thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

llama: end-to-end tests (#19802)

Browse source 
* tests: add end-to-end tests per model architecture

* fixup for rebase

* fix use-after-free in llama-model-loader.cpp

* fix CI

* fix WebGPU

* fix CI

* disable CI for macOS-latest-cmake-arm64

* use expert_weights_scale only if != 0.0f

* comments
This commit is contained in:
Johannes Gäßler 2026-03-08 12:30:21 +01:00 • committed by GitHub
parent a95047979a
commit a976ff081b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
33 changed files with 1607 additions and 633 deletions
Download patch file Download diff file Expand all files Collapse all files

3
src/llama-graph.cpp
View file

@ -509,6 +509,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
float * data = (float *) cross_kq_mask->data;
for (int i = 0; i < n_tokens; ++i) {
GGML_ASSERT(!cross->seq_ids_enc.empty() && "llama_encode must be called first");
for (int j = 0; j < n_enc; ++j) {
float f = -INFINITY;
@ -1150,6 +1151,7 @@ ggml_tensor * llm_graph_context::build_ffn(
return cur;
}
// TODO remove redundant scale_w argument
ggml_tensor * llm_graph_context::build_moe_ffn(
ggml_tensor * cur,
ggml_tensor * gate_inp,
@ -1607,6 +1609,7 @@ ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
// this need to be 1x1xN for broadcasting
cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens);
ggml_set_input(cur);
ggml_set_name(cur, "attn_scale");
res->add_input(std::move(inp));