- Remove dead #tensorrt==10.13.3.9 comment
- Add onnx and onnxsim for export scripts
- Pin minimum versions to ensure compatibility
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Replace bare except: with except Exception: in __del__ and reset() to avoid
catching KeyboardInterrupt/SystemExit
- Add response.raise_for_status() to fail clearly on HTTP errors
- Fix resolution log output to show HxW instead of CHW shape
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Advises users to disable cuda graph if experiencing high RAM usage
or errors with variable input resolutions.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- __del__: Add hasattr checks before del, iterate tensors dict to delete
each tensor individually, also clean up inputs/outputs
- reset: Same improvements plus reset cuda_graph_instance and graph
to None to prevent stale state on reuse
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Previously reported only new interpolated frames using a formula, not the
actual count. Now shows out_len which reflects true total frames output.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Without map_location, torch.load defaults to CPU then the model is moved
to GPU with .to(TORCH_DEVICE). This is inefficient and can cause issues on
systems where CPU tensors aren't properly set up.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Yuvraj Seegolam