upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-07 02:20:48 -05:00
Code Issues Packages Projects Releases Wiki Activity

metal : single-user mamba2 inference works

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-05-01 21:27:12 -04:00
parent 6def5cd729
commit 791998b42d
2 changed files with 8 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llama-model.cpp
View File

@@ -9009,7 +9009,7 @@ struct llm_build_mamba : public llm_graph_context {
ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], (d_inner + n_group*d_state)*ggml_element_size(xBC));
// {n_head, n_seq_tokens, n_seqs}
dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b);
dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b);
ggml_tensor * ssm_ids = ggml_view_1d(ctx0, state_copy, n_seqs, 0);
// TODO: use semistructured matrices to implement state-space duality