vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

metal : optimize FA kernels (#10171)

Browse source 
* ggml : add ggml_flash_attn_ext_get_prec

* metal : use F16 precision in FA kernels

ggml-ci

* metal : minor clean-up

* metal : compile-guard bf16 FA kernels

ggml-ci

* build : remove obsolete compile flag [no ci]

* metal : prevent int overflows [no ci]

* cuda : disable BF16 FA

ggml-ci

* metal : fix BF16 requirement for FA kernels

ggml-ci

* make : clean-up [no ci]
This commit is contained in:
Georgi Gerganov 2024-11-08 13:47:22 +02:00 • committed by GitHub
parent d05b3127bd
commit 841f27abdb
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 498 additions and 339 deletions
Download patch file Download diff file Expand all files Collapse all files

3
ggml/src/ggml-cuda.cu
View file

@ -3159,6 +3159,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
#ifndef FLASH_ATTN_AVAILABLE
return false;
#endif
if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
return false;
}
if (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) {
return true;
}