Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-12 21:37:19 +01:00
Code Issues Packages Projects Releases Wiki Activity

parallelize fattn compilation test

Browse Source
This commit is contained in:
slaren 2024-05-28 01:19:36 +02:00
parent f4003cfba1
commit 1ca802a3e0
6 changed files with 390 additions and 288 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Makefile
View File

@ -508,7 +508,7 @@ define NVCC_COMPILE
endef # NVCC_COMPILE
endif # JETSON_EOL_MODULE_DETECT
ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
ggml-cuda/%.o: ggml-cuda/%.cu ggml.h ggml-common.h ggml-cuda/common.cuh
$(NVCC_COMPILE)
ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)

378
ggml-cuda/fattn-common.cuh
View File

@ -49,7 +49,7 @@ typedef float (*vec_dot_KQ_f32_t)(
const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds);
template<typename T, int D>
static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0(
__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
@ -263,7 +263,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1(
}
template <typename T, int D>
static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0(
__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
@ -304,7 +304,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0(
}
template <typename T, int D>
static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_f16(
__device__ __forceinline__ T vec_dot_fattn_vec_KQ_f16(
const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds_v) {
const half2 * K_h2 = (const half2 *) K_c;
@ -393,7 +393,7 @@ typedef half (*dequantize_1_f16_t)(const void *, const int64_t);
typedef float (*dequantize_1_f32_t)(const void *, const int64_t);
template <typename T>
static __device__ __forceinline__ T dequantize_1_q4_0(const void * __restrict__ vx, const int64_t i) {
__device__ __forceinline__ T dequantize_1_q4_0(const void * __restrict__ vx, const int64_t i) {
const block_q4_0 * x = (const block_q4_0 *) vx;
const int64_t ib = i / QK4_0;
@ -485,7 +485,7 @@ static __device__ __forceinline__ T dequantize_1_q5_1(const void * __restrict__
}
template <typename T>
static __device__ __forceinline__ T dequantize_1_q8_0(const void * __restrict__ vx, const int64_t i) {
__device__ __forceinline__ T dequantize_1_q8_0(const void * __restrict__ vx, const int64_t i) {
const block_q8_0 * x = (const block_q8_0 *) vx;
const int64_t ib = i / QK8_0;
@ -504,7 +504,7 @@ static __device__ __forceinline__ T dequantize_1_q8_0(const void * __restrict__
}
template <typename T>
static __device__ __forceinline__ T dequantize_1_f16(const void * __restrict__ vx, const int64_t i) {
__device__ __forceinline__ T dequantize_1_f16(const void * __restrict__ vx, const int64_t i) {
const half * x = (const half *) vx;
return x[i];
@ -669,3 +669,369 @@ void launch_fattn(ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kern
(dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
CUDA_CHECK(cudaGetLastError());
}
template<int D, int ncols, int parallel_blocks, vec_dot_KQ_f16_t vec_dot_KQ, bool Q_q8_1, dequantize_1_f16_t dequantize_1_v> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
__global__ void flash_attn_vec_ext_f16(
const char * __restrict__ Q,
const char * __restrict__ K,
const char * __restrict__ V,
const char * __restrict__ mask,
float * __restrict__ dst,
float2 * __restrict__ dst_meta,
const float scale,
const float max_bias,
const float m0,
const float m1,
const uint32_t n_head_log2,
const int ne00,
const int ne01,
const int ne02,
const int ne03,
const int ne10,
const int ne11,
const int ne12,
const int ne13,
const int ne31,
const int nb31,
const int nb01,
const int nb02,
const int nb03,
const int nb11,
const int nb12,
const int nb13,
const int nb21,
const int nb22,
const int nb23,
const int ne0,
const int ne1,
const int ne2,
const int ne3) {
#if FP16_AVAILABLE
//In this kernel Q, K, V are matrices while i, j, k are matrix indices.
