All pull requests

Pull requests

• HIP: refuse --drafter until the engine runs DFlash2

  #1044·ThinkOffApp opened Oct 18, 2026Contributor·

  ·

• Benchmark rocm host identity

  #1043·ThinkOffApp opened Oct 18, 2026Contributor·

  ·

• docs: measure the ModelOpt checkpoint on one GB10

  #1042·akol1 opened Oct 18, 2026Contributor·9 tasks done·

  ·

• Nemotron 3.5 Bolt (nemotron_h_puzzle) on Metal — per-layer loader + latent MoE forward

  #1038·trickyfalcon opened Oct 18, 2026·

  ·

• docs: update contributor links to jrb-ai

  #1036·jrb-ai opened Oct 18, 2026·

  ·

• Windows: main's host tests build again on Windows hosts

  #1034·jayleaton opened Oct 18, 2026Contributor·7 of 9 tasks·

  ·

• fix(cuda): preserve first-token handoff through native host

  #1031·jrb-ai opened Oct 18, 2026·

  ·

• Sliding Weights lesson server: rejected lessons do not shift later training checks

  #1029·ThinkOffApp opened Oct 18, 2026Contributor·

  ·

• HIP: release cancelled prompt chunks at layer boundaries

  #1028·ThinkOffApp opened Oct 18, 2026Contributor·

  ·

• cuda/fp8: pack a rank's share of block-FP8 scales from inside a 128 block

  #1024·Bizuayeu opened Oct 18, 2026Contributor·9 tasks done·

  ·
  2

• core/modelopt: read MIXED_PRECISION's per-module recipes

  #1023·Bizuayeu opened Oct 18, 2026Contributor·9 tasks done·

  ·
  1

• native/cuda_memory: count every file the weight map names in weightBytes

  #1022·Bizuayeu opened Oct 18, 2026Contributor·9 tasks done·

  ·
  2

• core: decode and encode csf_32x2_v1 block scales on the host

  #1021·Bizuayeu opened Oct 18, 2026Contributor·9 tasks done·

  ·
  1

• aot: a launch's 1 may take the runtime-parameter variant after every folded candidate

  #1020·GustavBlack opened Oct 18, 2026Contributor·5 tasks done·

  ·
  1

  3

• Required tool calls in GLM's markup: xgrammar's glm_xml style

  #1018·Bizuayeu opened Oct 17, 2026Contributor·

  ·
  2

• GLM-5.3-Flash on CUDA: structured output (first_masked and held)

  #1017·Bizuayeu opened Oct 17, 2026Contributor·

  ·
  2

• GLM-5.3-Flash on CUDA: prompt_logprobs on /v1/completions

  #1016·Bizuayeu opened Oct 17, 2026Contributor·

  ·
  2

• GLM-5.3-Flash on two CUDA ranks: prompts admitted together (prefill_many)

  #1014·BobClawblaw opened Oct 17, 2026Contributor·

  ·
  2

• Windows ARM64 and NVIDIA RTX Spark N1X (sm_121) native CUDA support

  #1010·mweinbach opened Oct 17, 2026Contributor·

  ·
  1

• Qwen3.6-35B-A3B on Metal: routed experts in the 27B's engine (slice 1 of #931)

  #1009·cshintov opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·
  2

• Flash Next on Metal: a 1,048,576-token window with static YaRN x4 (stacked on #986)

  #1005·BobClawblaw opened Oct 17, 2026Contributor·

  ·
  2

• kolibri1: Kolibri-1 on CUDA through the lane host, drafts off (#481)

  #991·jschmied opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·
  8

• music3: MiniMax Music 3 lyrics to song on Metal, from Zig (stacked on #982)

  #987·drowzeys opened Oct 17, 2026Contributor·

  ·
  1

• Flash Next: images and video on native Metal, and warm passes that keep nothing skipped (stacked on #606)

  #986·BobClawblaw opened Oct 17, 2026Contributor·

  ·
  9

• drama: backbone attention runs the reference's 64-key softmax tiles

  #985·plotarmordev opened Oct 17, 2026Contributor·4 of 9 tasks·

  ·
  2