Fix compiled model export. #6568
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6568
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New FailuresAs of commit 289605a with merge base 461d61d ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
| @@ -28,7 +28,7 @@ class BackendDelegate { | |||
| // Max models cache size in bytes. | |||
| size_t max_models_cache_size = 10 * size_t(1024) * size_t(1024) * size_t(1024); | |||
| // If set to `true`, delegate pre-warms the most recently used asset. | |||
| bool should_prewarm_asset = true; | |||
| bool should_prewarm_asset = false; | |||
There was a problem hiding this comment.
Why do we need this change @cymbalrush?
There was a problem hiding this comment.
This feature was pre-warming the most recently used assets, which could impact the delegate initialization time and the first model load time. This optimization was beneficial only for the CPU backend and only when the subsequent asset load remained unchanged. So, disabling it by default.
There was a problem hiding this comment.
Thanks @cymbalrush, it's worth documenting the above behavior (on Line 30?)
There was a problem hiding this comment.
Let me create another PR for this.
|
@shoumikhin has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Exporting fails for a compiled model because it depends on
get_compiled_model_path, which requires the model to be loaded. This pull request also introduces a test and disables asset prewarming by default, as it has some impact on the delegate initialization time.