All pull requests

Pull requests

• Fix MiMo-V2.6 TP>4 serving: fused-fp8-QKV sharding, MTP loader, sm86 attention path

  #106·avtc opened Sep 23, 2026·

  ·
  3

• Warn when sliding-window KV capacity is pipeline-bound

  #104·kaka86mm opened Sep 23, 2026·

  ·

• Accept IEC suffixes (kiB/MiB/GiB/TiB) in human_readable_int

  #103·kaka86mm opened Sep 23, 2026·

  ·

• build(deps): bump the minor-update group across 1 directory with 192 updates

  dependencies
  #101·dependabot[bot] opened Sep 22, 2026Bot·

  ·

• Fixes GLM-5.3 (glm_moe_dsa, FP8 checkpoint) startup on Ampere. Refs #38.

  #100·noah80 opened Sep 22, 2026·

  ·

• feat(chat): support OpenAI "file" content parts by inlining documents as text

  #95·ming616 opened Sep 20, 2026·

  ·

• feat(qwen4exp): support fp8_e4m3 KV cache for QSA

  #90·me040526 opened Sep 17, 2026·

  ·
  1

  4

• Speculative-decoding verify attention: split-KV kernel for sm8x (+12% decode on the FA2 path)

  #82·Karl0007 opened Sep 14, 2026·

  ·
  2

• [Qwen4Exp][Bugfix] Slice the FlashInfer prefill query by the real (unpadded) token count

  #81·Karl0007 opened Sep 14, 2026·

  ·
  2