All pull requests

Pull requests

• feat(pi05): add FP32 CPU reference with BLAS-backed attention

  #176·wong-1994 opened Oct 26, 2026Contributor·

  ·
  1

• perf(qwen35): validate and optimize NVFP4 runtime on Thor-U

  #175·ChenxiHe231 opened Oct 26, 2026Contributor·

  ·

• feat(qwen35): Qwen3.5-35B-A3B NVFP4 runtime with native MTP and serving

  #173·ChenxiHe231 opened Oct 26, 2026Contributor·

  ·

• fix(cuda): reuse model-owned repacked weights in place

  #172·wong-1994 opened Oct 26, 2026Contributor·

  ·

• feat(audio): add file ASR and complete WAV TTS serving

  #171·qhy991 opened Oct 25, 2026Contributor·

  ·

• perf(tokenizer): avoid full vocabulary copies for unique template tokens

  #170·qhy991 opened Oct 25, 2026Contributor·

  ·

• perf(mlx): Qwen3 and MiniCPM5 prefill/decode performance on Metal

  #169·qhy991 opened Oct 25, 2026Contributor·

  ·

• perf(mlx): pipeline greedy decode with device-side token feedback

  #168·qhy991 opened Oct 25, 2026Contributor·

  ·

• perf(mlx): bounded long-context generation for Qwen3 and MiniCPM5

  #167·qhy991 opened Oct 25, 2026Contributor·

  ·

• feat(serving): add serving benchmarks and memory calibration tools

  #163·qhy991 opened Oct 23, 2026Contributor·

  ·

• feat(serving): supervise pinned Clef text decisions

  #162·qhy991 opened Oct 23, 2026Contributor·

  ·

• feat(serving): add the isolated MiniCPM vision provider

  #161·qhy991 opened Oct 23, 2026Contributor·

  ·

• feat(serving): serve native MiniCPM with explicit sampling and thinking profiles

  #160·qhy991 opened Oct 23, 2026Contributor·

  ·

• feat(serving): add the pinned MLX text provider and client integration

  #159·qhy991 opened Oct 23, 2026Contributor·

  ·

• feat(serving): supervise serial workers with bounded HTTP lifecycle

  #158·qhy991 opened Oct 23, 2026Contributor·

  ·

• feat(serving): define serial worker contracts and shared conformance cases

  #157·qhy991 opened Oct 23, 2026Contributor·

  ·

• Add native DreamZero-DROID BF16 joint video/action VLA

  #155·wong-1994 opened Oct 22, 2026Contributor·

  ·

• feat(cosmos-policy): add native LIBERO joint diffusion policy

  #154·wong-1994 opened Oct 21, 2026Contributor·

  ·

• perf(model): warm MiniCPM5's decode step once per KV capacity

  #153·qhy991 opened Oct 21, 2026Contributor·

  ·

• perf(model): fuse Qwen-Image image blocks on M4 Metal

  #152·qhy991 opened Oct 21, 2026Contributor·

  ·

• Add native Fast-WAM action inference on cuda-new

  #149·wong-1994 opened Oct 19, 2026Contributor·

  ·

• feat(model): run Clef on the GPU and the Neural Engine on the M4 (neuron split)

  #148·qhy991 opened Oct 18, 2026Contributor·

  ·

• feat(model): run Clef on the GPU and the Neural Engine without Python (column split)

  #147·qhy991 opened Oct 17, 2026Contributor·

  ·

• feat(coreml): write ML Programs in Rust and run them on the Neural Engine

  #146·qhy991 opened Oct 17, 2026Contributor·

  ·

• feat(mlx): add an optimized Clef GPU composition, selected by device

  #145·qhy991 opened Oct 17, 2026Contributor·

  ·
  1