[Misc] treewide: remove shm_threshold_bytes from deploy config

[NickCao]

Purpose

Since #5096, SharedMemoryConnector no longer reads shm_threshold_bytes, all payloads use shared memory.

Test Plan

vLLM Version: 0.29.0

vLLM-Omni Commit: 23d8700

N/A, only removal of dead code, this key is not referenced in any python code.

Test Result

N/A

BEFORE SUBMITTING: read CONTRIBUTING.md and run the precheck-pr skill with the code agent for a self-check against project conventions.
(anything written below this line will be removed by GitHub Actions)

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/vllm_omni_config.md, docs/design/module/entrypoints.md.

Module owners: @alex-jw-brooks @lishunyang12 @linyueqian

Routing: @alex-jw-brooks via module of the changed files, semantic router, CODEOWNERS; @lishunyang12 via module of the changed files, semantic router, CODEOWNERS; @linyueqian via module of the changed files

@NickCao, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[NickCao]

@vllm-omni-review-bot

[vllm-omni-review-bot]

Omni ReviewBot routing record

Assigned Strict under experiment vllm-omni-strict-5050-20260829.

[vllm-omni-review-bot]

Omni ReviewBot review

Scan:

Category	Result
Tests / verification	1 finding(s) below
Security	no finding reported
Docs / comments	1 finding(s) below
Behavior / compatibility	2 finding(s) below
Correctness	1 finding(s) below

Validated:

• [resolved] Concern 'YAML still gates small payloads to non-SHM': connector never branched on threshold since #5096; residual is CLI/examples only, not deploy
• [resolved] Concern 'sibling deploy YAMLs still pin 65536': at this head zero deploy files retain the key
• [resolved] Concern that docs still document the threshold: stage_configs.md:120 and shared_memory_connector.md:164-167 already describe always-SHM; residual is CLI/examples only.
• [claim-verified] SharedMemoryConnector no longer reads shm_threshold_bytes: shm_connector.py put/get always serialize via shm_write_bytes/shm_read_bytes; init only uses stage_id (vllm_omni/distributed/omni_connectors/connectors/shm_connector.py:27-61).
• [validated] Blast radius of YAML removal: audex_s2s(_30b), audex_tta(_30b), audex_tts(_30b), minimax_music3(_2gpu), nemotron_labs_voicechat(_duplex|streaming), personaplex — remaining extra keys (codec_streaming, connector_get*, codec_chunk_frames, etc.) intact; voicechat.yaml correctly drops empty extra:.
• [validated] personaplex.yaml:22-34 same pattern; remaining extras are chunk-adapter knobs, not SHM size thresholds.

Merged outcome keeps one minor finding: after removing inert shm_threshold_bytes from all 12 deploy YAMLs, the public --shm-threshold-bytes serve CLI flag and matching offline example/README surfaces remain and are silently dropped, contradicting the PR’s treewide/Python-unused claim—remove them here or explicitly track follow-up. All connector/docs/YAML structural validations and tombstone/sibling-residue candidates are dropped as non-issues or excluded polish.

Checked, no defect found:

• vllm_omni/deploy/audex_s2s.yaml:16 — After deleting shm_threshold_bytes, audex_s2s SharedMemoryConnector.extra is coherent (codec_streaming + wait/codec knobs only); SharedMemoryConnector.put has no size/threshold gate.
• vllm_omni/deploy/audex_s2s_30b.yaml:11 — 30B s2s connector.extra matches 2B after shm_threshold_bytes removal (both start at codec_streaming: true; threshold absent).
• vllm_omni/deploy/audex_tta.yaml:23 — After deleting shm_threshold_bytes, audex_tta.yaml SharedMemoryConnector.extra still keeps connector_get_* waits and audiocodec_*; audex_tta_30b.yaml mirrors the same keys.
• vllm_omni/deploy/audex_tts.yaml:28 — audex_tts.yaml and audex_tts_30b.yaml only drop shm_threshold_bytes; codec_streaming and codec_* extras remain (connector_get_* waits also retained).
• vllm_omni/deploy/minimax_music3.yaml:24 — minimax_music3 and minimax_music3_2gpu keep window/wait commentary after codec_streaming; only shm_threshold_bytes was removed.
• vllm_omni/deploy/nemotron_labs_voicechat.yaml:27 — Removing the sole extra key correctly yields a name-only SharedMemoryConnector; same shape as bagel/hunyuan/dynin deploy siblings.

