Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

mtmd/ggml: add ggml_build_forward_order (#26649)

Browse Source 
* ggml: add ggml_build_forward_order

ggml_build_forward_expand marks the tensor and all its ancestors for
compute, so using it as a pure ordering hint (keeping q, k and v
together) defeats ggml_build_forward_select: the unselected branch is
forced to run with inputs that were never uploaded. In the mtmd audio
graph this makes GEN_WAV calls execute the GEN_CODE branch with a
stale inp_code0, hitting the get_rows bound assert on CPU.

Add ggml_build_forward_order, which inserts nodes without the compute
flag; the flag is restored when the branch is actually selected.
Switch the q/k/v hints in clip_graph::build_attn to it.

* nit: reduce comments (AGENTS.md)
This commit is contained in:
Pascal
2026-08-06 00:47:59 +02:00
committed by GitHub
parent f9e832c10e
commit c8e03ce812
3 changed files with 14 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/clip.cpp
+4 -3
View File
@@ -708,9 +708,10 @@ ggml_tensor * clip_graph::build_attn(
ggml_tensor * sinks) const {
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
ggml_build_forward_expand(gf, q_cur);
ggml_build_forward_expand(gf, k_cur);
ggml_build_forward_expand(gf, v_cur);
// the order is fixed without the compute flag, so an unselected branch stays out of the compute set
ggml_build_forward_order(gf, q_cur);
ggml_build_forward_order(gf, k_cur);
ggml_build_forward_order(gf, v_cur);
ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
//cb(q, "q", il);