Issues

• [FEATURE]: Adapt the Cosmos MoT sequence packer to the two-way attention layout

  hpcaitech/ColossalAI#6504 · Filter by author bycainiu opened on Nov 3, 2026

• [FEATURE]: Add two-way flat attention for packed AR/DM batches

  hpcaitech/ColossalAI#6503 · Filter by author bycainiu opened on Nov 3, 2026

• [FEATURE]: Add segmented attention with per-segment visibility masks and a float64 reference

  hpcaitech/ColossalAI#6502 · Filter by author bycainiu opened on Nov 3, 2026

• [FEATURE]: Serve dense, packed TND and BSH layouts through ColoAttention on the Ascend NPU backend

  hpcaitech/ColossalAI#6501 · Filter by author bycainiu opened on Nov 3, 2026

• [BUG]: NPU fusion attention ignores the causal mask, rejects GQA and mis-sizes the two-way K/V segment

  hpcaitech/ColossalAI#6500 · Filter by author bycainiu opened on Nov 3, 2026

• [BUG]: _dispatch_kernel loads the DAO fallback even when the call will not use it

  hpcaitech/ColossalAI#6499 · Filter by author bycainiu opened on Nov 3, 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

• [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

• [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

• Quick survey: your experience as a PyTorch Ecosystem maintainer

  hpcaitech/ColossalAI#6470 · Filter by author michellearoth opened on Oct 16, 2026

• [BUG]: pipeline sequence classification picks the wrong token for left-padded inputs (Llama, Qwen2, OPT)

  hpcaitech/ColossalAI#6457 · Filter by author Arthur031221 opened on Sep 29, 2026

• fx tracer / rope init hard-code the device, so NPU runs fall back to CPU (or crash)

  hpcaitech/ColossalAI#6454 · Filter by author li-lizhe opened on Sep 26, 2026

• [BUG]: FP8 intranode detection ignores OpenMPI and Slurm local size values

  hpcaitech/ColossalAI#6437 · Filter by author tandede opened on Aug 20, 2026

• [BUG]: Q/K cross-wiring in fused_rotary_emb Triton kernel

  bug
  hpcaitech/ColossalAI#6428 · Filter by author truong-v opened on Jul 20, 2026