All pull requests

Pull requests

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

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

  ·

• 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·

  ·

• Fix flops profiler FLOPS and samples/s under gradient accumulation

  #8763·vineethsaivs opened Oct 6, 2026Contributor·

  ·

• [AutoSP] Rebuild query-sharded attention masks before SDPA

  #8761·jinyouzhi opened Oct 6, 2026Contributor·

  ·
  3

• [AutoSP] Fix tensor truthiness in get_node_shape_meta

  #8760·jinyouzhi opened Oct 6, 2026Contributor·

  ·

• Add an opt-in attention-output stash for Hugging Face reentrant checkpointing

  #8759·yh0903 opened Oct 6, 2026Contributor·

  ·

• Fix DataAnalyzer.run_map crashing when specific_threads is set

  #8758·Rakshit-gen opened Oct 6, 2026Contributor·

  ·
  1

• Reject an --exclude that leaves no slots to launch on

  #8755·Rakshit-gen opened Oct 5, 2026Contributor·

  ·
  1

• Record the initial lr in OneCycle so get_last_lr works before step

  #8753·Rakshit-gen opened Oct 5, 2026Contributor·

  ·
  1

• Keep each param group's beta2 in OneCycle

  #8751·Rakshit-gen opened Oct 5, 2026Contributor·

  ·
  1

  1

• [bug fix] ZeRO2: preserve fp32 grad accumulation precision in partition storage

  #8749·stas00 opened Oct 5, 2026Collaborator·

  ·
  9

• AutoEP: opt-in routing weights inside the expert activation

  #8748·yh0903 opened Oct 5, 2026Contributor·

  ·

• AutoEP: do not slice DeepEP buffers that already hold exactly the arrived rows

  #8747·yh0903 opened Oct 5, 2026Contributor·

  ·

• AutoEP: opt-in skip of the recomputed DeepEP combine

  #8745·yh0903 opened Oct 5, 2026Contributor·

  ·

• Add a pr-checklist agent skill and an AI contribution policy

  #8744·PKUWZP opened Oct 5, 2026Collaborator·

  ·
  19

• Return early from inference norm launchers when rows == 0

  #8743·arcusbuilds opened Oct 4, 2026·

  ·
  1

• Fix: [BUG] In step3, a runtime error will be thrown when inference_tp_size>

  #8740·goransh-walia opened Oct 4, 2026·

  ·

• Overlap AutoEP's expert weight gradients with the DeepEP backward combine

  #8739·yh0903 opened Oct 4, 2026Contributor·

  ·
  1

• Assert offload-state removal only when offload leaves the accelerator

  #8738·delock opened Oct 4, 2026Collaborator·

  ·

• Avoid circuler import caused by Hybrid Engine and OPT model

  #8737·tohtana opened Oct 3, 2026Collaborator·

  ·
  1

• Fix HybridEngine generation with inference tensor parallelism

  #8736·1fanwang opened Oct 3, 2026·1 task·

  ·
  1

• Count every pipeline stage in the fp32 gradient-clipping norm

  #8735·vineethsaivs opened Oct 3, 2026Contributor·

  ·

• Add an opt-in fused gate+up grouped GEMM for AutoEP experts

  #8733·yh0903 opened Oct 3, 2026Contributor·

  ·