All pull requests

Pull requests

• docs: a private pin is the one bump CI cannot check, and the three commands that are its evidence

  #2323·bong-water-water-bong opened Nov 13, 2026Collaborator·

  ·

• tools/ci/pr_review_queue.py: which open PRs still need a verdict at their CURRENT head

  Review effort 2/5
  #2322·bong-water-water-bong opened Nov 13, 2026Collaborator·

  ·
  4

• forward: Qwen3.6-35B-A3B-NPU2 answers -- [in, out] BF16 matrices and its shared-expert gate (refs #1786)

  Review effort 3/5
  #2318·bong-water-water-bong opened Nov 12, 2026Collaborator·

  ·
  7

• gpu-kernels: pin eb0f3839 — the 2^-6 prefill prescale (#2159), the routed GEMM fit (#572) and the registration guard

  Review effort 3/5
  #2316·bong-water-water-bong opened Nov 12, 2026Collaborator·

  ·
  2

  5

• scripts/pre-pr-checks.sh: run the six pre-merge checks as one command, one line per check

  Review effort 2/5
  #2313·bong-water-water-bong opened Nov 12, 2026Collaborator·

  ·
  6

• npu-pack: the import plan -- where each tensor bytes go, and proof that they add up (engine#1788, 3c)

  Review effort 3/5
  #2307·bong-water-water-bong opened Nov 12, 2026Collaborator·

  ·
  6

• npu-pack: a container payload widths and the shapes each role appears in, in info (engine#1788, 3c residue)

  Review effort 3/5
  #2304·bong-water-water-bong opened Nov 12, 2026Collaborator·

  ·
  10

• npu forward: Gemma4 verifies a draft in one batched pass; KV row width per layer (fixes #1780)

  Review effort 4/5
  #2303·bong-water-water-bong opened Nov 12, 2026Collaborator·

  ·
  2

  4

• recipes: a recipe -ub/-b does not replace LLAMA_ARG_UBATCH/LLAMA_ARG_BATCH (#572)

  Review effort 3/5
  #2292·bong-water-water-bong opened Nov 12, 2026Collaborator·

  ·
  6

• ci: the ratchet counts narrowing LINES per body, so a new one inside a known body is drift too

  Review effort 4/5
  #2244·bong-water-water-bong opened Nov 12, 2026Collaborator·

  ·
  25

• npu: whisper_transcribe — wav to token ids on the host, matching the reference decode (refs #1783)

  Review effort 4/5
  #2183·bong-water-water-bong opened Nov 11, 2026Collaborator·

  ·
  164

• npu-pack test: the importer's own read of use_qk_norm is pinned from JSON to the refusals (Refs #1889)

  Review effort 2/5
  #2100·bong-water-water-bong opened Nov 11, 2026Collaborator·

  ·
  8

• tools: pr_check_verdict.py says which red a PR's red is (Refs #2077)

  Review effort 2/5
  #2093·bong-water-water-bong opened Nov 11, 2026Collaborator·

  ·
  14

• hrx: compile the fork's HRX test targets, so a break in them is seen (Refs #2068)

  Review effort 3/5
  #2076·bong-water-water-bong opened Nov 11, 2026Collaborator·

  ·
  11

• npu: the speculative round's seam and bookkeeping, host-tested (engine#1781 item 1)

  Review effort 3/5
  #2059·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  17

• npu: the prefill launch decision as a pure function, pinned at the 96/97 crossover (engine#1959)

  #2055·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  15

• bench: one arm, with its memory trace, so a contaminated number says so (engine#572)

  #2049·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  9

• docs: the AXB35-02 USB limitation goes in the release notes, plus two audit facts (engine#823)

  #2046·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  9

• npu-pack: the in-block header cut scales with the blocks in the window (refs #1985) — fixture still owed

  #2021·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  10

• npu: the padded-head reader counts tile-rows in shape[0], so a [tiles, tile bytes] head dequantises (refs #1990, #1992)

  #2020·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  10

• gguf_to_q4nx: an unknown quantisation type refuses by file and number, not from inside an enum

  #2019·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  10

• census: MicroLlamaForCausalLM is report-only — its model_type is llama, which is already mapped (engine#1922)

  #2014·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  4

• npu-pack: a short epsilon vote no longer wins over the anchor residues (Qwen3-1.7B imports again, #1985)

  #1986·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  1

  30

• cmake: the Python tool self-tests get their own registration file (engine#1979, first family)

  #1981·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  6

• census_watch: say whether a pin bump fixes an uncovered class, or nobody upstream has it (engine#1922)

  #1967·bong-water-water-bong opened Nov 10, 2026Collaborator·

  ·
  12