[Feature] JoyAI-VL-Interaction streaming interaction serving layer by lishunyang12 · Pull Request #4575 · vllm-project/vllm-omni

lishunyang12 · 2026-06-19T20:06:21Z

The development of JoyAI-VL 8B was made possible through close collaboration within the JoyAI team

https://github.com/jd-opensource/JoyAI-VL-Interaction

chatgpt-codex-connector · 2026-06-19T20:06:29Z

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

Add vllm_omni/experimental/fullduplex/: a vLLM-Omni-native serving layer for proactive streaming video-language interaction (per-second speak / silence / delegate). Includes a generic full-duplex core (DuplexRuntime/Session/Adapter) and a JoyVL implementation: per-tick decision policy, 3-tier summary memory, OpenAI-compatible orchestrator, and pluggable ASR / TTS / delegation bridges (chat, text-to-image, image-edit, and a routing bridge). Ships the JD recipe, demo glue under examples/, and unit tests. Signed-off-by: lishunyang12 <125541396+lishunyang12@users.noreply.github.com>

linyueqian

Reviewed end to end. I reconstructed the package and ran it (this is the public version of the framework I looked at earlier).

Verified

tests/fullduplex 21/21 pass under the new experimental/ layout.
Serving stack imports and the FastAPI app builds. I drove InteractionSession.step both against a fake backend and against the real JoyAI-VL-Interaction-Preview model: forced-silence-before-query (skips the model call), response dedup -> silence, delegate submit+fold, async 3-tier memory, and skip_special_tokens=False all behave correctly. End to end the orchestrator reproduces the speak/silence decisions faithfully.
app.py and the rtsp/ scripts are clean and talk to the orchestrator over HTTP only.

Good calls

Moving under experimental/, and the new README.md "Scope" note that core/ + adapter.py are a demonstration (exercised by the tests), not the path the HTTP orchestrator currently drives. That clears up the main thing I would have flagged: the serving path is joyvl/serving/ -> decision/+memory/ directly, while DuplexRuntime/JoyVLDuplexAdapter are referenced only in test_runtime.py. Documenting it plainly is the right call.
SessionManager.reset() now calls session.reset(), so in-flight consolidation tasks are cancelled on reset/evict.
RTSP ingestion is a nice way around browser/WebRTC transport limits.

Minor

[low] _completion_response emits a non-standard top-level streamingharness key alongside interaction; it duplicates interaction.memory + timing. Consider renaming or dropping it.
[low] The PR description is a single line. For a 3.4k-line feature a short summary plus the reproduce steps (they already exist in the recipe) in the body would help reviewers; the PR body is the first thing people read.
[low, env-dependent] Deploy note worth adding to the recipe: on a host without nvcc/ninja, a plain vllm serve of the 8B can crash engine-core in the FlashInfer sampler JIT (FileNotFoundError: 'ninja') during profile_run. VLLM_USE_FLASHINFER_SAMPLER=0 (or installing ninja) works around it.

Overall this is in good shape and the logic is correct against the model. Nothing blocking from me.

Gaohan123

Great work! Let's unify the api-server and full-duplex runtime in the following PRs

Gaohan123 · 2026-06-20T01:20:46Z

@@ -0,0 +1,131 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+


I think later we can integrate asr and tts into stage_client of vllm-omni to better utillize features like async_chunk in follow-up PRs. Like PR #4257

Will keep in mind.

Gaohan123 · 2026-06-20T01:32:29Z

+  --limit-mm-per-prompt '{"image":256,"video":1}'
+
+# 2. Interaction orchestrator (OpenAI-compatible, :8070)
+python -m vllm_omni.experimental.fullduplex.joyvl.serving.server --port 8070 \


The transitional solution is ok. I suggest we need to integrate fulllduplex server into vllm-omni api-server in following PRs

Agreed — this is a transitional standalone orchestrator; folding the full-duplex server into the vLLM-Omni api-server is the right next step and I'll take it in a follow-up PR.

Gaohan123 · 2026-06-20T01:33:14Z

+orchestrator over its HTTP API only:
+
+```bash
+pip install vllm-omni[demo]          # gradio + opencv-python + requests


Better use uv with consistency

Gaohan123 · 2026-06-20T01:36:05Z

+
+from vllm_omni.experimental.fullduplex.joyvl.decision.output_parser import parse_action
+from vllm_omni.experimental.fullduplex.joyvl.memory.brain import InteractionBrain
+


The pytest marks is incorrect. Please refer to https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/ci/tests_markers/

Gaohan123 · 2026-06-20T01:36:39Z

+    OpenAIDelegationBridge,
+)
+from vllm_omni.experimental.fullduplex.joyvl.decision.policy import JoyVLPolicy, sample_frames
+


Fixed in 6984e11 — same core_model / cpu pytestmark added here.

