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

CUDA: fix Pascal FA, deq. KV to FP16 for batch > 8 (#7681)

Browse source
This commit is contained in:
Johannes Gäßler 2024-06-01 15:47:04 +02:00 • committed by GitHub
parent 9b596417af
commit 750f60c03e
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 73 additions and 29 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml-cuda/fattn-tile-f32.cu
View file

@ -275,13 +275,13 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
constexpr int D = 64;
constexpr int nwarps = 8;
fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
} break;
case 128: {
constexpr int D = 128;
constexpr int nwarps = 8;
fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
} break;
default: {
GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");