-
-
Notifications
You must be signed in to change notification settings - Fork 4.9k
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
[Core][Distributed] Refactor ipc buffer init in CustomAllreduce #10030
Conversation
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can do one of these:
🚀 |
e782d66
to
99a9c6e
Compare
cc @tlrmchlsmth |
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.
@hanzhi713 and @youkaichao could you share a few more details on what's going on?
Looks like this is related to #9815 -- is the idea to be more stable across Pytorch versions? Do you see any downsides to this?
Yes, we want to rely less on internal API to prevent future breaking. A downside to the current approach is that it doesn't support An alternative design I see is to do the one-time allocation of IPC-enabled buffers ourselves through CUDA C++ (i.e. cudaMalloc + ipc handle calls). |
#10064 will make this pr easier. we don't need to depend on pytorch's internal apis |
please merge main to use the functionality from #10064 |
Signed-off-by: Hanzhi Zhou <[email protected]>
99a9c6e
to
59b70cb
Compare
Signed-off-by: Hanzhi Zhou <[email protected]>
59b70cb
to
a1517a7
Compare
Done. @youkaichao PTAL, thanks! |
Signed-off-by: Hanzhi Zhou <[email protected]>
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 contribution! do you have some updated perf numbers? I assume it should not affect the performance.
nit: please fix the format. |
There're no perf changes from the C++. I can run the python side benchmark to be sure. |
Signed-off-by: Hanzhi Zhou <[email protected]>
@youkaichao I can confirm that there's no perf difference with |
there are some errors in the ci actually @hanzhi713 |
logger.info("Registering %d cuda graph addresses", len(offset)) | ||
all_data = [None] * dist.get_world_size(group=self.group) | ||
dist.all_gather_object(all_data, (handle, offset), group=self.group) |
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.
use broadcast as before in _gather_ipc_meta
?
code for reference:
for i, rank in enumerate(ranks):
dist.broadcast_object_list(all_data[i],
src=rank,
group=self.group,
device="cpu")
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.
Yeah it looks like we still need to use broadcast here
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.
to locally test it, run pytest -v -s tests/basic_correctness/test_basic_correctness.py::test_models_distributed[facebook/opt-125m-mp--A100]
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.
My old benchmark script had --enforce-eager
there which didn't catch this
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.
Should be fixed now.
Signed-off-by: Hanzhi Zhou <[email protected]>
@hanzhi713 thanks again for the great contribution! |
…-project#10030) Signed-off-by: Hanzhi Zhou <[email protected]> Signed-off-by: Isotr0py <[email protected]>
…-project#10030) Signed-off-by: Hanzhi Zhou <[email protected]> Signed-off-by: OmerD <[email protected]>
…-project#10030) Signed-off-by: Hanzhi Zhou <[email protected]> Signed-off-by: Loc Huynh <[email protected]>
…-project#10030) Signed-off-by: Hanzhi Zhou <[email protected]> Signed-off-by: Sumit Dubey <[email protected]>
…-project#10030) Signed-off-by: Hanzhi Zhou <[email protected]>
…-project#10030) Signed-off-by: Hanzhi Zhou <[email protected]> Signed-off-by: Maxime Fournioux <[email protected]>
…-project#10030) Signed-off-by: Hanzhi Zhou <[email protected]> Signed-off-by: Tyler Michael Smith <[email protected]>
…-project#10030) Signed-off-by: Hanzhi Zhou <[email protected]>
As discussed with @youkaichao, we use cuda API to share tensors instead of replying
_share_cuda_
, which won't break with expandable segment or future pytorch upgrade.Additional changes:
all_reduce
methods.vector<int64_t>
instead of pytorch tensor for handle during cuda graph ipc registration process. The use of Tensor was introduced in [Kernel][Misc] Use TORCH_LIBRARY instead of PYBIND11_MODULE for custom ops #5047. We sticked to non-tensor type to remove this complexity.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.Adding or changing kernels
Each custom kernel needs a schema and one or more implementations to be registered with PyTorch.
Tensors
require meta-functions. Meta-functions should be implemented and registered in python so that dynamic dims can be handled automatically. See above documents for a description of meta-functions.torch.libary.opcheck()
to test the function registration and meta-function for any registered ops. Seetests/kernels
for examples.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!