EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

cuda : Enable CUDA Graph usage for Nemotron Nano v2 (NemotronH) (#16328)

Browse Source 
* Fix Nemotron Nano v2 9B not executing as CUDA Graph on NVIDIA GPUs

* fix to ensure test-backend-ops check passes
This commit is contained in:
anavp-nvidia
2025-09-30 08:13:22 +00:00
committed by GitHub
parent 35fb82497e
commit a014310374
3 changed files with 20 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llama-model.cpp
View File

@@ -11751,6 +11751,7 @@ struct llm_graph_context_mamba : public llm_graph_context {
// TODO: skip computing output earlier for unused tokens
y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d));
cb(y, "mamba2_y_add_d", il);
y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
// grouped RMS norm
@@ -14705,6 +14706,7 @@ struct llm_build_nemotron_h : public llm_graph_context_mamba {
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
ggml_build_forward_expand(gf, inpL);
auto * inp = build_inp_mem_hybrid();
@@ -14736,7 +14738,7 @@ struct llm_build_nemotron_h : public llm_graph_context_mamba {
// add residual
cur = ggml_add(ctx0, cur, inpSA);
cb(cur, "block_out", il);
cb(cur, "nemotron_h_block_out", il);
// input for next layer
inpL = cur;