[ET-VK] Migrate ops to use DynamicDispatchNode
#11312
Merged
Conversation
This file contains hidden or 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
## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Jun 3, 2025
## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/) ghstack-source-id: 287884655 Pull Request resolved: #11312
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/11312
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ⏳ No Failures, 1 PendingAs of commit ac41e02 with merge base b5a6362 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D75878398 |
junpi3
approved these changes
Jun 3, 2025
## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/) [ghstack-poisoned]
This was referenced Jun 3, 2025
SS-JIA
added a commit
that referenced
this pull request
Jun 3, 2025
Pull Request resolved: #11312 ## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/) ghstack-source-id: 287945690
|
This pull request was exported from Phabricator. Differential Revision: D75878398 |
## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Jun 3, 2025
Pull Request resolved: #11312 ## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. ghstack-source-id: 287967830 Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/)
|
This pull request was exported from Phabricator. Differential Revision: D75878398 |
## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Jun 4, 2025
Pull Request resolved: #11312 ## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. ghstack-source-id: 288040418 Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/)
|
This pull request was exported from Phabricator. Differential Revision: D75878398 |
## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Jun 4, 2025
Pull Request resolved: #11312 ## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. ghstack-source-id: 288045732 Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/)
|
This pull request was exported from Phabricator. Differential Revision: D75878398 |
## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Jun 4, 2025
Pull Request resolved: #11312 ## Changes * Migrate operators that are used in the llama model to use `DynamicDispatchNode` instead of `DispatchNode` ## Motivation `DynamicDispatchNode` is a subclass of `DispatchNode` that allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary. Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches. ghstack-source-id: 288057588 Differential Revision: [D75878398](https://our.internmc.facebook.com/intern/diff/D75878398/)
|
This pull request was exported from Phabricator. Differential Revision: D75878398 |
facebook-github-bot
merged commit 2106ab9
into
gh/SS-JIA/235/base
97 of 98 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
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):
DynamicDispatchNode#11312Changes
DynamicDispatchNodeinstead ofDispatchNodeMotivation
DynamicDispatchNodeis a subclass ofDispatchNodethat allows dynamic selection of compute shaders, global and local work group sizing whenever the command buffer is encoded. This is critical for ensuring optimum performance when input shapes are dynamic, since it allows operators to select the best compute shader for the input conditions and also to adjust global work group sizing to launch the minimum number of work groups necessary.Without this change, performance of llama 3.2 1B with dynamic shapes enabled is terrible (< 1 tok/s) because global work group sizing is determined based on maximum tensor sizes, which is based on the maximum sequence length. In practice, the sequence length dimension of tensors (even during the prefill phase) will not approach the maximum. This results in a lot of inactive threads launched during compute shader dispatches.
Differential Revision: D75878398