[ET-VK][ez] Make squeeze insertion requirements more strict #9917
+18
−5
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
## Context Refactor the `SqueezeUnsqueezeInputs` pass to be more clear about its intention. For Llama models, input shapes to 4 bit linear will oftentimes have the shape `[1, seq_len, dim]`; under the current implementation of the pass, the input would be squeezed to `[seq_len, dim]` even though the squeeze is not necessary. The original intention of thispass was to squeeze inputs with shape `[batch_size, 1, dim]` to `[batch_size, dim]` before calling the 4-bit linear operator. ## Changes To avoid inserting unnecessary squeeze/unsqueezes, be more specific about when squeeze/unsqueeze should be added. I would like to consider refactoring this pass in the future, since the logic is currently a bit uninttuitive. Squeeze/unsqueeze is also inserted for gelu and relu, but this is to create a chain of unsqueeze/squeeze that will be eliminated by a later pass (see #8601 / D69673068). I think eventually it will be good to rewrite the pass to make shape management more explicit and self contained within the pass rather than inserting ops which are expected to be removed later on. Differential Revision: [D72480178](https://our.internmc.facebook.com/intern/diff/D72480178/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/9917
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit ae87928 with merge base 6adff9c ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D72480178 |
## Context Refactor the `SqueezeUnsqueezeInputs` pass to be more clear about its intention. For Llama models, input shapes to 4 bit linear will oftentimes have the shape `[1, seq_len, dim]`; under the current implementation of the pass, the input would be squeezed to `[seq_len, dim]` even though the squeeze is not necessary. The original intention of thispass was to squeeze inputs with shape `[batch_size, 1, dim]` to `[batch_size, dim]` before calling the 4-bit linear operator. ## Changes To avoid inserting unnecessary squeeze/unsqueezes, be more specific about when squeeze/unsqueeze should be added. I would like to consider refactoring this pass in the future, since the logic is currently a bit uninttuitive. Squeeze/unsqueeze is also inserted for gelu and relu, but this is to create a chain of unsqueeze/squeeze that will be eliminated by a later pass (see #8601 / D69673068). I think eventually it will be good to rewrite the pass to make shape management more explicit and self contained within the pass rather than inserting ops which are expected to be removed later on. Differential Revision: [D72480178](https://our.internmc.facebook.com/intern/diff/D72480178/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D72480178 |
## Context Refactor the `SqueezeUnsqueezeInputs` pass to be more clear about its intention. For Llama models, input shapes to 4 bit linear will oftentimes have the shape `[1, seq_len, dim]`; under the current implementation of the pass, the input would be squeezed to `[seq_len, dim]` even though the squeeze is not necessary. The original intention of thispass was to squeeze inputs with shape `[batch_size, 1, dim]` to `[batch_size, dim]` before calling the 4-bit linear operator. ## Changes To avoid inserting unnecessary squeeze/unsqueezes, be more specific about when squeeze/unsqueeze should be added. I would like to consider refactoring this pass in the future, since the logic is currently a bit uninttuitive. Squeeze/unsqueeze is also inserted for gelu and relu, but this is to create a chain of unsqueeze/squeeze that will be eliminated by a later pass (see #8601 / D69673068). I think eventually it will be good to rewrite the pass to make shape management more explicit and self contained within the pass rather than inserting ops which are expected to be removed later on. Differential Revision: [D72480178](https://our.internmc.facebook.com/intern/diff/D72480178/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D72480178 |
nathanaelsee
approved these changes
Apr 7, 2025
## Context Refactor the `SqueezeUnsqueezeInputs` pass to be more clear about its intention. For Llama models, input shapes to 4 bit linear will oftentimes have the shape `[1, seq_len, dim]`; under the current implementation of the pass, the input would be squeezed to `[seq_len, dim]` even though the squeeze is not necessary. The original intention of thispass was to squeeze inputs with shape `[batch_size, 1, dim]` to `[batch_size, dim]` before calling the 4-bit linear operator. ## Changes To avoid inserting unnecessary squeeze/unsqueezes, be more specific about when squeeze/unsqueeze should be added. I would like to consider refactoring this pass in the future, since the logic is currently a bit uninttuitive. Squeeze/unsqueeze is also inserted for gelu and relu, but this is to create a chain of unsqueeze/squeeze that will be eliminated by a later pass (see #8601 / D69673068). I think eventually it will be good to rewrite the pass to make shape management more explicit and self contained within the pass rather than inserting ops which are expected to be removed later on. Differential Revision: [D72480178](https://our.internmc.facebook.com/intern/diff/D72480178/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D72480178 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Context
Refactor the
SqueezeUnsqueezeInputspass to be more clear about its intention.For Llama models, input shapes to 4 bit linear will oftentimes have the shape
[1, seq_len, dim]; under the current implementation of the pass, the input would be squeezed to[seq_len, dim]even though the squeeze is not necessary.The original intention of thispass was to squeeze inputs with shape
[batch_size, 1, dim]to[batch_size, dim]before calling the 4-bit linear operator.Changes
To avoid inserting unnecessary squeeze/unsqueezes, be more specific about when squeeze/unsqueeze should be added.
I would like to consider refactoring this pass in the future, since the logic is currently a bit uninttuitive. Squeeze/unsqueeze is also inserted for gelu and relu, but this is to create a chain of unsqueeze/squeeze that will be eliminated by a later pass (see #8601 / D69673068). I think eventually it will be good to rewrite the pass to make shape management more explicit and self contained within the pass rather than inserting ops which are expected to be removed later on.
Differential Revision: D72480178