Upload app.py with huggingface_hub

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

Bump GPU duration budget for NVFP4 dequant overhead

Access _ATEN_OP_TABLE on NVFP4Tensor directly

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

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

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

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

Use NVFP4WeightOnlyConfig (correct torchao class name)

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

Bypass diffusers↔modelopt config builder; pass NVFP4_DEFAULT_CFG directly

Initial NVFP4 demo for Cosmos3-Super-Text2Image

initial commit