Spaces:
Running
Running
Commit History
-
Use HF dataset repo as source of truth for dataset.json
Unverified signaturePush dataset.json to HF repo after every label/preprocess batch so progress persists across ephemeral Space restarts. On download, pull the JSON back and restore labels before scanning. - Add upload_dataset_json_to_hf() and pull dataset.json in download - Remove hf_token param (use HF_TOKEN env var) - Reconcile stale .pt preprocessed flags on restart - Simplify Data Source UI: remove upload files, HF token, resume accordion - Add batch mode (max_count) for labeling and preprocessing - Add preprocessed flag to AudioSample, scan dedup via existing_paths - Make @spaces.GPU a no-op when running locally6c32e21
-
Fix meta tensor device mismatch by also patching torch.arange
Unverified signatureFSQ.__init__ calls torch.arange() through the torch module to build its implicit codebook. The meta device context set by newer transformers (which ignores low_cpu_mem_usage=False) redirects this to meta device, causing a device mismatch with our CPU-forced _levels buffer. Patch torch.arange globally during model loading to force device="cpu", alongside the existing tensor() patches for both VQ module namespaces.f266123 -
Patch both residual_fsq and finite_scalar_quantization tensor refs
Unverified signatureFSQ.__init__ also calls .prod().item() on a meta tensor created via its own 'from torch import tensor'. Patch both VQ module namespaces.011d30d -
Fix meta tensor: patch residual_fsq module's tensor reference, not torch.tensor
Unverified signatureResidualFSQ uses 'from torch import tensor' - a direct module-level reference that bypasses our torch.tensor patch. Now we replace the 'tensor' name in the residual_fsq module namespace directly, forcing CPU device for all tensor creation during FSQ init.0c47532 -
Fix meta tensor: monkey-patch ResidualFSQ to force CPU tensor creation
Unverified signatureZeroGPU's __torch_function__ hooks redirect ALL tensor creation to meta device, even with torch.device('cpu') context. The fix patches torch.tensor within ResidualFSQ.__init__ to explicitly pass device='cpu', so the levels assertion runs on real CPU tensors instead of meta tensors.bf4c82d -
Fix meta tensor crash: force CPU device context during model init
Unverified signatureZeroGPU redirects all tensor creation to meta device at the torch level, which breaks ResidualFSQ assertions. Wrapping from_pretrained in torch.device('cpu') overrides this redirection. Also removes boot diagnostics from app.py.b0a0560
-
Try _fast_init=False to prevent meta device initialization
ACE-Step Custom5b76ce1
-
Fix dtype parameter: use dtype instead of torch_dtype, remove device_map
ACE-Step Custom140e1e2 -
Fix meta tensor: add device_map=None and use torch_dtype parameter
ACE-Step Customa11c404 -
Fix meta tensor error: disable low_cpu_mem_usage for model loading
ACE-Step Custom23dc982 -
Deploy ACE-Step Custom Edition with bug fixes
ACE-Step Customa602628