mirrors/llama.cpp
0
0 Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-10-28 17:56:13 +08:00
Code Activity

glm5-next : cont q_absorbed for -sm tensor with non-unified KV

Browse Source
This commit is contained in:
Aman Gupta authored and Georgi Gerganov committed 2026-10-27 13:34:03 +02:00
1 parent e3db613f8e
commit 53db9b383f
1 file changed
+2 -1
Download Patch File Download Diff File Expand all files Collapse all files
src/models/glm5-next.cpp
+2 -1
View File
@@ -1072,7 +1072,8 @@ llama_model_glm5_next::graph::build_dsa_layer(
// absorb wk_b so the cache holds only the latent
ggml_tensor * q_absorbed = ggml_permute(ctx0, q, 0, 2, 1, 3);
q_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_absorbed);
q_absorbed = ggml_permute(ctx0, q_absorbed, 0, 2, 1, 3);
// FIXME: -sm tensor cannot split the per-stream view of a permuted q in build_attn_mha (non-unified KV)
q_absorbed = ggml_cont(ctx0, ggml_permute(ctx0, q_absorbed, 0, 2, 1, 3));
cb(q_absorbed, "q_absorbed", il);
ggml_tensor * kq_mask = inp_attn->get_kq_mask();