CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-29 08:41:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml : extend support for n_seq for soft_max and fattn

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-06-24 20:14:22 +03:00
parent 8c68219835
commit 1b74b9d73b
1 changed files with 9 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
ggml/src/ggml-cpu/ops.cpp
View File

@@ -4853,7 +4853,8 @@ static void ggml_compute_forward_soft_max_f32(
GGML_TENSOR_UNARY_OP_LOCALS GGML_TENSOR_UNARY_OP_LOCALS
//const int64_t ne11 = src1 ? src1->ne[1] : 1; const int64_t nb11 = src1 ? src1->nb[1] : 1;
const int64_t nb12 = src1 ? src1->nb[2] : 1;
// TODO: is this supposed to be ceil instead of floor? // TODO: is this supposed to be ceil instead of floor?
// https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370 // https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370
@@ -4878,6 +4879,10 @@ static void ggml_compute_forward_soft_max_f32(
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16); const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
for (int i1 = ir0; i1 < ir1; i1++) { for (int i1 = ir0; i1 < ir1; i1++) {
const int64_t i11 = (i1%ne01);
//const int64_t i12 = (i1/ne01)%ne02;
const int64_t i13 = (i1/ne01)/ne02;
// ALiBi // ALiBi
const uint32_t h = (i1/ne01)%ne02; // head const uint32_t h = (i1/ne01)%ne02; // head
const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f; const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
@@ -4886,8 +4891,8 @@ static void ggml_compute_forward_soft_max_f32(
float * dp = (float *)((char *) dst->data + i1*dst->nb[1]); float * dp = (float *)((char *) dst->data + i1*dst->nb[1]);
// broadcast the mask across rows // broadcast the mask across rows
ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data + i11*nb11 + i13*nb12) : NULL;
float * mp_f32 = src1 ? (float *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; float * mp_f32 = src1 ? (float *)((char *) src1->data + i11*nb11 + i13*nb12) : NULL;
ggml_vec_cpy_f32 (nc, wp, sp); ggml_vec_cpy_f32 (nc, wp, sp);
ggml_vec_scale_f32(nc, wp, scale); ggml_vec_scale_f32(nc, wp, scale);
@@ -7227,7 +7232,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
memset(VKQ32, 0, DV*sizeof(float)); memset(VKQ32, 0, DV*sizeof(float));
} }
const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL; const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + iq3*mask->nb[2]) : NULL;
// k indices // k indices
const int ik3 = iq3 / rk3; const int ik3 = iq3 / rk3;