[PERF] Symmetric memory allreduce

[ilmarkov]

Add an alternative to custom_allreduce and nccl on cuda - pytorch symmetric memory

Enabled by environment variable VLLM_ALLREDUCE_USE_SYMM_MEM=1.
Improves performance of TP allreduce for middle size input.

Bounds input sizes for custom allreduce as long as performance of two shot custom allreduce appears to be worse than nccl or pytorch symmetric memory based allreduce.

Max sizes for various world_sizes for custom allreduce and symmetric memory were chosen based on empirical results.

For world_sizes 2 and 4 pytorch two shot allreduce is used, for world sizes 6 and 8 pytorch multimem_all_reduce

Benchmark results:

Server: VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-70B-Instruct --disable-log-requests --no-enable-prefix-caching -tp $tp
Client:

DURATION_SECONDS=60; \
vllm bench serve \
--model meta-llama/Llama-3.1-70B-Instruct \
--dataset-name sonnet  \
--dataset-path benchmarks/sonnet.txt \
--request-rate "$qps" \
--num-prompts $((DURATION_SECONDS * qps))`

On Blackwell, B200

TP=4:

Baseline:

QPS	Mean TTFT (ms)	Median TTFT (ms)	Mean TPOT (ms)	Median TPOT (ms)
1	45.021	44.248	12.880	12.866
5	48.574	46.048	14.120	14.084
10	55.482	51.188	17.674	17.710

PR:

QPS	Mean TTFT (ms)	Median TTFT (ms)	Mean TPOT (ms)	Median TPOT (ms)
1	41.456	40.226	12.681	12.666
5	46.916	45.233	13.754	13.683
10	54.584	50.382	17.186	17.296

TP=8

Baseline:

QPS	Mean TTFT (ms)	Median TTFT (ms)	Mean TPOT (ms)	Median TPOT (ms)
1	42.838	41.157	10.286	10.240
5	48.772	46.882	13.583	13.924
10	54.444	51.688	16.513	16.592

PR:

QPS	Mean TTFT (ms)	Median TTFT (ms)	Mean TPOT (ms)	Median TPOT (ms)
1	37.005	34.039	10.254	10.223
5	44.329	44.387	13.353	13.762
10	50.864	48.993	15.442	15.404

Up to 8% TTFT speedup for TP=4
From 7 to 10% TTFT improvement, and up to 5-7% TPOT improvement for TP=8.

Validation:

VLLM_ALLREDUCE_USE_SYMM_MEM=1 lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-70B-Instruct,tensor_parallel_size=4 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto --limit 100

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  | 0.95|±  |0.0219|
|     |       |strict-match    |     5|exact_match|↑  | 0.90|±  |0.0302|

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

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 either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Summary of Changes

Hello @ilmarkov, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request integrates a new, performance-optimized allreduce implementation into vLLM, utilizing PyTorch's symmetric memory features for CUDA. The primary goal is to enhance the efficiency of tensor parallel allreduce operations, particularly for mid-sized inputs, by providing a faster alternative that can be conditionally enabled via an environment variable. The changes also involve adapting existing allreduce strategies to work synergistically with this new method.

Highlights

• New Allreduce Implementation: Introduced a SymmMemCommunicator class that leverages PyTorch's symmetric memory capabilities for efficient allreduce operations on CUDA devices, providing an alternative to existing CustomAllreduce and NCCL methods.
• Conditional Activation and Strategy Selection: The new symmetric memory allreduce can be enabled via the VLLM_ALLREDUCE_USE_SYMM_MEM environment variable. The implementation intelligently selects between 'two-shot' (for world sizes 2 and 4) and 'multimem' (for world sizes 6 and 8) allreduce strategies based on the world size, and is prioritized for suitable input tensor sizes.
• Performance Optimization and Benchmarking: This change aims to improve the performance of tensor parallel allreduce for middle-sized inputs, with benchmarks showing 7-10% improvement in Mean TTFT and up to 5-7% improvement in Mean TPOT for Llama-3.1-70B-Instruct with TP=4 and TP=8.
• Adaptive Custom Allreduce Bounds: The maximum supported input sizes for the existing CustomAllreduce implementation are adjusted when symmetric memory is enabled, ensuring that the most performant allreduce method is used for different input ranges.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces a new all-reduce implementation using PyTorch's symmetric memory, improving performance for medium-sized tensors. The code is well-structured, but I suggest increasing the flexibility of SymmMemCommunicator by allowing it to work with different dtypes and improving the accuracy of comments in custom_all_reduce.py.

[gemini-code-assist]

The SymmMemCommunicator is hardcoded to use torch.bfloat16, limiting its use with models using other dtypes. Consider initializing buffers based on the input tensor's dtype during the first all_reduce call to increase flexibility.

[gemini-code-assist]

The check not symm_mem_comm.disabled is redundant because should_use_symm_mem already performs this check. Removing the redundant check will make the code more concise.

Suggested change

	if symm_mem_comm is not None and not symm_mem_comm.disabled and \
	symm_mem_comm.should_use_symm_mem(input_):
	if symm_mem_comm is not None and \
	symm_mem_comm.should_use_symm_mem(input_):

[gemini-code-assist]

The comments for the sizes in the _MAX_SIZES dictionary are inaccurate. Update the comments to reflect the correct sizes.

Suggested change

	# Max sizes for each world size in case symmetric memory is available
	_MAX_SIZES = {
	2: 2 * MiB, # 1 MB
	4: 2 * MiB, # 1 MB
	6: MiB, # 512 KB
	8: MiB // 2, # 512 KB
	}
	# Max sizes for each world size in case symmetric memory is available
	_MAX_SIZES = {
	2: 2 * MiB, # 2 MiB
	4: 2 * MiB, # 2 MiB
	6: MiB, # 1 MiB
	8: MiB // 2, # 512 KiB
	}

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ilmarkov.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork