QUANT-QWEN38-27B-NVFP4-ARM: qwen36 compressed-tensors mixed-precision FP8 group_0 is refused, blocking an apples-to-apples vllm.cpp vs vLLM comparison

Blocker

/data/models/Qwen/Qwen3.6-27B-NVFP4 (the checkpoint nsk/vLLM serves) and its
-vllmcpp sibling declare quantization_config = compressed-tensors,
format: "mixed-precision", with two groups:

• group_0: FP8 float-quantized, per-output-channel weight_scale and
  DYNAMIC per-token activation quantization (input_activations.dynamic: true, strategy: token).
• group_1: NVFP4 nvfp4-pack-quantized.

Current main loads group_1 but refuses group_0 in
qwen3_5_dense_weights.cpp via compressed_tensors::RefusalForHfConfigRaw
(because LoadFp8Raw cannot represent a per-channel scale + dynamic activation).
So vllm-server exits fatal before reading a weight byte.

Why it matters

The vllm.cpp vs vLLM (nsk) performance comparison is currently cross-model
(vllm.cpp runs the AEON MTP checkpoint, vLLM runs qwen36). To make it
apples-to-apples, vllm.cpp must load the SAME qwen36 checkpoint vLLM serves.

Goal

Implement the compressed-tensors FP8 float-quantized per-channel + dynamic
activation arm so qwen36 loads, then re-run the identical evalscope random
matrix on both engines over the same checkpoint.

Owning row

QUANT-QWEN38-27B-NVFP4-ARM (the loader's compressed-tensors mixed-precision
resolution; spec .agents/specs/qwen38-27b-quant-arms.md)

Start in qwen3_5_dense_weights.cpp at compressed_tensors::RefusalForHfConfigRaw and inspect how LoadFp8Raw handles the checkpoint configuration. Read .agents/specs/qwen38-27b-quant-arms.md and compare the group_0 FP8 settings with the existing mixed-precision path. Done means the same qwen36 checkpoint loads in vllm-server and the identical evalscope random matrix can be rerun against vLLM.