama : fix save/load state * llama : do not print hparams for vocab-only models * llama : handle n_head == 0 * llama : use const ref for print_f and fix division by zero * llama : fix t5 uses of n_head and n_ff * llama : minor comment --------- Co-authored-by: Francis Couture-Harpin Co-authored-by: Georgi Gerganov --- convert_hf_to_gguf.py | 157 ++++++-- gguf-py/gguf/constants.py | 15 + gguf-py/gguf/gguf_writer.py | 21 +- gguf-py/gguf/tensor_mapping.py | 16 +- src/llama.cpp | 643 ++++++++++++++++++++++++++------- 5 files changed, 676 insertions(+), 176 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index a1d165023..bdc336644 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -13,7 +13,7 @@ import sys from enum import IntEnum from pathlib import Path from hashlib import sha256 -from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Sequence, TypeVar, cast +from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast import math import numpy as np @@ -677,6 +677,51 @@ class Model: special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) + def _set_vocab_builtin(self, model_name: Literal["gpt-neox", "llama-spm"], vocab_size: int): + tokenizer_path = Path(sys.path[0]) / "models" / f"ggml-vocab-{model_name}.gguf" + logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'") + vocab_reader = gguf.GGUFReader(tokenizer_path, "r") + + default_pre = "mpt" if model_name == "gpt-neox" else "default" + + field = vocab_reader.get_field(gguf.Keys.Tokenizer.MODEL) + assert field # tokenizer model + self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]).decode("utf-8")) + + field = vocab_reader.get_field(gguf.Keys.Tokenizer.PRE) + self.gguf_writer.add_tokenizer_pre(bytes(field.parts[-1]).decode("utf-8") if field else default_pre) + + field = vocab_reader.get_field(gguf.Keys.Tokenizer.LIST) + assert field # token list + self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size]) + + if model_name == "llama-spm": + field = vocab_reader.get_field(gguf.Keys.Tokenizer.SCORES) + assert field # token scores + self.gguf_writer.add_token_scores([field.parts[i].tolist()[0] for i in field.data][:vocab_size]) + + field = vocab_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE) + assert field # token types + self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size]) + + if model_name != "llama-spm": + field = vocab_reader.get_field(gguf.Keys.Tokenizer.MERGES) + assert field # token merges + self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data]) + + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)) is not None: + self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)) is not None: + self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)) is not None: + self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.PAD_ID)) is not None: + self.gguf_writer.add_pad_token_id(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_BOS)) is not None: + self.gguf_writer.add_add_bos_token(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_EOS)) is not None: + self.gguf_writer.add_add_eos_token(field.parts[-1].tolist()[0]) + @Model.register("GPTNeoXForCausalLM") class GPTNeoXModel(Model): @@ -2439,39 +2484,7 @@ class MambaModel(Model): self._set_vocab_sentencepiece() else: # Use the GPT-NeoX tokenizer when no tokenizer files are present - tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf" - logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'") - neox_reader = gguf.GGUFReader(tokenizer_path, "r") - - field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL) - self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]).decode("utf-8") if field else "gpt2") - - field = neox_reader.get_field(gguf.Keys.Tokenizer.PRE) - self.gguf_writer.add_tokenizer_pre(bytes(field.parts[-1]).decode("utf-8") if field else "mpt") - - field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST) - assert field - self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size]) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE) - assert field - self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size]) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES) - assert field - self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data]) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID) - self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0] if field else 1) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID) - self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0] if field else 0) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID) - self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0] if field else 0) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.PAD_ID) - self.gguf_writer.add_pad_token_id(field.parts[-1].tolist()[0] if field else 0) + self._set_vocab_builtin("gpt-neox", vocab_size) def set_gguf_parameters(self): d_model = self.find_hparam(["hidden_size", "d_model"]) @@ -2623,6 +2636,82 @@ class JinaBertV2Model(BertModel): self.gguf_writer.add_add_eos_token(True) +@Model.register("OpenELMForCausalLM") +class OpenELMModel(Model): + model_arch = gguf.MODEL_ARCH.OPENELM + + @staticmethod + def _make_divisible(v: float | int, divisor: int) -> int: + # ref: https://huggingface.co/apple/OpenELM-270M-Instruct/blob/eb111ff2e6724348e5b905984063d4064d4bc579/configuration_openelm.py#L34-L38 + new_v = max(divisor, int(v + divisor / 2) // divisor * divisor) + # Make sure that round down does not go down by more than 10%. + if new_v < 0.9 * v: + new_v += divisor + return new_v + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + ffn_multipliers: list[float] = self.hparams["ffn_multipliers"] + ffn_dim_divisor: int = self.hparams["ffn_dim_divisor"] + self._n_embd: int = self.hparams["model_dim"] + self._num_kv_heads: list[int] = self.hparams["num_kv_heads"] + self._num_query_heads: list[int] = self.hparams["num_query_heads"] + self._ffn_dims: list[int] = [ + OpenELMModel._make_divisible(multiplier * self._n_embd, ffn_dim_divisor) + for multiplier in ffn_multipliers + ] + assert isinstance(self._num_kv_heads, list) and isinstance(self._num_kv_heads[0], int) + assert isinstance(self._num_query_heads, list) and isinstance(self._num_query_heads[0], int) + + # Uses the tokenizer from meta-llama/Llama-2-7b-hf + def set_vocab(self): + try: + self._set_vocab_sentencepiece() + except FileNotFoundError: + self._set_vocab_builtin("llama-spm", self.hparams["vocab_size"]) + + def set_gguf_parameters(self): + n_embd = self._n_embd + head_dim = self.hparams["head_dim"] + rot_pct = 1.0 + assert self.block_count == len(self._num_kv_heads) + assert self.block_count == len(self._num_query_heads) + assert self.block_count == len(self._ffn_dims) + + self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) + self.gguf_writer.add_block_count(self.block_count) + self.gguf_writer.add_context_length(self.hparams["max_context_length"]) + self.gguf_writer.add_embedding_length(n_embd) + self.gguf_writer.add_feed_forward_length(self._ffn_dims) + self.gguf_writer.add_head_count(self._num_query_heads) + self.gguf_writer.add_head_count_kv(self._num_kv_heads) + self.gguf_writer.add_rope_freq_base(self.hparams["rope_freq_constant"]) + # https://huggingface.co/apple/OpenELM-270M-Instruct/blob/c401df2/modeling_openelm.py#L30 + self.gguf_writer.add_layer_norm_rms_eps(1e-6) + self.gguf_writer.add_rope_dimension_count(int(rot_pct * head_dim)) + self.gguf_writer.add_key_length(head_dim) + self.gguf_writer.add_value_length(head_dim) + self.gguf_writer.add_file_type(self.ftype) + + def find_hparam(self, keys: Iterable[str], optional: bool = False) -> Any: + if "n_layers" in keys: + return self.hparams["num_transformer_layers"] + + return super().find_hparam(keys, optional) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + + # split ff + if bid is not None and name == f"transformer.layers.{bid}.ffn.proj_1.weight": + ff_dim = self._ffn_dims[bid] + yield (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), data_torch[:ff_dim]) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), data_torch[ff_dim:]) + return + + yield (self.map_tensor_name(name), data_torch) + + @Model.register("ArcticForCausalLM") class ArcticModel(Model): model_arch = gguf.MODEL_ARCH.ARCTIC diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 419f10cee..d78fb1e44 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -160,6 +160,7 @@ class MODEL_ARCH(IntEnum): COMMAND_R = auto() DBRX = auto() OLMO = auto() + OPENELM = auto() ARCTIC = auto() DEEPSEEK2 = auto() BITNET = auto() @@ -285,6 +286,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.COMMAND_R: "command-r", MODEL_ARCH.DBRX: "dbrx", MODEL_ARCH.OLMO: "olmo", + MODEL_ARCH.OPENELM: "openelm", MODEL_ARCH.ARCTIC: "arctic", MODEL_ARCH.DEEPSEEK2: "deepseek2", MODEL_ARCH.BITNET: "bitnet", @@ -861,6 +863,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.OPENELM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], MODEL_ARCH.ARCTIC: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 75a8b2636..cf9554162 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -480,8 +480,11 @@ class GGUFWriter: def add_leading_dense_block_count(self, length: int) -> None: self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length) - def add_feed_forward_length(self, length: int) -> None: - self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) + def add_feed_forward_length(self, length: int | Sequence[int]) -> None: + if isinstance(length, int): + self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) + else: + self.add_array(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) def add_expert_feed_forward_length(self, length: int) -> None: self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length) @@ -495,11 +498,17 @@ class GGUFWriter: def add_decoder_start_token_id(self, id: int) -> None: self.add_uint32(Keys.LLM.DECODER_START_TOKEN_ID.format(arch=self.arch), id) - def add_head_count(self, count: int) -> None: - self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) + def add_head_count(self, count: int | Sequence[int]) -> None: + if isinstance(count, int): + self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) + else: + self.add_array(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) - def add_head_count_kv(self, count: int) -> None: - self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count) + def add_head_count_kv(self, count: int | Sequence[int]) -> None: + if isinstance(count, int): + self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count) + else: + self.add_array(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count) def add_key_length(self, length: int) -> None: self.add_uint32(Keys.Attention.KEY_LENGTH.format(arch=self.arch), length) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 20e28423b..6054f6159 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -24,6 +24,7 @@ class TensorNameMap: "backbone.embedding", # mamba "backbone.embeddings", # mamba-hf "transformer.in_out_embed", # Grok + "transformer.token_embeddings", # openelm "shared", # t5 ), @@ -37,6 +38,7 @@ class TensorNameMap: "word_embeddings_layernorm", # bloom "embeddings.LayerNorm", # bert "emb_ln", # nomic-bert + "transformer.norm", # openelm ), # Position embeddings @@ -69,6 +71,7 @@ class TensorNameMap: "model.norm_f", # mamba-qbert "backbone.norm_f", # mamba "transformer.rms_norm", # Grok + "transformer.norm", # openelm ), # Rope frequencies @@ -98,6 +101,7 @@ class TensorNameMap: "backbone.layers.{bid}.norm", # mamba "transformer.decoder_layer.{bid}.rms_norm", # Grok "transformer.blocks.{bid}.norm_attn_norm.norm_1", # dbrx + "transformer.layers.{bid}.attn_norm", # openelm ), # Attention norm 2 @@ -119,7 +123,8 @@ class TensorNameMap: "h.{bid}.attn.c_attn", # gpt2 "transformer.h.{bid}.mixer.Wqkv", # phi2 "encoder.layers.{bid}.attn.Wqkv", # nomic-bert - "model.layers.{bid}.self_attn.qkv_proj" # phi3 + "model.layers.{bid}.self_attn.qkv_proj", # phi3 + "transformer.layers.{bid}.attn.qkv_proj", # openelm ), # Attention query @@ -177,6 +182,7 @@ class TensorNameMap: "encoder.layers.{bid}.attn.out_proj", # nomic-bert "transformer.decoder_layer.{bid}.multi_head_attention.linear", # Grok "transformer.blocks.{bid}.norm_attn_norm.attn.out_proj", # dbrx + "transformer.layers.{bid}.attn.out_proj", # openelm ), # Attention output norm @@ -212,6 +218,7 @@ class TensorNameMap: "h.{bid}.ln_2", # gpt2 "model.layers.{bid}.ffn_norm", # internlm2 "transformer.decoder_layer.{bid}.rms_norm_2", # Grok + "transformer.layers.