ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-18 23:08:08 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

ggml : full ALiBi support (llama/7192)

Browse Source 
* ggml : full ALiBi support

* ggml : update ggml_soft_max_ext() CUDA, SYCL

* ggml : ggml_flash_attn_ext() support ALiBi (CPU)

* ggml : ggml_flash_attn_ext() support ALiBi (Metal)

* ggml : fix warning

* ggml : ggml_flash_attn_ext() support ALiBi (CUDA)

ggml-ci

* ggml : fix assert message

* vulkan : add dev notes

* ggml : require mask when using ALiBi

ggml-ci

* convert : fix convert for refact models
This commit is contained in:
Georgi Gerganov
2024-05-11 10:32:41 +03:00
parent 284fac39fb
commit e54329da7b
10 changed files with 261 additions and 622 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ggml-vulkan.cpp
View File

@ -3830,9 +3830,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return nullptr;
case GGML_OP_SOFT_MAX:
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32 || src2->type == GGML_TYPE_F16);
if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && (src2 == nullptr || src2->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_soft_max_f32;
}
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && src2->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
@ -4286,6 +4285,9 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx,
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
#pragma message("TODO: src2 is no longer used in soft_max - should be removed and ALiBi calculation should be updated")
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/7192")
ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_SOFT_MAX, {
ncols,
src1 != nullptr ? nrows_y : (uint32_t)0,