Skip to content

Commit 654a677

Browse files
ggerganovggerganov
authored
metal : add FA-vec kernel for head size 64 (#13583)
ggml-ci
1 parent 5364ae4 commit 654a677

File tree

2 files changed

+43
-1
lines changed

2 files changed

+43
-1
lines changed

‎ggml/src/ggml-metal/ggml-metal.m

Lines changed: 33 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -415,6 +415,13 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
415415
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
416416
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
417417
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512,
418+
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,
419+
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,
420+
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,
421+
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H64,
422+
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H64,
423+
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H64,
424+
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H64,
418425
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,
419426
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,
420427
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,
@@ -1362,6 +1369,13 @@ @implementation GGMLMetalClass
13621369
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm);
13631370
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, flash_attn_ext_q8_0_h256, has_simdgroup_mm);
13641371
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512, flash_attn_ext_q8_0_hk576_hv512, has_simdgroup_mm);
1372+
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64, flash_attn_ext_vec_f16_h64, has_simdgroup_reduction);
1373+
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64, flash_attn_ext_vec_bf16_h64, has_simdgroup_reduction && use_bfloat);
1374+
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64, flash_attn_ext_vec_q4_0_h64, has_simdgroup_reduction);
1375+
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H64, flash_attn_ext_vec_q4_1_h64, has_simdgroup_reduction);
1376+
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H64, flash_attn_ext_vec_q5_0_h64, has_simdgroup_reduction);
1377+
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H64, flash_attn_ext_vec_q5_1_h64, has_simdgroup_reduction);
1378+
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H64, flash_attn_ext_vec_q8_0_h64, has_simdgroup_reduction);
13651379
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96, flash_attn_ext_vec_f16_h96, has_simdgroup_reduction);
13661380
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96, flash_attn_ext_vec_bf16_h96, has_simdgroup_reduction && use_bfloat);
13671381
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96, flash_attn_ext_vec_q4_0_h96, has_simdgroup_reduction);
@@ -4358,7 +4372,7 @@ static bool ggml_metal_encode_node(
43584372
// TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
43594373
// for now avoiding mainly to keep the number of templates/kernels a bit lower
43604374
// these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
4361-
if (ne01 >= 20 || (ne00%128 != 0 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
4375+
if (ne01 >= 20 || (ne00%128 != 0 && ne00 != 64 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
43624376
switch (src1->type) {
43634377
case GGML_TYPE_F16:
43644378
{
@@ -4539,6 +4553,24 @@ static bool ggml_metal_encode_node(
45394553
use_vec_kernel = true;
45404554

45414555
switch (ne00) {
4556+
case 64:
4557+
{
4558+
switch (src1->type) {
4559+
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64].pipeline; break;
4560+
case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64].pipeline; break;
4561+
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64].pipeline; break;
4562+
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H64].pipeline; break;
4563+
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H64].pipeline; break;
4564+
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H64].pipeline; break;
4565+
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H64].pipeline; break;
4566+
default:
4567+
{
4568+
GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
4569+
GGML_LOG_ERROR("add template specialization for this type\n");
4570+
GGML_ABORT("add template specialization for this type");
4571+
}
4572+
}
4573+
} break;
45424574
case 96:
45434575
{
45444576
switch (src1->type) {

‎ggml/src/ggml-metal/ggml-metal.metal

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -4124,6 +4124,16 @@ kernel void kernel_flash_attn_ext_vec(
41244124

41254125
typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
41264126

4127+
template [[host_name("kernel_flash_attn_ext_vec_f16_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 64, 64, 8>;
4128+
#if defined(GGML_METAL_USE_BF16)
4129+
template [[host_name("kernel_flash_attn_ext_vec_bf16_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4, 1, dequantize_bf16_t4, bfloat4, 1, dequantize_bf16_t4, 64, 64, 8>;
4130+
#endif
4131+
template [[host_name("kernel_flash_attn_ext_vec_q4_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 8, dequantize_q4_0_t4, block_q4_0, 8, dequantize_q4_0_t4, 64, 64, 8>;
4132+
template [[host_name("kernel_flash_attn_ext_vec_q4_1_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 8, dequantize_q4_1_t4, block_q4_1, 8, dequantize_q4_1_t4, 64, 64, 8>;
4133+
template [[host_name("kernel_flash_attn_ext_vec_q5_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 8, dequantize_q5_0_t4, block_q5_0, 8, dequantize_q5_0_t4, 64, 64, 8>;
4134+
template [[host_name("kernel_flash_attn_ext_vec_q5_1_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 8, dequantize_q5_1_t4, block_q5_1, 8, dequantize_q5_1_t4, 64, 64, 8>;
4135+
template [[host_name("kernel_flash_attn_ext_vec_q8_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 8, dequantize_q8_0_t4, block_q8_0, 8, dequantize_q8_0_t4, 64, 64, 8>;
4136+
41274137
template [[host_name("kernel_flash_attn_ext_vec_f16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 96, 96, 4>;
41284138
#if defined(GGML_METAL_USE_BF16)
41294139
template [[host_name("kernel_flash_attn_ext_vec_bf16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4, 1, dequantize_bf16_t4, bfloat4, 1, dequantize_bf16_t4, 96, 96, 4>;

0 commit comments

Comments
 (0)