{bid}.ffn_norm", # openelm ), # Post feed-forward norm @@ -327,6 +334,7 @@ class TensorNameMap: "encoder.layers.{bid}.mlp.fc2", # nomic-bert "model.layers.{bid}.mlp.c_proj", # starcoder2 "encoder.layer.{bid}.mlp.wo", # jina-bert-v2 + "transformer.layers.{bid}.ffn.proj_2", # openelm "model.layers.{bid}.residual_mlp.w2", # arctic "encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2 ), @@ -348,7 +356,8 @@ class TensorNameMap: "model.layers.{bid}.self_attn.q_layernorm", # persimmon "model.layers.{bid}.self_attn.q_norm", # cohere "transformer.blocks.{bid}.attn.q_ln", # sea-lion - "encoder.layer.{bid}.attention.self.layer_norm_q" # jina-bert-v2 + "encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2 + "transformer.layers.{bid}.attn.q_norm", # openelm ), MODEL_TENSOR.ATTN_K_NORM: ( @@ -356,7 +365,8 @@ class TensorNameMap: "model.layers.{bid}.self_attn.k_layernorm", # persimmon "model.layers.{bid}.self_attn.k_norm", # cohere "transformer.blocks.{bid}.attn.k_ln", # sea-lion - "encoder.layer.{bid}.attention.self.layer_norm_k" # jina-bert-v2 + "encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2 + "transformer.layers.{bid}.attn.k_norm", # openelm ), MODEL_TENSOR.ROPE_FREQS: ( diff --git a/src/llama.cpp b/src/llama.cpp index 1fe2b9f79..721b8f4e5 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -104,8 +104,10 @@ #define LLAMA_ATTRIBUTE_FORMAT(...) #endif +// bump if necessary #define LLAMA_MAX_NODES 8192 -#define LLAMA_MAX_EXPERTS 160 +#define LLAMA_MAX_LAYERS 256 +#define LLAMA_MAX_EXPERTS 160 // DeepSeekV2 // // logging @@ -224,6 +226,7 @@ enum llm_arch { LLM_ARCH_COMMAND_R, LLM_ARCH_DBRX, LLM_ARCH_OLMO, + LLM_ARCH_OPENELM, LLM_ARCH_ARCTIC, LLM_ARCH_DEEPSEEK2, LLM_ARCH_BITNET, @@ -266,6 +269,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_COMMAND_R, "command-r" }, { LLM_ARCH_DBRX, "dbrx" }, { LLM_ARCH_OLMO, "olmo" }, + { LLM_ARCH_OPENELM, "openelm" }, { LLM_ARCH_ARCTIC, "arctic" }, { LLM_ARCH_DEEPSEEK2, "deepseek2" }, { LLM_ARCH_BITNET, "bitnet" }, @@ -1134,6 +1138,22 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_OPENELM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_ARCTIC, { @@ -2052,8 +2072,10 @@ enum e_model { MODEL_160M, MODEL_220M, MODEL_250M, + MODEL_270M, MODEL_335M, MODEL_410M, + MODEL_450M, MODEL_770M, MODEL_780M, MODEL_0_5B, @@ -2108,19 +2130,20 @@ struct llama_hparams { uint32_t n_vocab; uint32_t n_ctx_train; // context size the model was trained on uint32_t n_embd; - uint32_t n_head; - uint32_t n_head_kv; uint32_t n_layer; uint32_t n_rot; uint32_t n_swa = 0; // sliding window attention (SWA) uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head - uint32_t n_ff; uint32_t n_expert = 0; uint32_t n_expert_used = 0; uint32_t n_vocab_type = 0; // for BERT-style token types uint32_t n_rel_attn_bkts = 0; + std::array n_head_arr; + std::array n_head_kv_arr; + std::array n_ff_arr; + uint32_t n_layer_dense_lead = 0; uint32_t n_lora_q = 0; uint32_t n_lora_kv = 0; @@ -2168,17 +2191,18 @@ struct llama_hparams { if (this->n_vocab != other.n_vocab) return true; if (this->n_ctx_train != other.n_ctx_train) return true; if (this->n_embd != other.n_embd) return true; - if (this->n_head != other.n_head) return true; - if (this->n_head_kv != other.n_head_kv) return true; if (this->n_layer != other.n_layer) return true; if (this->n_rot != other.n_rot) return true; if (this->n_swa != other.n_swa) return true; if (this->n_embd_head_k != other.n_embd_head_k) return true; if (this->n_embd_head_v != other.n_embd_head_v) return true; - if (this->n_ff != other.n_ff) return true; if (this->n_expert != other.n_expert) return true; if (this->n_expert_used != other.n_expert_used) return true; + if (this->n_head_arr != other.n_head_arr) return true; + if (this->n_head_kv_arr != other.n_head_kv_arr) return true; + if (this->n_ff_arr != other.n_ff_arr) return true; + if (this->n_rel_attn_bkts != other.n_rel_attn_bkts) return true; if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true; if (this->n_lora_q != other.n_lora_q) return true; @@ -2210,18 +2234,53 @@ struct llama_hparams { return false; } - uint32_t n_gqa() const { + uint32_t n_head(uint32_t il = 0) const { + if (il < n_layer) { + return n_head_arr[il]; + } + + GGML_ASSERT(false); + return 0; + } + + uint32_t n_head_kv(uint32_t il = 0) const { + if (il < n_layer) { + return n_head_kv_arr[il]; + } + + GGML_ASSERT(false); + return 0; + } + + uint32_t n_ff(uint32_t il = 0) const { + if (il < n_layer) { + return n_ff_arr[il]; + } + + GGML_ASSERT(false); + return 0; + } + + uint32_t n_gqa(uint32_t il = 0) const { + const uint32_t n_head = this->n_head(il); + const uint32_t n_head_kv = this->n_head_kv(il); + if (n_head_kv == 0) { return 0; } + return n_head/n_head_kv; } - uint32_t n_embd_k_gqa() const { // dimension of key embeddings across all k-v heads + uint32_t n_embd_k_gqa(uint32_t il = 0) const { // dimension of key embeddings across all k-v heads + const uint32_t n_head_kv = this->n_head_kv(il); + return n_embd_head_k * n_head_kv; } - uint32_t n_embd_v_gqa() const { // dimension of value embeddings across all k-v heads + uint32_t n_embd_v_gqa(uint32_t il = 0) const { // dimension of value embeddings across all k-v heads + const uint32_t n_head_kv = this->n_head_kv(il); + return n_embd_head_v * n_head_kv; } @@ -2238,6 +2297,8 @@ struct llama_hparams { } }; +static_assert(std::is_trivially_copyable::value, "llama_hparams must be trivially copyable"); + struct llama_cparams { uint32_t n_ctx; // context size used during inference uint32_t n_batch; @@ -2857,9 +2918,7 @@ static bool llama_kv_cache_init( const struct llama_hparams & hparams = model.hparams; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); - const int64_t n_layer = hparams.n_layer; + const int64_t n_layer = hparams.n_layer; cache.has_shift = false; @@ -2867,13 +2926,6 @@ static bool llama_kv_cache_init( cache.recurrent = model.arch == LLM_ARCH_MAMBA; cache.v_trans = !cparams.flash_attn; - // TODO: support mixed recurrent Transformer architectures - // NOTE: (!a || b) is a logical implication (a -> b) - GGML_ASSERT(!cache.recurrent || n_embd_k_gqa == hparams.n_embd_k_s()); - GGML_ASSERT(!cache.recurrent || n_embd_v_gqa == hparams.n_embd_v_s()); - GGML_ASSERT( cache.recurrent || n_embd_k_gqa == hparams.n_embd_k_gqa()); - GGML_ASSERT( cache.recurrent || n_embd_v_gqa == hparams.n_embd_v_gqa()); - cache.head = 0; cache.size = kv_size; cache.used = 0; @@ -2923,6 +2975,9 @@ static bool llama_kv_cache_init( cache.v_l.reserve(n_layer); for (int i = 0; i < (int) n_layer; i++) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s(); + struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front(); ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size); ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size); @@ -3798,9 +3853,9 @@ struct llama_model_loader { bool get_arr(const std::string & key, std::vector & result, const bool required = true) { const int kid = gguf_find_key(meta, key.c_str()); - if (kid < 0) { + if (kid < 0 || gguf_get_kv_type(meta, kid) != GGUF_TYPE_ARRAY) { if (required) { - throw std::runtime_error(format("key not found in model: %s", key.c_str())); + throw std::runtime_error(format("array key not found in model: %s", key.c_str())); } return false; } @@ -3808,22 +3863,53 @@ struct llama_model_loader { struct GGUFMeta::ArrayInfo arr_info = GGUFMeta::GKV::get_kv(meta, kid); - if (arr_info.gt != GGUF_TYPE_FLOAT32 && arr_info.gt != GGUF_TYPE_INT32) { - throw std::runtime_error(format("%s is not a float32 or int32 array", key.c_str())); + switch (arr_info.gt) { + case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; + case GGUF_TYPE_INT32: GGML_ASSERT( + (std::is_same::value) || + (std::is_same::value)); break; + default: + throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str())); } - // GGML_ASSERT(gguf_type_size(arr_info.gt) == sizeof(T)); - GGML_ASSERT((arr_info.gt != GGUF_TYPE_FLOAT32 || std::is_same::value)); - GGML_ASSERT((arr_info.gt != GGUF_TYPE_INT32 || std::is_same::value)); - result.resize(arr_info.length); result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length); return true; } + template + bool get_arr(const std::string & key, std::array & result, const bool required = true) { + const int kid = gguf_find_key(meta, key.c_str()); + + if (kid < 0 || gguf_get_kv_type(meta, kid) != GGUF_TYPE_ARRAY) { + if (required) { + throw std::runtime_error(format("array key not found in model: %s", key.c_str())); + } + return false; + } + + struct GGUFMeta::ArrayInfo arr_info = + GGUFMeta::GKV::get_kv(meta, kid); + + switch (arr_info.gt) { + case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; + case GGUF_TYPE_INT32: GGML_ASSERT( + (std::is_same::value) || + (std::is_same::value)); break; + default: + throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str())); + } + + GGML_ASSERT(arr_info.length <= N_MAX); + + std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin()); + + return true; + } + template - bool get_arr(const enum llm_kv kid, T& result, const bool required = true) { + bool get_arr(const enum llm_kv kid, T & result, const bool required = true) { return get_arr(llm_kv(kid), result, required); } @@ -3848,6 +3934,50 @@ struct llama_model_loader { return get_key(llm_kv(kid), result, required); } + // get array of n <= N_MAX elements, or a single element repeated n times + template + bool get_key_or_arr(const std::string & key, std::array & result, uint32_t n, const bool required = true) { + GGML_ASSERT(n <= N_MAX); + + const int kid = gguf_find_key(meta, key.c_str()); + + if (kid < 0) { + if (required) { + throw std::runtime_error(format("key not found in model: %s", key.c_str())); + } + return false; + } + + if (gguf_get_kv_type(meta, kid) == GGUF_TYPE_ARRAY) { + struct GGUFMeta::ArrayInfo arr_info = + GGUFMeta::GKV::get_kv(meta, kid); + + if (n != arr_info.length) { + throw std::runtime_error(format("key %s has wrong array length; expected %u, got %u", key.c_str(), n, (uint32_t) arr_info.length)); + } + + return get_arr(key, result, required); + } else { + T value; + + bool ok = get_key(key, value, required); + if (!ok) { + return false; + } + + for (uint32_t i = 0; i < n; i++) { + result[i] = value; + } + + return true; + } + } + + template + bool get_key_or_arr(const enum llm_kv kid, T & result, uint32_t n, const bool required = true) { + return get_key_or_arr(llm_kv(kid), result, n, required); + } + std::string get_arch_name() const { return arch_name; } @@ -4340,8 +4470,10 @@ static const char * llama_model_type_name(e_model type) { case MODEL_160M: return "160M"; case MODEL_220M: return "220M"; case MODEL_250M: return "250M"; + case MODEL_270M: return "270M"; case MODEL_335M: return "335M"; case MODEL_410M: return "410M"; + case MODEL_450M: return "450M"; case MODEL_770M: return "770M"; case MODEL_780M: return "780M"; case MODEL_0_5B: return "0.5B"; @@ -4425,20 +4557,18 @@ static void llm_load_hparams( ml.get_key(LLM_KV_GENERAL_NAME, model.name, false); // get hparams kv - ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab); + ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab); // everything past this point is not vocab-related if (hparams.vocab_only) { return; } - ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); - ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); - ml.get_key(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff); - ml.get_key(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head); - ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer); - ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false); - ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); + ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); + ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); + ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer); + ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false); + ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS); GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert); @@ -4448,9 +4578,18 @@ static void llm_load_hparams( GGML_ASSERT(hparams.n_expert_used == 0); } + // zero-out the per-layer hparams + std::fill(hparams.n_head_arr.begin(), hparams.n_head_arr.end(), 0); + std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0); + std::fill(hparams.n_ff_arr.begin(), hparams.n_ff_arr.end(), 0); + + ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, hparams.n_layer); + ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer); + // n_head_kv is optional, default to n_head - hparams.n_head_kv = hparams.n_head; - ml.get_key(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv, false); + hparams.n_head_kv_arr = hparams.n_head_arr; + + ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv_arr, hparams.n_layer, false); bool rope_finetuned = false; ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false); @@ -4478,27 +4617,32 @@ static void llm_load_hparams( ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false); - // sanity check for n_rot (optional) - { - hparams.n_rot = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; + // non-transformer models do not have attention heads + if (hparams.n_head() > 0) { + // sanity check for n_rot (optional) + hparams.n_rot = hparams.n_embd / hparams.n_head(); ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false); if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) { - if (hparams.n_rot != hparams.n_embd / hparams.n_head) { - throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head)); + if (hparams.n_rot != hparams.n_embd / hparams.n_head()) { + throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head())); } } // gpt-neox n_rot = rotary_pct * (n_embd / n_head) // gpt-j n_rot = rotary_dim + + hparams.n_embd_head_k = hparams.n_embd / hparams.n_head(); + ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false); + + hparams.n_embd_head_v = hparams.n_embd / hparams.n_head(); + ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false); + } else { + hparams.n_rot = 0; + hparams.n_embd_head_k = 0; + hparams.n_embd_head_v = 0; } - hparams.n_embd_head_k = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; - ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false); - - hparams.n_embd_head_v = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; - ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false); - // arch-specific KVs switch (model.arch) { case LLM_ARCH_LLAMA: @@ -4521,7 +4665,7 @@ static void llm_load_hparams( case 40: model.type = e_model::MODEL_13B; break; case 48: model.type = e_model::MODEL_34B; break; case 60: model.type = e_model::MODEL_30B; break; - case 80: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_65B : e_model::MODEL_70B; break; + case 80: model.type = hparams.n_head() == hparams.n_head_kv() ? e_model::MODEL_65B : e_model::MODEL_70B; break; default: model.type = e_model::MODEL_UNKNOWN; } } @@ -4690,7 +4834,7 @@ static void llm_load_hparams( switch (hparams.n_layer) { case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break; case 32: model.type = e_model::MODEL_7B; break; - case 40: model.type = hparams.n_head == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break; + case 40: model.type = hparams.n_head() == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break; case 80: model.type = e_model::MODEL_70B; break; default: model.type = e_model::MODEL_UNKNOWN; } @@ -4882,46 +5026,58 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_OPENELM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 16: model.type = e_model::MODEL_270M; break; + case 20: model.type = e_model::MODEL_450M; break; + case 28: model.type = e_model::MODEL_1B; break; + case 36: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_GPTNEOX: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); ml.get_key(LLM_KV_USE_PARALLEL_RESIDUAL, hparams.use_par_res); switch (hparams.n_layer) { case 6: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 512: model.type = e_model::MODEL_14M; break; case 2048: model.type = e_model::MODEL_70M; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 12: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 3072: model.type = e_model::MODEL_160M; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 16: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 8192: model.type = e_model::MODEL_1B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 24: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 4096: model.type = e_model::MODEL_410M; break; case 8192: model.type = e_model::MODEL_1_4B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 32: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 10240: model.type = e_model::MODEL_2_8B; break; case 16384: model.type = e_model::MODEL_6_9B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 36: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 20480: model.type = e_model::MODEL_12B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 44: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 24576: model.type = e_model::MODEL_20B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; @@ -4984,13 +5140,13 @@ static void llm_load_hparams( case 6: model.type = e_model::MODEL_60M; break; // t5-small case 8: model.type = e_model::MODEL_80M; break; // flan-t5-small case 12: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 3072: model.type = e_model::MODEL_220M; break; // t5-base case 2048: model.type = e_model::MODEL_250M; break; // flan-t5-base default: model.type = e_model::MODEL_UNKNOWN; } break; case 24: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 4096: model.type = e_model::MODEL_770M; break; // t5-large case 2816: model.type = e_model::MODEL_780M; break; // flan-t5-large case 16384: model.type = e_model::MODEL_3B; break; // t5-3b @@ -5533,44 +5689,78 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train); + auto print_f = [](const std::function & f, uint32_t n) { + bool is_var = false; + + std::vector v; + for (uint32_t i = 0; i < n; ++i) { + v.push_back(f(i)); + if (v[i] != v[0]) { + is_var = true; + } + } + + std::stringstream ss; + + if (is_var) { + ss << "["; + for (uint32_t i = 0; i < n; ++i) { + ss << v[i]; + if (i < n - 1) { + ss << ", "; + } + } + ss << "]"; + } else { + ss << v[0]; + } + + return ss.str(); + }; + // hparams LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(ml.fver)); LLAMA_LOG_INFO("%s: arch = %s\n", __func__, LLM_ARCH_NAMES.at(model.arch)); LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, llama_model_vocab_type_name(vocab.type)); LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, hparams.n_vocab); LLAMA_LOG_INFO("%s: n_merges = %u\n", __func__, (int) vocab.bpe_ranks.size()); - LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train); - LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd); - LLAMA_LOG_INFO("%s: