All pull requests

Pull requests

• [SpecDecode][EXL3] deepseek_mtp: normalize rank-sliced weight names (fixes MTP trellis overlay loading)

  #11·malaiwah opened Jul 25, 2026·

  ·
  2

• [Local] Carry current Triton MLA tuning

  #10·voipmonitor opened Apr 22, 2026Owner·

  ·

• [Local] Add PCIe custom allreduce opt-in

  #9·voipmonitor opened Apr 22, 2026Owner·

  ·

• [Local] Add seq-length gate for speculative decode

  #8·voipmonitor opened Apr 22, 2026Owner·

  ·

• [SpecDecode] Add local argmax helper for Llama Eagle3

  #7·voipmonitor opened Apr 22, 2026Owner·

  ·

• [SpecDecode] Add local argmax helpers for Qwen3 drafts

  #6·voipmonitor opened Apr 22, 2026Owner·

  ·

• [SpecDecode] Allow draft-specific attention backend and KV dtype

  #5·voipmonitor opened Apr 22, 2026Owner·

  ·

• [SpecDecode] Fix async proposer synchronization

  #4·voipmonitor opened Apr 22, 2026Owner·

  ·

• [Core] Enable FP8 KV cache with DCP for MLA

  #3·voipmonitor opened Apr 22, 2026Owner·

  ·

• Tracking: Kimi MLA upstream split simulation

  #2·voipmonitor opened Apr 22, 2026Owner·

  ·
  2

• [MTP][Runtime] Reuse draft attention metadata across MTP steps

  #1·voipmonitor opened Apr 12, 2026Owner·

  ·
  1