Add NemotronH_Omni_Reasoning_V3 as a supported Nemotron architecture (#52929)
Signed-off-by: Nave Assaf <nassaf@nvidia.com>
Nave Assaf committed
Aug 19, 2026 at 15:34 UTC
58de6cbdc2fac255ca6fe6855f90c65f80a76d57
3 files changed
+9
-1
tests/models/registry.py
+4
@@ -1250,6 +1250,10 @@ _MULTIMODAL_EXAMPLE_MODELS = {
1250
"NemotronH_Super_Omni_Reasoning_V3": _HfExamplesInfo(
1251
"nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16", is_available_online=False
1252
),
1253
+ # TODO: Change repo id once pertinent archs are public.
1254
+ "NemotronH_Omni_Reasoning_V3": _HfExamplesInfo(
1255
+ "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16", is_available_online=False
1256
+ ),
1257
"OpenCUAForConditionalGeneration": _HfExamplesInfo(
1258
"xlangai/OpenCUA-7B",
1259
trust_remote_code=True,
vllm/config/speculative.py
+4
-1
@@ -490,7 +490,10 @@ class SpeculativeConfig:
490
{"n_predict": n_predict, "architectures": ["ErnieMTPModel"]}
491
)
492
493
- if hf_config.architectures[0] == "NemotronH_Super_Omni_Reasoning_V3":
493
+ if hf_config.architectures[0] in (
494
+ "NemotronH_Super_Omni_Reasoning_V3",
495
+ "NemotronH_Omni_Reasoning_V3",
496
+ ):
497
# Promote VLM's text_config so MTP detection below fires correctly
498
hf_config = hf_config.text_config
499
vllm/model_executor/models/registry.py
+1
@@ -535,6 +535,7 @@ _MULTIMODAL_MODELS = {
535
"NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
536
"NemotronH_Nano_Omni_Reasoning_V3": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
537
"NemotronH_Super_Omni_Reasoning_V3": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
538
+ "NemotronH_Omni_Reasoning_V3": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
539
"NVLM_D": ("nvlm_d", "NVLM_D_Model"),
540
"MuseGlimmerForConditionalGeneration": ("muse_glimmer", "MuseGlimmerForCausalLM"),
541
"OpenCUAForConditionalGeneration": ("opencua", "OpenCUAForConditionalGeneration"),