n_head = %u\n", __func__, hparams.n_head); - LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv); - LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); - LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); - LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa); - LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k); - LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v); - LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa()); - LLAMA_LOG_INFO("%s: n_embd_k_gqa = %u\n", __func__, hparams.n_embd_k_gqa()); - LLAMA_LOG_INFO("%s: n_embd_v_gqa = %u\n", __func__, hparams.n_embd_v_gqa()); - LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_eps); - LLAMA_LOG_INFO("%s: f_norm_rms_eps = %.1e\n", __func__, hparams.f_norm_rms_eps); - LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv); - LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias); - LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale); - LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff); - LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert); - LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used); - LLAMA_LOG_INFO("%s: causal attn = %d\n", __func__, hparams.causal_attn); - LLAMA_LOG_INFO("%s: pooling type = %d\n", __func__, hparams.pooling_type); - LLAMA_LOG_INFO("%s: rope type = %d\n", __func__, hparams.rope_type); - LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type); - LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train); - LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); - LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn); - LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); - LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); - LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner); - LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state); - LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank); + LLAMA_LOG_INFO("%s: vocab_only = %d\n", __func__, hparams.vocab_only); + + if (!hparams.vocab_only) { + LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train); + LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd); + LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); + LLAMA_LOG_INFO("%s: n_head = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); + LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa); + LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k); + LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v); + LLAMA_LOG_INFO("%s: n_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_embd_k_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_embd_k_gqa(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_embd_v_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_embd_v_gqa(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_eps); + LLAMA_LOG_INFO("%s: f_norm_rms_eps = %.1e\n", __func__, hparams.f_norm_rms_eps); + LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv); + LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias); + LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale); + LLAMA_LOG_INFO("%s: n_ff = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert); + LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used); + LLAMA_LOG_INFO("%s: causal attn = %d\n", __func__, hparams.causal_attn); + LLAMA_LOG_INFO("%s: pooling type = %d\n", __func__, hparams.pooling_type); + LLAMA_LOG_INFO("%s: rope type = %d\n", __func__, hparams.rope_type); + LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type); + LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train); + LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); + LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn); + LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); + LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); + LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner); + LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state); + LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank); + } + LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str()); if (ml.n_elements >= 1e12) { @@ -5765,13 +5955,13 @@ static bool llm_load_tensors( // create tensors for the weights { const int64_t n_embd = hparams.n_embd; - const int64_t n_embd_head = (hparams.n_head == 0) ? 0 : n_embd / hparams.n_head; + const int64_t n_embd_head = hparams.n_head() > 0 ? n_embd / hparams.n_head() : 0; const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_vocab = hparams.n_vocab; const int64_t n_vocab_type = hparams.n_vocab_type; - const int64_t n_ff = hparams.n_ff; + const int64_t n_ff = hparams.n_ff(); const int64_t n_expert = hparams.n_expert; if (n_expert > 0 && hparams.n_expert_used == 0) { @@ -6292,8 +6482,8 @@ static bool llm_load_tensors( layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); // optional q and k layernorms, present in StableLM 2 12B - layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head()}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv()}, llama_model_loader::TENSOR_NOT_REQUIRED); // optional FFN norm, not present in StableLM 2 12B which uses parallel residual layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); @@ -6664,7 +6854,7 @@ static bool llm_load_tensors( model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading - const int64_t n_ff = hparams.n_ff; + const int64_t n_ff = hparams.n_ff(); const int64_t n_embd_head_k = hparams.n_embd_head_k; const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); @@ -6677,10 +6867,10 @@ static bool llm_load_tensors( layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head}); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head()}); layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head(), n_embd}); layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); @@ -6696,7 +6886,7 @@ static bool llm_load_tensors( model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading - const int64_t n_ff = hparams.n_ff; + const int64_t n_ff = hparams.n_ff(); const int64_t n_embd_head_k = hparams.n_embd_head_k; const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); @@ -6709,10 +6899,10 @@ static bool