Gaohan123 · 2026-06-20T01:36:55Z

+from vllm_omni.experimental.fullduplex.core.runtime import DuplexRuntime
+from vllm_omni.experimental.fullduplex.core.session import DuplexSession, DuplexSessionConfig
+from vllm_omni.experimental.fullduplex.joyvl.adapter import JoyVLDuplexAdapter
+


Fixed in 6984e11

Gaohan123 · 2026-06-20T01:51:57Z

+
+## Scope
+
+The runnable serving path is `joyvl/serving/` driving `decision/` + `memory/` directly.


Important: left for following PRs

…io app Reference the upstream JoyAI-VL-Interaction WebUI (services/webui) in front of the orchestrator instead of shipping a separate Gradio app. Remove app.py and gradio from the demo extra. Signed-off-by: lishunyang12 <125541396+lishunyang12@users.noreply.github.com>

hsliuustc0106

Posted the inline review notes we discussed.

hsliuustc0106 · 2026-06-20T14:02:27Z

+
+    async def _start_response(self, prev: asyncio.Task | None, emit: Emit) -> asyncio.Task:
+        if prev is not None and not prev.done():
+            await prev


[blocking] _start_response() waits for the previous response to finish before the input loop can continue. That means a second response trigger can delay later response.cancel / barge-in events until the old response has fully streamed, which breaks the full-duplex cancellation contract. I reproduced this with a slow adapter: the first response emitted all deltas before the cancel was handled. This should cancel/epoch-stale the active response, or otherwise keep the input loop able to process cancel events immediately.

Fixed in 7abcc07. _start_response no longer awaits the previous response's full stream — it stales the active response (barge-in) and cancels it, then starts the new one, so the input loop can service cancel/barge-in events immediately. Added test_runtime_new_response_supersedes_inflight_without_blocking (slow adapter, second trigger supersedes the first; only the latest response completes).

hsliuustc0106 · 2026-06-20T14:02:27Z

+            )
+        if url:
+            return chat_bridge(url)
+        return StubDelegationBridge()


[blocking] With the current defaults, omitting --delegation-backend-url still enables delegation through StubDelegationBridge(). That contradicts the recipe guidance that leaving the URL unset keeps delegation off, and in real use it can fold fake stub answers into session memory. The default should probably be no delegation unless a real backend is configured, with the stub limited to tests/demo-only paths.

Fixed in 7abcc07. _build_delegation now returns None (delegation off) unless a real backend is configured; StubDelegationBridge is opt-in only via --delegation-kind stub. So omitting --delegation-backend-url truly keeps delegation off — no stub answers folded into memory — matching the recipe. Added test_build_delegation_off_unless_backend_configured.

hsliuustc0106 · 2026-06-20T14:02:27Z

+    async def submit(self, question: str, note: str, frames: list[tuple[str, str]]) -> str:
+        self._counter += 1
+        task_id = f"deleg-{self._counter}"
+        self._tasks[task_id] = asyncio.create_task(self._solve(question, note, frames))


[non-blocking] These background tasks are removed only when poll() is called after completion. If a session is reset/evicted or the client disconnects before polling again, the shared bridge keeps the finished task/result around indefinitely. Consider a done callback or explicit reset/cancel path so long-lived demo servers do not accumulate abandoned delegation tasks.

Fixed in 7abcc07. Added cancel(task_id) to every bridge (and the protocol); brain.reset() now cancels all pending delegations, so reset/evicted sessions don't leave finished-but-unpolled tasks on the shared bridge. Added test_delegation_cancel_drops_task.

hsliuustc0106 · 2026-06-20T14:14:40Z

+
+
+def create_app(config: InteractionConfig) -> FastAPI:
+    app = FastAPI(title="vLLM-Omni Interaction Server")


No shutdown/lifespan hook, so the AsyncOpenAI/httpx clients and pending background tasks leak on shutdown (and on every uvicorn reload).

create_app only registers routes and main() just runs uvicorn. OpenAIBackend._client and the AsyncOpenAI held by OpenAIDelegationBridge are never closed, and the asyncio.create_tasks in delegation._tasks and session._consolidating are never cancelled — so on graceful stop you get "Unclosed client session" warnings and orphaned tasks.

