EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

llama : add gpt-oss (#15091)

Browse Source 
* oai moe

* compat with new checkpoint

* add attn sink impl

* add rope scaling yarn

* logits match with latest transformers code

* wip chat template

* rm trailing space

* use ggml_scale_bias

* rm redundant is_swa_all

* convert interleaved gate_up

* graph : fix activation function to match reference (#7)

* vocab : handle o200k_harmony special tokens

* ggml : add attention sinks support (#1)

* llama : add attn sinks

* ggml : add attn sinks

* cuda : add attn sinks

* vulkan : add support for sinks in softmax

remove unnecessary return

* ggml : add fused swiglu_oai op (#11)

* ggml : add fused swiglu_oai op

* Update ggml/src/ggml-cpu/ops.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* update CUDA impl

* cont : metal impl

* add vulkan impl

* test-backend-ops : more test cases, clean up

* llama : remove unfused impl

* remove extra lines

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

* repack mxfp4 upon conversion

* clean up a bit

* enable thinking

* add quick hack to render only some special tokens

* fix bf16 conversion

* remove vocab hack

* webui ok

* support chat parsing for gpt-oss

* fix webui

* direct mapping mxfp4, FINALLY

* force using mxfp4

* properly use lazy tensor

* ggml : add mxfp4

ggml : use e8m0 conversion instead of powf

Co-authored-by: Diego Devesa <slarengh@gmail.com>

change kvalues_mxfp4 table to match e2m1 (#6)

metal : remove quantization for now (not used)

cuda : fix disabled CUDA graphs due to ffn moe bias

vulkan : add support for mxfp4

cont : add cm2 dequant

* ggml : add ggml_add_id (#13)

* ggml : add ggml_add_id

* add cuda impl

* llama : add weight support check for add_id

* perf opt

* add vulkan impl

* rename cuda files

* add metal impl

* allow in-place ggml_add_id

* llama : keep biases on CPU with --cpu-moe

* llama : fix compile error

ggml-ci

* cuda : add fallback for __nv_cvt_e8m0_to_bf16raw

ggml-ci

* cleanup

ggml-ci

* sycl : fix supports_op for MXFP4

ggml-ci

* fix Unknown reasoning format

* ggml-cpu : fix AVX build

ggml-ci

* fix hip build

ggml-ci

* cuda : add mxfp4 dequantization support for cuBLAS

ggml-ci

* ggml-cpu : fix mxfp4 fallback definitions for some architectures

ggml-ci

* cuda : fix version required for __nv_cvt_e8m0_to_bf16raw

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Georgi Gerganov
2025-08-05 22:10:36 +03:00
committed by GitHub
parent f324a3b715
commit fd1234cb46
83 changed files with 2942 additions and 227 deletions
Download Patch File Download Diff File Expand all files Collapse all files

