tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 20:25:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: FA support for Deepseek (Ampere or newer) (#13306)

Browse Source 
* CUDA: FA support for Deepseek (Ampere or newer)

* do loop unrolling via C++ template
This commit is contained in:
Johannes Gäßler
2025-05-09 13:34:58 +02:00
committed by GitHub
parent 27ebfcacba
commit 0cf6725e9f
33 changed files with 852 additions and 547 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
src/llama-graph.cpp
View File

@ -1227,8 +1227,19 @@ ggml_tensor * llm_graph_context::build_attn_mha(
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
if (v_mla) {
#if 0
// v_mla can be applied as a matrix-vector multiplication with broadcasting across dimension 3 == n_tokens.
// However, the code is optimized for dimensions 0 and 1 being large, so this is ineffient.
cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
cur = ggml_mul_mat(ctx0, v_mla, cur);
#else
// It's preferable to do the calculation as a matrix-matrix multiplication with n_tokens in dimension 1.
// The permutations are noops and only change how the tensor data is interpreted.
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_mul_mat(ctx0, v_mla, cur);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_cont(ctx0, cur); // Needed because ggml_reshape_2d expects contiguous inputs.
#endif
}
cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);