llm_load_tensors( layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head}); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head()}); layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head(), n_embd}); layer.attn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}); layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); @@ -6861,8 +7051,8 @@ static bool llm_load_tensors( layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); if (n_layer >= 64){ - layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}); - layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}); + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head()}); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv()}); } layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); @@ -6904,6 +7094,42 @@ static bool llm_load_tensors( layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); } } break; + case LLM_ARCH_OPENELM: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + // init output from the input tok embed + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + const int64_t n_head = hparams.n_head(i); + const int64_t n_head_qkv = 2*hparams.n_head_kv(i) + n_head; + const int64_t n_embd_head = hparams.n_embd_head_k; + + const int64_t n_ff = hparams.n_ff(i); + + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_head_qkv*n_embd_head}); + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head}); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head*n_embd_head, n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; case LLM_ARCH_GPTNEOX: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -7011,13 +7237,13 @@ static bool llm_load_tensors( if (!is_lite) { layer.wq_a = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}); - layer.wq_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.n_head * hparams.n_embd_head_k}); + layer.wq_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.n_head() * hparams.n_embd_head_k}); } else { layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); } layer.wkv_a_mqa = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}); - layer.wkv_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, hparams.n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {hparams.n_head * hparams.n_embd_head_v, n_embd}); + layer.wkv_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, hparams.n_head() * (n_embd_head_qk_nope + hparams.n_embd_head_v)}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { hparams.n_head() * ( hparams.n_embd_head_v), n_embd}); layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); @@ -7104,7 +7330,7 @@ static bool llm_load_tensors( auto & layer = model.layers[i]; layer.attn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_ATTN_NORM, "weight", i), {n_embd}); - layer.attn_rel_b_enc = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ENC_ATTN_REL_B, "weight", i), {hparams.n_head, hparams.n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_rel_b_enc = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ENC_ATTN_REL_B, "weight", i), {hparams.n_head(), hparams.n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); layer.wq_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); layer.wk_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); @@ -7117,7 +7343,7 @@ static bool llm_load_tensors( layer.ffn_up_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_UP, "weight", i), {n_embd, n_ff}); layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_ATTN_NORM, "weight", i), {n_embd}); - layer.attn_rel_b = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_ATTN_REL_B, "weight", i), {hparams.n_head, hparams.n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_rel_b = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_ATTN_REL_B, "weight", i), {hparams.n_head(), hparams.n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); @@ -7126,7 +7352,7 @@ static bool llm_load_tensors( layer.attn_norm_cross = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_CROSS_ATTN_NORM, "weight", i), {n_embd}); // this tensor seems to be unused in HF transformers implementation - layer.attn_rel_b_cross = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "weight", i), {hparams.n_head, hparams.n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_rel_b_cross = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "weight", i), {hparams.n_head(), hparams.n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); layer.wq_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); layer.wk_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); @@ -7455,8 +7681,8 @@ static void llm_build_kv_store( int64_t il) { const int64_t n_ctx = cparams.n_ctx; - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); GGML_ASSERT(kv.size == n_ctx); @@ -7749,12 +7975,12 @@ static struct ggml_tensor * llm_build_kqv( const llm_build_cb & cb, int il) { const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_head = hparams.n_head(il); + const int64_t n_head_kv = hparams.n_head_kv(il); const int64_t n_embd_head_k = hparams.n_embd_head_k; - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); const int64_t n_embd_head_v = hparams.n_embd_head_v; - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3); cb(q, "q", il); @@ -7961,8 +8187,8 @@ struct llm_build_context { n_layer (hparams.n_layer), n_rot (hparams.n_rot), n_ctx (cparams.n_ctx), - n_head (hparams.n_head), - n_head_kv (hparams.n_head_kv), + n_head (hparams.n_head()), + n_head_kv (hparams.n_head_kv()), n_embd_head_k (hparams.n_embd_head_k), n_embd_k_gqa (hparams.n_embd_k_gqa()), n_embd_head_v (hparams.n_embd_head_v), @@ -8034,6 +8260,8 @@ struct llm_build_context { ggml_set_input(lctx.inp_K_shift); for (int il = 0; il < n_layer; ++il) { + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); struct ggml_tensor * rope_factors = build_rope_factors(il); struct ggml_tensor * tmp = // we rotate only the first n_rot dimensions @@ -8093,6 +8321,9 @@ struct llm_build_context { } for (int il = 0; il < n_layer; ++il) { + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); + ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self.k_l[il], n_embd_k_gqa, nm, ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), @@ -11974,6 +12205,131 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_openelm() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + