Qwen3.6-27B NVFP4 checkpoints fail because dense loader requires BF16 `lm_head`

[rohitpaul]

Problem

The Qwen3.6 dense loader fails to load two common 27B NVFP4 checkpoints:

• nvidia/Qwen3.6-27B-NVFP4
• unsloth/Qwen3.6-27B-NVFP4

Both fail with:

server: fatal: vt: dense loader: expected BF16 for lm_head.weight
at include/vllm/model_executor/models/dense_weight_loaders.h:83

Cause

LoadQwen3_5Dense() always loads an explicit lm_head.weight using the BF16-only LoadBf16Transposed() path.

However, these checkpoints use different quantized output heads:

• NVIDIA: ModelOpt NVFP4, U8 [248320, 2560], with weight_scale and weight_scale_2
• Unsloth: FP8, F8_E4M3 [248320, 5120], with weight_scale

Model revisions tested:

• NVIDIA: 0893e1606ff3d5f97a441f405d5fc541a6bdf404
• Unsloth: ccdaab7e68af2409599b8949a8f2685703c9bae5

Environment

• RTX 5090 (sm_120a)
• CUDA 13.0.2
• vllm.cpp commit 0e3bf3c0

Expected behavior

Support the quantized lm_head formats used by these NVFP4 checkpoints, or document the exact supported checkpoint/revision and return a clearer unsupported-format error.