Spaces:
Sleeping
Sleeping
Commit History
-
fix: remap bare Sequential keys to net.* when loading ddg_da.pt
Checkpoint was saved when DriftPredictorMLP used a bare nn.Sequential, producing keys like '0.weight'. Current class wraps it as self.net, expecting 'net.0.weight'. Remap on load so old .pt files load correctly.ef2e4d9 -
fix: patch torchmetrics._apply using CPU probe tensor instead of self.device
The root cause of 'Torch not compiled with CUDA enabled' is that torchmetrics.Metric._apply does fn(torch.zeros(1, device=self.device)) where self.device='cuda:0' from the checkpoint — even after map_location moves all tensors to CPU, because _device is an attribute not a tensor. Fix: replace the CUDA probe with fn(torch.zeros(1, device='cpu')) so the destination device is inferred safely without touching CUDA hardware. Patch is applied before pytorch_forecasting is imported so it is in place when Lightning restores metric state during load_from_checkpoint.ec4688a -
fix: use callable map_location to skip Lightning model.to() — avoids torchmetrics CUDA assertion
49e908a -
cleanup: remove debug tracing, restore clean health endpoint
5c3e365 -
fix: patch raw checkpoint to replace all CUDA refs before load
ae54b1b -
fix: patch torchmetrics._apply for CPU-only load of GPU-trained checkpoint
c2a87a5 -
debug: global exception handler returns JSON traceback
3cbc2e6 -
debug: expose full traceback on predict errors
86ba7a9 -
debug: attempt model download in /health and surface errors
81e7718 -
debug: add model status to /health endpoint
b39262b -
fix: download models from HF Hub before existence check
90e1547 -
fix: track app/models/ package, ignore only top-level /models/
7af21f9 -
fix: move Optional import to top of feature_engineer.py
65a23c3 -
Add Space README config
21122fb -
StockPro ML backend with pytorch-forecasting TFT
9334ec6