const int64_t n_head = hparams.n_head(il); + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_head_qkv = 2*n_head_kv + n_head; + + cur = inpL; + struct ggml_tensor * residual = cur; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0)); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head)); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv))); + cb(Vcur, "Vcur", il); + + Qcur = llm_build_norm(ctx0, Qcur, hparams, + model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(Qcur, "Qcur", il); + + Kcur = llm_build_norm(ctx0, Kcur, hparams, + model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(Kcur, "Kcur", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + Vcur = ggml_reshape_2d(ctx0, Vcur, n_embd_head * n_head_kv, n_tokens); + cb(Qcur, "Vcur", il); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, NULL, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + inpL = cur; + } + + cur = inpL; + + // norm + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + struct ggml_cgraph * build_gptneox() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -13242,6 +13598,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_olmo(); } break; + case LLM_ARCH_OPENELM: + { + result = llm.build_openelm(); + } break; case LLM_ARCH_GPTNEOX: { result = llm.build_gptneox(); @@ -18772,6 +19132,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_GEMMA: case LLM_ARCH_GEMMA2: case LLM_ARCH_STARCODER2: + case LLM_ARCH_OPENELM: case LLM_ARCH_GPTNEOX: return LLAMA_ROPE_TYPE_NEOX; @@ -19365,8 +19726,6 @@ static void llama_state_get_data_internal(struct llama_context * ctx, llama_data const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); // NOTE: kv_size and kv_buf_size are mostly used for sanity checks const uint32_t kv_head = llama_kv_cache_cell_max(kv_self); @@ -19386,6 +19745,9 @@ static void llama_state_get_data_internal(struct llama_context * ctx, llama_data std::vector tmp_buf; for (int il = 0; il < (int) n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head); tmp_buf.resize(k_size); @@ -19518,8 +19880,6 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) { const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); size_t kv_buf_size; uint32_t kv_head; @@ -19551,6 +19911,9 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) { GGML_ASSERT(kv_self.total_size() >= kv_buf_size); for (int il = 0; il < (int) n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head); ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size); @@ -19713,8 +20076,6 @@ size_t llama_state_seq_get_size(struct llama_context* ctx, llama_seq_id seq_id) const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); for (uint32_t i = 0; i < kv_self.size; ++i) { const auto & cell = kv_self.cells[i]; @@ -19725,6 +20086,9 @@ size_t llama_state_seq_get_size(struct llama_context* ctx, llama_seq_id seq_id) } for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // types of keys and values s_cell_data_size += sizeof(int32_t) * 2; // k_size_row and v_size_el values of layer @@ -19799,14 +20163,15 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); // Write the layer count data_ctx.write(&n_layer, sizeof(n_layer)); - // Write n_embd_v_gqa - data_ctx.write(&n_embd_v_gqa, sizeof(n_embd_v_gqa)); + // Write n_embd_v_gqa (reference value) + { + const uint32_t n_embd_v_gqa_ref = hparams.n_embd_v_gqa() + hparams.n_embd_k_s(); + data_ctx.write(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref)); + } // Iterate the ranges and write all the pos (this is the token position in the prompt) for (const auto & range : cell_ranges) { @@ -19820,6 +20185,8 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam // Get whole range at a time std::vector tmp_buf; for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + // Write key type const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type; data_ctx.write(&k_type_i, sizeof(k_type_i)); @@ -19840,6 +20207,8 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam // TODO: simplify, reduce copy-paste if (!kv_self.v_trans) { for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // Write value type const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; data_ctx.write(&v_type_i, sizeof(v_type_i)); @@ -19860,6 +20229,8 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam // For the values, they are transposed, so we also need the element size and get the element ranges from each row const uint32_t kv_size = kv_self.size; for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // Write value type const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; data_ctx.write(&v_type_i, sizeof(v_type_i)); @@ -19928,14 +20299,14 @@ size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, // Sanity check model compatibility const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); + if (n_layer != n_layer_ref) { LLAMA_LOG_ERROR("%s: mismatched n_layer (%d != %d)\n", __func__, n_layer, n_layer_ref); return 0; } - if (n_embd_v_gqa != n_embd_v_gqa_ref) { - LLAMA_LOG_ERROR("%s: mismatched n_embd_v_gqa (%d != %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref); + + if (hparams.n_embd_v_gqa() != n_embd_v_gqa_ref) { + LLAMA_LOG_ERROR("%s: mismatched n_embd_v_gqa (%d != %d)\n", __func__, hparams.n_embd_v_gqa(), n_embd_v_gqa_ref); return 0; } @@ -19975,6 +20346,8 @@ size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, // For each layer, read the keys for each cell, one row is one cell, read as one contiguous blo for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + // Read type of key int32_t k_type_i_ref; memcpy(&k_type_i_ref, inp, sizeof(k_type_i_ref)); @@ -20007,6 +20380,8 @@ size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, // TODO: simplify, reduce copy-paste if (!kv_self.v_trans) { for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // Read type of value int32_t v_type_i_ref; memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref)); @@ -20038,6 +20413,8 @@ size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, } else { // For each layer, read the values for each cell (transposed) for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // Read type of value int32_t v_type_i_ref; memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref));