Bump vLLM version for DSV4 B200 disagg

[hjjq]

Note

Low Risk
Config-only version bump in a single benchmark recipe; main risk is benchmark/runtime compatibility with the pinned Dynamo wheel, not application code paths.

Overview
Updates the DeepSeek V4 B200 disaggregated low-latency Slurm recipe (disagg-b200-low-latency.yaml) from vLLM v0.20.1 / 0.20.0 to v0.23.0.

The bump is applied consistently on model.container, identity.container.image, and identity.frameworks.vllm so the launched container and recorded benchmark identity match the new stack. No changes to Dynamo wheel, vLLM CLI flags, Slurm layout, or benchmark parameters.

^{Reviewed by Cursor Bugbot for commit e821fa1. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit e821fa1. Configure here.}

[cursor]

Recipe image master mismatch

Medium Severity

This bump sets model.container and identity.container.image to vllm/vllm-openai:v0.23.0, but dsv4-fp4-b200-dynamo-vllm in nvidia-master.yaml still uses vllm/vllm-openai:v0.20.1. Those fields are meant to match, so sweep metadata and launcher container aliases can disagree with the image srtctl pulls from the recipe.

Additional Locations (1)

• benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency.yaml#L133-L137

 

^{Reviewed by Cursor Bugbot for commit e821fa1. Configure here.}

[RohitNagraj]

Superseded by #1899, which carries the same change on a branch in this repository so CI can run on it. Closing in favor of #1899.