[Bug]: DGX Spark playbook supported models fail in trtllm-serve: Nemotron-3 Nano Omni and GPT-OSS rejected as unsupported due to MODEL_MAP registry mismatch

[evanrossd]

System Info

Environment

DGX Spark (128 GB)
Container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc13
Also reproduced on: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20

Models

nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4
openai/gpt-oss-20b

Command

trtllm-serve serve
--backend tensorrt
--trust_remote_code
nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4

Expected

Model is supported according to the DGX Spark playbook.

Actual

NotImplementedError:
The given huggingface model architecture
NemotronH_Nano_Omni_Reasoning_V3
is not supported in TRT-LLM yet

Investigation

• AutoConfig.from_pretrained(..., trust_remote_code=True) succeeds.
• Architecture returned:
  NemotronH_Nano_Omni_Reasoning_V3
• The architecture is registered in:
  tensorrt_llm._torch.models.modeling_utils.MODEL_CLASS_MAPPING
• trtllm-serve / llm_utils imports:
  tensorrt_llm.models.automodel.AutoModelForCausalLM
• AutoModelForCausalLM.get_trtllm_model_class() looks up MODEL_MAP.
• MODEL_MAP does not contain NemotronH_Nano_Omni_Reasoning_V3.
• This appears to be a registry mismatch between the legacy automodel path and the newer _torch model registry.

Question

Is the DGX Spark playbook intended to use a different code path than trtllm-serve, or is this a regression in the released containers?

Who can help?

trtllm-serve serve
--backend tensorrt
--trust_remote_code
--host 0.0.0.0
--port 8000
nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4

---> Identical command for gpt-oss <---

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Launch the official TensorRT-LLM container from the DGX Spark playbook (reproduced on both 1.3.0rc13 and 1.3.0rc20).

Set a valid Hugging Face token:

export HF_TOKEN=<your_token>

Start the container with GPU access:

docker run --rm -it
--gpus all
--ipc=host
--network=host
--ulimit memlock=-1
--ulimit stack=67108864
-e HF_TOKEN=$HF_TOKEN
-v $HOME/.cache/huggingface:/root/.cache/huggingface
nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
bash

Verify the TensorRT-LLM version:

python3 -c "import tensorrt_llm; print(tensorrt_llm.version)"

Launch a model listed as supported in the DGX Spark playbook:

trtllm-serve serve
--backend tensorrt
--trust_remote_code
--host 0.0.0.0
--port 8000
nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4

Observe the failure:

NotImplementedError:
The given huggingface model architecture
NemotronH_Nano_Omni_Reasoning_V3
is not supported in TRT-LLM yet
Repeat with another model listed as supported (e.g. openai/gpt-oss-20b) and observe a similar architecture-not-supported failure.

Expected behavior

The models should serve properly.

actual behavior

NotImplementedError: The given huggingface model architecture NemotronH_Nano_Omni_Reasoning_V3 is not supported in TRT-LLM yet

additional notes

I get this:

Skipping import of cpp extensions due to incompatible torch version 2.11.0a0+eb65b36914.nv26.02 for torchao version 0.15.0 Please see pytorch/ao#2919 for more info
/usr/local/lib/python3.12/dist-packages/modelopt/torch/init.py:36: UserWarning: transformers version 5.5.4 is incompatible with nvidia-modelopt and may cause issues. Please install recommended version with pip install nvidia-modelopt[hf] if working with HF models.
_warnings.warn(
[TensorRT-LLM] TensorRT LLM version: 1.3.0rc20
/usr/local/lib/python3.12/dist-packages/tensorrt_llm/serve/openai_protocol.py:145: UserWarning: Field name "schema" in "ResponseFormat" shadows an attribute in parent "OpenAIBaseModel"
class ResponseFormat(OpenAIBaseModel):
[07/05/2026-01:22:02] [TRT-LLM] [I] [llmapi] Using LLM with TensorRT backend
[07/05/2026-01:22:02] [TRT-LLM] [I] [plugin] Set PluginConfig.nccl_plugin to None.
Loading Model: [1/3] Downloading HF model
[07/05/2026-01:22:02] [TRT-LLM] [I] [llmapi] Downloading model nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 from HuggingFace
[07/05/2026-01:22:02] [TRT-LLM] [I] [llmapi] Finished downloading model nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 from HuggingFace
Downloaded model to /root/.cache/huggingface/hub/models--nvidia--Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4/snapshots/dc5f0b0bfddf8b6e0f5891475be9af05b80126fe
Time: 0.310s
Loading Model: [2/3] Loading HF model to memory
Traceback (most recent call last):
File "/usr/local/bin/trtllm-serve", line 6, in
sys.exit(main())
^^^^^^
File "/usr/local/lib/python3.12/dist-packages/click/core.py", line 1485, in call
return self.main(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/click/core.py", line 1406, in main
rv = self.invoke(ctx)
^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/click/core.py", line 1873, in invoke
return _process_result(sub_ctx.command.invoke(sub_ctx))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/click/core.py", line 1269, in invoke
return ctx.invoke(self.callback, **ctx.params)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/click/core.py", line 824, in invoke
return callback(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/commands/serve.py", line 1154, in serve
_serve_llm()
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/commands/serve.py", line 1126, in _serve_llm
launch_server(
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/commands/serve.py", line 385, in launch_server
llm = LLM(**llm_args)
^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm.py", line 1519, in init
super().init(model, tokenizer, tokenizer_mode, skip_tokenizer_init,
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm.py", line 349, in init
self._build_model()
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm.py", line 1553, in _build_model
super()._build_model()
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm.py", line 1381, in _build_model
self._engine_dir, self._hf_model_dir = model_loader()
^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm_utils.py", line 818, in call
return self._build_model(), self._hf_model_dir
^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm_utils.py", line 959, in _build_model
build_task(self.get_engine_dir())
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm_utils.py", line 913, in build_task
model_loader(engine_dir=engine_dir)
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm_utils.py", line 265, in call
pipeline()
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm_utils.py", line 206, in call
self.step_forward()
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm_utils.py", line 235, in step_forward
self.step_handlersself.counter
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/llmapi/llm_utils.py", line 529, in _load_model_from_hf
model_cls = AutoModelForCausalLM.get_trtllm_model_class(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/tensorrt_llm/models/automodel.py", line 85, in get_trtllm_model_class
raise NotImplementedError(
NotImplementedError: The given huggingface model architecture NemotronH_Nano_Omni_Reasoning_V3 is not supported in TRT-LLM yet

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

[waynehacking8]

The playbook never passes --backend tensorrt - its Step 8 commands run trtllm-serve with the default pytorch backend (serve.py defaults backend="pytorch"), and that flag is what routes model loading into the legacy engine flow whose MODEL_MAP doesn't include these architectures (tensorrt_llm/models/automodel.py, the exact NotImplementedError you hit). Both models are registered on the pytorch side (_torch/models/modeling_nemotron_nano.py registers NemotronH_Nano_Omni_Reasoning_V3, modeling_gpt_oss.py registers GptOssForCausalLM), so dropping --backend tensorrt should fix both failures. The tensorrt backend is being wound down anyway (#15810 is removing its tests), so I wouldn't expect the legacy registry to grow new entries - verified by source read on current main and the playbook repo, not on Spark hardware.

[bo-nv]

please drop --backend tensorrt

[evanrossd]

thank you