const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
Q += nb02* blockIdx.y + nb01*ic0;
K += nb12*(blockIdx.y / gqa_ratio);
V += nb22*(blockIdx.y / gqa_ratio);
const half * maskh = (const half *) mask + ne11*ic0;
const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
const half slopeh = __float2half(slopef);
static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
constexpr int nwarps = D / WARP_SIZE;
const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
__builtin_assume(tid < D);
__shared__ half KQ[ncols*D];
half2 * KQ2 = (half2 *) KQ;
half kqmax[ncols];
#pragma unroll
for (int j = 0; j < ncols; ++j) {
kqmax[j] = -HALF_MAX_HALF;
}
half kqsum[ncols] = {0.0f};
__shared__ half kqmax_shared[ncols][WARP_SIZE];
__shared__ half kqsum_shared[ncols][WARP_SIZE];
#pragma unroll
for (int j = 0; j < ncols; ++j) {
if (threadIdx.y == 0) {
kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
kqsum_shared[j][threadIdx.x] = 0.0f;
}
}
__syncthreads();
// Convert Q to half2 (f16 K) or q8_1 (quantized K) and store in registers:
half2 Q_h2[ncols][D/(2*WARP_SIZE)];
int Q_i32[ncols][D/(sizeof(int)*QK8_1) == 0 ? 1 : D/(sizeof(int)*QK8_1)];
half2 Q_ds[ncols][D/QK8_1 == 0 ? 1 : D/QK8_1];
if (Q_q8_1) {
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
const int j = j0 + threadIdx.y;
if (j0 + nwarps > ncols && j >= ncols) {
break;
}
// Reuse KQ as temporary storage for converting Q to q8_1:
int * tmp_q_i32 = (int *) &KQ[j*D];
half2 * tmp_q_ds = (half2 *) (tmp_q_i32 + D/sizeof(int));
// Set memory to zero if out of bounds:
if (ncols > 2 && ic0 + j >= ne01) {
#pragma unroll
for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
tmp_q_i32[i] = 0;
}
if (threadIdx.x < D/QK8_1) {
tmp_q_ds[threadIdx.x] = make_half2(0.0f, 0.0f);
}
continue;
}
const float * Q_f = (const float *) (Q + j*nb01);
#pragma unroll
for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
quantize_q8_1_to_shared<half2>(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds);
}
}
__syncthreads();
#pragma unroll
for (int j = 0; j < ncols; ++j) {
int * tmp_q_i32 = (int *) &KQ[j*D];
half2 * tmp_q_ds = (half2 *) (tmp_q_i32 + D/sizeof(int));
#pragma unroll
for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i];
Q_ds[j][i0/WARP_SIZE] = tmp_q_ds[i/QI8_1];
}
}
__syncthreads();
} else {
#pragma unroll
for (int j = 0; j < ncols; ++j) {
const float2 * Q_f2_j = (const float2 *) (Q + j*nb01);
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const float2 tmp = ncols <= 2 || ic0 + j < ne01 ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
}
}
}
#pragma unroll
for (int j = 0; j < ncols; ++j) {
KQ[j*D + tid] = -HALF_MAX_HALF;
}
half2 VKQ[ncols] = {{0.0f, 0.0f}};
const int k_start = parallel_blocks == 1 ? 0 : ip*D;
for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
// Calculate KQ tile and keep track of new maximum KQ values:
// For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
// see https://github.com/ggerganov/llama.cpp/pull/7061 .
// Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
half kqmax_new = kqmax[0];
half kqmax_new_arr[ncols];
#pragma unroll
for (int j = 0; j < ncols; ++j) {
kqmax_new_arr[j] = kqmax[j];
}
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
const int i_KQ = i_KQ_0 + threadIdx.y;
if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
break;
}
#pragma unroll
for (int j = 0; j < ncols; ++j) {
half sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_h2[j], Q_i32[j], Q_ds[j]);
sum = warp_reduce_sum(sum);
sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
if (ncols == 1) {
kqmax_new = ggml_cuda_hmax(kqmax_new, sum);
} else {
kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
}
if (threadIdx.x == 0) {
KQ[j*D + i_KQ] = sum;
}
}
}
#pragma unroll
for (int j = 0; j < ncols; ++j) {
half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
kqmax_new_j = warp_reduce_max(kqmax_new_j);
if (threadIdx.x == 0) {
