You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When replacing the view nodes with reshape nodes, the metadata of the original view nodes, are assigned to the reshape nodes in the wrong order.
For example, assume we have 2 view nodes, view_1, view_2. So we will have a list, [metadata1, metadata2], from this code
I haven't dive deep yet, but after torch.fx.subgraph_rewriter.replace_pattern(gm, orig, replacement), in the new graph, the view nodes are replaced with reshape nodes, however, the order of reshape nodes can be reshape_default_2, reshape_default_1. In this case, when we set the metadata back by set_metadata(gm, replacement_op, metadata), the metadata can mismatch
To Reproduce
I am sorry, I cannot share the graph of the model, I will try my best to have e toy model to reproduce the error.
Steps to reproduce the behavior:
Expected behavior
Environment
Build information about Torch-TensorRT can be found by turning on debug messages
Torch-TensorRT Version (e.g. 1.0.0): 2.4.0
PyTorch Version (e.g. 1.0): 2.4.1
CPU Architecture: x86_64
OS (e.g., Linux): linux
How you installed PyTorch (conda, pip, libtorch, source): pip
Build command you used (if compiling from source):
Are you using local sources or building from archives: no
Python version: 3.11.9
CUDA version:12.6
GPU models and configuration: NVIDIA L4
Any other relevant information:
Additional context
The text was updated successfully, but these errors were encountered:
I check the source code, seems we have converter support for torch.ops.aten.view, so probably the best way to solve this is to delete the view_to_reshape pass
Bug Description
When replacing the view nodes with reshape nodes, the metadata of the original view nodes, are assigned to the reshape nodes in the wrong order.
For example, assume we have 2 view nodes, view_1, view_2. So we will have a list, [metadata1, metadata2], from this code
I haven't dive deep yet, but after
torch.fx.subgraph_rewriter.replace_pattern(gm, orig, replacement)
, in the new graph, the view nodes are replaced with reshape nodes, however, the order of reshape nodes can be reshape_default_2, reshape_default_1. In this case, when we set the metadata back byset_metadata(gm, replacement_op, metadata)
, the metadata can mismatchTo Reproduce
I am sorry, I cannot share the graph of the model, I will try my best to have e toy model to reproduce the error.
Steps to reproduce the behavior:
Expected behavior
Environment
conda
,pip
,libtorch
, source): pipAdditional context
The text was updated successfully, but these errors were encountered: