Zenas
Zenas
### 🐛 Describe the bug Certain standard operations such as `torch.prod()` or `torch.special.entr()` cause a CUDA runtime error **only** when executed on GPU. The same operations work fine on CPU....
### Expected behavior A tiny Transformer-like block exported via torch.export crashes TVM when importing with tvm.relax.frontend.torch.from_exported_program(ep). Before the crash, PyTorch emits warnings that torch.export inserted a get_attr node without a...
### Summary Importing a torch.exported STFT + RNN toy model fails in the TVM Relax Torch frontend with: ``` AssertionError: Unsupported function types ['rnn_tanh.input', 'real.default', 'unfold.default', 'imag.default', 'fft_fft.default'] ``` This...
### Summary Importing a torch.exported program into TVM Relax triggers a segmentation fault inside FFI during construction of a Relax Tuple. The minimal model performs a 4D advanced indexing write...
### Actual behavior Importing an exported PyTorch program fails early in TVM Relax Torch frontend with: ``` AssertionError: Unsupported function types ['randn.default'] ``` The model is intentionally small and uses...
### Summary In a Python process where PyTorch and Transformers are imported first, importing TVM immediately segfaults before any TVM API is used. The backtrace points to an LLVM static...
## Summary When using `nn.CrossEntropyLoss`, if the target tensor has the wrong shape (e.g., `[N, C]` instead of `[N]`), OneFlow does not raise a clear Python exception. ## Code to...
## Summary When `nn.MaxUnpool2d` is called with invalid indices (e.g., all values set to `-1`), OneFlow does not raise a Python error. Instead, the process aborts due to a C++...
## Summary Passing a wrongly shaped target y to nn.CrossEntropyLoss (e.g., logits with shape (N, C) but y has shape (1,) instead of (N,)) triggers an internal assertion: ``` oneflow::CHECK...
## Summary Accessing a buffer that was registered with the name "0" (a numeric string) through module._buffers["0"] inside forward() works in eager mode, but segfaults in Graph mode during execution...