Issues

• [BUG]: Sharded optimizer loading retains old momentum absent from the checkpoint

  hpcaitech/ColossalAI#6549 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: DistributedCAME corrupts optimizer moment buffers after updates

  hpcaitech/ColossalAI#6546 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: OneCycleLR silently ignores three_phase and epoch-based schedule options

  hpcaitech/ColossalAI#6544 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: MixedPrecisionOptimizer cannot step after adding a parameter group

  hpcaitech/ColossalAI#6541 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: LinearWarmupLR becomes negative after the configured training horizon

  hpcaitech/ColossalAI#6540 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: Optimizer closures run with autograd disabled in four no-grad step implementations

  hpcaitech/ColossalAI#6538 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: Adafactor never refreshes the parameter RMS used by default learning-rate scaling

  hpcaitech/ColossalAI#6536 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: Adafactor discards a configured manual learning rate and fails on step

  hpcaitech/ColossalAI#6534 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: Warmup and delay schedulers recompute the decay rate at the phase boundary

  hpcaitech/ColossalAI#6532 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: Sharded checkpoint loading silently ignores parameter shape errors

  hpcaitech/ColossalAI#6531 · Filter by author Afloat16 opened on Nov 12, 2026

• [BUG]: LR scheduler stops driving the optimizer after load_optimizer when a plugin casts it to a distributed optimizer

  bug
  hpcaitech/ColossalAI#6523 · Filter by author yuxuan-lou opened on Nov 11, 2026

• [BUG]: TorchFSDPPlugin writes no config.json in a sharded model checkpoint, so from_pretrained cannot load it

  hpcaitech/ColossalAI#6498 · Filter by author JazyJiang opened on Nov 3, 2026

• [BUG]: GeminiPlugin cannot load a sharded model checkpoint: GeminiDDP.named_parameters() rejects remove_duplicate

  hpcaitech/ColossalAI#6497 · Filter by author JazyJiang opened on Nov 3, 2026

• [BUG]: GeminiPlugin with gradient accumulation computes wrong gradients on more than one GPU

  hpcaitech/ColossalAI#6496 · Filter by author JazyJiang opened on Nov 3, 2026

• [BUG]: TorchFSDPPlugin cannot boost torch.optim.AdamW: AdamW.__init__() rejects decoupled_weight_decay

  hpcaitech/ColossalAI#6495 · Filter by author JazyJiang opened on Nov 3, 2026

• [BUG]: Booster(mixed_precision="bf16") fails with a TypeError: BF16MixedPrecision is an empty stub

  hpcaitech/ColossalAI#6493 · Filter by author JazyJiang opened on Nov 3, 2026

• [BUG]: TorchDDPPlugin with fp16: loading a sharded optimizer checkpoint fails, the optimizer is unwrapped only once

  hpcaitech/ColossalAI#6492 · Filter by author JazyJiang opened on Nov 3, 2026

• [BUG]: loading a sharded optimizer checkpoint fails on CPU: CpuAccelerator.synchronize raises

  hpcaitech/ColossalAI#6491 · Filter by author JazyJiang opened on Nov 3, 2026

• [BUG]: bf16 training does not resume exactly: fp32 master weights are rebuilt from bf16 weights on load

  hpcaitech/ColossalAI#6490 · Filter by author JazyJiang opened on Nov 3, 2026

• [test] Track Gemma 2C2 requirements for shared numerical validation

  hpcaitech/ColossalAI#6488 · Filter by author richardoo-707 opened on Nov 1, 2026
  2

• [BUG] Default EPMoELayer group is rejected by get_ep_size

  hpcaitech/ColossalAI#6485 · Filter by author richardoo-707 opened on Oct 27, 2026

• [FEATURE]: Improve the bug report issue template and add a regression template

  enhancement
  hpcaitech/ColossalAI#6483 · Filter by author JazyJiang opened on Oct 26, 2026

• [booster] TorchFSDPPlugin hardcodes torch.cuda.current_device(), breaking non-CUDA accelerators

  bugin-progressmodule: booster/plugin
  hpcaitech/ColossalAI#6480 · Filter by author JazyJiang opened on Oct 26, 2026
  1

• [BUG] [Security] Automated static analysis findings for hpcaitech/ColossalAI (academic research)

  bug
  hpcaitech/ColossalAI#6475 · Filter by author fcas opened on Oct 21, 2026

• [FEATURE]: Support Transformers 5.17 in ShardFormer

  enhancement
  hpcaitech/ColossalAI#6473 · Filter by author Icarus1411 opened on Oct 21, 2026