ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

CUDA: fix non-cont. inputs for batched mat mul (#13155)

Browse source
This commit is contained in:
Johannes Gäßler 2025-04-29 16:00:27 +02:00 • committed by GitHub
parent 7d3af70b08
commit cdf76586b2
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 94 additions and 42 deletions
Download patch file Download diff file Expand all files Collapse all files

1
src/llama-model.cpp
View file

@ -10195,7 +10195,6 @@ struct llm_build_deepseek2 : public llm_graph_context {
// {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
ggml_mul_mat_set_prec(q_nope_absorbed, GGML_PREC_F32);
cb(q_nope_absorbed, "q_nope_absorbed", il);
// {kv_lora_rank, n_head, n_tokens}