Add Continuous Decoding support in GQA #21523
Merged
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
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
BowenBao
added a commit
to BowenBao/onnxruntime-genai
that referenced
this pull request
Aug 2, 2024
Results are validated with model-generate.py by using a int4 quantized model as the original model's assistant. The output sequence is the same and increased tps is observed. NOTE: Only MHA decoder only models, batch size 1, CPU, greedy select top is supported in this initial version. GQA needs microsoft/onnxruntime#21523 to support seqlen > 1 in token phase. * Updated builder.py to produce MHA graph that supports seqlen > 1 in token phase. * Introduce speculative decoding currently through a separate Generator class. This can be merged with existing Generator potentially on either API level or implementation level. * Extended various components for functionalities to support speculative search. Previously most methods are hardcoded assuming seqlen == 1 for token phase.
yufenglee
reviewed
Sep 9, 2024
yufenglee
reviewed
Sep 9, 2024
yufenglee
reviewed
Sep 9, 2024
yufenglee
reviewed
Sep 10, 2024
onnxruntime/contrib_ops/cuda/bert/group_query_attention_impl.cu
Outdated
Show resolved
Hide resolved
yufenglee
reviewed
Sep 10, 2024
onnxruntime/contrib_ops/cuda/bert/group_query_attention_impl.cu
Outdated
Show resolved
Hide resolved
yufenglee
reviewed
Sep 10, 2024
aciddelgado
changed the title
yufenglee
reviewed
Sep 11, 2024
onnxruntime/contrib_ops/cuda/bert/group_query_attention_impl.cu
Outdated
Show resolved
Hide resolved
tianleiwu
reviewed
Sep 11, 2024
tianleiwu
reviewed
Sep 11, 2024
tianleiwu
reviewed
Sep 11, 2024
onnxruntime/contrib_ops/cpu/bert/group_query_attention_helper.h
Outdated
Show resolved
Hide resolved
tianleiwu
reviewed
Sep 11, 2024
onnxruntime/contrib_ops/cuda/bert/group_query_attention_helper.h
Outdated
Show resolved
Hide resolved
|
Please fix PREfast warnings. |
|
Seems that you also need change this line: |
|
Python format failed. Please run lintrunner |
tianleiwu
reviewed
Sep 12, 2024
tianleiwu
reviewed
Sep 13, 2024
tianleiwu
approved these changes
Sep 13, 2024
BowenBao
added a commit
to BowenBao/onnxruntime-genai
that referenced
this pull request
Oct 15, 2024
Results are validated with model-generate.py by using a int4 quantized model as the original model's assistant. The output sequence is the same and increased tps is observed. NOTE: Only MHA decoder only models, batch size 1, CPU, greedy select top is supported in this initial version. GQA needs microsoft/onnxruntime#21523 to support seqlen > 1 in token phase. * Updated builder.py to produce MHA graph that supports seqlen > 1 in token phase. * Introduce speculative decoding currently through a separate Generator class. This can be merged with existing Generator potentially on either API level or implementation level. * Extended various components for functionalities to support speculative search. Previously most methods are hardcoded assuming seqlen == 1 for token phase.
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.
Description
This PR will add support for Continuous Decoding for batch_size = 1 input. From now on, GQA can take arbitrary length input using seqlens_k as total_sequence_length - 1 and the sequence length of qkv as new_sequence_length.
This change will not affect the default behavior of GQA
Motivation and Context
Prior to this change it was impossible to support sequence_length > 1 inputs when past context was given. This use case is essential to making continuous decoding work, which is one of our current efforts in ORT-GenAI.