fix(vllm): migrate GLM-Image NIXL config to the stages deploy schema

[GuanLuo]

Overview:

GLM-Image AR→DiT disaggregation over NIXL dies on the first
/v1/images/generations request. The AR stage's engine core exits, the request
returns HTTP 500, and the deployment does not recover:

File "vllm_omni/worker/gpu_model_runner.py", line 749, in _update_states
    resumed_token_ids = req_data.all_token_ids[req_id]
KeyError: '<uuid>-<8 hex>'

followed by Stage-0 replica-0 is dead and [AsyncOmni] Engine dead.

Summary

stage_configs/glm_image_nixl.yaml was the last config on vllm-omni's legacy
stage_args schema. That schema bypasses the deploy-config resolution in
vllm_omni/config/stage_config.py, which is what keeps the AR scheduler class
and the async_scheduling engine arg consistent: _resolve_scheduler() returns
OmniARScheduler for sync and OmniARAsyncScheduler for async, and the caller
derives the engine arg from whichever it selected.

Bypassing that, the config pinned scheduler_cls: OmniARScheduler — the
synchronous scheduler — and left async_scheduling unset, so the engine ran
async scheduling against a sync scheduler. The sync scheduler does not populate
CachedRequestData.all_token_ids, but the model runner's async-scheduling
resume path indexes it unconditionally. First request, KeyError, dead engine
core.

The scheduler_cls line dates to #10071 and was plausible when written: the
sync/async split in omni_ar_scheduler.py landed separately, keeping the old
name for the sync variant while async_scheduling defaults on. The only other
place that line appears upstream is a platforms.npu block, where it is
harmless because the NPU model runner has no all_token_ids resume path.

This migrates the file to the stages schema and drops scheduler_cls, so the
matched scheduler / async_scheduling pair is resolved by vllm-omni rather than
hand-declared.

Details:

#12709 fixed an earlier crash in this same file and named this cause exactly —
"the legacy stage_args config format applies no schema defaults", and "the SHM
variant is unaffected because it consumes vllm-omni's own packaged config in the
stages format, whose parser supplies the default". That fix supplied the one
missing default; this removes the reason defaults go missing at all.

Topology that GLM_IMAGE_PIPELINE already freezes is dropped rather than
restated: stage and model types, model/tokenizer subdirs, input sources, the
AR→DiT input processor, and final-output flags. That retires the
engine_input_source: [] workaround from #12709, since input sources now come
from the pipeline.

async_chunk: false is now explicit — DeployConfig defaults it to True and
GLM-Image declares no async-chunk next-stage processor. This matches vllm-omni's
own deploy/glm_image.yaml.

Deployment knobs, env-var device/TP selection, and sampling params carry over
unchanged. Top-level connectors:/edges: are the stages-schema spelling of
the previous runtime.connectors/runtime.edges; the connector initializer
normalizes both to the same internal shape. The dropped runtime.defaults
(window_size, max_inflight) has no consumer in vllm-omni.

The launch script needs no change: --stage-configs-path detects a stages
document and routes it to the deploy-config loader.

Validation

• pre-commit run --files examples/backends/vllm/launch/stage_configs/glm_image_nixl.yaml — clean.
• Config resolves through vllm-omni's deploy loader to the intended matched
  pair: scheduler_cls=OmniARAsyncScheduler with async_scheduling=True.
• Manual 2-GPU single-node run via disagg_omni_glm_image_nixl.sh: both stages
  start, /v1/images/generations returns an image, and the AR engine core no
  longer dies on the first request.

Where should the reviewer start?

Single file, -73/+46. The load-bearing change is the absence of
scheduler_cls — everything else is schema translation or removal of topology
the pipeline already owns.

One thing carried over unchanged and worth a second opinion: stage 1's
output_connectors: {to_stage_router: dit_to_router}. The to_stage_router key
appears nowhere in either this repo or vllm-omni, so the DiT→router hop may not
actually be on NIXL. Left as-is here to keep this change to the crash fix;
worth its own issue.

🤖 Generated with Claude Code

---

Summary by CodeRabbit

• Refactor
  • Updated the GLM image deployment configuration to the current deployment schema.
  • Simplified pipeline and stage configuration settings for improved consistency.
  • Consolidated connector, scheduling, caching, and execution settings at the deployment level.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: fe9c45e2-f898-4ec0-ae9b-20db043d0ac8

📥 Commits

Reviewing files that changed from the base of the PR and between bb5b141 and bd962a9.

📒 Files selected for processing (1)

• examples/backends/vllm/launch/stage_configs/glm_image_nixl.yaml

---

Walkthrough

The GLM-Image Nixl configuration now uses the deployment configuration schema. It defines deployment-level settings, connectors, and stage edges. The AR and DiT stages use direct deployment-stage fields.

Changes

GLM-Image deployment configuration

Layer / File(s)	Summary
Deployment topology and connector configuration examples/backends/vllm/launch/stage_configs/glm_image_nixl.yaml	The configuration references the frozen glm_image pipeline and defines deployment settings, Nixl connectors, and the AR-to-DiT edge.
Direct AR and DiT stage fields examples/backends/vllm/launch/stage_configs/glm_image_nixl.yaml	The AR and DiT stages move from nested legacy fields to direct device, resource, scheduling, connector, and sampling fields. Legacy model, runtime, processor, input-source, and final-output declarations are removed.

Estimated code review effort: 3 (Moderate) | ~20 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description covers the incident, technical cause, changes, review focus, and validation, but it omits the required Related Issues section.	Add the required Related Issues section and either link the relevant issue or confirm that no related issue exists.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the migration of the GLM-Image NIXL configuration to the stages deploy schema.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}

[devin-ai-integration]

Devin Review found 4 potential issues.

[ayushag-nv]

looks good