All pull requests

Pull requests

• [Core][Megakernel] Move megakernel_runtime into megakernel/adapter

  #2961·Nanfang-Li opened Nov 8, 2026Collaborator·7 of 13 tasks·

  ·

• [Perf][Kernel][DeepSeek-V4.1] Window the FP8 candidate metadata past the SC VMEM budget and skip unused top-k in sampling

  #2960·johy0330 opened Nov 8, 2026Collaborator·6 of 13 tasks·

  ·

• [Bugfix][DeepSeek-V4.1] Add two-phase entry/exit barriers to push_fused_moe cross-chip collectives

  ready
  #2959·weikuo0506 opened Nov 8, 2026Collaborator·

  ·

• [Bugfix][DeepSeek-V4.1] Gather Engram HBM lookup hashes with an all-to-all

  #2958·wenqinI opened Nov 8, 2026Collaborator·

  ·
  1

• [Perf][DeepSeek-V4.1] Reuse SWA metadata across layers and decode steps

  ready
  #2956·CarmenDong opened Nov 8, 2026Collaborator·1 task done·

  ·
  2

• [Perf][CI] Evaluate xla_jf_program_hbm_alignment_in_kib=1 in perf pipeline

  #2955·ghpvnist opened Nov 8, 2026·

  ·

• [Core] Raiden offload: opt-in partial tails for hybrid models under DCP

  ready
  #2954·azhao155 opened Nov 8, 2026Collaborator·3 of 6 tasks·

  ·
  4

• [Perf][Core] Copy sampled-token logprobs with the async output copy, not per step

  #2952·ethannnnnn opened Nov 8, 2026Collaborator·4 of 13 tasks·

  ·
  2

• [Perf][Core] Skip top-k/top-p in the sampler without speculative decoding when no request sets them

  #2951·ethannnnnn opened Nov 8, 2026Collaborator·4 of 13 tasks·

  ·
  2

• [CI/Build] Bump tpu-sync-torch to 0.0.1.dev20261029081917

  ready
  #2950·sixiang-google opened Nov 7, 2026Collaborator·4 of 13 tasks·

  ·
  2

• [Megakernel] Add CPU and multi-rank plugin test kit

  #2949·weinan1997 opened Nov 7, 2026Collaborator·

  ·
  7

• [Megakernel] Add StepRouting per-step inspection and routing

  #2948·weinan1997 opened Nov 7, 2026Collaborator·

  ·
  7

• [CI/Build] Run pre-commit in Buildkite for merge-queue builds and after cutover

  readyskip-tpu-ci
  #2947·theminghuang opened Nov 7, 2026Collaborator·1 task done·

  ·
  1

• [CI/Build] List the /ci precheck results one per line

  readyskip-tpu-ci
  #2944·theminghuang opened Nov 7, 2026Collaborator·1 task done·

  ·
  1

• [Bugfix] Reset the same-step prefix-hit guard from KVCacheManager.new_step_starts

  ready
  #2943·theyueli opened Nov 7, 2026Collaborator·6 tasks done·

  ·
  1

• [Core] PrefixAwareScheduler: make the tier-probe cap configurable

  #2938·PrimaldevX opened Nov 7, 2026·

  ·
  6

• Write the boundary GDN state into decode-filled mamba blocks under MTP

  ready
  #2937·yinlin09 opened Nov 7, 2026Collaborator·

  ·
  4

• [Kernel][Spec Decode] Support non-causal attention in batched RPA (#2882)

  #2936·Aravind0403 opened Nov 7, 2026·

  ·
  1

  9

• [Kernel][Bugfix] Align RPA v3 fetch block size to compute block size for non-power-of-two page sizes

  #2934·Aravind0403 opened Nov 7, 2026·5 of 13 tasks·

  ·
  1

  1

• [glm-agentx-pd] GLM-5.2 AgentX P/D prefill: rows split on expert-parallel MoE, MTP cache-only pass, row-paired log-sum-exps, q absorb blocks #xClimb

  #2932·bigeye opened Nov 7, 2026Collaborator·

  ·
  2

• [glm-agentx-pd] GLM-5.2 AgentX P/D prefill: short prompts beside long prefills #xClimb

  #2931·bigeye opened Nov 7, 2026Collaborator·

  ·
  3

• [glm-agentx-pd] GLM-5.2 AgentX P/D prefill: deferred step choice, shared tokenizer cache and KV reuse #xClimb

  #2930·bigeye opened Nov 7, 2026Collaborator·

  ·
  4

• [glm-agentx-pd] GLM-5.2 AgentX: host request path reuses the previous turn #xClimb

  #2929·bigeye opened Nov 7, 2026Collaborator·

  ·
  6

• [glm-agentx-pd] GLM-5.2 AgentX P/D: bring PR #2681's decode, prefill and host optimizations #xClimb

  #2928·bigeye opened Nov 7, 2026Collaborator·

  ·
  3

• [CI/Build] Separate operational tools from user examples

  #2926·LJX2017 opened Nov 7, 2026·3 of 5 tasks·

  ·
  1

  5