[Bug]: Qwen3.8-27B NVFP4 cannot load with the PyTorch backend

[mihai-chiorean]

System Info

• TensorRT-LLM: current main
• Backend: PyTorch
• Model: unsloth/Qwen3.8-27B-NVFP4
• Reproduced on: DGX Spark / SM121

Who can help?

No response

Information

• The official example scripts
• The transformers checkpoint loader
• The PyTorch backend

Reproduction

Attempt to load the checkpoint through LLM(...) or trtllm-serve with the PyTorch backend and text-only multimodal execution (disable_mm_encoder=True).

The checkpoint is a dense Qwen3.8/Qwen3.5-family VLM with mixed compressed-tensors quantization: packed NVFP4 dense MLP weights, FP8 attention/GDN/lm_head weights, FP8 KV cache, and one MTP layer.

Expected behavior

The checkpoint should select Qwen3_5VLModel, preserve each config group's quantization, map packed compressed-tensors names to TRT-LLM parameters, construct a rowwise-FP8 lm_head, and generate coherent text with and without MTP.

Actual behavior

Current main has several independent blockers:

1. Generic dense Qwen3.6/3.8 VLM configs are classified as Qwen-Image-Bench configs.
2. The compressed-tensors parser rejects or loses multiple config_groups.
3. Packed NVFP4 names/scales are not normalized for the Qwen mapper.
4. Dense MLP quant-config keys do not match the _DenseMlpAdapter runtime module path.
5. The explicit FP8 lm_head entry is dropped, which would cast the FP8 weight to bf16 and discard its scale.

Validation

With focused fixes for those concerns, an SM121 full-model smoke loaded 23.91 GiB of model weights, selected native trtllm::nvfp4_gemm, constructed lm_head as FP8_PER_CHANNEL_PER_TOKEN, and generated coherent 64-token responses for two prompts. The same stack also generated coherently with the checkpoint MTP layer enabled.

The changes will be proposed as small dependent PRs so each concern can be reviewed independently.

[mihai-chiorean]

Draft PR stack:

1. [#17723][fix] Keep Qwen3.8 dense VLM on the generic model #17724 - correct dense Qwen3.6/3.8 VLM routing (independent)
2. [#17723][fix] Normalize packed compressed-tensors NVFP4 names #17726 - normalize packed compressed-tensors NVFP4 names (independent)
3. [#17723][feat] Parse multiple compressed-tensors config groups #17725 - parse multiple compressed-tensors config groups (independent generic foundation)
4. [#17723][fix] Re-path dense Qwen quantization entries #17727 - re-path dense Qwen quant entries (functionally consumes [#17723][feat] Parse multiple compressed-tensors config groups #17725 output)
5. [#17723][fix] Preserve rowwise FP8 Qwen lm_head weights #17728 - preserve rowwise FP8 Qwen lm_head (functionally consumes [#17723][feat] Parse multiple compressed-tensors config groups #17725 output)

The full five-PR stack has been smoke-tested on SM121 with unsloth/Qwen3.8-27B-NVFP4, with coherent generation both without MTP and with the checkpoint MTP layer.