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

CUDA: tuned mul_mat_q kernels (#2546)

Browse source
This commit is contained in:
Johannes Gäßler 2023-08-09 09:42:34 +02:00 • committed by GitHub
parent f5bfea0580
commit 25d43e0eb5
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 676 additions and 386 deletions
Download patch file Download diff file Expand all files Collapse all files

1
README.md
View file

@ -406,7 +406,6 @@ Building the program with BLAS support may lead to some performance improvements
--->
| Option | Legal values | Default | Description |
|-------------------------|------------------------|---------|-------------|
| LLAMA_CUDA_MMQ_Y | Positive integer >= 32 | 64 | Tile size in y direction when using the custom CUDA kernels for prompt processing. Higher values can be faster depending on the amount of shared memory available. Power of 2 heavily recommended. |
| LLAMA_CUDA_FORCE_DMMV | Boolean | false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
| LLAMA_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| LLAMA_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |