All issues

AutoTP loads module weights by hardcoded class name, so 219 of 233 transformers norm classes are silently left on meta

#8447

· alanhuangyoo opened

on Sep 6, 2026

ZeRO-3 applies Muon's Newton-Schulz once per micro-batch, so gradient accumulation changes the optimizer

#8443

· alanhuangyoo opened

on Sep 6, 2026

Muon runs no Newton-Schulz at ZeRO stage 0, which is the default: the plainest Muon config trains with SGD

#8441

· alanhuangyoo opened

on Sep 6, 2026

Default gradient_clipping divides every Muon update by its own norm, shrinking the step by a model-sized factor

#8439

· alanhuangyoo opened

on Sep 6, 2026

Muon under AutoTP orthogonalizes each shard, so the update depends on the TP degree

#8437

· alanhuangyoo opened

on Sep 6, 2026

[BUG] Muon + fp16 does not train under ZeRO 1/2: the first loss-scale overflow permanently poisons the momentum buffer

#8432

· alanhuangyoo opened

on Sep 6, 2026

[Per-Head Muon] Head geometry that lives outside num_attention_heads/head_dim: Kimi-K3 KDA layers and GLM-5.2's DSA indexer

#8420

· alanhuangyoo opened

on Sep 5, 2026

ZeRO skips the optimizer step on gradient overflow without logging anything

#8415

· sfc-gh-abkulkarni opened

on Sep 4, 2026

`_cast_module_mixed_precision` casts every dtype torch calls floating point, corrupting FP8/MX/NVFP4 parameters

#8414

· sfc-gh-abkulkarni opened

on Sep 4, 2026

[BUG] Curriculum data sampler checkpoints the global numpy RNG, not its own, so resume replays the sampling stream

#8405

· ebarkhordar opened

on Sep 3, 2026

ZeRO-3 HybridMegatron QKV layout mutations are not persisted

#8391

· LiRunGuo opened

on Sep 2, 2026

DeepCompile offload_activation skips saved views when the aliased base is not saved

#8387

· sfc-gh-truwase opened

on Sep 1, 2026