All pull requests

Pull requests

• cuda: tree attention and DFlash block attention for every family, checked against float64

  #746·jschmied opened Oct 12, 2026Contributor·8 tasks done·

  ·

• GLM: run the fragment check at load

  #745·harrisonfriia opened Oct 12, 2026·8 tasks done·

  ·

• core: make the fragment check catch a transposed operand

  #744·harrisonfriia opened Oct 12, 2026·8 tasks done·

  ·

• native: the CUDA lane backend forwards every entry a family's backend gives

  #742·Bizuayeu opened Oct 12, 2026Contributor·8 tasks done·

  ·

• cuda: EXL3 routed experts behind experts.Grouped (#593, stacked on #740)

  #741·BHCC2025 opened Oct 12, 2026Contributor·

  ·

• cuda: EXL3 projections behind qlinear (#593)

  #740·BHCC2025 opened Oct 12, 2026Contributor·

  ·

• AX300: self-developed host-staged all-gather with fp8 payloads for PCIe-only nodes

  #739·liumorrisclaw opened Oct 12, 2026Contributor·

  ·
  1

• Dashboard: prefill queue progress and a working Memory tile

  #738·heitke opened Oct 12, 2026Contributor·

  ·

• perf(qwen CUDA): enable tuned stream planning on GB10 (sm_121)

  #737·0xWhiteMage opened Oct 12, 2026·

  ·

• cuda: RMS norm rows and token rows for every family, Nemotron's embed the first user

  #736·jschmied opened Oct 12, 2026Contributor·8 tasks done·

  ·

• Prompt cache: a state a conversation resumes again is a branch point, not superseded

  #735·BadAd84 opened Oct 12, 2026·6 of 8 tasks·

  ·

• cuda/tree_round: core/tree_round's verify and commit step on CUDA, shared by every family

  #734·jschmied opened Oct 12, 2026Contributor·8 tasks done·

  ·

• native cuda: refuse --drafter, --speed-up and --prompt-cache-over-cap instead of ignoring them

  #733·jschmied opened Oct 12, 2026Contributor·8 tasks done·

  ·

• Flash Next keeps its buffers wired between requests: a residency set, and a keepalive tick the GPU runs

  #732·sethforprivacy opened Oct 12, 2026·

  ·
  3

• lanes: the GPU sampling rule once (gpu_rule.Rule): Metal's cfg and CUDA's Rule filled from it

  #731·jschmied opened Oct 12, 2026Contributor·8 tasks done·

  ·

• core/quant: FP8 and NVFP4 values and float64 references in one place

  #730·jschmied opened Oct 12, 2026Contributor·8 tasks done·

  ·

• EXL3: share decoded weights across prefill rows

  #729·mrv777 opened Oct 12, 2026·

  ·

• MiMo-V2.6-Flash on CUDA: EXL3, DFlash, two GPUs (stacked on #703)

  #728·ajensenwaud opened Oct 12, 2026·

  ·

• DeepSeek-V4.1: performance kernels, family images and device bench runners

  #727·jayleaton opened Oct 12, 2026Contributor·3 of 8 tasks·

  ·

• Qwen3.8-27B on CUDA: image input and --kv-cache fp8 (256K with vision on a 5090)

  #726·BobClawblaw opened Oct 12, 2026Contributor·

  ·

• Nemotron CUDA: prefill each reply in the background for the next turn

  #723·plotarmordev opened Oct 12, 2026Contributor·7 of 8 tasks·

  ·

• Nemotron CUDA: kept prompt states hold each attention layer's first rows only

  #722·plotarmordev opened Oct 12, 2026Contributor·8 tasks done·

  ·

• Nemotron CUDA: keep prompt states where messages end, staged inside a chunk without an extra pass

  #721·plotarmordev opened Oct 12, 2026Contributor·8 tasks done·

  ·
  2

• Nemotron CUDA: prompt passes of up to 4,096 rows between kept points

  #720·plotarmordev opened Oct 12, 2026Contributor·8 tasks done·

  ·

• Flash Next on Metal: sampled decoding (temperature, top_k, top_p, min_p, seed) on the host

  #719·BobClawblaw opened Oct 12, 2026Contributor·6 of 8 tasks·

  ·