kqmax_shared[j][threadIdx.y] = kqmax_new_j;
}
}
__syncthreads();
#pragma unroll
for (int j = 0; j < ncols; ++j) {
half kqmax_new_j = kqmax_shared[j][threadIdx.x];
kqmax_new_j = warp_reduce_max(kqmax_new_j);
const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
kqmax[j] = kqmax_new_j;
const half val = hexp(KQ[j*D + tid] - kqmax[j]);
kqsum[j] = kqsum[j]*KQ_max_scale + val;
KQ[j*D + tid] = val;
VKQ[j] *= __half2half2(KQ_max_scale);
}
__syncthreads();
#pragma unroll
for (int k0 = 0; k0 < D; k0 += 2) {
if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
break;
}
half2 V_k;
reinterpret_cast<half&>(V_k.x) = dequantize_1_v(V + (k_VKQ_0 + k0 + 0)*nb21, tid);
reinterpret_cast<half&>(V_k.y) = dequantize_1_v(V + (k_VKQ_0 + k0 + 1)*nb21, tid);
#pragma unroll
for (int j = 0; j < ncols; ++j) {
VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
}
}
__syncthreads();
}
#pragma unroll
for (int j = 0; j < ncols; ++j) {
kqsum[j] = warp_reduce_sum(kqsum[j]);
if (threadIdx.x == 0) {
kqsum_shared[j][threadIdx.y] = kqsum[j];
}
}
__syncthreads();
#pragma unroll
for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
if (ncols > 2 && ic0 + j_VKQ >= ne01) {
break;
}
kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
if (parallel_blocks == 1) {
dst_val /= kqsum[j_VKQ];
}
const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
}
if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
}
#else
NO_DEVICE_CODE;
#endif // FP16_AVAILABLE
}
#define DECL_FATTN_VEC_F16_INST(D, ncols, parallel_blocks, vec_dot_KQ, Q_q8_1, dequantize_1_v) \
template __global__ void flash_attn_vec_ext_f16<D, ncols, parallel_blocks, vec_dot_KQ, Q_q8_1, dequantize_1_v>( \
const char * __restrict__ Q, \
const char * __restrict__ K, \
const char * __restrict__ V, \
const char * __restrict__ mask, \
float * __restrict__ dst, \
float2 * __restrict__ dst_meta, \
const float scale, \
const float max_bias, \
const float m0, \
const float m1, \
const uint32_t n_head_log2, \
const int ne00, \
const int ne01, \
const int ne02, \
const int ne03, \
const int ne10, \
const int ne11, \
const int ne12, \
const int ne13, \
const int ne31, \
const int nb31, \
const int nb01, \
const int nb02, \
const int nb03, \
const int nb11, \
const int nb12, \
const int nb13, \
const int nb21, \
const int nb22, \
const int nb23, \
const int ne0, \
const int ne1, \
const int ne2, \
const int ne3)
extern DECL_FATTN_VEC_F16_INST(64, 1, 4, (vec_dot_fattn_vec_KQ_f16<half, 64>), false, dequantize_1_f16<half>);
extern DECL_FATTN_VEC_F16_INST(128, 1, 4, (vec_dot_fattn_vec_KQ_f16<half, 128>), false, dequantize_1_f16<half>);
extern DECL_FATTN_VEC_F16_INST(256, 1, 4, (vec_dot_fattn_vec_KQ_f16<half, 256>), false, dequantize_1_f16<half>);
#define DECL_FATTN_VEC_INST(type_VKQ, D, cols_per_block, parallel_blocks, type_suffix_K, type_suffix_V) \
template __global__ void flash_attn_vec_ext_##type_VKQ< \
(D), cols_per_block, parallel_blocks, \
vec_dot_fattn_vec_KQ_##type_suffix_K<type_VKQ, (D)>, ggml_type_##type_suffix_K != GGML_TYPE_F16, dequantize_1_##type_suffix_V<type_VKQ>>( \
const char * __restrict__ Q, \
const char * __restrict__ K, \
const char * __restrict__ V, \
const char * __restrict__ mask, \
float * __restrict__ dst, \
float2 * __restrict__ dst_meta, \
const float scale, \
const float max_bias, \
const float m0, \
const float m1, \
const uint32_t n_head_log2, \
const int ne00, \
const int ne01, \
const int ne02, \
const int ne03, \
const int ne10, \
const int ne11, \
const int ne12, \
const int ne13, \
const int ne31, \
const int nb31, \
const int nb01, \
const int nb02, \
const int nb03, \
const int nb11, \
const int nb12, \
const int nb13, \
const int nb21, \
const int nb22, \
const int nb23, \
const int ne0, \
const int ne1, \
const int ne2, \
const int ne3)
extern DECL_FATTN_VEC_INST(f16, 128, 1, 4, q4_0, q4_0);
extern DECL_FATTN_VEC_INST(f16, 128, 1, 4, q8_0, q8_0);

7
ggml-cuda/fattn-vec-f16-f16.cu Normal file
View File

@ -0,0 +1,7 @@
#include "common.cuh"
#include "fattn-common.cuh"
#include "fattn-vec-f16.cuh"
DECL_FATTN_VEC_F16_INST(64, 1, 4, (vec_dot_fattn_vec_KQ_f16<half, 64>), false, dequantize_1_f16<half>);
DECL_FATTN_VEC_F16_INST(128, 1, 4, (vec_dot_fattn_vec_KQ_f16<half, 128>), false, dequantize_1_f16<half>);
DECL_FATTN_VEC_F16_INST(256, 1, 4, (vec_dot_fattn_vec_KQ_f16<half, 256>), false, dequantize_1_f16<half>);

5
ggml-cuda/fattn-vec-f16-q4_0-q4_0.cu Normal file
View File

@ -0,0 +1,5 @@
#include "common.cuh"
#include "fattn-common.cuh"
#include "fattn-vec-f16.cuh"
DECL_FATTN_VEC_INST(f16, 128, 1, 4, q4_0, q4_0);

5
ggml-cuda/fattn-vec-f16-q8_0-q8_0.cu Normal file
View File

@ -0,0 +1,5 @@
#include "common.cuh"
#include "fattn-common.cuh"
#include "fattn-vec-f16.cuh"
DECL_FATTN_VEC_INST(f16, 128, 1, 4, q8_0, q8_0);

