feat(qwen3_5_moe): support compressed-tensors NVFP4 experts with block-fp8 dense side

[JIAQI13]

Fixes #263
Also addresses #236 (uncensored/abliterated Qwen3.6-35B-A3B NVFP4 support). Related: #165, #252, #274.

Problem

Community abliterated/uncensored Qwen3.6-35B-A3B NVFP4 checkpoints exported with llm-compressor (quant_method: compressed-tensors, format: mixed-precision) cannot be converted:

• routed experts: NVFP4 stored as weight_packed (uint8) + weight_scale (fp8-e4m3 block) + scalar weight_global_scale
• dense side (attn/GDN + shared expert): 128×128 block FP8 (.weight_scale, W8A8)

FreeToken only recognized the nvidia-modelopt NVFP4 expert layout (weight / weight_scale / weight_scale_2), so:

• the expert bank spec matched zero layers (the ValueError: Missing MoE expert source layers reported in [Bug] ft checkpoint fails on Qwen3.6-35B-A3B abliterated NVFP4 checkpoints: Missing MoE expert source layers #263 on v0.2.0-beta.14), and
• on current main, the dense block-FP8 weights hit the per-tensor FP8 dequant path and crash with RuntimeError: The size of tensor a (2048) must match the size of tensor b (16) (a 128×128 block scale fed to the per-tensor dequant).

The official nvidia/Qwen3.6-35B-A3B-NVFP4 (modelopt layout) and block-FP8 checkpoints are unaffected.

Changes

• models/nvfp4_banks.py: add load_nvfp4_ct_expert_source_banks / ..._parallel — a compressed-tensors NVFP4 expert source builder for the offload banks. Parses weight_packed / weight_scale / weight_global_scale (the stored global is the quant-side scale; the dequant global is its reciprocal, broadcast per output row — same convention vLLM uses), and builds the single-file safetensors weight map from the safetensors header when no model.safetensors.index.json is present.
• models/qwen3_5_moe/config.py: add _compressed_tensors_mixed_fp8block() detection (8-bit block group + 4-bit group_size=16 tensor_group group in config_groups); when matched, route experts to NVFP4 banks and the dense side (attn/GDN/shared expert) to native 128×128 block FP8 (attn_quant/dense_quant = "fp8_block").
• models/qwen3_5_moe/weight.py: register the CT expert key pattern/source spec; route the dense pass of these checkpoints through the existing block-FP8 loader (_iter_weights_fp8), renaming the block scale suffix .weight_scale → .weight_scale_inv and skipping dynamic-activation scales.
• models/qwen3_5_moe/moe.py, gdn.py: enable the block-FP8 modules for the shared expert and GDN when dense_quant/attn_quant == "fp8_block" (previously gated on the experts being block-FP8).
• tests/models/test_qwen3_5_moe_ct_mixed_detect.py: detection tests — mixed CT flagged, pure-CT-NVFP4 / per-channel-FP8 / modelopt mixed / no-quant not flagged.

Pure modelopt NVFP4, block-FP8, dense NVFP4, and pure compressed-tensors NVFP4 paths are unchanged.

Tested on (real hardware)

• GPU: RTX 3080 Laptop, 16 GB VRAM (sm_86), driver 610.57.04; CPU: Intel i7-10870H, 32 GB RAM + swap; CachyOS
• FreeToken 0.1.2, torch 2.11.0+cu130
• Checkpoint: kyaky/Qwen3.6-35B-A3B-Uncensored-NVFP4 (compressed-tensors mixed-precision, single 23 GB model.safetensors)
• A/B with the same command
  ft checkpoint --model <HF dir> --out <out> --dtype bfloat16 --moe-backend offload --shard-gib 8:
  • main: conversion fails (dense block-FP8 dequant RuntimeError: tensor size 2048 vs 16; [Bug] ft checkpoint fails on Qwen3.6-35B-A3B abliterated NVFP4 checkpoints: Missing MoE expert source layers #263 reports the equivalent bank-build ValueError: Missing MoE expert source layers on the Desktop build)
  • this branch: converts in 33 s — 20.19 GiB FTW, 3 shards, 653 weight + 240 experts_bank tensors
• ft serve --model-path <FTW> --moe-backend offload --moe-cache-auto: loads via the FTW fast path (3.25 GiB dense + 16.9 GiB expert banks), API server is ready, /v1/chat/completions returns correct generations.
• Detection unit tests pass standalone (5/5, no weights required).

Scope note

This covers the compressed-tensors expert layout (also used by sakamakismile/Huihui-...-abliterated-NVFP4 in #263). The other failing checkpoint in that report, joshebbs/qwen3.6-35b-abliterated-nvfp4-modelopt, uses an unusual doubled key prefix (language_model.model.layers.N.mlp.experts.*) that neither the existing modelopt regex nor this change matches — that looks like a separate export bug and is not handled here.