Suggestion: add a lifespan to FastAPI(...) (or @app.on_event("shutdown")) that awaits an aclose() on the manager, and give OpenAIBackend / the delegation bridges an async def aclose() that closes the client and cancels + awaits the pending tasks.

hsliuustc0106

Adding two follow-up lifecycle/race findings after another pass.

hsliuustc0106 · 2026-06-20T14:14:42Z

+
+def create_app(config: InteractionConfig) -> FastAPI:
+    app = FastAPI(title="vLLM-Omni Interaction Server")
+    manager = SessionManager(config)


[should-fix] This app never registers a shutdown/lifespan hook, so the shared async resources owned by SessionManager are left for process teardown: OpenAIBackend._client, OpenAIDelegationBridge._client, delegation _tasks, and each session's _consolidating tasks. On uvicorn shutdown/reload this can leave unclosed httpx clients and pending tasks. Please add a manager aclose() and call it from FastAPI lifespan/shutdown; the backends/bridges should close their clients and cancel+await pending tasks.

Fixed in 1ee7489. create_app now registers a FastAPI lifespan that calls SessionManager.aclose() on shutdown/reload: it cancels+awaits pending delegation and consolidation tasks and closes the AsyncOpenAI/httpx clients. Added aclose() to ModelBackend/OpenAIBackend, Summarizer, and every DelegationBridge (+protocol). Test: test_manager_aclose_closes_without_error (idempotent).

hsliuustc0106 · 2026-06-20T14:14:42Z

+        async with self._locks[session_id]:
+            return await session.step(frames, query)
+
+    def reset(self, session_id: str) -> None:


[should-fix] reset() mutates and drops the session without acquiring the per-session lock. A concurrent /v1/chat/completions can be inside session.step() under that lock while /reset calls session.reset() and removes the lock/session from the manager, so the in-flight request continues on a reset/discarded session. Making reset async and acquiring the same lock before resetting/removing the session would make reset semantics deterministic. InteractionSession.reset() should also await cancelled consolidation tasks instead of clearing the set immediately.

Fixed in 1ee7489. SessionManager.reset/set_persona are now async and acquire the per-session lock, so they wait for any in-flight step() on that session instead of mutating/dropping it mid-request; step captures session+lock together (no await between) so a concurrent reset can't swap them out. InteractionSession.reset()/aclose() now await asyncio.gather(*consolidating, return_exceptions=True) after cancelling, instead of clearing the set immediately. Test: test_session_reset_cancels_and_awaits_consolidation.

hsliuustc0106 · 2026-06-20T14:15:07Z

+        async with self._locks[session_id]:
+            return await session.step(frames, query)
+
+    def reset(self, session_id: str) -> None:


reset() / set_persona() mutate session state without the per-session lock, so they race an in-flight step() on the same session.

manager.reset() pops _sessions / _locks and calls session.reset() while another coroutine may be awaiting session.step() under that lock. session.reset() also does task.cancel() on the consolidation tasks without awaiting, so they keep running (against a discarded brain) until the next loop tick. Impact is bounded — the session is dropped — but the in-flight request continues on a reset session and its lock is already gone from the dict.

Suggestion: make manager.reset / set_persona acquire self._locks[session_id] (i.e. make them async), and in session.reset() do await asyncio.gather(*self._consolidating, return_exceptions=True) after cancelling.

hsliuustc0106 · 2026-06-20T14:19:12Z

+            )
+        if url:
+            return chat_bridge(url)
+        return StubDelegationBridge()


Defaulting delegation on with the stub contradicts the recipe and folds fake text into memory.

enable_delegation defaults to True and --no-delegation is off by default, so a deployment that omits --delegation-backend-url lands here: StubDelegationBridge. That stub's poll() returns a ready digest of "(background result for: …) — stub digest; wire a real brain here.", which fold_delegations then appends to qa_history and feeds into the memory prefix for every later turn. But the recipe tells users: "Omit --delegation-backend-url to keep delegation off."

Suggestion: when no backend URL is configured, keep delegation off (return None / don't build a bridge, equivalent to --no-delegation), and reserve StubDelegationBridge for tests/demo.

hsliuustc0106 · 2026-06-20T14:19:15Z

+    async def submit(self, question: str, note: str, frames: list[tuple[str, str]]) -> str:
+        self._counter += 1
+        task_id = f"deleg-{self._counter}"
+        self._tasks[task_id] = asyncio.create_task(self._solve(question, note, frames))


Submitted delegation tasks live in self._tasks until poll() removes them, so abandoned tasks accumulate for the life of the process.

The bridge is a singleton on the SessionManager, shared across all sessions. poll() only runs while a session keeps calling fold_delegations; if a session is reset/evicted (which clears brain._pending_delegations but not this dict) or a client disconnects before folding, the entry — and the still-running create_task — stays here forever.

