Fix Cuda out of memory issue for eager runner

[helunwencser]

Stack from ghstack (oldest at bottom):

• -> #6866

This PR updates the eager runner to disable grad and save memory usage.

It also update the prompt format to not include bos.

Differential Revision: D65962743

[pytorch-bot[bot]]

:link: Helpful Links

:test_tube: See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6866

• :page_facing_up: Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

:heavy_exclamation_mark: 2 Active SEVs

There are 2 currently active SEVs. If your PR is affected, please view them below:

• [DomainsOnly] Jobs fail with GLIBC version not found
• Failure to deploy Self-Hosted runners - High Queue Times and canceled jobs

:white_check_mark: No Failures

As of commit 5c9dbfb6f4365c49f35b80eaec70bbb116c95f76 with merge base e95f171316421ccca5583ce0e4a31743dc9a58c1 (): :green_heart: Looks good so far! There are no failures yet. :green_heart:

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[facebook-github-bot]

This pull request was exported from Phabricator. Differential Revision: D65962743

[facebook-github-bot]

This pull request was exported from Phabricator. Differential Revision: D65962743

[facebook-github-bot]

This pull request was exported from Phabricator. Differential Revision: D65962743

[facebook-github-bot]

This pull request was exported from Phabricator. Differential Revision: D65962743

[facebook-github-bot]

This pull request was exported from Phabricator. Differential Revision: D65962743