Julia Bazińska

Issues in XLA compilation while running on TPU

1

Hi, I am using your code to train Conformer using Cloud TPU VM on GCP. Have you managed to run the conformer code on TPU and maybe you could share...

Incorrect parsing of parentheses in DaCe programs

1

It looks like parentheses don't get parsed correctly. ``` import dace import numpy as np N, F_in, F_out = 2, 3, 4 @dace.program def fn(a, b, c): c[:] = (a.T...

Add a library node for np.transpose for >2 dims and a cuTENSOR implementation

2

Adds - a library node for np.transpose for >2 dims - a cuTENSOR implementation - pure implementation (copy-pasted what was in replacements.py before) - cutensor environment What I don't know:...

Memlet WCR propagation bug

1

**Describe the bug** WCR edges are not propagated correctly. As a side effect, applying greedy fusion to an erroneous graph results in a graph that gives incorrect results. **To Reproduce**...

Support nested SDFGs in `ScheduleType.GPU_ThreadBlock_Dynamic`

1

**Describe the bug** Encountered when implementing a Graph Attentional Operator. When using `ScheduleType.GPU_ThreadBlock_Dynamic`, incorrect code gets generated. The error is: ``` (dace14) [jbazinsk@ault24 bugs]$ python compile_error.py Keeping schedule sequential for...

Implement a node module replacements for GNN support.

1

Implementing module replacements: * torch.Module becomes a placeholder torch fn, then a placeholder op in ONNX graph, which gets expanded to a replacement implementation. * main replacement mechanism in `daceml/onnx/nodes/replacement.py`...

Support for pathlib.Path in datasets 2.19.0

1

### Describe the bug After the recent update of `datasets`, Dataset.save_to_disk does not accept a pathlib.Path anymore. It was supported in 2.18.0 and previous versions. Is this intentional? Was it...