[ET-VK][Llama] Apply XNNPACK partitoner as well when lowering to Vulkan #6830
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
## Context The final logit linear layer in the Transformer architecture has extremely large tensors, since the output and weight tensors will have a tensor with dim equal to the vocabulary size, which may be extremely large. Because of this, image textures cannot be used to execute the op when running with the Vulkan delegate, so an implementation using buffer based tensors must be used. Unfortunately, Vulkan does not have a performant implementation of linear with buffer based tensors at the moment. As a result, if this final linear layer is executed in Vulkan, model inference is extremely slow. ## Changes The below diff will prevent the final logit linear layer from being delegated to Vulkan by enforcing a GPU buffer limit. This diff modifies the export llama script to apply the XNNPACK partitioner after the Vulkan partitioner if lowering to Vulkan, to ensure that remaining ops will be accelerated with XNNPACK. 4 bit quantization will also apply an additional Quantizer after applying the Vulkan quantizer (which will skip the final logit linear layer) so that the final logit linear can be quantized as well. ## Long Term This is a temporary measure while an optimized buffer based linear implementation is developed. Once the Vulkan implementation achieves parity with XNNPACK, the final logit linear will be delegated to Vulkan once more. Differential Revision: [D65899827](https://our.internmc.facebook.com/intern/diff/D65899827/) [ghstack-poisoned]
pytorch-bot
bot
added
ciflow/periodic
module: vulkan
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6830
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: ❌ 1 New FailureAs of commit 5dcfa4f with merge base ecdc007 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
SS-JIA
added a commit
that referenced
this pull request
Nov 13, 2024
## Context The final logit linear layer in the Transformer architecture has extremely large tensors, since the output and weight tensors will have a tensor with dim equal to the vocabulary size, which may be extremely large. Because of this, image textures cannot be used to execute the op when running with the Vulkan delegate, so an implementation using buffer based tensors must be used. Unfortunately, Vulkan does not have a performant implementation of linear with buffer based tensors at the moment. As a result, if this final linear layer is executed in Vulkan, model inference is extremely slow. ## Changes The below diff will prevent the final logit linear layer from being delegated to Vulkan by enforcing a GPU buffer limit. This diff modifies the export llama script to apply the XNNPACK partitioner after the Vulkan partitioner if lowering to Vulkan, to ensure that remaining ops will be accelerated with XNNPACK. 4 bit quantization will also apply an additional Quantizer after applying the Vulkan quantizer (which will skip the final logit linear layer) so that the final logit linear can be quantized as well. ## Long Term This is a temporary measure while an optimized buffer based linear implementation is developed. Once the Vulkan implementation achieves parity with XNNPACK, the final logit linear will be delegated to Vulkan once more. Differential Revision: [D65899827](https://our.internmc.facebook.com/intern/diff/D65899827/) ghstack-source-id: 253391362 Pull Request resolved: #6830
|
This pull request was exported from Phabricator. Differential Revision: D65899827 |
…ing to Vulkan" ## Context The final logit linear layer in the Transformer architecture has extremely large tensors, since the output and weight tensors will have a tensor with dim equal to the vocabulary size, which may be extremely large. Because of this, image textures cannot be used to execute the op when running with the Vulkan delegate, so an implementation using buffer based tensors must be used. Unfortunately, Vulkan does not have a performant implementation of linear with buffer based tensors at the moment. As a result, if this final linear layer is executed in Vulkan, model inference is extremely slow. ## Changes The below diff will prevent the final logit linear layer from being delegated to Vulkan by enforcing a GPU buffer limit. This diff modifies the export llama script to apply the XNNPACK partitioner after the Vulkan partitioner if lowering to Vulkan, to ensure that remaining ops will be accelerated with XNNPACK. 4 bit quantization will also apply an additional Quantizer after applying the Vulkan quantizer (which will skip the final logit linear layer) so that the final logit linear can be quantized as well. ## Long Term This is a temporary measure while an optimized buffer based linear implementation is developed. Once the Vulkan implementation achieves parity with XNNPACK, the final logit linear will be delegated to Vulkan once more. Differential Revision: [D65899827](https://our.internmc.facebook.com/intern/diff/D65899827/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Nov 13, 2024
Pull Request resolved: #6830 ## Context The final logit linear layer in the Transformer architecture has extremely large tensors, since the output and weight tensors will have a tensor with dim equal to the vocabulary size, which may be extremely large. Because of this, image textures cannot be used to execute the op when running with the Vulkan delegate, so an implementation using buffer based tensors must be used. Unfortunately, Vulkan does not have a performant implementation of linear with buffer based tensors at the moment. As a result, if this final linear layer is executed in Vulkan, model inference is extremely slow. ## Changes The below diff will prevent the final logit linear layer from being delegated to Vulkan by enforcing a GPU buffer limit. This diff modifies the export llama script to apply the XNNPACK partitioner after the Vulkan partitioner if lowering to Vulkan, to ensure that remaining ops will be accelerated with XNNPACK. 4 bit quantization will also apply an additional Quantizer after applying the Vulkan quantizer (which will skip the final logit linear layer) so that the final logit linear can be quantized as well. ## Long Term This is a temporary measure while an optimized buffer based linear implementation is developed. Once the Vulkan implementation achieves parity with XNNPACK, the final logit linear will be delegated to Vulkan once more. Differential Revision: [D65899827](https://our.internmc.facebook.com/intern/diff/D65899827/) ghstack-source-id: 253425727
|
This pull request was exported from Phabricator. Differential Revision: D65899827 |
mcr229
approved these changes
Nov 13, 2024
