All pull requests

Pull requests

• 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·

  ·
  1

• 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·

  ·
  1

• 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

  1

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

  #1018·Bizuayeu opened Oct 17, 2026Contributor·

  ·
  1

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

  #1017·Bizuayeu opened Oct 17, 2026Contributor·

  ·
  1

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

  #1016·Bizuayeu opened Oct 17, 2026Contributor·

  ·
  1

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

  #1014·BobClawblaw opened Oct 17, 2026Contributor·

  ·
  2

• fix(lanes): allocate trims a grafted window's held count and branch start with its rows

  #1013·chaog992 opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·

• nemotron: Super's MTP drafts on the draft vocabulary's FP8 rows while a stream's text is covered, same replies (stacked on #1006)

  #1011·plotarmordev opened Oct 17, 2026Contributor·9 tasks done·

  ·

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

  #1010·mweinbach opened Oct 17, 2026·

  ·

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

  #1009·cshintov opened Oct 17, 2026Contributor·7 of 9 tasks·

  ·
  1

• nemotron: Super's checkpoint prompt passes run their FP8 layers, NVFP4 experts and router on wide tiles, same bits (stacked on #999, #1002)

  #1008·plotarmordev opened Oct 17, 2026Contributor·9 tasks done·

  ·

• glm: say why a kept prompt state restore fails, and spare a state an outstanding lookup still restores

  #1007·heitke opened Oct 17, 2026Contributor·

  ·
  1

  2

• nemotron: Super verifies windows of up to 32 rows on CUDA, same replies (stacked on #999, #997)

  #1006·plotarmordev opened Oct 17, 2026Contributor·9 tasks done·

  ·

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

  #1005·BobClawblaw opened Oct 17, 2026Contributor·

  ·
  1

• WebP images through a pinned libwebp 1.6.0, to Pillow's pixels

  #1004·BobClawblaw opened Oct 17, 2026Contributor·

  ·

• nemotron: Super's latent shared expert plans once at load and, on one GPU, runs beside the routed experts (stacked on #996, #999)

  #1002·plotarmordev opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·
  4

• nemotron: a job's first message carries one row, so a decode step sends one message, not 16 rows of padding

  #1000·tovereast opened Oct 17, 2026·

  ·

• nemotron: decode forms of Super's latent top-k and NVFP4 experts on CUDA, same bits

  #999·plotarmordev opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·
  4

• Test unbuffered refusal bodies with an explicit sender gate

  #998·cwschroeder opened Oct 17, 2026Contributor·

  ·

• nemotron: --mtp-drafts and --mtp-confidence set a lone CUDA stream's MTP depth

  #997·plotarmordev opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·
  5

• nemotron: a prompt segment's buffers read the owner's shared-expert plans and pair ids

  #996·plotarmordev opened Oct 17, 2026Contributor·8 of 9 tasks·

  ·
  4

• GLM-5.3-Flash on two CUDA ranks: linears packed to 4, 6 or 8 bits

  #995·BobClawblaw opened Oct 17, 2026Contributor·

  ·
  12