118
ggml/src/ggml-metal/ggml-metal.m
View File

@@ -195,6 +195,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_ROW_C4,
GGML_METAL_KERNEL_TYPE_DIV,
GGML_METAL_KERNEL_TYPE_DIV_ROW_C4,
GGML_METAL_KERNEL_TYPE_ADD_ID,
GGML_METAL_KERNEL_TYPE_REPEAT_F32,
GGML_METAL_KERNEL_TYPE_REPEAT_F16,
GGML_METAL_KERNEL_TYPE_REPEAT_I32,
@@ -234,6 +235,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,
GGML_METAL_KERNEL_TYPE_GET_ROWS_MXFP4,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,
@@ -286,6 +288,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_MXFP4_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4,
@@ -310,6 +313,10 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_2,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_3,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_4,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_5,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3,
GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4,
@@ -351,6 +358,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_MXFP4_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,
@@ -373,6 +381,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,
@@ -397,6 +406,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MXFP4_F16,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16,
@@ -579,6 +589,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_REGLU,
GGML_METAL_KERNEL_TYPE_GEGLU,
GGML_METAL_KERNEL_TYPE_SWIGLU,
GGML_METAL_KERNEL_TYPE_SWIGLU_OAI,
GGML_METAL_KERNEL_TYPE_GEGLU_ERF,
GGML_METAL_KERNEL_TYPE_GEGLU_QUICK,
GGML_METAL_KERNEL_TYPE_SUM_ROWS,
@@ -1199,6 +1210,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW_C4, mul_row_c4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW_C4, div_row_c4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ID, add_id, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32, repeat_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16, repeat_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32, repeat_i32, true);
@@ -1238,6 +1250,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1, get_rows_q5_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0, get_rows_q8_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_MXFP4, get_rows_mxfp4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K, get_rows_q2_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K, get_rows_q3_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K, get_rows_q4_K, true);
@@ -1290,6 +1303,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_MXFP4_F32, mul_mv_mxfp4_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2, mul_mv_ext_f16_f32_r1_2, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3, mul_mv_ext_f16_f32_r1_3, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4, mul_mv_ext_f16_f32_r1_4, has_simdgroup_reduction);
@@ -1314,6 +1328,10 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3, mul_mv_ext_q8_0_f32_r1_3, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4, mul_mv_ext_q8_0_f32_r1_4, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5, mul_mv_ext_q8_0_f32_r1_5, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_2, mul_mv_ext_mxfp4_f32_r1_2, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_3, mul_mv_ext_mxfp4_f32_r1_3, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_4, mul_mv_ext_mxfp4_f32_r1_4, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_5, mul_mv_ext_mxfp4_f32_r1_5, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2, mul_mv_ext_q4_K_f32_r1_2, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3, mul_mv_ext_q4_K_f32_r1_3, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4, mul_mv_ext_q4_K_f32_r1_4, has_simdgroup_reduction);
@@ -1355,6 +1373,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_MXFP4_F32, mul_mv_id_mxfp4_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, has_simdgroup_reduction);
@@ -1377,6 +1396,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32, mul_mm_mxfp4_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32, mul_mm_mxfp4_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, has_simdgroup_mm);
@@ -1401,6 +1422,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16, mul_mm_id_q5_0_f16, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16, mul_mm_id_q5_1_f16, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16, mul_mm_id_q8_0_f16, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MXFP4_F16, mul_mm_id_mxfp4_f16, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16, mul_mm_id_q2_K_f16, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16, mul_mm_id_q3_K_f16, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16, mul_mm_id_q4_K_f16, has_simdgroup_mm);
@@ -1583,6 +1605,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REGLU, reglu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU, geglu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SWIGLU, swiglu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SWIGLU_OAI, swiglu_oai, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_ERF, geglu_erf, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_QUICK, geglu_quick, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true);
@@ -1774,6 +1797,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_SWIGLU_OAI:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
return ggml_is_contiguous_1(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
@@ -1791,6 +1815,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_OP_SUB:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_ADD_ID:
return op->src[0]->type == GGML_TYPE_F32;
case GGML_OP_ACC:
case GGML_OP_REPEAT:
@@ -2042,6 +2067,7 @@ static int ggml_metal_encode_node(
const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT;
const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
size_t offs_src0 = 0;
@@ -2291,6 +2317,38 @@ static int ggml_metal_encode_node(
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
}
} break;
case GGML_OP_ADD_ID:
{
GGML_ASSERT(src0t == GGML_TYPE_F32);
GGML_ASSERT(src1t == GGML_TYPE_F32);
GGML_ASSERT(src2t == GGML_TYPE_I32);
GGML_ASSERT(dstt == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous_rows(src0));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ID].pipeline;
ggml_metal_kargs_add_id args = {
/*.ne0 =*/ ne0,
/*.ne1 =*/ ne1,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb11 =*/ nb11,
/*.nb21 =*/ nb21,
};
[encoder setComputePipelineState:pipeline];
[encoder setBytes:&args length:sizeof(args) atIndex:0];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