Verdict: COMMENT

Findings

• [P2] This diff deletes shm_threshold_bytes: 65536 from 12 deploy YAMLs because Sha… — vllm_omni/entrypoints/cli/serve.py
  This diff deletes shm_threshold_bytes: 65536 from 12 deploy YAMLs because SharedMemoryConnector no longer honors a size gate (put always serialize_obj → shm_write_bytes). The same dead knob remains as public --shm-threshold-bytes in vllm_omni/entrypoints/cli/serve.py:396-399 (default 65536; help still describes an SHM size threshold) and in eight offline end2end.py parsers plus examples/offline_inference/covo_audio/README.md:50. OrchestratorArgs has no such field (unchanged by this diff); underscore-form shm_threshold_bytes is absent treewide, so the hyphen CLI flag cannot affect the connector. Please remove the serve CLI arg and the offline/example mirrors in this PR, or explicitly narrow the treewide claim and track follow-up—otherwise any Test Plan assertion that the key is not referenced in Python is false and the most user-visible dead surface survives.

Evidence: unchanged by this diff, present in the PR-time tree: serve.py:396-399 "--shm-threshold-bytes", / type=int, / default=65536, / help="The threshold for the shared memory size.",; shm_connector.py:45-51 payload = self.serialize_obj(data) then meta = shm_write_bytes(payload, name=put_key) with no size-threshold branch; OrchestratorArgs (arg_utils.py:490-610) defines cross-stage fields through tokenizer with no shm_threshold_bytes; tree grep: zero shm_threshold_bytes literals; hyphen-form remains in serve.py + examples/offline_inference/{qwen3_omni,qwen2_5_omni,qwen3_tts,ming_tts,minicpmo,step_audio2,mimo_audio,covo_audio}/end2end.py and covo_audio/README.md:50 | --shm-threshold-bytes| |65536 | Shared memory threshold (bytes) |.

• [P2] Confirmed: SharedMemoryConnector does not branch on size — put() always SHM-w… — ``
  Confirmed: SharedMemoryConnector does not branch on size — put() always SHM-writes (shm_connector.py:45-55). Residual is dead user-facing advertising only: `--shm-threshold-bytes` in serve.py:396-399 (and offline end2end parsers); no `shm_threshold_bytes` consumer remains in vllm_omni (unchanged by this deploy-YAML cleanup diff).

Evidence: unchanged by this diff, present in the PR-time tree: shm_connector.py:45-55 payload = self.serialize_obj(data) / size = len(payload) / meta = shm_write_bytes(payload, name=put_key) / metadata = {"shm": meta, "size": size} — no threshold compare before SHM write; serve.py:396-399 --shm-threshold-bytes / default=65536 / help="The threshold for the shared memory size." — CLI advertising only; repo-wide shm_threshold_bytes identifier grep empty (no transport consumer).

• [P2] Resolved: deploy YAML shm_threshold_bytes never gated transport—SharedMemoryCon… — ``
  Resolved: deploy YAML shm_threshold_bytes never gated transport—SharedMemoryConnector always SHM-writes (shm_connector.py put path). Residual --shm-threshold-bytes is CLI/examples-only and unconsumed; not deploy.

Evidence: unchanged by this diff, present in the PR-time tree: vllm_omni/distributed/omni_connectors/connectors/shm_connector.py:45-51 payload = self.serialize_obj(data) / size = len(payload) / … / meta = shm_write_bytes(payload, name=put_key) — no threshold branch, always SHM. Residual CLI only: vllm_omni/entrypoints/cli/serve.py:396-399 --shm-threshold-bytes / default=65536 / help="The threshold for the shared memory size." — argparse definition with no vllm_omni consumer of shm_threshold_bytes. Deploy: repo-wide grep for shm_threshold_bytes is empty after this PR's YAML removals.

