All pull requests

Pull requests

• feat(qwen4): add Qwen3.8 Flash-Next support and shared prefill

  #774·fivetide opened Sep 23, 2026·

  ·

• ENH(qwen35): MTP speculative decode on dense TP through the shared draft and accept path

  #769·alpineQ opened Sep 16, 2026·

  ·

• MQ4-Lloyd (qt=52): per-tensor Lloyd codebook in the MQ4V2 container, -15% KLD at 4 bpw; mq4-pro prefill fix; AWQ rmsnorm rewrite

  ratchet-raise
  #768·Kaden-Schutt opened Sep 15, 2026·

  ·

• fix(mtp): chunk scalar Q8 batched GEMM above 64 rows

  #767·HUSRCF opened Sep 15, 2026·

  ·
  3

• perf(gfx1201): split packed-Q8 verifier KV scans — +22.6% DFlash E2E at 21.5K context

  #760·HUSRCF opened Sep 14, 2026·

  ·

• perf(dflash): enable validated S1-S6 launch fusions on gfx1201

  #759·HUSRCF opened Sep 14, 2026·

  ·

• perf(gfx1151): validate R4/R8 Q8 KV-sharing for long-context DFlash

  #756·HUSRCF opened Sep 14, 2026·

  ·
  3

• refactor: execute sealed MoE through shared steps

  ratchet-raise
  #755·fivetide opened Sep 12, 2026·

  ·
  3

• Escha-W2 port: 2-bit trellis decoded in the GEMV — Qwen3.6-35B-A3B and Qwen3.8-27B

  meta:wip
  #694·nwoolmer opened Sep 4, 2026·

  ·
  2

• feat(dispatch): seal collective-aware MoE step execution

  #677·fivetide opened Sep 1, 2026·

  ·

• fix(vmm): retry access per physical mapping on Windows

  #643·HUSRCF opened Aug 26, 2026·

  ·
  1

• Fixed character encoding for arch-specific generate loops.

  #595·fgheorghe opened Aug 16, 2026·

  ·
  1

• Fix tool calls being silently dropped on DeepSeek V4 and Cohere2MoE

  #593·fgheorghe opened Aug 16, 2026·

  ·
  1

• feat(ds4): add DeepSeek-V4-Flash-0731 — routed-expert paging, +28.8% cache budget, and an experimental prefetch adapter

  meta:wip
  #563·nwoolmer opened Aug 3, 2026·

  ·
  6

• [Superseded] refactor: unify device mesh, model parallelism, and generation dispatch

  #527·fivetide opened Jul 12, 2026·

  ·
  7