-
Notifications
You must be signed in to change notification settings - Fork 116
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
CUDA graph compilation #154
Merged
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
4 tasks
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Outdated suggestions cannot be applied.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.
This PR adds support for compiling the model into a static CUDA graph. See Accelerating PyTorch with CUDA Graphs for more details on CUDA graphs and how they can reduce latency.
To enable this (experimental) feature:
There is a tradeoff to be aware of when using CUDA graphs, namely that it increases memory overhead by 3-10GB depending on model size. However, the observed decrease in latency can be as much as 50%, so if you don't need to run with very large batch sizes and are more latency constrained than throughput, this is a very compelling feature to enable.
In practice, CUDA graphs are most useful in cases where there are excess GPU flops available, such as during decoding. As such, we do not use the compiled version of the model during prefill, only during the decoding steps. Which means in practice that the benefits of enabling compilation will be most pronounced when generating longer sequences (for which more time is spent during decoding).
Current limitations:
If any of these conditions are not met, then LoRAX will fallback to using eager execution for the batch.
Thanks to folks on the Punica team for updating kernels to support graph tracing. Additionally, we modified kernels to support padding with -1 (necessary for CUDA graph's requirement that input shapes be constant across batches).
Comparison:
gpt2-medium, time to generate 100 tokens:
no adapter
1 adapter (rank 16)