[Fix] Remove MoRI-IO patches from vLLM Disagg benchmarks

[simondanielsson]

These patches were upstreamed in vllm-project/vllm#40344 so we can use the nightly image instead.

Switching to nightly also requires us to:

• rename a2a backend from mori to mori_low_latency
• change MORI_READ_MODE=1 envvar to a read_mode=1 flag.

Run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27334107848

Results from this run are very similar to the existing Kimi vllm disagg results, as expected

---

Note

Medium Risk
Changes affect multi-node prefill/decode KV transfer and MoE all2all for production benchmark paths; behavior now depends on a nightly image rather than battle-tested in-tree patches, though validation noted similar results to prior runs.

Overview
MI355X vLLM disaggregated Kimi K2.5 and MiniMax M2.5 configs now use a pinned ROCm nightly image that includes upstream MoRI-IO fixes (vllm#40344), with a TODO to move to a stable release later. Scenario additional-settings no longer pass VLLM_MORIIO_CONNECTOR_READ_MODE.

setup_deps.sh drops the large set of runtime vLLM/MoRI-IO and scheduler patches (save/load KV timeouts, transfer timeouts, read-mode scheduler fixes, idle KV reaper); vLLM-disagg startup is now recipe deps, amd-quark, and UCX/RIXL paths only.

Serving wiring changes for the new stack: --all2all-backend mori → mori_low_latency in models_vllm.yaml; read_mode: true in MoRIIO kv_connector_extra_config in server_vllm.sh instead of the env var; job.slurm stops injecting that env and bumps the default vllm-router image. submit.sh no longer exports the read-mode env. perf-changelog.yaml documents the config-key updates.

^{Reviewed by Cursor Bugbot for commit 556d926. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[functionstackx]

can u have an specific night hash pinned here here instead of just the generic "nightly"

[simondanielsson]

We can! The caveat (and why I didn't pin it here) was that the pinned nightlies are pruned from docker hub after 14 days so the config will then cease working in 2 weeks. WDYT?

[functionstackx]

can u have an specific night hash pinned here here instead of just the generic "nightly"

[claude]

Claude finished @simondanielsson's task in 2m 25s —— View job

---

PR Review: Remove MoRI-IO patches from vLLM Disagg benchmarks

• Gather context and read changed files
• Validate master config changes and perf-changelog
• Check for bugs, logic errors, and breaking changes
• Post review findings

Summary

Clean removal of ~555 lines of MoRI-IO monkey patches now that they've been upstreamed in vllm#40344. The three changes (backend rename mori → mori_low_latency, env var → read_mode config flag, patch removal) are consistent and the CI run confirms they work.

One blocking issue: perf-changelog.yaml must be updated since amd-master.yaml was modified — see inline comment.

The unpinned nightly image tags are already under discussion with the reviewer.

[simondanielsson]

Run is succeeding, and perf is very similar to the existing Kimi vllm-disagg results (with some noise of course)

[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 3c94a6f. Configure here.}

[functionstackx]

@simondanielsson is there is validated run? if so, lgtm

[simondanielsson]

Before merge I will

1. update the pinned vllm nightlies again (since vllm imges are pruned from docker hub every 2 weeks)
2. Re-run CI
3. Post link here

EDIT: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27334107848/job/80753697924

[functionstackx]

thanks @simondanielsson

[functionstackx]

appreipcate your help on getting first class MoRI collective working @simondanielsson ! <3 with the high conc hanging fixed now, the experience for end users will be much better now!

[simondanielsson]

Run succeeded: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27334107848

cc: @chunfangamd should be ready for merge