Commit History

3 days ago

  1. Fix OOM and extreme lag by bypassing unused reference attention in single-human mode and chunking multi-human reference attention Running

    Daniel
    03a50fa
  2. Accelerate flash_attention using xformers Cutlass and contiguous SDPA kernels

    Daniel
    23defbb
  3. Optimize ZeroGPU duration to 300s, set recommended 15 steps, and remove torch.compile lag

    Daniel
    d05af4e
  4. Disable layer-by-layer VRAM offloading to run DiT natively on GPU

    Daniel
    5045d7c
  5. Fix tensor device mismatch between cuda and cpu in SDPA fallback

    Daniel
    85175b0

4 days ago

  1. Refine SDPA attention masking and adjust ZeroGPU duration

    Daniel
    63e3991
  2. Implement safe_memory_efficient_attention with PyTorch SDPA fallback

    Verified signature
    5edc824
  3. Disable empty_cache during diffusion loop to prevent dynamo handler collision

    Verified signature
    a8ab186
  4. Add PyTorch native SDPA fallback in flash_attention

    Verified signature
    67b2c08
  5. Direct layer hidden_states collection in Wav2Vec2

    Verified signature
    5547fec
  6. Set attn_implementation to eager for Wav2Vec2

    Verified signature
    b7c2342
  7. Fix tokenizer path fallback to HF Hub model ids

    Verified signature
    ca5c28f
  8. Fix utils init import get_gpu_manager

    Verified signature
    5e918f1
  9. Fix CUDA re-init in forked subprocess: defer model imports & lazy xformers

    Verified signature
    e579ffb
  10. Upgrade Gradio SDK to 6.14.0 and patch schema validation

    Verified signature
    1dc8a76
  11. Fix ZeroGPU compatibility: mock gcnArchName and decouple xfuser

    Verified signature
    fcb6a26
  12. Deploy InfiniteTalk Gradio Demo with ZeroGPU support

    Verified signature
    13c793c
  13. initial commit

    Verified signature
    70af39c