MiniMax H3 full INT8/BF16 are unstable in ComfyUI; pruned is fine
还没有人认领这个 Issue。
评估
- 难度
- 3/5
- 预计耗时
- 1-2 天
- 新手友好度
- 74/100
- Issue 类型
- 缺陷
- 描述清晰度
- 描述清楚
- 活跃度
- 活跃
- 技术栈
- python
调研方向
Start in comfy/ldm/minimax/model.py and inspect the timestep/modulation path around the reported cast of t_emb. Reproduce with the full INT8 and BF16 checkpoints, comparing against the pruned checkpoint and checking frames at the reported 17-frame cadence. Done means the full checkpoints no longer show periodic brightness dips or flashing, while the pruned checkpoint remains unaffected.
由索引模型根据 Issue 内容生成。
描述
I found a reproducible issue with MiniMax H3 in ComfyUI.
Problem
These full checkpoints show periodic brightness dips / flashing / temporal instability:
- minimax_h3_fl2va_int8_convrot.safetensors
- minimax_h3_fl2va_bf16.safetensors
This checkpoint does not: - minimax_h3_fl2va_pruned_int8_convrot.safetensors
I also verified this is not the text encoder.
Cause
The issue appears related to this cast in comfy/ldm/minimax/model.py: t_emb = self.time_embedder(t_vals).to(dtype)
Changing it to:
t_emb = self.time_embedder(t_vals) fixed the issue locally for both full INT8 and full BF16.
Notes
I compared full vs pruned and found all 829 non-modulation tensors were byte-for-byte identical, so the problem appears to be in the modulation / AdaLN / time-embedding path.
The dips/spikes were periodic at a 17-frame cadence. At 24 fps, that is roughly every 0.708 seconds. In our affected runs, the artifact appeared at or very near frames 17, 34, 51, 68, 85, 102, 119, 136, 153, and 170. In the continuation workflow, the behavior also lined up with H3’s temporal structure: with a 39-frame context, the first visible dip appeared 12 newly generated frames later, which is exactly the next 17-frame boundary relative to the model’s 17n+5 temporal geometry. That regular cadence is a useful reproduction clue and should make it easy to inspect frame-by-frame.
Please investigate whether the full MiniMax H3 checkpoints are being downcast incorrectly in the timestep/modulation path.
- 主要语言
- Python
- 星标
- 9.8k
- 派生
- 738
- PR 合并指标
- 30 天内没有已合并 PR
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
MiniMax-AI/MiniMax-H3 的其他 Issue
-
loss of init.py未关闭
难度 3/5 半天 新手友好度 15/100
MiniMax-AI/MiniMax-H3#92 · 1 条评论 ·
-
难度 5/5 一周以上 新手友好度 5/100
MiniMax-AI/MiniMax-H3#89 ·
-
难度 5/5 一周以上 新手友好度 5/100
MiniMax-AI/MiniMax-H3#88 ·
-
،her未关闭
难度 5/5 一周以上 新手友好度 10/100
MiniMax-AI/MiniMax-H3#87 ·
-
难度 5/5 一周以上 新手友好度 25/100
MiniMax-AI/MiniMax-H3#86 ·
查看 MiniMax-AI/MiniMax-H3 的全部 Issue
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 72/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 76/100
omegaup/omegaup#10353 · 1 条评论 ·
维护者通常 5 天内回复
-
bot-triaged module: LrScheduler module: optimizer triaged
难度 2/5 1-3 小时 新手友好度 70/100
维护者通常 1 天内回复
-
documentation glm5 llm-model
难度 2/5 1-3 小时 新手友好度 78/100
vllm-project/vllm-ascend#19062 ·
维护者通常 1 天内回复
-
author-page correction
难度 2/5 1-3 小时 新手友好度 68/100
acl-org/acl-anthology#10496 · 1 条评论 ·
维护者通常 1 天内回复