fix(vllm): migrate GLM-Image NIXL config to the stages deploy schema (cherry-pick #12984)

[GuanLuo]

Summary

• Cherry-pick of fix(vllm): migrate GLM-Image NIXL config to the stages deploy schema #12984 to release/1.4.0.
• GLM-Image AR→DiT disaggregation over NIXL dies on the first
  /v1/images/generations request: the AR stage's engine core exits with
  KeyError on the request id in gpu_model_runner._update_states, the request
  returns HTTP 500, and the deployment does not recover.
• stage_configs/glm_image_nixl.yaml was the last config on vLLM-Omni's legacy
  stage_args schema, which applies no schema defaults. It pinned
  scheduler_cls: OmniARScheduler (the synchronous scheduler) while leaving
  async_scheduling unset, so the engine ran async scheduling against a sync
  scheduler. The sync scheduler does not populate
  CachedRequestData.all_token_ids, which the async resume path indexes
  unconditionally.
• Migrating to the stages schema and dropping scheduler_cls lets vLLM-Omni
  resolve the matched scheduler / async_scheduling pair. Also retires the
  engine_input_source: [] workaround cherry-picked in fix(vllm): declare entry-stage engine_input_source in GLM-Image NIXL config (cherry-pick #12709) #12719, since input
  sources now come from the frozen pipeline.

Fixes DYN-3802

Original PR

• Main PR: fix(vllm): migrate GLM-Image NIXL config to the stages deploy schema #12984
• Merge commit: bed27ef
• Cherry-pick applied cleanly with no conflicts; the resulting file is
  byte-identical to the version merged to main.

Test plan

• CI passes
• 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

🤖 Generated with Claude Code

---

[devin-ai-integration]

Devin Review found 1 potential issue.

[devin-ai-integration]

🔍 oc.decode on AR_TP relies on lazy interpolation when AR_TP is unset (DiT node)

tensor_parallel_size now goes through ${oc.decode:...}. OmegaConf's oc.decode raises TypeError when handed a non-string (the unquoted literal 1 in an interpolation default is parsed as an int, not a string), so this expression only survives if either the env var is set or the value is never resolved. On the DiT node, examples/backends/vllm/launch/disagg_omni_glm_image_nixl.sh does not export AR_TP, so stage 0's field would resolve to the int default if the loader eagerly resolves the whole file. The identical pattern already existed for num_gpus: ${oc.decode:${oc.env:DIT_TP,1}} on the AR node before this PR (and the author reports a successful 2-GPU run), which is the reason this is not flagged as a bug — but a cross-node run where AR_TP is unset on the DiT machine is worth an explicit check, and quoting the defaults (${oc.env:AR_TP,'1'}) would make it unconditionally safe.

---

Was this helpful? React with 👍 or 👎 to provide feedback.