Suggestion: add a done_callback that drops the entry once the task finishes, and evict by id on reset; or track tasks per-session and cancel+evict on reset.

hsliuustc0106 · 2026-06-20T14:19:16Z

+            await task
+        self.session.state = DuplexState.CLOSED
+
+    async def _start_response(self, prev: asyncio.Task | None, emit: Emit) -> asyncio.Task:


_start_response does await prev inside the input loop, so a second response trigger blocks all later events — including RESPONSE_CANCEL / barge-in — until the previous response finishes streaming. That breaks the full-duplex contract (you can't interrupt a response): the first slow response drains all its deltas before a cancel is processed.

This is the core/ demonstration path (driven only by test_runtime.py, not the HTTP serving layer), so it's not on the shipped path today — but it's a real bug worth fixing before this runtime informs the api-server merge. Fix: cancel/replace the in-flight response instead of await prev, or move the response stream off the input loop.

…off by default, cancel leaked tasks - runtime: a new response now stales+cancels the in-flight one instead of awaiting its full stream, so the input loop can service cancel/barge-in immediately - serving: _build_delegation returns None unless a real backend is configured; StubDelegationBridge is opt-in via --delegation-kind stub (no more fake answers folded into memory) - delegation: add cancel(task_id) to all bridges; brain.reset() cancels pending delegations so reset/evicted sessions don't leak tasks - tests: supersede-without-blocking, delegation-off-by-default, cancel-drops-task Signed-off-by: lishunyang12 <125541396+lishunyang12@users.noreply.github.com>

lishunyang12 · 2026-06-20T14:22:06Z

Thanks @hsliuustc0106

…eset/set_persona against in-flight step - create_app now registers a FastAPI lifespan that calls SessionManager.aclose(): closes AsyncOpenAI/httpx clients (backend, summarizer, OpenAIDelegationBridge) and cancels+awaits pending delegation/consolidation tasks on shutdown/reload - SessionManager.reset/set_persona are async and acquire the per-session lock, so they no longer race an in-flight step(); reset removes the session only after the lock is held - InteractionSession.reset()/aclose() cancel AND await consolidation tasks instead of clearing the set immediately - add aclose() to ModelBackend/OpenAIBackend, Summarizer, and all DelegationBridges (+protocol) - tests: delegation aclose, session reset awaits consolidation, manager aclose idempotent Signed-off-by: lishunyang12 <125541396+lishunyang12@users.noreply.github.com>

lishunyang12 · 2026-06-20T14:32:51Z

Thanks again @hsliuustc0106 — both follow-up lifecycle findings addressed in 1ee7489:

lishunyang12 · 2026-06-20T14:39:55Z

Note on the docs/readthedocs.org:vllm-omni check failure — it's a transient network timeout in the RTD build environment, not caused by this PR.

Signed-off-by: lishunyang <lishunyang12@163.com>

princepride

LGTM

@Gaohan123

… async_chunk TTS Implements the remaining review follow-ups from vllm-project#4575 (@Gaohan123). B — unify the api-server and full-duplex transport: - joyvl/serving exposes a mountable router (build_interaction_router) and mount_fullduplex(app, config), so the interaction layer runs standalone or is mounted onto the vLLM-Omni OpenAI api-server (opt-in via VLLM_OMNI_FULLDUPLEX_BACKEND_URL). - New WebSocket /v1/fullduplex/stream (DuplexStreamHandler) drives the core DuplexRuntime over the event protocol -> streaming response deltas with barge-in. HTTP (one decision/request) and WS (streaming) share the same InteractionSession engine. C — consume the staged async_chunk TTS: - bridges/speech.py factors the vLLM-Omni staged speech-stream protocol (/v1/audio/speech/stream, qwen3_tts.yaml async_chunk) into a reusable, tested relay; the example tts_bridge uses it and forwards each audio chunk as it arrives. ASR documented as one-shot (staged streaming ASR is the remaining deeper integration). Docs: recipe + fullduplex README updated for the WS transport, the api-server mount, and the staged async_chunk TTS. Tests: tests/fullduplex 34 passed (HTTP route, WS streaming, mount, speech relay). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: lishunyang12 <125541396+lishunyang12@users.noreply.github.com>

