Hugging Face's logo

Spaces:
multimodalart
/
Cosmos3-Super-Text2Image
Running on Zero

App Files Community
Cosmos3-Super-Text2Image

Commit History

  1. Upload app.py with huggingface_hub Running

    multimodalart HF Staff
    Verified signature
    fb0c49e

  2. Calibrate GPU duration to ~18s/step at 1024² (measured 13-15s)

    multimodalart HF Staff
    Verified signature
    1b2d4d4

  3. Bump GPU duration budget for NVFP4 dequant overhead

    multimodalart HF Staff
    Verified signature
    b98d01b

  4. Access _ATEN_OP_TABLE on NVFP4Tensor directly

    multimodalart HF Staff
    Verified signature
    ac4a7e1

  5. Wrap NVFP4 matmul handlers to upcast non-NVFP4 inputs to weight.orig_dtype

    multimodalart HF Staff
    Verified signature
    fbc83a5

  6. Also register inner-tensor cuda_aliases so packer captures NVFP4 inner storages

    multimodalart HF Staff
    Verified signature
    c406a24

  7. Patch ZeroGPU empty_fake to recurse into wrapper subclasses (NVFP4Tensor pack)

    multimodalart HF Staff
    Verified signature
    86bd9cc

  1. Register aten.empty_like on NVFP4Tensor so ZeroGPU can pack quantized weights

    multimodalart HF Staff
    Verified signature
    0d3980c

  2. Use NVFP4WeightOnlyConfig (correct torchao class name)

    multimodalart HF Staff
    Verified signature
    921f3f9

  3. Switch NVFP4 backend to torchao (diffusers↔modelopt cfg dict↔list drift)

    multimodalart HF Staff
    Verified signature
    c77828b

  4. Bypass diffusers↔modelopt config builder; pass NVFP4_DEFAULT_CFG directly

    multimodalart HF Staff
    Verified signature
    db9d038

  5. Initial NVFP4 demo for Cosmos3-Super-Text2Image

    multimodalart HF Staff
    Verified signature
    046dae3

  6. initial commit

    multimodalart HF Staff
    Verified signature
    1753094