• [P2] Resolved: at this head, zero files retain shm_threshold_bytes (repo-wide grep… — ``
  Resolved: at this head, zero files retain shm_threshold_bytes (repo-wide grep empty); sibling deploy YAMLs do not still pin 65536 for that key.

Evidence: PR-time tree grep shm_threshold_bytes / shm_threshold → 0 matches (unchanged absence outside the deletion diff). Representative post-removal connectors: vllm_omni/deploy/audex_s2s.yaml:16-18 extra: / codec_streaming: true / connector_get_sleep_s: 0.01 — no shm_threshold_bytes; vllm_omni/deploy/nemotron_labs_voicechat.yaml:26-29 connector_of_shared_memory: / name: SharedMemoryConnector then blank line then stages: — entire extra: (incl. former shm_threshold_bytes) removed.

• **[P2] Resolved: connector docs already describe always-SHM (stage_configs.md:120 None** — stage_configs.md:120Resolved: connector docs already describe always-SHM (stage_configs.md:120None. All payloads use shared memory.; shared_memory_connector.md:164-167 no inline path/size threshold). No in-tree shm_threshold_bytes keys; residual is CLI/examples --shm-threshold-bytes` (e.g. serve.py) plus a stale dynin_omni_multiconnector.yaml:31 comment.

Evidence: unchanged by this diff, present in the PR-time tree: docs/configuration/stage_configs.md:120 | SharedMemoryConnector | Same-host KV transfer between stages (default for bundled YAMLs). | None. All payloads use shared memory. |; docs/design/feature/omni_connectors/shared_memory_connector.md:164-167 #### 6.1 All Payloads Use Shared Memory / put() writes every serialized payload to shared memory. The connector has no / inline-payload path or size threshold.; residual outside connector docs: vllm_omni/entrypoints/cli/serve.py:396-399 --shm-threshold-bytes / help="The threshold for the shared memory size."

[linyueqian]

I added ready to get a lane on this, and build 15413 at 23d87005 failed only Simple · Other Test after 9m21s, with every other step green. That step does not read deploy profiles, and the failure shape (single step, nine to ten minutes) is the NIXL connector timeout flake tracked in #7615 that has hit four other builds today, so I do not read it as anything this PR did. Buildkite will not rebuild a finished build at the same commit, and this branch is also 69 commits behind main, so the quickest path is a rebase onto current main and a push; the ready label is already on, which means the push itself starts no build, so ping here and I will re-fire the lane. The review is done on my side (the key is dead since #5096 and the merge result against main is exactly the twelve deletions), so this merges as soon as a lane at the new head is green.

[NickCao]

Rebased, check that no shm_threshold_bytes reference is left.

[linyueqian]

Approving at 23d87005. This removes the shm_threshold_bytes: 65536 line from twelve deploy profiles (and the extra: block it was the only child of, in one of them). I checked the claim rather than the description: since #5096 nothing under vllm_omni/distributed or anywhere else in the package reads shm_threshold_bytes, so the key was inert configuration that would only mislead the next person tuning a connector, and after this PR the string no longer appears anywhere in the tree. Pure YAML removal, no behaviour change. Validation: static read of the diff and a tree-wide search on the PR head; no PR code executed.

Rebased head 57b24021 carries the same twelve-line, config-only removal as the reviewed 23d87005 (patches are byte-identical) and now sits on current main with zero commits behind, so the earlier read stands. Approving once the general lane is green on this head, then merging.

[linyueqian]

Checked as asked: no shm_threshold_bytes reference remains anywhere under vllm_omni/, tests/ or docs/ at 57b24021. The red on build 15515 is Tiny Model Tests · Diffusion (base) only; those tests read no deploy profiles, and the same step passed on build 15505 for a branch on the same main, so it is not this change. Buildkite will not rebuild the same commit from a label cycle, so either push an empty commit for a fresh lane or merge directly on your side; the approval is on the content.