Set 0d shape to python constant tensors #1918
Conversation
titaiwangms
added
the
module: torchlib
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
onnxscript/function_libs/torch_lib/graph_building/_graph_building_torch.py
Fixed
Show fixed
Hide fixed
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #1918 +/- ##
==========================================
- Coverage 75.49% 75.49% -0.01%
==========================================
Files 252 252
Lines 27346 27355 +9
Branches 3172 3176 +4
==========================================
+ Hits 20645 20651 +6
Misses 5740 5740
- Partials 961 964 +3 ☔ View full report in Codecov by Sentry. |
| start = op.Reshape(start, op.Constant(value_ints=[-1])) | ||
| length = op.Reshape(length, op.Constant(value_ints=[-1])) | ||
| if IsScalar(dim): | ||
| dim = op.Reshape(dim, op.Constant(value_ints=[-1])) |
There was a problem hiding this comment.
I suggest changing the -1 to 1 ... since it is going to have the shape [1]. Ensures better shape inference with all of the various shape-inference implementations around.
| dim = op.Reshape(dim, op.Constant(value_ints=[-1])) | ||
|
|
||
| if IsScalar(start): | ||
| start = op.Reshape(start, op.Constant(value_ints=[-1])) |
There was a problem hiding this comment.
Same as above, here and below
| attributes=dict(value=constant_tensor), | ||
| )[0] | ||
| value.setDebugName(_rename_intermediate_value(value.debugName())) | ||
| if shape is not None: |
There was a problem hiding this comment.
I suggest not touching this file as it is deprecated.
This may be low priority. It all starts #1916 that I re-wrote aten_narrow to adapt the change that I moved YracedONNXFunction to promote python constants to tensors before function call. With doing that, IsScalar loses its functionnality because we didn't assign shape to tensors. That's why I create this PR. |
|
Leave it as it is for now. For torchlib, we keep the implementation, and CI tests will be deprecated soon with OpSignature PR. |
3 participants
From #1916, it is found that the undefined shape disables the functionality of IsScalar op:
onnxscript/onnxscript/function_libs/torch_lib/graph_building/_graph_building_torch.py
Line 357 in 561a600