All pull requests

Pull requests

• fix(metrics): report prefill work after each completed step

  #1001·ktsaou opened Oct 7, 2026·

  ·
  1

• perf(glm): B12X fused all-reduce + RMSNorm at GLM-5.3 DSA sites (port of joninco)

  #992·voipmonitor opened Oct 6, 2026·

  ·
  1

• perf(glm): L2 prefetch of GLM-5.3 attention weights during decode (port of joninco)

  #991·voipmonitor opened Oct 6, 2026·

  ·
  1

• fix(glm): start GLM-5.3 (744B) on B12X

  #990·voipmonitor opened Oct 6, 2026·

  ·
  1

• [Worker] Load weights without the split limit under expandable segments (canonical twin of #980)

  #981·voipmonitor opened Oct 4, 2026·

  ·
  1

• [KK] NVFP4-CSF: expand the next layer's scales during its attention

  #979·voipmonitor opened Oct 4, 2026·

  ·
  1

• [KK] qwen4_exp: resolve the PLE storage dtype from an FP4-CSF checkpoint root

  #975·voipmonitor opened Oct 4, 2026·

  ·
  1

• [KK] Load lossless MXFP4-CSF DeepSeek-V4-Flash and Vision-Exp checkpoints

  #973·voipmonitor opened Oct 4, 2026·

  ·
  1

• [KK] b12x MoE: opt-in NVFP4-activation prefill for W4A16 experts, decode rows stay W4A16

  #972·voipmonitor opened Oct 4, 2026·

  ·
  6

• Load exact GLM Spark projections and NVFP4 MTP from the QAD source

  #968·voipmonitor opened Oct 3, 2026·

  ·
  1

• [KK] Qwen3.8-Flash-Next: share one host-RAM PLE table between same-host processes

  #962·voipmonitor opened Oct 1, 2026·

  ·
  4

• fix(scheduler): give prefill lanes to queued requests only while admissible

  #960·voipmonitor opened Oct 1, 2026·

  ·
  2

• feat(moe): native B12X MXFP8 MoE for ModelOpt MTP experts

  #957·voipmonitor opened Oct 1, 2026·

  ·
  2

• Own FP4-CSF loading and preserve native DS4.1 expert precision

  #956·voipmonitor opened Oct 1, 2026·

  ·
  1

• [KK] Release dense MLA priming buffers after preparation

  #953·voipmonitor opened Sep 30, 2026·

  ·
  1

• chore(release): wheels require CUTLASS DSL 4.7.1 with B12X (canonical twin of #951)

  #952·voipmonitor opened Sep 30, 2026·

  ·
  3

• [KK] Load DS4.1 and Kimi-K3 lossless X4T checkpoints

  #949·voipmonitor opened Sep 29, 2026·

  ·
  1

• fix(scheduler): let decodes preempt in a prefill turn that scheduled nothing (canonical twin of #946)

  #948·voipmonitor opened Sep 29, 2026·

  ·
  1

• fix(release): drop Python staged by other branches from the shared wheel build cache (canonical twin of #939)

  #940·voipmonitor opened Sep 29, 2026·

  ·
  1

• [KK] Prepare B12X Kimi KDA prefill and expose MLA partial precision

  #937·voipmonitor opened Sep 29, 2026·

  ·
  1

• fix(model-state): stage per-step prefill-flag uploads

  #923·FujitsuPolycom opened Sep 27, 2026·

  ·
  1

• feat(worker): GPU stall watchdog; RoCEnante health on every rank

  #916·tobymao opened Sep 27, 2026·

  ·
  6

• perf(glm5next): serve native block-FP8 DSA q_b_proj and o_proj on Marlin

  #915·tobymao opened Sep 27, 2026·

  ·
  6

• fix: stop request-path Triton kernels recompiling per batch

  #914·tobymao opened Sep 27, 2026·

  ·
  6

• fix(shm_broadcast): never drop a write notification

  #913·tobymao opened Sep 27, 2026·

  ·
  5