Issues

• --allgather-cp gives wrong log-probs and NaN grads on stock Qwen3 attention

  radixark/miles#4469 · Filter by author HJSang opened on Oct 27, 2026

• Load DeepSeek-V4.1-Flash's FP8/FP4 release directly instead of converting it to BF16 offline

  radixark/miles#4445 · Filter by author micahtyong opened on Oct 26, 2026

• [RFC] R3: keep first-computed MoE routes when a session re-prefills

  radixark/miles#4341 · Filter by author syeehyn opened on Oct 24, 2026

• Fixes from running fully-async agentic RL with miles on AMD

  radixark/miles#4271 · Filter by author Arist12 opened on Oct 21, 2026

• FSDP2 model offload corrupts uneven shards despite final synchronization

  radixark/miles#4238 · Filter by author Palind-Rome opened on Oct 21, 2026

• [Roadmap] Miles AMD Q4 Development

  radixark/miles#4206 · Filter by author XinyuJiangCMU opened on Oct 19, 2026

• Score centering: torch backend's tail ratio drifts from 1 under filtered sampling

  radixark/miles#4204 · Filter by author HJSang opened on Oct 19, 2026
  1

• [Bug] Installed SGLang parser plugins are not loaded by the launcher, engine-argument rendering or rollout workers

  radixark/miles#4163 · Filter by author mmbajo opened on Oct 18, 2026

• Quick survey: your experience as a PyTorch Ecosystem maintainer

  radixark/miles#4139 · Filter by author michellearoth opened on Oct 16, 2026

• rollout /generate client: httpx pool maintenance stalls the event loop at hundreds of in-flight requests

  radixark/miles#4134 · Filter by author yyming009 opened on Oct 16, 2026
  1

• CI: stage-a-cpu HF downloads fail with 403 "credentials expired"

  radixark/miles#4115 · Filter by author yueming-yuan opened on Oct 15, 2026

• [RFC] Log-probs from the LM-head input, without materialising the logits

  radixark/miles#4108 · Filter by author omri-manolo opened on Oct 15, 2026

• Step-1 training hang (multi-node, Qwen3.5-35B-A3B, TP2/PP2/EP4 + MTP + full recompute): a few ranks stuck in Triton autotune / Inductor compile / MoE dispatcher D2H sync while all peers wait in the DP grad reduce-scatter

  radixark/miles#4046 · Filter by author yyming009 opened on Oct 13, 2026
  1

• [Question] DeepSeek-V4-Flash-0731: long/repetitive thinking after SFT and in an RL QAT baseline

  radixark/miles#4024 · Filter by author li126com opened on Oct 12, 2026
  1

• [Bug] Uniform fractional rewards can produce nonzero GRPO advantages through float32 rounding

  radixark/miles#4019 · Filter by author zhoutong-hai opened on Oct 11, 2026

• [Bug] Colocated update_weights segfaults on every rank: uneven ExpertGather all_gather hits NCCL 2.29/2.30 AllGatherV with a tuner plugin

  radixark/miles#4003 · Filter by author luohao opened on Oct 10, 2026

• [Perf] [Bug] Bump Transformer Engine version to 2.18 for THD+A2A context parallelism performance

  radixark/miles#3994 · Filter by author charliechenye opened on Oct 9, 2026

• [RFC] Honor FP8 modules_to_not_convert during Megatron weight sync

  radixark/miles#3979 · Filter by author leviking98z-rgb opened on Oct 8, 2026

• Session server and /samples ingest performance (tracking)

  radixark/miles#3976 · Filter by author guapisolo opened on Oct 8, 2026
  

• [Roadmap] Sparse attention TopK

  radixark/miles#3970 · Filter by author zianglih opened on Oct 7, 2026

• [Feature] YAML-configured multi-dataset mixing for RL training with per-step prompt ratios

  radixark/miles#3947 · Filter by author Shi-Dong opened on Oct 6, 2026
  1

• [DeepSeek-V4.1] Clarify official RL/SFT support on Hopper GPUs and the recommended x86 image

  radixark/miles#3862 · Filter by author li126com opened on Sep 29, 2026
  2

• [Tracker] NVFP4 weight update enhancements

  radixark/miles#3734 · Filter by author zianglih opened on Sep 27, 2026

• Deterministic TP2 serving can fail NCCL weight transfer with mismatched channel limits

  radixark/miles#3714 · Filter by author ashtonchew opened on Sep 25, 2026

• Bridge drops SwiGLU and GELU activation fusion flags

  radixark/miles#3713 · Filter by author ashtonchew opened on Sep 25, 2026