All pull requests

Pull requests

• perf(npu-coop): CPU-free GPU→NPU→GPU ring 271 → 128 µs/run on Halo (dark)

  #818·fivetide opened Oct 5, 2026Collaborator·5 of 12 tasks·

  ·

• perf(gfx1100): speed up long-context DFlash verify 13.4% with FA2 split-KV and retain HipGraph/Redline PM4

  #817·HUSRCF opened Oct 5, 2026Contributor·8 of 15 tasks·

  ·
  1

• perf(gfx1100): add opt-in MQ4V2 K5120 gate/up decode specialization

  #816·HUSRCF opened Oct 5, 2026Contributor·12 of 14 tasks·

  ·
  1

• feat(qwen4): source-precision KLD teacher (kld_teacher.py)

  #815·fivetide opened Oct 4, 2026Collaborator·

  ·

• hip-bridge: keep pageable weight copies out of reclaim on every load

  #814·aldrouil opened Oct 3, 2026Contributor·8 of 13 tasks·

  ·

• refactor(gemma4): run Gemma 4 layers as declarative Steps; calibration tools on the step program (carries #795)

  #813·fivetide opened Oct 3, 2026Collaborator·

  ·

• serve: multi-slot thinking-on — read the assistant framing from the Jinja render

  #811·ghazni101 opened Oct 3, 2026Contributor·4 tasks done·

  ·

• fix(decode): MQ6G256 SwiGLU-residual dispatch + retained-PM4 tape coverage for legacy 6-bit dense

  #810·aldrouil opened Oct 3, 2026Contributor·11 of 18 tasks·

  ·

• feat(offload): memory.offload_exec=passback — scheduled co-inference of the spilled layers

  #809·aldrouil opened Oct 2, 2026Contributor·11 of 16 tasks·

  ·

• serve: opt-in embedded chat UI at /ui

  #808·ghazni101 opened Oct 2, 2026Contributor·6 of 8 tasks·

  ·
  1

• serve: completion timings no longer report wall tok/s as decode tok/s

  #807·ghazni101 opened Oct 2, 2026Contributor·3 tasks done·

  ·
  1

• serve: cold-reset the next request after a mid-generation failure

  #806·ghazni101 opened Oct 2, 2026Contributor·4 of 5 tasks·

  ·
  1

• fix(ci): make beta CI green (layering, clippy deny, pinned-LLVM gates)

  ratchet-raise
  #804·fivetide opened Oct 2, 2026Collaborator·

  ·
  1

• fix(qwen35): grow the VMM KV prefix in the mrope (VL) forward entries

  #803·fivetide opened Oct 2, 2026Collaborator·

  ·
  1

• fix(rdna-compute): require pre-grown FA2 Q16 scratch only for wide-only batches

  #802·fivetide opened Oct 2, 2026Collaborator·

  ·
  1

• fix(gfx1151): keep the GDN chunk-scan twin's output WMMAs at full EXEC

  #797·fivetide opened Sep 30, 2026Collaborator·

  ·

• refactor(qwen35): run Qwen3.5 layers as declarative Steps (stacked on #774)

  #795·fivetide opened Sep 30, 2026Collaborator·

  ·

• feat(offload): partial GPU offload for dense qwen3.5 — host-RAM spill + optional CPU execution

  #793·aldrouil opened Sep 29, 2026Contributor·8 of 14 tasks·

  ·
  1

• feat(decide): Jev-compatible decide endpoint (/v1/systemone) with session mode

  #790·nwoolmer opened Sep 29, 2026Collaborator·12 tasks done·

  ·
  6

• feat(serve): publish resident load-ack metadata and stop advertising unservable ids

  #787·aldrouil opened Sep 28, 2026Contributor·8 of 11 tasks·

  ·
  4

• Feat/gguf qwen35 hybrid ingest

  #785·smashdog opened Sep 28, 2026Contributor·14 tasks·

  ·
  2

• dflash adaptive-B: upstream reconciliation (ctx gate + PM4 exclusion) | SSE prefill keepalive | loader-wedge autoheal doc

  #784·samuelishida opened Sep 27, 2026·11 of 16 tasks·

  ·
  3

• fix(serve): return slot admission when a session leaves the GPU

  #776·bohemian-starling opened Sep 24, 2026Contributor·6 of 14 tasks·

  ·

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

  #774·fivetide opened Sep 23, 2026Collaborator·

  ·
  17

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

  #769·alpineQ opened Sep 16, 2026Contributor·13 of 17 tasks·

  ·
  5