…ing to Vulkan" ## Context The final logit linear layer in the Transformer architecture has extremely large tensors, since the output and weight tensors will have a tensor with dim equal to the vocabulary size, which may be extremely large. Because of this, image textures cannot be used to execute the op when running with the Vulkan delegate, so an implementation using buffer based tensors must be used. Unfortunately, Vulkan does not have a performant implementation of linear with buffer based tensors at the moment. As a result, if this final linear layer is executed in Vulkan, model inference is extremely slow. ## Changes The below diff will prevent the final logit linear layer from being delegated to Vulkan by enforcing a GPU buffer limit. This diff modifies the export llama script to apply the XNNPACK partitioner after the Vulkan partitioner if lowering to Vulkan, to ensure that remaining ops will be accelerated with XNNPACK. 4 bit quantization will also apply an additional Quantizer after applying the Vulkan quantizer (which will skip the final logit linear layer) so that the final logit linear can be quantized as well. ## Long Term This is a temporary measure while an optimized buffer based linear implementation is developed. Once the Vulkan implementation achieves parity with XNNPACK, the final logit linear will be delegated to Vulkan once more. Differential Revision: [D65899827](https://our.internmc.facebook.com/intern/diff/D65899827/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Nov 14, 2024
Pull Request resolved: #6830 ## Context The final logit linear layer in the Transformer architecture has extremely large tensors, since the output and weight tensors will have a tensor with dim equal to the vocabulary size, which may be extremely large. Because of this, image textures cannot be used to execute the op when running with the Vulkan delegate, so an implementation using buffer based tensors must be used. Unfortunately, Vulkan does not have a performant implementation of linear with buffer based tensors at the moment. As a result, if this final linear layer is executed in Vulkan, model inference is extremely slow. ## Changes The below diff will prevent the final logit linear layer from being delegated to Vulkan by enforcing a GPU buffer limit. This diff modifies the export llama script to apply the XNNPACK partitioner after the Vulkan partitioner if lowering to Vulkan, to ensure that remaining ops will be accelerated with XNNPACK. 4 bit quantization will also apply an additional Quantizer after applying the Vulkan quantizer (which will skip the final logit linear layer) so that the final logit linear can be quantized as well. ## Long Term This is a temporary measure while an optimized buffer based linear implementation is developed. Once the Vulkan implementation achieves parity with XNNPACK, the final logit linear will be delegated to Vulkan once more. ghstack-source-id: 253568942 Differential Revision: [D65899827](https://our.internmc.facebook.com/intern/diff/D65899827/)
|
This pull request was exported from Phabricator. Differential Revision: D65899827 |
facebook-github-bot
merged commit b6da372
into
gh/SS-JIA/147/base
78 of 81 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
SS-JIA
added a commit
that referenced
this pull request
Nov 14, 2024
…an (#6857) * [ET-VK] Enforce GPU buffer limit when partitioning Pull Request resolved: #6829 ## Context In Vulkan, there is a limit on the number of elements a GPU buffer can have. If a GPU buffer exceeds this limit, then the API will either produce an error or undefined behaviour will ensue. ## Changes Along with `texture_limits`, introduce a configurable `buffer_limit` entry in the partitioner configuration. ghstack-source-id: 253568943 Differential Revision: [D65899828](https://our.internmc.facebook.com/intern/diff/D65899828/) * [ET-VK][Llama] Apply XNNPACK partitoner as well when lowering to Vulkan Pull Request resolved: #6830 ## Context The final logit linear layer in the Transformer architecture has extremely large tensors, since the output and weight tensors will have a tensor with dim equal to the vocabulary size, which may be extremely large. Because of this, image textures cannot be used to execute the op when running with the Vulkan delegate, so an implementation using buffer based tensors must be used. Unfortunately, Vulkan does not have a performant implementation of linear with buffer based tensors at the moment. As a result, if this final linear layer is executed in Vulkan, model inference is extremely slow. ## Changes The below diff will prevent the final logit linear layer from being delegated to Vulkan by enforcing a GPU buffer limit. This diff modifies the export llama script to apply the XNNPACK partitioner after the Vulkan partitioner if lowering to Vulkan, to ensure that remaining ops will be accelerated with XNNPACK. 4 bit quantization will also apply an additional Quantizer after applying the Vulkan quantizer (which will skip the final logit linear layer) so that the final logit linear can be quantized as well. ## Long Term This is a temporary measure while an optimized buffer based linear implementation is developed. Once the Vulkan implementation achieves parity with XNNPACK, the final logit linear will be delegated to Vulkan once more. ghstack-source-id: 253568942 Differential Revision: [D65899827](https://our.internmc.facebook.com/intern/diff/D65899827/) --------- Co-authored-by: Stephen Jia <ssjia@meta.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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
The final logit linear layer in the Transformer architecture has extremely large tensors, since the output and weight tensors will have a tensor with dim equal to the vocabulary size, which may be extremely large. Because of this, image textures cannot be used to execute the op when running with the Vulkan delegate, so an implementation using buffer based tensors must be used.
Unfortunately, Vulkan does not have a performant implementation of linear with buffer based tensors at the moment. As a result, if this final linear layer is executed in Vulkan, model inference is extremely slow.
Changes
The below diff will prevent the final logit linear layer from being delegated to Vulkan by enforcing a GPU buffer limit.
This diff modifies the export llama script to apply the XNNPACK partitioner after the Vulkan partitioner if lowering to Vulkan, to ensure that remaining ops will be accelerated with XNNPACK. 4 bit quantization will also apply an additional Quantizer after applying the Vulkan quantizer (which will skip the final logit linear layer) so that the final logit linear can be quantized as well.
Long Term
This is a temporary measure while an optimized buffer based linear implementation is developed. Once the Vulkan implementation achieves parity with XNNPACK, the final logit linear will be delegated to Vulkan once more.
Differential Revision: D65899827