[Diffusion] Load serialized Comfy W4A8 checkpoints

[mickqian]

Summary

• auto-detect Comfy asym_w4a8_int8 metadata from safetensors headers
• validate packed INT4 weights, FP8 group scales, FP32 channel scales, codebooks, and optional correction tensors before model construction
• load serialized W4A8 linears through Comfy Kitchen without requantization or an explicit --quantization flag
• preserve TP/offload support when ConvRot group boundaries remain valid, and reject FSDP consistently with other Comfy layer-marked formats
• document the MiniMax-H3 FL2VA/Ref2VA DiT override flow

This is checkpoint-driven only: users select the W4A8 file with --transformer-weights-path; --quantization remains reserved for online quantization of unquantized weights.

Validation

• pre-commit run --all-files
• added focused unit coverage for global Comfy metadata resolution and packed W4A8 parameter construction
• inspected the remote safetensors headers for:
  • starsfriday/MiniMax-H3-w4a8 FL2VA (200 W4A8 layers)
  • Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI FL2VA and Ref2VA (200 W4A8 layers each)
• all three headers match the implemented weight/scale/codebook contract

Local unit tests were not run; they are left to CI per the diffusion development workflow.

Size

• existing production files: +114 / -2
• new production files: 281 lines
• tests and docs: 91 lines

---

CI States

Latest PR Test (Base): ✅ Run #32680535211
Latest PR Test (Extra): ❌ Run #32680535066
Latest PR Test (AMD ROCm 7.2): ⏳ Run #32680535183

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
lmsysorg	🟢 Ready	View Preview	Aug 23, 2026, 5:47 AM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[mickqian]

/tag-and-rerun-ci