HuMo 17B, MXFP8 for ComfyUI
This is an MXFP8 build of HuMo 17B, ByteDance's Apache-2.0 model, for ComfyUI's mixed-precision ops. It was made from Comfy-Org's fp16 repack (Comfy-Org/HuMo, humo_17B_fp16.safetensors) by Crackio. Crackio's Lip Sync uses it for its "Best quality" mode.
What was done
- Quantized: every Linear weight of the 40 transformer blocks, 560 layers in all. They are MXFP8: e4m3 values with one e8m0 scale per 32 weights, made with ComfyUI's own quantizer (comfy_kitchen).
- Kept exactly as in the fp16 source: everything else, meaning biases, norms, modulation, embeddings, the head and the audio projection.
- Size: 17.9 GB.
- Weight error: the quantized weights' relative RMS error against fp16 is about 2.7%. A plain e4m3 cast of the same weights is 3.3–3.7%.
- How it differs from Comfy-Org's file:
humo_17B_fp8_e4m3fncasts every tensor to e4m3 with no scales, biases and norms included. Some of its biases lose 20–99% of their value. - Script: the conversion script is
to_mxfp8_stream.py. It streams the fp16 file from Hugging Face tensor by tensor and needs a ComfyUI with comfy_kitchen MXFP8.
Use
Put the file in ComfyUI/models/diffusion_models/ and load it with the standard Load Diffusion Model node (weight_dtype: default). ComfyUI reads the per-layer comfy_quant metadata itself, which needs a recent ComfyUI with comfy_kitchen.
- RTX 50 series (Blackwell): the tensor cores run MXFP8 natively. In Crackio's Lip Sync tests it drew 1.7× faster than
humo_17B_fp8_e4m3fn, with the same pictures (two faces, fixed inputs). - Other GPUs: ComfyUI unpacks the weights as it runs.
Licence
Apache-2.0, like HuMo itself and Comfy-Org's repack. These are modified (quantized) copies of ByteDance's HuMo 17B weights.
Checksum
e8901000e61c89197aea5c8c76afceb8415293440d305bcba3aaa212d15726ff humo_17B_mxfp8_crackio.safetensors