Post-training quantization script (hf_ptq.py) fails on Llama-4-Scout-17B due to missing architectures key in config.json

[rajulshakya777]

The generic post-training quantization script, examples/llm_ptq/hf_ptq.py, fails when attempting to process and export a checkpoint for the meta-llama/Llama-4-Scout-17B-16E-Instruct model.

The script works correctly for previous generation models like Llama 3, but it encounters a TypeError with Llama 4.

Root cause: the script expects the model's config.json to have an architectures key, which it uses to identify the model type.
The Llama 4 Scout model's config file does not contain this key, causing the script to crash when it tries to access model.config.architectures[0].

This is a blocker as it prevents the use of the standard post-training quantization workflow for a major new model release.

Error Log:
```
Traceback (most recent call last):
File "/mnt/TensorRT-Model-Optimizer/examples/llm_ptq/hf_ptq.py", line 772, in
main(args)
File "/mnt/TensorRT-Model-Optimizer/examples/llm_ptq/hf_ptq.py", line 616, in main
export_tensorrt_llm_checkpoint(
File "/mnt/TensorRT-Model-Optimizer/modelopt/torch/export/model_config_export.py", line 553, in export_tensorrt_llm_checkpoint
raise e
File "/mnt/TensorRT-Model-Optimizer/modelopt/torch/export/model_config_export.py", line 487, in export_tensorrt_llm_checkpoint
for (
File "/mnt/TensorRT-Model-Optimizer/modelopt/torch/export/model_config_export.py", line 154, in torch_to_tensorrt_llm_checkpoint
architecture = model.config.architectures[0]
~~~~~~~~~~~~~~~~~~~~~~~~~~^^^
TypeError: 'NoneType' object is not subscriptable
```

---

Steps/Code to reproduce bug

1. Set up an environment using an NVIDIA TensorRT-LLM container.
2. Download the Llama 4 Scout model from Hugging Face:
   ```bash
   huggingface-cli download meta-llama/Llama-4-Scout-17B-16E-Instruct
   --local-dir /path/to/Llama-4-Scout-17B-16E-Instruct
   ```
3. Run the PTQ script:
   ```bash
   python3 examples/llm_ptq/hf_ptq.py
   --model_dir /path/to/Llama-4-Scout-17B-16E-Instruct
   --output_dir /tmp/quantized_llama4
   --calib_dataset cnn_dailymail
   --num_calib_size 32
   --dtype float16
   --qformat fp8
   ```

---

Expected behavior

The hf_ptq.py script should successfully quantize the Llama 4 model and export a TensorRT-LLM compatible checkpoint, just as it does for Llama 3 models.

---

System information

• Container used (if applicable): Official NVIDIA TensorRT-LLM container
• OS (e.g., Ubuntu 22.04, CentOS 7, Windows 10): Ubuntu 24.04
• CPU architecture (x86_64, aarch64): x86_64
• GPU name (e.g. H100, A100, L40S): H100
• GPU memory size: 80 GB
• Number of GPUs: 8
• Library versions (if applicable):
  • Python: 3.12
  • ModelOpt version or commit hash: latest inside container
  • CUDA: 12.9
  • PyTorch: 2.8.0a0+5228986c39.nv25.5
  • Transformers: (container version)
  • TensorRT-LLM: 0.21.0
  • ONNXRuntime: (container version)
  • TensorRT: (container version)
• Any other details that may help: Bug only occurs with Llama 4 Scout models (config.json missing architectures key).