All pull requests

Pull requests

• Accept warmup_num_steps=0 in the warmup LR schedulers again

  #8804·vineethsaivs opened Oct 10, 2026Contributor·

  ·

• Use C++20 for XPU op builders on torch 2.12+

  #8803·jinyouzhi opened Oct 10, 2026Contributor·

  ·

• Warn when the Ulysses SP DataLoader is not sharded across the SP group

  #8802·varun-projects opened Oct 10, 2026·

  ·
  4

• Guard the deprecated tensor model parallel mpu fallbacks

  #8801·varun-projects opened Oct 10, 2026·

  ·
  1

  4

• [AutoEP] Plan expert transfers from the live runtime placement

  #8800·jinyouzhi opened Oct 10, 2026Contributor·

  ·

• Add KLXPU (KUNLUNXIN XPU) accelerator

  #8799·GuCarpenter opened Oct 10, 2026·

  ·
  9

• comms_logger: honor in-section enabled and bound per-op timer history

  #8798·hgdytzyy opened Oct 10, 2026·

  ·
  1

  8

• Skip accelerator sync for unprofiled comm ops

  #8795·guptaishaan opened Oct 10, 2026·

  ·
  1

  2

• Average each loss of a tuple in pipeline eval_batch

  #8793·vineethsaivs opened Oct 9, 2026Contributor·

  ·

• Add vocabulary-parallel fused linear cross entropy

  #8792·jinyouzhi opened Oct 9, 2026Contributor·4 tasks·

  ·

• Add vocabulary-parallel logprobs, entropy and KL ops

  #8791·jinyouzhi opened Oct 9, 2026Contributor·

  ·
  11

• Restore the data sampler's own RNG when resuming from a checkpoint

  #8788·Rakshit-gen opened Oct 9, 2026Contributor·

  ·
  1

  6

• perf(rollout): select generation and prefill buckets with a tunable cost model

  #8786·SuperGoodGame opened Oct 9, 2026·

  ·
  1

  24

• [Draft] Support HiFloat8 training for Dense Linear and AutoEP on NPU

  #8784·hazelduan opened Oct 9, 2026·

  ·
  6

• Fix numerical stability in Triton masked softmax

  #8782·arcusbuilds opened Oct 9, 2026·

  ·
  1

  5

• [Bug Fix] Fix CPUAdagrad updates with nonzero weight decay

  #8781·hwchen2017 opened Oct 9, 2026Contributor·

  ·
  6

• Keep single-rank barrier() free of torch's device_id warning

  #8779·usmanmateen opened Oct 9, 2026·

  ·
  1

  9

• [fix] NPU: RoPE correctness, inverse-frequency caching, and KV cache preallocation in the inference op

  #8776·VenusTZZ opened Oct 8, 2026Contributor·2 tasks done·

  ·
  1

• Fix ZenFlow CPU affinity on Slurm masks and multi-node runs

  #8774·Aravind-11 opened Oct 8, 2026·

  ·
  1

• Average sparse gradients by world size regardless of gradient_predivide_factor

  #8772·vineethsaivs opened Oct 7, 2026Contributor·

  ·

• Skip SuperOffload's eager CPU step on non-finite gradients

  #8771·vineethsaivs opened Oct 7, 2026Contributor·

  ·

• Add an opt-in route_norm flag for the legacy MoE gate

  #8769·Aravind-11 opened Oct 6, 2026·

  ·
  4

• Fix autotuner latency ranking and the FLOPS metric name

  #8767·Aravind-11 opened Oct 6, 2026·

  ·
  1

• Detect NEON when py-cpuinfo reports no CPU flags

  #8765·vineethsaivs opened Oct 6, 2026Contributor·

  ·

• Round to nearest in NPUQuantizer, matching the CUDA quantizer

  #8764·vineethsaivs opened Oct 6, 2026Contributor·

  ·
  2