[Misc] Rev DeepEP

[varun-sundar-rabindranath]

Purpose

Update DeepEP commit to DeepEP main (73b6ea4) to pick up additional hidden-size support that is useful for gpt-oss low-latency DP/EP

Test Plan

Run unit test test_deepep_moe.py locally

e2e:
Server : VLLM_ALL2ALL_BACKEND="deepep_high_throughput" VLLM_USE_DEEP_GEMM=1 canhazgpu run -g2 -- vllm serve Qwen/Qwen3-30B-A3B-FP8 --trust-remote-code --enable-expert-parallel --data-parallel-size 2 -no-enable-prefix-caching

Server: VLLM_ALL2ALL_BACKEND="deepep_low_latency" VLLM_USE_DEEP_GEMM=1 canhazgpu run -g2 -- vllm serve Qwen/Qwen3-30B-A3B-FP8 --trust-remote-code --enable-expert-parallel --data-parallel-size 2 --port 9010 --no-enable-prefix-caching

lm_eval: lm_eval --model local-completions --tasks gsm8k --model_args model=Qwen/Qwen3-30B-A3B-FP8,base_url=http://127.0.0.1:9010/v1/completions,num_concurrent=30,max_retries=3 --limit 100

Test Result

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  | 0.94|±  |0.0239|
|     |       |strict-match    |     5|exact_match|↑  | 0.94|±  |0.0239|

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  | 0.86|±  |0.0349|
|     |       |strict-match    |     5|exact_match|↑  | 0.93|±  |0.0256|

[mergify]

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

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

[varun-sundar-rabindranath]

cc @mgoin @tlrmchlsmth @bnellnm @yewentao256 @LucasWilkinson @SageMoore

[gemini-code-assist]

Code Review

This pull request updates the DeepEP dependency to a newer commit, enabling support for a hidden size of 3072. The changes are straightforward, involving an update to the commit hash in the installation script and adding the new size to the supported sizes list. My review includes a performance-related suggestion to use a frozenset for the list of supported hidden sizes, which is more efficient for membership checks in a low-latency context.

[gemini-code-assist]

For improved performance, consider using a frozenset for SUPPORTED_HIDDEN_SIZES. Membership testing (in) is O(1) on average for sets, while it is O(n) for lists. Given that this check is in a performance-sensitive path for low-latency kernels, this optimization is beneficial. A frozenset is suitable here as it's an immutable constant.

Suggested change

	SUPPORTED_HIDDEN_SIZES = [2048, 2560, 3072, 4096, 5120, 6144, 7168]
	SUPPORTED_HIDDEN_SIZES = frozenset([2048, 2560, 3072, 4096, 5120, 6144, 7168])

[mgoin]

Nice :) Please fix the conflict and we can ready

[yewentao256]

LGTM, thanks for the work!