CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages 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

25
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -4,6 +4,7 @@
#include "ggml-cuda/common.cuh"
#include "ggml-cuda/acc.cuh"
#include "ggml-cuda/add-id.cuh"
#include "ggml-cuda/arange.cuh"
#include "ggml-cuda/argmax.cuh"
#include "ggml-cuda/argsort.cuh"
@@ -2259,6 +2260,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_OP_ADD1: // TODO: more efficient implementation
ggml_cuda_op_add(ctx, dst);
break;
case GGML_OP_ADD_ID:
ggml_cuda_op_add_id(ctx, dst);
break;
case GGML_OP_SUB:
ggml_cuda_op_sub(ctx, dst);
break;
@@ -2333,6 +2337,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_GLU_OP_SWIGLU:
ggml_cuda_op_swiglu(ctx, dst);
break;
case GGML_GLU_OP_SWIGLU_OAI:
ggml_cuda_op_swiglu_oai(ctx, dst);
break;
case GGML_GLU_OP_GEGLU_ERF:
ggml_cuda_op_geglu_erf(ctx, dst);
break;
@@ -2607,6 +2614,9 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
const std::string ffn_moe_gate_bias_prefix = "ffn_moe_gate_biased";
const std::string ffn_moe_up_bias_prefix = "ffn_moe_up_biased";
const std::string ffn_moe_down_bias_prefix = "ffn_moe_down_biased";
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
@@ -2629,7 +2639,13 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
#endif
}
if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1 && (node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) && (node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true)) {
if (node->op == GGML_OP_ADD &&
node->src[1] && node->src[1]->ne[1] > 1 &&
(node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) &&
(node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true) &&
strncmp(node->name, ffn_moe_gate_bias_prefix.c_str(), ffn_moe_gate_bias_prefix.size()) != 0 &&
strncmp(node->name, ffn_moe_up_bias_prefix.c_str(), ffn_moe_up_bias_prefix.size()) != 0 &&
strncmp(node->name, ffn_moe_down_bias_prefix.c_str(), ffn_moe_down_bias_prefix.size()) != 0) {
// disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
// by means of matching node names. See
// https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
@@ -3227,6 +3243,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
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]);
@@ -3277,6 +3294,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -3423,6 +3441,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE:
case GGML_OP_ADD:
case GGML_OP_ADD_ID:
case GGML_OP_ADD1:
case GGML_OP_SUB:
case GGML_OP_MUL:
@@ -3503,6 +3522,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];
return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0;
}
// TODO: more general-purpose attention sink support [TAG_ATTN_SINKS]
if (op->src[4] && op->src[0]->ne[0] != 64 && op->src[0]->ne[0] != 128) { // currently only sinks for head_size 64 and 128 are supported
return false;
}
if (op->src[0]->ne[0] == 192) {
return false;
}