minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix DSpark loading for hybrid DSV4 NVFP4 (#33276)

Browse Source
This commit is contained in:
Mohammad Miadh Angkad
2026-08-02 18:05:09 +00:00
committed by GitHub
parent 1a3bea77f2
commit ec741e4161
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_loader/loader.py
+3 -2
View File
@@ -263,10 +263,11 @@ def _get_quantization_config(
ModelOptFp4Config, ModelOptFp4Config,
) )
# MTP MoE layers (model.decoder.*) are not NVFP4 quantized. # Draft experts serialized under mtp.* remain source MXFP4.
# NextN exposes them as model.decoder.*; DSpark as stages.*.
nvfp4_exclude_modules = list( nvfp4_exclude_modules = list(
nvfp4_meta.get("exclude_modules") or [] nvfp4_meta.get("exclude_modules") or []
) + ["model.decoder.*"] ) + ["model.decoder.*", "stages.*"]
nvfp4_config = ModelOptFp4Config( nvfp4_config = ModelOptFp4Config(
is_checkpoint_nvfp4_serialized=True, is_checkpoint_nvfp4_serialized=True,
group_size=int(nvfp4_meta["group_size"]), group_size=int(nvfp4_meta["group_size"]),