All pull requests

Pull requests

• fix(kimi-k3): give the bf16 f_b_proj a real copy of f_a at one token

  #2626·gbyu-amd opened Oct 19, 2026Contributor·

  ·
  1

• fix(atomesh-ci): TW Spur weekly failures (Mesh cache, PENDING dispatch, amd-tw account)

  #2625·junyyang-amd opened Oct 19, 2026Contributor·

  ·
  1

• [WIP][feat]: bound dense MHA/GQA cached-prefill gather workspace

  #2624·MengqingCao opened Oct 19, 2026Contributor·

  ·
  1

• feat(offload): run lmcache_mp under multi-node DP with one loopback server per host

  #2621·zejunchen-zejun opened Oct 17, 2026Collaborator·

  ·
  1

• [Kimi-K3 DSpark] Pad the draft's block decode to 96 query heads on gfx1250

  #2620·zejunchen-zejun opened Oct 17, 2026Collaborator·

  ·
  1

• fix(atomesh): keep dashboard metrics finite without dropping zero

  #2619·kvnloo opened Oct 16, 2026Contributor·

  ·
  1

  

• docs(index): correct ATOM acronym

  #2618·peterjunpark opened Oct 16, 2026Contributor·1 task·

  ·
  1

  2

  

• [gfx1250] DeepSeek-V4.1: run on gfx1250

  #2617·yhl-amd opened Oct 16, 2026Contributor·4 of 5 tasks·

  ·
  1

• fix(moe): preserve MoRI v2 live-row contract

  #2616·XiaobingSuper opened Oct 16, 2026Contributor·

  ·
  2

• fix(blockscale): gate row-major group32 GEMM by arch

  #2615·XiaobingSuper opened Oct 16, 2026Contributor·

  ·
  2

• fix(minimax-m3): enable per-token FP8 KV on gfx1250

  #2614·XiaobingSuper opened Oct 16, 2026Contributor·

  ·
  3

• [FlyDSL] [gfx1250] Add opt-in MXFP4 linear path via AITER gemm_a4w4 with 32x4 scales

  #2613·aoli26 opened Oct 16, 2026Contributor·1 task done·

  ·
  1

• feat(qwen): add experimental Mono decode and TP8 serving reproduction

  #2610·yifehuan opened Oct 16, 2026Contributor·

  ·
  72

• (recipe) Add Qwen3.8-Flash-Next AgentX Recipe

  #2608·zhuyuhua-v opened Oct 16, 2026Collaborator·1 task·

  ·
  1

• fix(dsa): chunk the indexer decode logits into a shared workspace

  ci:atom
  #2607·HaonanWang98 opened Oct 16, 2026Contributor·1 task·

  ·
  18

  

• fix(qwen4_exp): load unscaled Quark MTP weights as BF16

  #2606·ZLkanyo009 opened Oct 16, 2026Contributor·2 of 3 tasks·

  ·
  1

  

• [Scheduler] Keep DP-attention ranks in one phase during prefill

  #2605·karverma-amd opened Oct 16, 2026·1 task·

  ·
  6

  

• fix(cudagraph): keep each decode graph's parallel stream off the launch queue

  #2604·nehmathe2 opened Oct 16, 2026·1 task·

  ·
  9

  

• Profiling graph capture of the draft runner

  #2601·mohbasit opened Oct 15, 2026Contributor·1 task done·

  ·
  1

  

• scheduler: add DP prefill-sync mode to avoid group-wide eager decode

  #2600·karverma-amd opened Oct 15, 2026·1 task·

  ·
  1

  

• DeepSeek-V4.1-Flash: LMCache tier offload through the MP server (vLLM plugin)

  #2598·PerryZhang01 opened Oct 15, 2026Contributor·

  ·
  2

  

• add asm mla qh64 kvsf

  #2597·feifei14119 opened Oct 15, 2026Contributor·1 task·

  ·
  6

  

• [DP] Decouple prefill cost balancing from PrefillDelayer, round costs up

  #2596·zejunchen-zejun opened Oct 15, 2026Collaborator·

  ·
  2

• fix(minimax-m3): enable gfx1250 and fix batched accuracy

  #2594·XiaobingSuper opened Oct 15, 2026Contributor·

  ·
  5

• [Recipe] Qwen3.5 AgentX on MI355X

  #2592·ZhiweiYan-96 opened Oct 15, 2026Contributor·

  ·
  3