281
ggml-cuda/fattn-vec-f16.cu
View File

@ -2,287 +2,6 @@
#include "fattn-common.cuh"
#include "fattn-vec-f16.cuh"
template<int D, int ncols, int parallel_blocks, vec_dot_KQ_f16_t vec_dot_KQ, bool Q_q8_1, dequantize_1_f16_t dequantize_1_v> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_vec_ext_f16(
const char * __restrict__ Q,
const char * __restrict__ K,
const char * __restrict__ V,
const char * __restrict__ mask,
float * __restrict__ dst,
float2 * __restrict__ dst_meta,
const float scale,
const float max_bias,
const float m0,
const float m1,
const uint32_t n_head_log2,
const int ne00,
const int ne01,
const int ne02,
const int ne03,
const int ne10,
const int ne11,
const int ne12,
const int ne13,
const int ne31,
const int nb31,
const int nb01,
const int nb02,
const int nb03,
const int nb11,
const int nb12,
const int nb13,
const int nb21,
const int nb22,
const int nb23,
const int ne0,
const int ne1,
const int ne2,
const int ne3) {
#if FP16_AVAILABLE
//In this kernel Q, K, V are matrices while i, j, k are matrix indices.
const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
Q += nb02* blockIdx.y + nb01*ic0;
K += nb12*(blockIdx.y / gqa_ratio);
V += nb22*(blockIdx.y / gqa_ratio);
const half * maskh = (const half *) mask + ne11*ic0;
const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
const half slopeh = __float2half(slopef);
static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
constexpr int nwarps = D / WARP_SIZE;
const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
__builtin_assume(tid < D);
__shared__ half KQ[ncols*D];
half2 * KQ2 = (half2 *) KQ;
half kqmax[ncols];
#pragma unroll
for (int j = 0; j < ncols; ++j) {
kqmax[j] = -HALF_MAX_HALF;
}
half kqsum[ncols] = {0.0f};
__shared__ half kqmax_shared[ncols][WARP_SIZE];
__shared__ half kqsum_shared[ncols][WARP_SIZE];
#pragma unroll
for (int j = 0; j < ncols; ++j) {
if (threadIdx.y == 0) {
kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
kqsum_shared[j][threadIdx.x] = 0.0f;
}
}
__syncthreads();
// Convert Q to half2 (f16 K) or q8_1 (quantized K) and store in registers:
half2 Q_h2[ncols][D/(2*WARP_SIZE)];
int Q_i32[ncols][D/(sizeof(int)*QK8_1) == 0 ? 1 : D/(sizeof(int)*QK8_1)];
half2 Q_ds[ncols][D/QK8_1 == 0 ? 1 : D/QK8_1];
if (Q_q8_1) {
#pragma unroll
for (int j0 = 0; j0 < ncols; j0 += nwarps) {
const int j = j0 + threadIdx.y;
if (j0 + nwarps > ncols && j >= ncols) {
break;
}
// Reuse KQ as temporary storage for converting Q to q8_1:
int * tmp_q_i32 = (int *) &KQ[j*D];
half2 * tmp_q_ds = (half2 *) (tmp_q_i32 + D/sizeof(int));
// Set memory to zero if out of bounds:
if (ncols > 2 && ic0 + j >= ne01) {
#pragma unroll
for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
tmp_q_i32[i] = 0;
}
if (threadIdx.x < D/QK8_1) {
tmp_q_ds[threadIdx.x] = make_half2(0.0f, 0.0f);
}
continue;
}
const float * Q_f = (const float *) (Q + j*nb01);
#pragma unroll
for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
quantize_q8_1_to_shared<half2>(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds);
}
}
__syncthreads();
#pragma unroll
for (int j = 0; j < ncols; ++j) {
int * tmp_q_i32 = (int *) &KQ[j*D];
half2 * tmp_q_ds = (half2 *) (tmp_q_i32 + D/sizeof(int));
#pragma unroll
for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i];
Q_ds[j][i0/WARP_SIZE] = tmp_q_ds[i/QI8_1];
}
}
__syncthreads();
} else {
#pragma unroll
for (int j = 0; j < ncols; ++j) {
const float2 * Q_f2_j = (const float2 *) (Q + j*nb01);
#pragma unroll
for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
const float2 tmp = ncols <= 2 || ic0 + j < ne01 ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
}
}
}
#pragma unroll
for (int j = 0; j < ncols; ++j) {
KQ[j*D + tid] = -HALF_MAX_HALF;
}
half2 VKQ[ncols] = {{0.0f, 0.0f}};
const int k_start = parallel_blocks == 1 ? 0 : ip*D;
for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
// Calculate KQ tile and keep track of new maximum KQ values:
// For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
// see https://github.com/ggerganov/llama.cpp/pull/7061 .
// Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
half kqmax_new = kqmax[0];
half kqmax_new_arr[ncols];
#pragma unroll
for (int j = 0; j < ncols; ++j) {
kqmax_new_arr[j] = kqmax[j];
}
#pragma unroll
for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
const int i_KQ = i_KQ_0 + threadIdx.y;
if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
break;
}
#pragma unroll
for (int j = 0; j < ncols; ++j) {
half sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_h2[j], Q_i32[j], Q_ds[j]);
sum = warp_reduce_sum(sum);
sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
if (ncols == 1) {
kqmax_new = ggml_cuda_hmax(kqmax_new, sum);
} else {
kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
}
if (threadIdx.x == 0) {
KQ[j*D + i_KQ] = sum;
}
}
}
#pragma unroll
for (int j = 0; j < ncols; ++j) {
half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
kqmax_new_j = warp_reduce_max(kqmax_new_j);
if (threadIdx.x == 0) {
kqmax_shared[j][threadIdx.y] = kqmax_new_j;
}
}
__syncthreads();
#pragma unroll
for (int j = 0; j < ncols; ++j) {
half kqmax_new_j = kqmax_shared[j][threadIdx.x];
kqmax_new_j = warp_reduce_max(kqmax_new_j);
const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
kqmax[j] = kqmax_new_j;
const half val = hexp(KQ[j*D + tid] - kqmax[j]);
kqsum[j] = kqsum[j]*KQ_max_scale + val;
KQ[j*D + tid] = val;
VKQ[j] *= __half2half2(KQ_max_scale);
}
__syncthreads();
#pragma unroll
for (int k0 = 0; k0 < D; k0 += 2) {
if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
break;
}
half2 V_k;
reinterpret_cast<half&>(V_k.x) = dequantize_1_v(V + (k_VKQ_0 + k0 + 0)*nb21, tid);
reinterpret_cast<half&>(V_k.y) = dequantize_1_v(V + (k_VKQ_0 + k0 + 1)*nb21, tid);
#pragma unroll
for (int j = 0; j < ncols; ++j) {
VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
}
}
__syncthreads();
}
#pragma unroll
for (int j = 0; j < ncols; ++j) {
kqsum[j] = warp_reduce_sum(kqsum[j]);
if (threadIdx.x == 0) {
kqsum_shared[j][threadIdx.y] = kqsum[j];
}
}
__syncthreads();
#pragma unroll
for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
if (ncols > 2 && ic0 + j_VKQ >= ne01) {
break;
}
kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
if (parallel_blocks == 1) {
dst_val /= kqsum[j_VKQ];
}
const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
}
if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
}
#else
NO_DEVICE_CODE;
#endif // FP16_AVAILABLE
}
void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_tensor * KQV = dst;
ggml_tensor * Q = dst->src[0];