Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-28 15:18:26 +01:00
Code Issues Packages Projects Releases Wiki Activity

metal : specialize for head size

Browse Source
This commit is contained in:
Georgi Gerganov 2024-01-21 12:01:55 +02:00
parent 52ae085750
commit b97325800a
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
2 changed files with 179 additions and 122 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
ggml-metal.m
View File

@ -147,7 +147,9 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,
GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
@ -518,7 +520,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16, flash_attn_ext_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true);
@ -2172,6 +2176,7 @@ static bool ggml_metal_graph_compute(
} break;
case GGML_OP_FLASH_ATTN_EXT:
{
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(src0->type == GGML_TYPE_F16);
struct ggml_tensor * src2 = gf->nodes[i]->src[2];
@ -2202,7 +2207,19 @@ static bool ggml_metal_graph_compute(
float scale;
memcpy(&scale, dst->op_params, sizeof(float));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16].pipeline;
id<MTLComputePipelineState> pipeline = nil;
switch (ne00) {
case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break;
case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
default:
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
GGML_METAL_LOG_ERROR("add template specialization for this size\n");
GGML_ASSERT(false && "add template specialization for this size");
}
}
// TODO: extend if necessary
[encoder setComputePipelineState:pipeline];

42
ggml-metal.metal
View File

@ -1959,6 +1959,43 @@ kernel void kernel_leaky_relu_f32(
dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * slope;
}
typedef void (flash_attn_ext_f16_t)(
device const char * q,
device const char * k,
device const char * v,
device const char * mask,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne31,
constant uint64_t & nb31,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant float & scale,
threadgroup half * shared,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]);
template<int64_t D> // head size
kernel void kernel_flash_attn_ext_f16(
device const char * q,
device const char * k,
@ -2002,7 +2039,6 @@ kernel void kernel_flash_attn_ext_f16(
return;
}
const int64_t D = ne00;
const int64_t D4 = D/4;
// TODO: can we move this to the stack?
@ -2097,6 +2133,10 @@ kernel void kernel_flash_attn_ext_f16(
}
}
template [[host_name("kernel_flash_attn_ext_f16_h64" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<64>;
template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<80>;
template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<128>;
kernel void kernel_cpy_f16_f16(
device const half * src0,
device half * dst,