[DML EP] Support partial rotary embedding#22417
Merged
PatriceVignola merged 3 commits intomainfrom Oct 16, 2024
Merged
Conversation
BLSharda
added a commit
to BLSharda/onnxruntime-genai
that referenced
this pull request
Oct 12, 2024
Patrice has raised PR that fixed RotaryEmbedding for partial factor in ORT so workaround is not needed anymore. microsoft/onnxruntime#22417
Contributor
It would change the existing style of the test that we didn't add. All we did was re-enabling a DML test.
sumitsays
approved these changes
Oct 16, 2024
BLSharda
added a commit
to BLSharda/onnxruntime-genai
that referenced
this pull request
Oct 17, 2024
Patrice has raised PR that fixed RotaryEmbedding for partial factor in ORT so workaround is not needed anymore. microsoft/onnxruntime#22417
guschmue
pushed a commit
that referenced
this pull request
Oct 18, 2024
### Description This adds support for partial RotaryEmbedding to DML. Essentially, partial RotaryEmbedding simply consists of doing the rotary embedding calculation on a subregion of the input tensor of as if its head size was `rotary_embedding_dim`, while leaving the second part of the tensor (i.e. `head_size - rotary_embedding_dim`) alone. To achieve this, all we need to do is follow the following steps: 1. Split the tensor into 2 parts 2. Run the rotary embedding algorithm on the first part, just like we were doing before on the entire tensor 3. Join the 2 parts back together Since we're leaving the middle part intact, the RotaryEmbedding fusion will still be done within DML. Also, the concat at the end is essentially free because DML optimizes it out and directly allocate the result of RotaryEmbedding at the right place. The only overhead here is the splitting of the tensor at the beginning, which we should eventually make part of the RotaryEmbedding fusion within DML. ### Motivation and Context This fix allows us to correctly run models that have a `partial_rotary_factor` setting in huggingface, including Nvidia's Nemotron: https://huggingface.co/nvidia/Nemotron-Mini-4B-Instruct
apsonawane
pushed a commit
that referenced
this pull request
Oct 21, 2024
### Description This adds support for partial RotaryEmbedding to DML. Essentially, partial RotaryEmbedding simply consists of doing the rotary embedding calculation on a subregion of the input tensor of as if its head size was `rotary_embedding_dim`, while leaving the second part of the tensor (i.e. `head_size - rotary_embedding_dim`) alone. To achieve this, all we need to do is follow the following steps: 1. Split the tensor into 2 parts 2. Run the rotary embedding algorithm on the first part, just like we were doing before on the entire tensor 3. Join the 2 parts back together Since we're leaving the middle part intact, the RotaryEmbedding fusion will still be done within DML. Also, the concat at the end is essentially free because DML optimizes it out and directly allocate the result of RotaryEmbedding at the right place. The only overhead here is the splitting of the tensor at the beginning, which we should eventually make part of the RotaryEmbedding fusion within DML. ### Motivation and Context This fix allows us to correctly run models that have a `partial_rotary_factor` setting in huggingface, including Nvidia's Nemotron: https://huggingface.co/nvidia/Nemotron-Mini-4B-Instruct
Contributor
|
This PR has been cherry-picked into the |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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.
Description
This adds support for partial RotaryEmbedding to DML. Essentially, partial RotaryEmbedding simply consists of doing the rotary embedding calculation on a subregion of the input tensor of as if its head size was
rotary_embedding_dim, while leaving the second part of the tensor (i.e.head_size - rotary_embedding_dim) alone.To achieve this, all we need to do is follow the following steps:
Since we're leaving the middle part intact, the RotaryEmbedding fusion will still be done within DML. Also, the concat at the end is essentially free because DML optimizes it out and directly allocate the result of RotaryEmbedding at the right place. The only overhead here is the splitting of the tensor at the beginning, which we should eventually make part of the RotaryEmbedding fusion within DML.
Motivation and Context
This fix allows us to correctly run models that have a
partial_rotary_factorsetting in huggingface, including Nvidia's Nemotron: https://huggingface.co/nvidia/Nemotron-Mini-4B-Instruct