[encoder setBuffer:id_dst offset:offs_dst atIndex:4];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_REPEAT:
{
id<MTLComputePipelineState> pipeline;
@@ -2710,6 +2768,9 @@ static int ggml_metal_encode_node(
case GGML_GLU_OP_SWIGLU:
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SWIGLU].pipeline;
break;
case GGML_GLU_OP_SWIGLU_OAI:
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SWIGLU_OAI].pipeline;
break;
case GGML_GLU_OP_GEGLU_ERF:
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GEGLU_ERF].pipeline;
break;
@@ -2720,7 +2781,9 @@ static int ggml_metal_encode_node(
GGML_ABORT("fatal error");
}
const int32_t swp = ((const int32_t *) dst->op_params)[1];
const int32_t swp = ggml_get_op_params_i32(dst, 1);
const float alpha = ggml_get_op_params_f32(dst, 2);
const float limit = ggml_get_op_params_f32(dst, 3);
const int32_t i00 = swp ? ne0 : 0;
const int32_t i10 = swp ? 0 : ne0;
@@ -2734,6 +2797,8 @@ static int ggml_metal_encode_node(
/*.nb1 =*/ nb1,
/*.i00 =*/ src1 ? 0 : i00,
/*.i10 =*/ src1 ? 0 : i10,
/*.alpha=*/ alpha,
/*.limit=*/ limit
};
[encoder setComputePipelineState:pipeline];
@@ -2992,8 +3057,13 @@ static int ggml_metal_encode_node(
} else {
[encoder setBuffer:h_src0 offset:offs_src0 atIndex:1];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&args length:sizeof(args) atIndex:3];
if (id_src2) {
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
} else {
[encoder setBuffer:h_src0 offset:offs_src0 atIndex:2];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:3];
[encoder setBytes:&args length:sizeof(args) atIndex:4];
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
@@ -3291,6 +3361,7 @@ static int ggml_metal_encode_node(
src0t == GGML_TYPE_Q5_0 ||
src0t == GGML_TYPE_Q5_1 ||
src0t == GGML_TYPE_Q8_0 ||
src0t == GGML_TYPE_MXFP4 ||
src0t == GGML_TYPE_IQ4_NL ||
false) && (ne11 >= 2 && ne11 <= 8)
) ||
@@ -3383,6 +3454,14 @@ static int ggml_metal_encode_node(
case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5].pipeline; break;
default: GGML_ABORT("not implemented");
} break;
case GGML_TYPE_MXFP4:
switch (r1ptg) {
case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_2].pipeline; break;
case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_3].pipeline; break;
case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_4].pipeline; break;
case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_5].pipeline; break;
default: GGML_ABORT("not implemented");
} break;
case GGML_TYPE_Q4_K:
switch (r1ptg) {
case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2].pipeline; break;
@@ -3481,6 +3560,7 @@ static int ggml_metal_encode_node(
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32 ].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32 ].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32 ].pipeline; break;
case GGML_TYPE_MXFP4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32 ].pipeline; break;
case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32 ].pipeline; break;
case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32 ].pipeline; break;
case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32 ].pipeline; break;
@@ -3623,6 +3703,13 @@ static int ggml_metal_encode_node(
nr0 = N_R0_Q8_0;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
} break;
case GGML_TYPE_MXFP4:
{
nsg = N_SG_MXFP4;
nr0 = N_R0_MXFP4;
smem = 32*sizeof(float);
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_MXFP4_F32].pipeline;
} break;
case GGML_TYPE_Q2_K:
{
nsg = N_SG_Q2_K;
@@ -3756,8 +3843,6 @@ static int ggml_metal_encode_node(
case GGML_OP_MUL_MAT_ID:
{
// src2 = ids
const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
GGML_ASSERT(src2t == GGML_TYPE_I32);
GGML_ASSERT(!ggml_is_transposed(src0));
@@ -3883,6 +3968,7 @@ static int ggml_metal_encode_node(
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16 ].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16 ].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16 ].pipeline; break;
case GGML_TYPE_MXFP4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MXFP4_F16 ].pipeline; break;
case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16 ].pipeline; break;
case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16 ].pipeline; break;
case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16 ].pipeline; break;
@@ -4018,6 +4104,13 @@ static int ggml_metal_encode_node(
nr0 = N_R0_Q8_0;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32].pipeline;
} break;
case GGML_TYPE_MXFP4:
{
nsg = N_SG_MXFP4;
nr0 = N_R0_MXFP4;
smem = 32*sizeof(float);
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_MXFP4_F32].pipeline;
} break;
case GGML_TYPE_Q2_K:
{
nsg = N_SG_Q2_K;
@@ -4170,6 +4263,7 @@ static int ggml_metal_encode_node(
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0 ].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1 ].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0 ].pipeline; break;
case GGML_TYPE_MXFP4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_MXFP4 ].pipeline; break;
case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K ].pipeline; break;
case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K ].pipeline; break;
case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K ].pipeline; break;
@@ -4980,11 +5074,14 @@ static int ggml_metal_encode_node(
GGML_ASSERT(ne11 == ne21);
GGML_ASSERT(ne12 == ne22);
struct ggml_tensor * src3 = node->src[3];
struct ggml_tensor * src3 = node->src[3]; // mask
struct ggml_tensor * src4 = node->src[4]; // sinks
size_t offs_src3 = 0;
size_t offs_src4 = 0;
id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
id<MTLBuffer> id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil;
GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16);
GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
@@ -5000,8 +5097,6 @@ static int ggml_metal_encode_node(
const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32);
const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33);
const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
float scale;
float max_bias;
float logit_softcap;
@@ -5389,7 +5484,12 @@ static int ggml_metal_encode_node(
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:4];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:5];
if (id_src4) {
[encoder setBuffer:id_src4 offset:offs_src4 atIndex:5];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:5];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:6];
if (!use_vec_kernel) {
// half8x8 kernel