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

ggml-cuda : add rope f16, restore performance with parallel decoding (#3272)

Browse source 
* ggml-cuda : add rope f16, restore performance

* offload KQ_mask with all models

* fix rope shift

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
slaren 2023-09-20 13:00:28 +02:00 • committed by GitHub
parent db0fc2da06
commit e04dc51988
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
4 changed files with 110 additions and 67 deletions
Download patch file Download diff file Expand all files Collapse all files

1
ggml-cuda.h
View file

@ -31,6 +31,7 @@ GGML_API void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tens
GGML_API void ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
GGML_API void ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_set_main_device(int main_device);
GGML_API void ggml_cuda_set_mul_mat_q(bool mul_mat_q);