[Bug]: qwen3_5_weight_mapper rejects ModelOpt W4A16_AWQ checkpoints.

System Info

• GPU: NVIDIA RTX A6000
• GPU RAM: 48GB
• tensorrt_llm == 1.3.0rc15 (container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc15)
• nvidia-modelopt == 0.45.0.dev114+g2b02ab9a1
• Model: Qwen/Qwen3.5-4B
• Backend: PyTorch (trtllm-serve serve ... --backend pytorch

Who can help?

No response

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

# 1. Quantize Qwen3.5-4B with ModelOpt
python3 Model-Optimizer/examples/llm_ptq/hf_ptq.py \
  --pyt_ckpt_path /path/to/Qwen3.5-4B \
  --qformat int4_awq \
  --output_dir /path/to/qwen35-awq-ckpt \
  --export_fmt hf \
  --calib_size 512

# 2. Resulting hf_quant_config.json:
# {
#   "producer": {"name": "modelopt", "version": "0.45.0.dev114"},
#   "quantization": {
#     "quant_algo": "W4A16_AWQ",
#     "kv_cache_quant_algo": "FP8",
#     "group_size": 128,
#     "has_zero_point": false,
#     "pre_quant_scale": true
#   }
# }

# Try to Serve
trtllm-serve serve /path/to/qwen35-awq-ckpt \
  --backend pytorch \
  --host 0.0.0.0 --port 8080

Expected behavior

Server loads the checkpoint. TRT-LLM correctly identifies the format.

actual behavior

File "tensorrt_llm/_torch/models/checkpoints/hf/qwen3_5_weight_mapper.py",
     line 160, in _split_qkv_scale_tensor
    assert tensor.shape[0] == expected_total_blocks, (
AssertionError: Expected packed qkv scale tensor with leading dim 64,
                got torch.Size([8192, 20])

additional notes

In qwen3_5_weight_mapper, _split_qkv_scale_tensor hardcodes a /128 block factor along the output dimension:

def _split_qkv_scale_tensor(
        self, tensor: torch.Tensor, expected_q: int, expected_v: int
    ) -> tuple[torch.Tensor, ...]:
        expected_q_blocks = math.ceil(expected_q / 128)
        expected_v_blocks = math.ceil(expected_v / 128)
        expected_total_blocks = expected_q_blocks * 2 + expected_v_blocks
        assert tensor.shape[0] == expected_total_blocks, (
            f"Expected packed qkv scale tensor with leading dim {expected_total_blocks}, "
            f"got {tensor.shape}"
        )
        return torch.split(tensor, [expected_q_blocks, expected_q_blocks, expected_v_blocks], dim=0)

But standard AWQ stores per-output-channel scales:

• ModelOpt W4A16_AWQ with group_size=128: (out_dim, in_dim/128) = (8192, 20)

In this case, the leading dim is the full output channel count (8192), not the block count (64).

I'm not sure whether this should be classified as a bug or whether TensorRT-LLM simply hasn't added int4_awq support for Qwen3.5 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.

Start in tensorrt_llm/_torch/models/checkpoints/hf/qwen3_5_weight_mapper.py, especially _split_qkv_scale_tensor, and reproduce the failure with the provided ModelOpt W4A16_AWQ Qwen3.5 checkpoint and PyTorch serving command. Compare the mapper's expected packed scale shape with the reported (8192, 20) tensor; done means the checkpoint loads successfully while existing supported formats remain correct.