drozbay/Wan2.1-HuMo-14B-int8-convrot

Wan2.1 HuMo 14B — INT8 + ConvRot (ComfyUI)

INT8 quantization of HuMo 14B (human-centric video generation built on Wan2.1-T2V-14B) for ComfyUI's native int8_tensorwise quantized-weight path (comfy-kitchen). 17.35 GB vs 34.1 GB fp16.

Why INT8 + ConvRot instead of fp8

ConvRot applies a QuaRot-style block-Hadamard rotation (groupsize 256) to weights offline and to activations online, spreading outliers so per-channel INT8 stays accurate:

format	weight reconstruction error
int8 + ConvRot (this repo)	0.88% mean, 0.95% max
plain int8 per-channel	1–3%
fp8 e4m3	~2.6% (mantissa floor)

INT8 tensor-core GEMMs are also at least as fast as fp8 — in my testing this checkpoint runs noticeably faster than the fp8 equivalent in ComfyUI, with no visible quality loss vs fp16. The runtime picks CUTLASS or cuBLAS int8 kernels per-GPU automatically, so the same file works on Ampere/Ada/Blackwell.

What's quantized

All 560 per-token block linears across the 40 transformer blocks: self_attn.{q,k,v,o}, cross_attn.{q,k,v,o}, audio_cross_attn.{q,k,v,o}, ffn.{0,2} — int8 [N,K] weights with per-channel float32 [N,1] scales and an embedded per-layer comfy_quant config (format: int8_tensorwise, convrot: true, convrot_groupsize: 256).

Kept in fp16 (small-M / quality-sensitive): patch_embedding, text_embedding, time_embedding, time_projection, audio_proj.*, head.

Per-layer relative error: mean 0.878%, min 0.808%, max 0.952%; cosine similarity ≥ 0.99995 on every layer (verified against the fp16 source at conversion time).

Usage

Drop into ComfyUI/models/diffusion_models/ and load with the standard Load Diffusion Model node — quantization is detected from the embedded layer configs, no special loader needed. Requires a ComfyUI version with comfy-kitchen int8 support. Use exactly as you would the fp16 HuMo model (same workflows, audio conditioning, etc.).

Conversion

Converted with comfy-model-tools quant_int8_convrot.py: fp32 upcast → block-Hadamard rotation (via comfy_kitchen's own kernels, so stored weights match the runtime exactly) → per-channel absmax int8, with per-layer reconstruction verification.

Tools used:

• comfy-model-tools — the quantization script
• comfy-kitchen — Hadamard rotation / int8 kernels (the same code that runs at inference time)
• The conversion was planned, executed, and format-verified with the help of Claude (Claude Fable 5, via Claude Code)

Related quantization tooling: ComfyUI-QuantOps, convert_to_quant

Source: Wan2_1-HuMo-14B_fp16.safetensors (ComfyUI repack of HuMo 14B).

Credits

• HuMo — ByteDance / Tsinghua University
• Wan2.1 — Alibaba
• ConvRot int8 path in ComfyUI/comfy-kitchen; rotation approach based on QuaRot