-
-
Notifications
You must be signed in to change notification settings - Fork 4.3k
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
PP comm optimization: replace send with partial send + allgather #6695
PP comm optimization: replace send with partial send + allgather #6695
Conversation
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge). To run full CI, you can do one of these:
🚀 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for the PR! I have few questions:
- Do you have benchmark results?
- At the first glance it seems always better to enable this mechanism. Have you observed performance regression with this PR in certain circumstances? Like when the tensor size is small the allgather overhead is not negligible (just guess)?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think this is a good approach. This is really similar to the optimization from Section 4.1 of https://arxiv.org/abs/2104.04473, except that the sender pipeline stage will still do an AllReduce instead of a ReduceScatter. Is that right?
I think the ReduceScatter -> Send to the next pipeline stage -> AllGather is an optimization that we should explore along these lines in the future but it's going to require a lot more software engineering than this :)
LGTM from a first pass. Thanks for the contribution! |
Hey! @youkaichao any blockers for merge here? |
This one should have better improvement when your inter-node bandwidth is not high. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Does it make sense to add an integration test? vLLM does have some 4 GPU CI runs in buildkite, so it should be possible to add one. Otherwise this code won't be exercised and may break.
Other than that, LGTM
+1. Should be just a matter of adding a parameter to existing pipeline tests |
Sorry for the delay, just notice this PR and the message. The idea looks good to me, we use Implementation wise, can we use this by default? As long as vllm/vllm/distributed/parallel_state.py Line 569 in 9f0e69b
def send_tensor_dict(
self,
tensor_dict: Dict[str, Union[torch.Tensor, Any]],
dst: Optional[int] = None,
+ all_gather_group=Optional[GroupCoordinator],
)
...
if tensor.is_cpu:
# use metadata_group for CPU tensors
- torch.distributed.send(tensor,
+ torch.distributed.send(tensor[slice],
dst=self.ranks[dst],
group=metadata_group)
+ # add allgather with `all_gather_group.cpu_group`
else:
# use group for GPU tensors
- torch.distributed.send(tensor,
+ torch.distributed.send(tensor[slice],
dst=self.ranks[dst],
group=group)
+ # add allgather with `all_gather_group.device_group`
# change `recv_tensor_dict` accordingly And update this line vllm/vllm/worker/worker_base.py Line 279 in 9f0e69b
to |
Done and tested |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for the great work!
merge as failing tests are unrelated. |
…m-project#6695) Co-authored-by: Aurick Qiao <aurick.qiao@snowflake.com>
Submitting on behalf of @zhisbug as he is on break. Communication optimization for pipeline parallelism, we observed 5% improvement in throughput for llama 3.1 405b.
BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE
PR Checklist (Click to Expand)
Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.
PR Title and Classification
Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:
[Bugfix]
for bug fixes.[CI/Build]
for build or continuous integration improvements.[Doc]
for documentation fixes and improvements.[Model]
for adding a new model or improving an existing model. Model name should appear in the title.[Frontend]
For changes on the vLLM frontend (e.g., OpenAI API server,LLM
class, etc.)[Kernel]
for changes affecting CUDA kernels or other compute kernels.[Core]
for changes in the core vLLM logic (e.g.,LLMEngine
,AsyncLLMEngine
,Scheduler
, etc.)[Hardware][Vendor]
for hardware-specific changes. Vendor name should appear in the prefix (e.g.,[Hardware][AMD]
).[Misc]
for PRs that do not fit the above categories. Please use this sparingly.Note: If the PR spans more than one category, please include all relevant prefixes.
Code Quality
The PR need to meet the following code quality standards:
format.sh
to format your code.docs/source/
if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.Notes for Large Changes
Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with
rfc-required
and might not go through the PR.What to Expect for the Reviews
The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:
action-required
label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.Thank You
Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!