Issues

• [AI][Performance][KunlunXin] Qwen3.5-VL 2B forward is 11.6× slower: SDPA math kernel 48%, addmm in cpu_fallback 24%

  flagos-ai/Torch-FL#668 · Filter by author tonyh168 opened on Oct 27, 2026

• [AI][Performance][Iluvatar] Qwen3.5-VL 2B forward is 31× slower than vendor torch: SDPA never leaves the aten math kernel, 25% of time in cpu_fallback

  flagos-ai/Torch-FL#667 · Filter by author tonyh168 opened on Oct 27, 2026

• [AI][BUG][MetaX] torch.compile(backend="flagos") returns a wrong score for Qwen3.5-VL; sdpa path fails in fake-tensor propagation

  flagos-ai/Torch-FL#662 · Filter by author tonyh168 opened on Oct 26, 2026

• [AI][BUG][DCU] torch.compile with FlagGems: torch_device_fn stubs raise on the CPU-only torch

  flagos-ai/Torch-FL#661 · Filter by author tonyh168 opened on Oct 26, 2026

• [AI][BUG][DCU] Track FlagGems round rejecting strided tensors

  flagos-ai/Torch-FL#659 · Filter by author luojung opened on Oct 26, 2026

• [AI][BUG][DCU] Track FlagGems remainder errors for infinite divisors

  flagos-ai/Torch-FL#658 · Filter by author luojung opened on Oct 26, 2026

• [AI][BUG][DCU] Track FlagGems clamp/clamp_min losing NaN propagation

  flagos-ai/Torch-FL#657 · Filter by author luojung opened on Oct 26, 2026

• [AI][BUG][DCU] Track FlagGems rand/rand_like returning 1.0 in BF16/FP16

  flagos-ai/Torch-FL#656 · Filter by author luojung opened on Oct 26, 2026

• [AI][BUG][DCU] Boxing-only index_put/index_put_/_index_put_impl_ round-trip through CPU

  flagos-ai/Torch-FL#633 · Filter by author luojung opened on Oct 22, 2026
  1

• [AI][BUG][Iluvatar] Non-deterministic results across repeated identical forward passes (vision tower reproduces it standalone; NaN on some samples)

  flagos-ai/Torch-FL#617 · Filter by author tonyh168 opened on Oct 21, 2026
  1

• [AI] DTensor/FSDP2 all-gather & reduce-scatter fail over flagcx backend (missing coalesced virtuals)

  ai-generated
  flagos-ai/Torch-FL#607 · Filter by author xufeng102403 opened on Oct 21, 2026

• [AI][Performance][DCU] R5 host-dispatch patch silently skipped with pinned FlagGems: missing _descriptor_cache_key

  flagos-ai/Torch-FL#586 · Filter by author luojung opened on Oct 19, 2026

• [AI][DCU] Track native_layer_norm CPU-parity regression and boxing control for FlagGems #7247

  flagos-ai/Torch-FL#585 · Filter by author luojung opened on Oct 19, 2026

• [AI][Performance][DCU] R5 same-device DAS/FlagOS/boxing survey: 29 cases and classified follow-ups

  flagos-ai/Torch-FL#582 · Filter by author luojung opened on Oct 18, 2026

• [AI][Performance][DCU] Track FlagGems repeat_interleave output_size readback and R5 integration

  flagos-ai/Torch-FL#581 · Filter by author luojung opened on Oct 18, 2026

• [AI][Performance][DCU] Track FlagGems Hygon int64 sort multistage latency on R5

  flagos-ai/Torch-FL#580 · Filter by author luojung opened on Oct 18, 2026

• [AI][Performance][DCU] Integer indexing wrappers add aminmax and two host scalar readbacks

  flagos-ai/Torch-FL#579 · Filter by author luojung opened on Oct 18, 2026

• [AI] [TsingMicro] boxing-ON registration set is unasserted: test_conf_registration_consistency.py excludes tsingmicro

  ai-generated
  flagos-ai/Torch-FL#575 · Filter by author lvyufeng opened on Oct 17, 2026

• [AI][BUG][Kunlunxin] libxpytorch.so cannot bind torch 2.10: c10_cuda_check_implementation ABI signature break (…ib vs …jb), first flagos op hangs forever

  ai-generatedbugP1
  flagos-ai/Torch-FL#573 · Filter by author lvyufeng opened on Oct 16, 2026
  2

• [AI][BUG][TsingMicro] Host kernel driver is not loaded: no /dev nodes, module version mismatch (0.29842.5 vs V0.28791.2) — bert is unmeasurable on this chip

  ai-generatedbugP1
  flagos-ai/Torch-FL#572 · Filter by author lvyufeng opened on Oct 16, 2026

• [AI] redirect_gloo_requests leaves no way to get a real Gloo group: DCP checkpoint finalization hangs forever on flagos

  ai-generated
  flagos-ai/Torch-FL#562 · Filter by author lvyufeng opened on Oct 16, 2026

• [AI][BUG] cuda->flagos alias hardcodes device index 0: Tensor.cuda() and Module.to('cuda') ignore set_device() in multi-rank jobs

  ai-generated
  flagos-ai/Torch-FL#561 · Filter by author lvyufeng opened on Oct 16, 2026

• [AI][BUG][CUDA] Five flaggems_python override dispatch tests are unsatisfiable on a boxing-off wheel, and no CI lane runs them

  ai-generated
  flagos-ai/Torch-FL#555 · Filter by author lvyufeng opened on Oct 15, 2026

• [AI] [BUG][Kunlunxin] Two-device scatter/gather returns uninitialized rows: the cross-card D2D copy never reaches its destination

  ai-generated
  flagos-ai/Torch-FL#544 · Filter by author lvyufeng opened on Oct 15, 2026
  1

• [AI] DCP checkpoint save aborts on flagos: _OverlappingCpuLoader can't match a flagos tensor (_get_available_device_type answers cuda, same cause as #384)

  ai-generated
  flagos-ai/Torch-FL#533 · Filter by author lvyufeng opened on Oct 15, 2026