NVFP4 support for Qwen3_5MoeExperts (fused MoE): quantizer registration + export serializer

[frischeDaten]

Summary

mtq.quantize silently skips the fused experts of transformers Qwen3_5MoeExperts (Qwen3.5/3.6-VL-MoE) — they're 3-D nn.Parameters (gate_up_proj [E,2I,H], down_proj [E,H,I]) with a per-expert F.linear loop, and there's no QuantModuleRegistry entry. Result: the ~50GB expert bulk stays bf16, and export_hf_checkpoint has no serializer for it (layer_utils.get_experts_list only handles per-expert nn.Linear, Mixtral/DBRX-style).

Proposed

1. Register a _QuantQwen3_5MoeExperts(QuantModule) (analogous to _QuantLlama4TextExperts but with F.linear/(out,in) layout — no transpose). Reference impl that works today as an external registration is included in the HF repo below.
2. Add a fused-3-D export path so export_hf_checkpoint emits w13/w2 NVFP4 (packed uint8 + fp8 block scale + fp32 per-shard global + input scale).

Workaround shipped

Custom registration + manual NVFP4 packing via NVFP4QTensor (global amax/(6·448), block amax/(6·global)), writing the vLLM-ready checkpoint directly (also avoids export_hf_checkpoint OOM on unified-memory boxes).

Artifacts

Full quantize.py + resulting model (67GB→22GB, vision intact): https://huggingface.co/frischeDaten/Qwen3.6-VL-35B-A3B-NVFP4-DGX-Spark-VisionSafe

[meenchen]

@frischeDaten can you try the main branch and see if that works for you?

[frischeDaten]

Thanks @meenchen — confirmed, main resolves the quantizer side. 🎉

register_fused_experts_on_the_fly / _QuantFusedExperts now auto-detects and wraps the fused Qwen3_5MoeExperts with no custom registration. Verified against the real transformers module (modelopt 0.46.0.dev193+gd984de379, transformers 5.14.1):

experts type BEFORE: Qwen3_5MoeExperts       | is QuantModule: False
experts type AFTER : QuantQwen3_5MoeExperts   | is QuantModule: True   # auto-detected & wrapped
gate_up per-expert weight quantizers: 8 (expect 8)
calibrated amax present on experts   : True (8/8)
quantized forward finite: True | mean rel delta vs fp: 0.0444          # weights actually quantized

The per-expert-index-from-storage-offset approach is also nicer than the whole-tensor workaround I'd been using locally — thanks for the thorough fix (and the Mixtral/DeepSeek/MiniMax/Jamba coverage).

One note on scope: I validated the detection + quantization path on CPU with INT8_DEFAULT_CFG, since NVFP4 fake-quant asserts amax must be a CUDA tensor and my GB10 was busy. The interception path is format-independent, so this confirms the experts are no longer skipped — but I haven't yet run a full NVFP4 export end-to-end (mtq.quantize → export_hf_checkpoint → load in vLLM) on the real 67B checkpoint.

Quick question before this can be closed: is the export side (the second half of this issue — serializing the fused experts into the standard NVFP4 checkpoint layout) also expected to work on main now, or is that still pending? I'm happy to run the full NVFP4 export on GB10 and report back.