…llm-project#4575) Signed-off-by: lishunyang12 <125541396+lishunyang12@users.noreply.github.com> Signed-off-by: Yueqian Lin <linyueqian@outlook.com> Signed-off-by: lishunyang <lishunyang12@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Yueqian Lin <linyueqian@outlook.com>

lishunyang12 requested review from Gaohan123, david6666666, hsliuustc0106, yenuo26 and ywang96 as code owners June 19, 2026 20:06

lishunyang12 force-pushed the joyvl-fullduplex-serving branch 3 times, most recently from e17bc73 to fd77975 Compare June 19, 2026 20:25

lishunyang12 added ready label to trigger buildkite CI merge-test label to trigger buildkite merge test CI labels Jun 19, 2026

lishunyang12 force-pushed the joyvl-fullduplex-serving branch 6 times, most recently from f62f13b to 93df450 Compare June 19, 2026 20:58

lishunyang12 force-pushed the joyvl-fullduplex-serving branch from 93df450 to ba01d34 Compare June 19, 2026 21:01

hsliuustc0106 added the high priority high priority issue, needs to be done asap label Jun 20, 2026

Merge branch 'main' into joyvl-fullduplex-serving

f2442f7

hsliuustc0106 added ready label to trigger buildkite CI merge-test label to trigger buildkite merge test CI and removed ready label to trigger buildkite CI merge-test label to trigger buildkite merge test CI labels Jun 20, 2026

linyueqian reviewed Jun 20, 2026

View reviewed changes

lishunyang12 added merge-test label to trigger buildkite merge test CI and removed ready label to trigger buildkite CI merge-test label to trigger buildkite merge test CI labels Jun 20, 2026

Gaohan123 reviewed Jun 20, 2026

View reviewed changes

lishunyang12 force-pushed the joyvl-fullduplex-serving branch from a12a8e8 to 5ef25db Compare June 20, 2026 10:34

lishunyang12 enabled auto-merge (squash) June 20, 2026 10:52

lishunyang12 disabled auto-merge June 20, 2026 10:52

Merge branch 'main' into joyvl-fullduplex-serving

c139dc1

hsliuustc0106 reviewed Jun 20, 2026

View reviewed changes

[Docs] Exclude experimental package from API docs build

f598022

Signed-off-by: lishunyang <lishunyang12@163.com>

princepride approved these changes Jun 21, 2026

View reviewed changes

lishunyang12 enabled auto-merge (squash) June 21, 2026 06:53

princepride added the ready label to trigger buildkite CI label Jun 21, 2026

lishunyang12 merged commit e241005 into vllm-project:main Jun 21, 2026
8 checks passed

This was referenced Jun 22, 2026

[Doc] JoyVL recipe: describe model as retrained Qwen3-VL, not vanilla #4622

Closed

[Fix] JoyVL serving: align with reference engine (bounded long-term memory, timestamps, max_pixels) + recipe updates #4623

Merged

yuxinyuan mentioned this pull request Jul 15, 2026

[Bug]: Regression on dependency declaration #5123

Closed

1 task done

This was referenced Jul 28, 2026

Clean stage proc rebase chickeyton/vllm-omni#8

Open

Clean stage proc rebase chickeyton/vllm-omni#7

Open

Sy0307 mentioned this pull request Aug 22, 2026

[RFC][Tracking] Full-Duplex Serving Roadmap and PR Status #6494

Open

		@@ -0,0 +1,131 @@
		# SPDX-License-Identifier: Apache-2.0
		# SPDX-FileCopyrightText: Copyright contributors to the vLLM project


		from vllm_omni.experimental.fullduplex.joyvl.decision.output_parser import parse_action
		from vllm_omni.experimental.fullduplex.joyvl.memory.brain import InteractionBrain


		## Scope

		The runnable serving path is `joyvl/serving/` driving `decision/` + `memory/` directly.



		def create_app(config: InteractionConfig) -> FastAPI:
		app = FastAPI(title="vLLM-Omni Interaction Server")

Conversation

lishunyang12 commented Jun 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chatgpt-codex-connector Bot commented Jun 19, 2026

Uh oh!

linyueqian left a comment

Choose a reason for hiding this comment

Uh oh!

Gaohan123 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lishunyang12 Jun 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lishunyang12 Jun 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lishunyang12 Jun 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lishunyang12 Jun 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hsliuustc0106 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hsliuustc0106 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lishunyang12 commented Jun 19, 2026 •

edited

Loading

lishunyang12 Jun 20, 2026 •

edited

Loading

lishunyang12 Jun 20, 2026 •

edited

Loading

lishunyang12 Jun 20, 2026 •

edited

Loading

lishunyang12 Jun 20, 2026 •

edited

Loading

lishunyang12 commented Jun 20, 2026 •

edited

Loading

lishunyang12 commented Jun 20, 2026 •

edited

Loading

lishunyang12 commented Jun 20, 2026 •

edited

Loading