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

Dflash support for nemotron-3.5 (#26905)

Browse Source 
* conversion: skip untrained DFlash embeddings

* Add Nemotron DFlash support

* Add DFlash NVFP4 support

* Address review comments

* add missing output_s for nvfp4

* Include change for keeping residual for last layer also if requested in future dflash models

* Update conversion/qwen.py

Defensive check, not needed

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Fixing bug introduced by merge conflict

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
This commit is contained in:
lnigam
2026-08-11 18:46:26 +05:30
committed by GitHub
co-authored by Sigbjørn Skjæret
parent 6e62ba5384
commit cc078b45b6
6 changed files with 42 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/nemotron-h.cpp
+11 -1
View File
@@ -177,8 +177,11 @@ llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_
auto * inp = build_inp_mem_hybrid();
ggml_tensor * inp_out_ids = build_inp_out_ids();
const bool extract_final_inp = (size_t) n_layer < cparams.embeddings_layer_inp.size() && cparams.embeddings_layer_inp[n_layer];
for (int il = 0; il < n_layer; ++il) {
res->t_layer_inp[il] = inpL;
struct ggml_tensor * inpSA = inpL;
// norm
@@ -195,7 +198,7 @@ llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_
cur = build_ffn_layer(cur, model, il);
}
if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked && !extract_final_inp) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
@@ -209,6 +212,13 @@ llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_
}
cur = inpL;
if (extract_final_inp) {
res->t_layer_inp[n_layer] = cur;
if (inp_out_ids && cparams.embeddings_nextn_masked) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
}
}
cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);