All pull requests

Pull requests

• [AMD] Qwen4-Exp TP2: pipeline prefill all-reduces over SDMA and FP8/BF16

  #78·yicyang opened Oct 30, 2026Collaborator·

  ·

• feat(allreduce): Integrate aiter FP8 QR for prefill

  #77·apinge opened Oct 29, 2026Owner·5 tasks·

  ·

• [ROCm] Add decode swizzled GEMV and safe caller-owned outputs

  #76·TianHao65 opened Oct 28, 2026Collaborator·

  ·

• [Backport] Support CUDA IPC on ROCm for multimodal data in Qwen3.8 branch

  #75·sammysun0711 opened Oct 28, 2026Collaborator·5 tasks·

  ·
  7

• feat(multimodal): decode JPEG inputs with rocJPEG on ROCm

  #74·bobofang11235 opened Oct 28, 2026·

  ·
  4

• [ROCm] Add guarded swizzled GEMV and direct outputs for Qwen3.5 Decode

  #72·TianHao65 opened Oct 28, 2026Collaborator·

  ·

• feat(engine): opt-in early forkserver for scheduler processes

  #71·bobofang11235 opened Oct 27, 2026·

  ·

• feat(model_loader): chunked fastsafetensors loading with an all_local option

  #70·bobofang11235 opened Oct 27, 2026·

  ·
  1

• perf(eplb): cache rows of logical_to_all_physical_map_cpu

  #69·bobofang11235 opened Oct 27, 2026·

  ·

• fix(qwen4_exp): free replaced params after load and look up expert mappings by expert id

  #68·bobofang11235 opened Oct 27, 2026·

  ·

• Fix R-Fork weight loading on MI308X

  #67·yicyang opened Oct 27, 2026Collaborator·

  ·
  13

• feat(ssm): support non-MTP ReplaySSM with radix cache on/off

  #66·apinge opened Oct 26, 2026Owner·

  ·

• [CI] Add Unified Qwen3.5 & Qwen3.8 CI

  #65·sammysun0711 opened Oct 25, 2026Collaborator·5 tasks·

  ·
  6

• [ROCm] Port PR #8 hipBLASLt linears to qwen4_v0.5.21

  #63·TianHao65 opened Oct 22, 2026Collaborator·

  ·
  2

• feat(gdn) Add opt-in AITER FlyDSL GDN prefill on gfx942 for Qwen3.8 Flash Next

  #62·apinge opened Oct 22, 2026Owner·5 tasks·

  ·
  32

• [GDN]Tune gfx942 GDN prefill and chain verification

  #52·sammysun0711 opened Oct 20, 2026Collaborator·5 tasks·

  ·
  1

• perf(rocm): separate QSA KV preparation and use packed PR40 attention

  #44·TianHao65 opened Oct 13, 2026Collaborator·5 tasks done·

  ·
  5

  

• perf(rocm): accelerate Qwen3.8 QSA decode path

  #38·TianHao65 opened Sep 24, 2026Collaborator·

  ·

• bugfix(MTP): Fix HIP non-greedy EAGLE verification for MTP

  #36·apinge opened Sep 18, 2026Owner·5 tasks·

  ·
  1

• feat(hc): split the HC mix kernel to drop its grid barriers

  #35·yicyang opened Sep 18, 2026Collaborator·5 tasks·

  ·
  1

• feat(qsa): give the sparse-GQA decode split-KV parallelism on gfx942

  #34·yicyang opened Sep 18, 2026Collaborator·5 tasks·

  ·
  8

• feat(qsa): add a Triton paged decode MQA kernel

  #33·yicyang opened Sep 18, 2026Collaborator·5 tasks·

  ·
  2

• Add GR Read token parallelism for prefill

  #29·apinge opened Sep 13, 2026Owner·5 tasks·

  ·

• Support Qwen3.8 Quark MXFP4 Model on MI308X

  #28·sammysun0711 opened Sep 3, 2026Collaborator·5 tasks·

  ·
  4

• Support vectorized_5d kv cache for HybridLinearKVPool in Qwen3.5

  #24·apinge opened Aug 22, 2026Owner·5 tasks·

  ·
  5