All pull requests

Pull requests

• cuda: GLM, MiniMax and Nemotron placed from one launch, by flags or aliases (T1b of #2560)

  #2919·laudney opened Nov 11, 2026Contributor·7 of 8 tasks·

  ·

• glm cuda: abort the communicator through its owner (T1c of #2560)

  #2918·laudney opened Nov 11, 2026Contributor·6 tasks done·

  ·

• fabric/mesh: one launch description, --tp up to 64 and the rank aliases (T1a of #2560)

  #2917·laudney opened Nov 11, 2026Contributor·4 tasks done·

  ·

• qwen27 Metal (M1-M4): solo prompt passes take the frame's rows, not 128 (#2907)

  #2916·ashhart opened Nov 11, 2026Owner·

  ·
  1

• qwen27 Metal prompts: drop the GDN tails buffer nothing reads (#2905)

  #2915·ashhart opened Nov 11, 2026Owner·

  ·
  1

• cuda: the release server refuses a GPU it has no kernels for, by name, at startup (#2904)

  #2914·ashhart opened Nov 11, 2026Owner·

  ·
  1

• qwen27 Metal (M1-M4): projections no longer barrier internally, and XS sums compile out before tensor units (#2900)

  #2913·ashhart opened Nov 11, 2026Owner·

  ·
  1

• docs(xpu): the B70 prefill row predates 1.1.0's precise prompt DeltaNet (#2859)

  #2906·ashhart opened Nov 11, 2026Owner·

  ·

• native: the CUDA kept-state identity is written once through any hasher, its keys unchanged

  #2903·kotevcode opened Nov 11, 2026Contributor·4 tasks done·

  ·
  1

• cuda source check: declare the SDK names main's kernels use

  #2902·Bizuayeu opened Nov 11, 2026Contributor·4 tasks done·

  ·
  1

• server: --execution-witness serves a read-scoped snapshot of what runs, or why it is unknown

  #2901·kotevcode opened Nov 11, 2026Contributor·7 of 8 tasks·

  ·

• cuda aot: refuse a captured spec whose launch PDL disagrees with its PDL constant (#2715)

  #2899·ashhart opened Nov 11, 2026Owner·

  ·

• flashnext: pool lanes keep their states and draft as pieces on whole pages, no run (#2444 P3, F2)

  #2898·jschmied opened Nov 11, 2026Contributor·

  ·

• core: Admit and Ledger move to engine_admission.zig, unchanged

  #2896·kotevcode opened Nov 11, 2026Contributor·4 tasks done·

  ·
  1

• xpu: a sampled Qwen3.8-27B row draws from a top-k kept on the card

  #2895·Mirrdhyn opened Nov 11, 2026Contributor·

  ·
  1

  3

• glm cuda: NVFP4_NZ5 checkpoints, block 32's experts with E2M1's code 0b1000 read as +5

  #2894·Bizuayeu opened Nov 11, 2026Contributor·8 tasks done·

  ·
  3

• drama3: one codec schedule for CUDA and Metal, each backend binds only its kernels (#2118)

  #2893·jschmied opened Nov 11, 2026Contributor·

  ·
  3

• core: one block-FP8 reader and exact scale widening (#2460 item 7)

  #2892·jschmied opened Nov 11, 2026Contributor·

  ·
  3

• ci: run the Metal engine host tests (test-native) in the macOS nightly

  #2891·ashhart opened Nov 11, 2026Owner·

  ·
  1

• llama on Metal: Qwen2 checkpoints, Bimo-7B-CyberAgent's bf16 release with its q, k and v biases and YaRN at HF's rounding points

  #2890·BobClawblaw opened Nov 11, 2026Contributor·

  ·
  3

• test: Qwen27 reply history follows #2699's newest-state retention

  #2889·ashhart opened Nov 11, 2026Owner·

  ·
  1

• prompt radix: pages carry their reuse identity, and a request's chunk starts decide its resumes (#2444 P2, C6)

  #2888·laudney opened Nov 11, 2026Contributor·6 of 8 tasks·

  ·

• fix(qwen27): warn when Metal startup exceeds live free memory

  #2886·garvrao80 opened Nov 11, 2026Contributor·4 of 5 tasks·

  ·
  2

• server: serve --config FILE.toml reads flags from a file (stacked on #2883)

  #2885·laudney opened Nov 11, 2026Contributor·4 tasks done·

  ·
  1

  4

• server: the last --compact-at wins (auto then a fraction kept auto)

  #2883·laudney opened Nov 11, 2026Contributor·4 tasks done·

  ·
  3