Bug: minimax_h3_ref2va_pruned_fp8_scaled.safetensors has corrupted comfy_quant metadata
- Dominant language
- Python
- Stars
- 133k
- Forks
- 15.7k
- Avg merge
- 1d 6h
- Merged PRs (30d)
- 155
Description
Model: https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_ref2va_pruned_fp8_scaled.safetensorsIssue: 42 out of 200 `comfy_quant` tensors contain zero values instead of valid JSON metadata, causing loading to fail with:```UnicodeDecodeError: 'utf-32-be' codec can't decode bytes in position 24-26: truncated data```at `comfy/ops.py:1138` when parsing `layer_conf = json.loads(layer_conf.numpy().tobytes())`.Affected layers (comfy_quant keys with zero values):- blocks.7.mlp.fc1/fc2- blocks.8.attn.out_proj/qkv_proj, mlp.fc1/fc2- blocks.9.attn.out_proj/qkv_proj, mlp.fc1/fc2- blocks.32.mlp.fc1/fc2- blocks.33-36.attn and mlp layers- blocks.43-47.attn and mlp layersVerification:- File size matches: 20,958,205,608 bytes- MD5: `3dc7a1e3ed880630a841c56a514af11e` (matches what's served from HF)- fl2va fp8_scaled works fine (200/200 comfy_quant valid)- ref2va int8_convrot works fine (200/200 comfy_quant valid)Workaround: Use `minimax_h3_ref2va_pruned_int8_convrot.safetensors` instead.
Contributor guide
Research direction
Start at comfy/ops.py:1138 and reproduce loading the linked Hugging Face safetensors file, comparing its comfy_quant metadata with the working fl2va and int8_convrot files. Done means the affected model has valid JSON metadata for all 200 tensors and loads without the reported UnicodeDecodeError, or the artifact is confirmed as an upstream file issue.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Needs clarification
- Newbie friendliness
- 30/100