CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-06 09:46:50 +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

216
ggml/src/ggml-cpu/ops.cpp
View File

@@ -8,6 +8,7 @@
#include "vec.h"
#include <float.h>
#include <algorithm>
// ggml_compute_forward_dup
@@ -1283,6 +1284,7 @@ void ggml_compute_forward_add(
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:
@@ -1309,6 +1311,77 @@ void ggml_compute_forward_add(
}
}
// ggml_compute_forward_add_id
static void ggml_compute_forward_add_id_f32(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(src2->type == GGML_TYPE_I32);
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT(src1->nb[0] == sizeof(float));
const int ith = params->ith;
const int nth = params->nth;
const int nr = ggml_nrows(src0);
GGML_TENSOR_TERNARY_OP_LOCALS
GGML_ASSERT( nb0 == sizeof(float));
GGML_ASSERT(nb10 == sizeof(float));
// rows per thread
const int dr = (nr + nth - 1)/nth;
// row range for this thread
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int ir = ir0; ir < ir1; ++ir) {
// src0 indices
const int i3 = ir/(ne2*ne1);
const int i2 = (ir - i3*ne2*ne1)/ne1;
const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
// src1 indices
const int i11 = *(int32_t *) ((char *) src2->data + i1*nb20 + i2*nb21);
GGML_ASSERT(i11 >= 0 && i11 < ne11);
ggml_vec_add_f32(ne0,
(float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ),
(float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
(float *) ((char *) src1->data + i11*nb11));
}
}
void ggml_compute_forward_add_id(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_add_id_f32(params, dst);
} break;
default:
{
GGML_ABORT("unsupported type for ggml_compute_forward_add_id: %s", ggml_type_name(src0->type));
}
}
}
// ggml_compute_forward_add1
static void ggml_compute_forward_add1_f32(
@@ -1660,6 +1733,7 @@ void ggml_compute_forward_add1(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -1787,6 +1861,7 @@ void ggml_compute_forward_acc(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -3614,6 +3689,93 @@ static void ggml_compute_forward_swiglu(
}
}
// ggml_compute_forward_swiglu_oai
static void ggml_compute_forward_swiglu_oai_f32(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
char * src0_d = (char *) src0->data;
char * src1_d = (char *) (src1 ? src1->data : src0->data);
const size_t src0_o = src0->nb[1];
const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));
if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
}
const int ith = params->ith;
const int nth = params->nth;
const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);
GGML_ASSERT(dst->ne[0] == nc);
GGML_ASSERT(ggml_nrows(dst) == nr);
const int32_t swapped = 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);
// rows per thread
const int dr = (nr + nth - 1)/nth;
// row range for this thread
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int i1 = ir0; i1 < ir1; i1++) {
float * src0_p = (float *) (src0_d + i1*src0_o);
float * src1_p = (float *) (src1_d + i1*src1_o);
float * dst_p = (float *) ((char *) dst->data + i1*(dst->nb[1]));
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
for (int k = 0; k < nc; k++) {
const float x = std::min(src0_p[k], limit);
const float y = std::clamp(src1_p[k], -limit, limit);
const float out_glu = x / (1.f + expf(alpha * (-x)));
dst_p[k] = out_glu * (y + 1.f);
}
#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
const float x = dst_p[k];
GGML_UNUSED(x);
assert(!isnan(x));
assert(!isinf(x));
}
#endif
}
}
static void ggml_compute_forward_swiglu_oai(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_swiglu_oai_f32(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}
// ggml_compute_forward_geglu_erf
static void ggml_compute_forward_geglu_erf_f32(
@@ -4599,6 +4761,7 @@ void ggml_compute_forward_out_prod(
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:
@@ -4873,6 +5036,7 @@ void ggml_compute_forward_set(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -5134,6 +5298,7 @@ void ggml_compute_forward_get_rows(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -5523,6 +5688,7 @@ static void ggml_compute_forward_soft_max_f32(
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
assert(ggml_is_contiguous(dst));
assert(ggml_are_same_shape(src0, dst));
@@ -5557,6 +5723,9 @@ static void ggml_compute_forward_soft_max_f32(
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
// sinks
const float * sk = src2 ? (float *)((char *) src2->data) : nullptr;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
@@ -5599,9 +5768,18 @@ static void ggml_compute_forward_soft_max_f32(
float max = -INFINITY;
ggml_vec_max_f32(ne00, &max, wp);
// if we have sinks, make a correction as if they were included in the softmax
if (sk) {
max = MAX(max, sk[i02]);
}
ggml_float sum = ggml_vec_soft_max_f32(ne00, dp, wp, max);
assert(sum > 0.0);
if (sk) {
sum += (ggml_float) expf(sk[i02] - max);
}
sum = 1.0/sum;
ggml_vec_scale_f32(ne00, dp, sum);
@@ -5836,6 +6014,7 @@ void ggml_compute_forward_clamp(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -7989,12 +8168,14 @@ void ggml_compute_forward_argsort(
static void ggml_compute_forward_flash_attn_ext_f16(
const ggml_compute_params * params,
const ggml_tensor * q,
const ggml_tensor * k,
const ggml_tensor * v,
const ggml_tensor * mask,
ggml_tensor * dst) {
const ggml_tensor * q = dst->src[0];
const ggml_tensor * k = dst->src[1];
const ggml_tensor * v = dst->src[2];
const ggml_tensor * mask = dst->src[3];
const ggml_tensor * sinks = dst->src[4];
GGML_TENSOR_LOCALS(int64_t, neq, q, ne)
GGML_TENSOR_LOCALS(size_t, nbq, q, nb)
GGML_TENSOR_LOCALS(int64_t, nek, k, ne)
@@ -8189,6 +8370,23 @@ static void ggml_compute_forward_flash_attn_ext_f16(
}
}
// sinks
if (sinks) {
const float s = ((float *)((char *) sinks->data))[h];
float ms = 1.0f;
float vs = 1.0f;
if (s > M) {
ms = expf(M - s);
ggml_vec_scale_f32(DV, VKQ32, ms);
} else {
vs = expf(s - M);
}
S = S*ms + vs;
}
// V /= S
const float S_inv = 1.0f/S;
ggml_vec_scale_f32(DV, VKQ32, S_inv);
@@ -8208,17 +8406,13 @@ static void ggml_compute_forward_flash_attn_ext_f16(
void ggml_compute_forward_flash_attn_ext(
const ggml_compute_params * params,
const ggml_tensor * q,
const ggml_tensor * k,
const ggml_tensor * v,
const ggml_tensor * mask,
ggml_tensor * dst) {
switch (dst->op_params[3]) {
case GGML_PREC_DEFAULT:
case GGML_PREC_F32:
{
// uses F32 accumulators
ggml_compute_forward_flash_attn_ext_f16(params, q, k, v, mask, dst);
ggml_compute_forward_flash_attn_ext_f16(params, dst);
} break;
default:
{
@@ -9080,6 +9274,10 @@ void ggml_compute_forward_glu(
{
ggml_compute_forward_swiglu(params, dst);
} break;
case GGML_GLU_OP_SWIGLU_OAI:
{
ggml_compute_forward_swiglu_oai(params, dst);
} break;
case GGML_GLU_OP_GEGLU_ERF:
{
ggml_compute_forward_geglu_erf(params, dst);