Skip to content

Always compile Edge engines and dual-profile PI05 language I/O - #4691

Open
micwill755 wants to merge 1 commit into
exporters-parityfrom
exporters-dual-profile
Open

Always compile Edge engines and dual-profile PI05 language I/O#4691
micwill755 wants to merge 1 commit into
exporters-parityfrom
exporters-dual-profile

Conversation

@micwill755

Copy link
Copy Markdown
Collaborator

Summary

  • Drop dry-run: every export compiles TensorRT. Patches stay around compile only; eager remains unpatched HF/LeRobot.
  • prepare returns all component bundles in one call. Language uses dual-profile torch_tensorrt.Input specs (inputs_embeds / ds_stack / KV) so seq is not frozen to compact length.
  • Compare TRT vs unpatched eager on bundle.parity_output (language uses lm_hidden_states).

Stacked on #4685.

Test plan

  • python tools/hf/run_pi05_export.py builds vision/language/action engines (language dual-profile)
  • python tools/hf/run_groot_export.py
  • python -m pytest tools/hf/exporters/tests/test_edge_exporter.py

@meta-cla meta-cla Bot added the cla signed label Sep 8, 2026
@github-actions github-actions Bot added the documentation Improvements or additions to documentation label Sep 8, 2026
@github-actions
github-actions Bot requested a review from narendasan September 8, 2026 07:08
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

cla signed documentation Improvements or additions to documentation

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant