Skip to content

Commit 4efd5a8

Browse files
ggerganovggerganov
authored
metal : fix checks for available FA kernels (ggml-org#15700)
* metal : fix checks for available FA kernels ggml-ci * cont : fix comment [no ci]
1 parent 2749662 commit 4efd5a8

File tree

2 files changed

+14
-49
lines changed

2 files changed

+14
-49
lines changed

‎ggml/src/ggml-metal/ggml-metal.m

Lines changed: 11 additions & 39 deletions
Original file line numberDiff line numberDiff line change
@@ -523,13 +523,6 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
523523
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
524524
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
525525
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512,
526-
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40,
527-
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40,
528-
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40,
529-
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40,
530-
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40,
531-
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40,
532-
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40,
533526
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,
534527
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,
535528
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,
@@ -1562,13 +1555,6 @@ @implementation GGMLMetalClass
15621555
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);
15631556
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, flash_attn_ext_q8_0_h256, has_simdgroup_mm);
15641557
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);
1565-
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40, flash_attn_ext_vec_f16_h40, has_simdgroup_reduction);
1566-
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40, flash_attn_ext_vec_bf16_h40, has_simdgroup_reduction && use_bfloat);
1567-
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40, flash_attn_ext_vec_q4_0_h40, has_simdgroup_reduction);
1568-
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40, flash_attn_ext_vec_q4_1_h40, has_simdgroup_reduction);
1569-
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40, flash_attn_ext_vec_q5_0_h40, has_simdgroup_reduction);
1570-
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40, flash_attn_ext_vec_q5_1_h40, has_simdgroup_reduction);
1571-
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40, flash_attn_ext_vec_q8_0_h40, has_simdgroup_reduction);
15721558
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64, flash_attn_ext_vec_f16_h64, has_simdgroup_reduction);
15731559
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);
15741560
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);
@@ -1909,9 +1895,15 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
19091895
case GGML_OP_ARANGE:
19101896
return true;
19111897
case GGML_OP_FLASH_ATTN_EXT:
1912-
if (op->src[0]->ne[0] == 32) {
1913-
// head size == 32 (e.g. bert-bge-small)
1914-
// TODO: not sure if it is worth adding kernels for this size
1898+
// for new head sizes, add checks here
1899+
if (op->src[0]->ne[0] != 40 &&
1900+
op->src[0]->ne[0] != 64 &&
1901+
op->src[0]->ne[0] != 80 &&
1902+
op->src[0]->ne[0] != 96 &&
1903+
op->src[0]->ne[0] != 112 &&
1904+
op->src[0]->ne[0] != 128 &&
1905+
op->src[0]->ne[0] != 192 &&
1906+
op->src[0]->ne[0] != 256) {
19151907
return false;
19161908
}
19171909
if (op->src[0]->ne[0] == 576) {
@@ -5138,10 +5130,8 @@ static int ggml_metal_encode_node(
51385130

51395131
bool use_vec_kernel = false;
51405132

5141-
// TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
5142-
// for now avoiding mainly to keep the number of templates/kernels a bit lower
5143-
// these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
5144-
if (ne01 >= 20 || (ne00%128 != 0 && ne00 != 64 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
5133+
// use non-vec kernel if the batch size is large or if the vec-kernel is not supported for this head size
5134+
if (ne01 >= 20 || (ne00 == 40 || ne00 == 80 || ne00 == 112)) {
51455135
switch (src1->type) {
51465136
case GGML_TYPE_F16:
51475137
{
@@ -5329,24 +5319,6 @@ static int ggml_metal_encode_node(
53295319
use_vec_kernel = true;
53305320

53315321
switch (ne00) {
5332-
case 40:
5333-
{
5334-
switch (src1->type) {
5335-
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40].pipeline; break;
5336-
case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40].pipeline; break;
5337-
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40].pipeline; break;
5338-
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40].pipeline; break;
5339-
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40].pipeline; break;
5340-
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40].pipeline; break;
5341-
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40].pipeline; break;
5342-
default:
5343-
{
5344-
GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
5345-
GGML_LOG_ERROR("add template specialization for this type\n");
5346-
GGML_ABORT("add template specialization for this type");
5347-
}
5348-
}
5349-
} break;
53505322
case 64:
53515323
{
53525324
switch (src1->type) {

‎ggml/src/ggml-metal/ggml-metal.metal

Lines changed: 3 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -4803,6 +4803,9 @@ kernel void kernel_flash_attn_ext_vec(
48034803
ushort3 ntg[[threads_per_threadgroup]],
48044804
ushort tiisg[[thread_index_in_simdgroup]],
48054805
ushort sgitg[[simdgroup_index_in_threadgroup]]) {
4806+
static_assert(DK % 32 == 0, "DK must be divisible by 32");
4807+
static_assert(DV % 32 == 0, "DV must be divisible by 32");
4808+
48064809
const short nsg = ntg.y; // number of simdgroups
48074810
const short iwg = tgpig[2]%nwg;
48084811

@@ -5160,16 +5163,6 @@ kernel void kernel_flash_attn_ext_vec(
51605163

51615164
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;
51625165

5163-
template [[host_name("kernel_flash_attn_ext_vec_f16_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 40, 40, 8>;
5164-
#if defined(GGML_METAL_USE_BF16)
5165-
template [[host_name("kernel_flash_attn_ext_vec_bf16_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4, 1, dequantize_bf16_t4, bfloat4, 1, dequantize_bf16_t4, 40, 40, 8>;
5166-
#endif
5167-
template [[host_name("kernel_flash_attn_ext_vec_q4_0_h40")]] 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, 40, 40, 8>;
5168-
template [[host_name("kernel_flash_attn_ext_vec_q4_1_h40")]] 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, 40, 40, 8>;
5169-
template [[host_name("kernel_flash_attn_ext_vec_q5_0_h40")]] 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, 40, 40, 8>;
5170-
template [[host_name("kernel_flash_attn_ext_vec_q5_1_h40")]] 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, 40, 40, 8>;
5171-
template [[host_name("kernel_flash_attn_ext_vec_q8_0_h40")]] 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, 40, 40, 8>;
5172-
51735166
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>;
51745167
#if defined(GGML_METAL_USE_BF16)
51755168
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>;

0 commit comments

Comments
 (0)