Skip to content

✨[Feature] Preserve Higher-Level Operators in Torch-TRT Dynamo Paths #2729

Description

@gs-olive

Is your feature request related to a problem? Please describe.
In a few cases, torch.nn.Module objects are decomposed by AOT further than desired. For instance, see #2683 (comment). In that case, torch.nn.functional.interpolate should be decomposed into aten.upsample_* or left as-is, for direct conversion to TRT.

Describe the solution you'd like
Use a pre-AOT pass to replace such nn operators with a custom op or explicitly exclude these operators from AOT decomposition (needs investigation as to whether this is possible for both Dynamo paths). See the related issues and work below.

Additional context
Related to #1894, #1979

Metadata

Metadata

Assignees

Labels

feature requestNew feature or requeststory: Dynamo Frontend & Partitioningtorch.compile, torch.export, FX graph tracing, graph partitioner, graph breaks, and the Dynamo-to-TR

Type

No type

Projects

No projects

Relationships

None yet

Development

No branches or pull requests

Issue actions