f131882299
- eererdan/memory: MEMO-013 作废更正(7.14非Windows版号)、新增 MEMO-014/015/016 - eererdan/experience: EXPER-029/030 修正(H3本地部署+马赛克根因) - memory/eed + skill(hldp-context-protect等) 同步 - 注: 新家 origin token 失效暂推旧家 old-origin 保底
48 lines
2.3 KiB
Markdown
48 lines
2.3 KiB
Markdown
# MiniMax H3 Issue 报告（提给 Comfy-Org/ComfyUI）
|
||
|
||
> 提交地址: https://github.com/Comfy-Org/ComfyUI/issues/new
|
||
> 生成: 2026-08-05 · 耳耳蛋 EED-EXPER-029
|
||
|
||
---
|
||
|
||
## 标题
|
||
|
||
**MiniMax H3 produces pure noise/mosaic on AMD RX 7900 XTX (RDNA3/gfx1100) with official template — all quantization/backend combos affected**
|
||
|
||
## 正文
|
||
|
||
### Environment
|
||
- GPU: AMD Radeon RX 7900 XTX (gfx1100, RDNA3), 24GB VRAM
|
||
- RAM: 32GB + 106GB swap
|
||
- OS: Ubuntu 24.04, ROCm 7.2.4
|
||
- PyTorch: 2.13.0+rocm7.2
|
||
- ComfyUI: master @ 1868372d, comfy-kitchen 0.2.26, comfy-aimdo 0.4.11
|
||
- Launch args: `--disable-pinned-memory --enable-dynamic-vram --fast-disk --vram-headroom 3 --async-offload` (+ `ROCR_VISIBLE_DEVICES=0` to hide iGPU)
|
||
|
||
### Symptom
|
||
MiniMax H3 official T2V template (and every variant tried) generates pure noise / mosaic ("snow"). Video VAE zero-latent test decodes fine (smooth grey), model files verify complete (DiT 932 keys, TE 2054 keys), GPU runs at 100% + 92°C (it computes, but produces garbage). Different seeds give different noise (not fixed corruption).
|
||
|
||
### Verified / excluded (all still noise)
|
||
- TE: nvfp4_awq AND int8_convrot (both native & emulated quant paths)
|
||
- Sampler: res_multistep AND euler
|
||
- Resolution: 640x352 / 832x480 / 1344x768
|
||
- Backend: HIP (comfy-kitchen) AND eager-only (pure PyTorch)
|
||
- Memory: swap increased 56G→106G (rules out OOM for T2V)
|
||
- Attention: applied `low_precision_attention=False` to H3 DiT (issue #15263 fix) — still noise
|
||
- Model files: safetensors integrity OK
|
||
|
||
### Additional data
|
||
- I2V (first_frame) path: ComfyUI SIGABRT crash, /var/crash report VmPeak=66.9GB (separate OOM issue)
|
||
- HIP backend registers correctly: `registered HIP backend for gfx1100 (with WMMA)`
|
||
- Native ops: int8_tensorwise, convrot_w4a4 | emulated: nvfp4, fp8, mxfp8
|
||
|
||
### Hypothesis
|
||
RDNA3 (gfx1100) specific operator issue in the quantized DiT forward (convrot_w4a4 / int8 attention path). A user on AMD RX 7800 XT (RDNA4) + Windows ROCm 7.14 reports video generation works fine (#15283), suggesting newer ROCm / RDNA4 fixes it, while RDNA3 on ROCm 7.2.4 does not.
|
||
|
||
### Files used (official)
|
||
- minimax_h3_fl2va_pruned_int8_convrot.safetensors (21G)
|
||
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (15.7G)
|
||
- minimax_h3_video_vae_fp16 / minimax_h3_audio_vae_fp32
|
||
|
||
Happy to provide logs/sidecars or run specific tests if helpful.
|