All pull requests

Pull requests

• fix(cosmos3): default trust_remote_code to False in tokenizer load

  #1571·AmirF194 opened Sep 29, 2026·

  ·

• Refactor contiguous block offload and simplify model launchers

  #1570·Chernobyllight opened Sep 29, 2026Contributor·

  ·

• Add reusable contiguous CPU block offload across models and accelerators

  #1565·Chernobyllight opened Sep 28, 2026Contributor·

  ·

• feat(minimax-h3): add Radeon core-sw attention path

  #1563·Shan2L opened Sep 28, 2026·

  ·

• Add RealtimeWAM inference and repository-local benchmark evaluation（support libero，libero plus and robotwin）

  #1562·Charles2530 opened Sep 27, 2026Contributor·

  ·

• add lightvae for h3

  #1558·gushiqiao opened Sep 24, 2026Contributor·

  ·

• Fix NPU Weight Corruption During CPU Offload and Simplify Layout Configurations

  #1556·Chernobyllight opened Sep 24, 2026Contributor·

  ·

• feat: support pipefusion for MiniMax H3

  #1554·Qin-sx opened Sep 23, 2026Contributor·

  ·

• Support H3-world inference

  #1550·Musisoul opened Sep 22, 2026Collaborator·

  ·

• feat(minimax-h3): add a single-GPU AMD ROCm recipe (aiter attention, 4-step Turbo LoRA), measured on MI355X

  #1534·stayinalive181 opened Sep 19, 2026·

  ·

• feat(musa): add MiniMax H3 Moorcat sparse attention

  #1529·Watebear opened Sep 18, 2026Collaborator·

  ·

• feat(minimax-h3): add VDN inference and parallel

  #1528·Watebear opened Sep 17, 2026Collaborator·

  ·

• Unify CPU Block Offload Launchers and Host/NUMA Weight Sharing

  #1527·Chernobyllight opened Sep 17, 2026Contributor·

  ·

• feat(offload): add CPU block offload and host/NUMA weight sharing across Hunyuan Image 3.0, MiniMax-H3, Qwen-Image, and Wan

  #1524·Chernobyllight opened Sep 17, 2026Contributor·

  ·

• feat(minimax-h3): add ComfyUI FL4 inference mode

  #1513·Watebear opened Sep 11, 2026Collaborator·

  ·

• Ar ws

  #1512·huochaitiantang opened Sep 11, 2026Collaborator·

  ·

• Integrate OpenPI π0.5 LIBERO Support with the Latest LightX2V

  #1504·Chernobyllight opened Sep 10, 2026Contributor·

  ·

• Refactor offload

  #1501·Watebear opened Sep 9, 2026Collaborator·

  ·

• feat(openpi): add aligned LIBERO training support

  #1500·Chernobyllight opened Sep 9, 2026Contributor·

  ·

• feat: integrate OpenPI π0.5-LIBERO inference and ROS2 evaluation

  #1491·Chernobyllight opened Sep 7, 2026Contributor·

  ·

• fix(minimax-h3): align Ref2AV output

  #1490·Watebear opened Sep 6, 2026Collaborator·

  ·
  9

• Accelerate The Inference Speed Of Wan2.2 On NVIDIA Thor T5000

  #1489·Lunemere opened Sep 6, 2026Contributor·

  ·

• feat(musa): add sparse attention settings

  #1488·Watebear opened Sep 5, 2026Collaborator·

  ·

• fix(audio_io): remove unreachable duplicate except in torchcodec probe

  #1487·Anai-Guo opened Sep 5, 2026·

  ·

• feat: integrate OpenPI π0.5-LIBERO inference and ROS2 evaluation

  #1476·Chernobyllight opened Sep 4, 2026Contributor·

  ·