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

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

@@ -1,49 +1,64 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
-# Stage config for running GLM-Image with AR -> DiT disaggregation
+# Deploy config for running GLM-Image with AR -> DiT disaggregation
 # using NixlConnector for inter-stage transfer.
 #
 # Works for both single-node and cross-node deployments:
 # - Single-node: Use CUDA_VISIBLE_DEVICES to assign different GPUs to each stage process
 # - Cross-node: Each stage runs on a separate physical machine with independent GPU numbering
 # Both stages use devices: "0" so they always use the first visible GPU in their environment.
+#
+# Topology (stage types, model architectures, tokenizer/model subdirs, input
+# sources, and the AR->DiT input processor) is frozen in vLLM-Omni's
+# GLM_IMAGE_PIPELINE and must not be restated here. This file carries
+# deployment knobs and connector wiring only.
+pipeline: glm_image
+
+# DeployConfig defaults async_chunk to True, but GLM-Image declares no
+# async-chunk next-stage processor. Matches vLLM-Omni's own glm_image.yaml.
+async_chunk: false
+
+trust_remote_code: true
+enable_prefix_caching: false
+distributed_executor_backend: "mp"
 
-stage_args:
+# Connector definitions referenced by input_connectors/output_connectors below.
+connectors:
+  ar_to_dit_nixl:
+    name: NixlConnector
+    extra:
+      # Set transport backend(s) as needed for your environment, e.g. ["UCX"], ["LIBFABRIC"].
+      # If omitted, connector defaults apply.
+      # backends: ["UCX"]
+      # Optional tuning knobs; adjust/remove based on your NIXL deployment.
+      enforce_handshake_compat: false
+  dit_to_router:
+    # Final DIT stage -> Router connector.
+    # Enables multi-node: router and DIT on different machines (no SHM needed).
+    # Remove to fall back to single-node SHM.
+    name: NixlConnector
+    extra:
+      enforce_handshake_compat: false
+
+edges:
+  - from: 0
+    to: 1
+
+stages:
   # Stage 0: AR model (token generation)
   # Note: tensor_parallel_size controlled via AR_TP env var (default: 1).
   # Devices controlled via AR_GPUS env var (default: "0").
   - stage_id: 0
-    stage_type: llm
-    runtime:
-      process: true
-      devices: ${oc.env:AR_GPUS,0}
-      requires_multimodal_data: true
-    engine_args:
-      model_stage: ar
-      worker_type: ar
-      max_num_seqs: 1
-      model_arch: GlmImageForConditionalGeneration
-      model_subdir: vision_language_encoder
-      tokenizer_subdir: processor
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      tensor_parallel_size: ${oc.env:AR_TP,1}
-      gpu_memory_utilization: 0.85
-      enforce_eager: true
-      trust_remote_code: true
-      engine_output_type: token_ids
-      distributed_executor_backend: "mp"
-      enable_prefix_caching: false
-      max_num_batched_tokens: 32768
-    # Entry stage: input comes from the request, not from another stage.
-    # Declared explicitly because the stage_args format applies no schema
-    # defaults, and vLLM-Omni reads this field unconditionally.
-    engine_input_source: []
+    devices: ${oc.env:AR_GPUS,0}
+    tensor_parallel_size: ${oc.decode:${oc.env:AR_TP,1}}
+    max_num_seqs: 1
+    gpu_memory_utilization: 0.85
+    enforce_eager: true
+    max_num_batched_tokens: 32768
     # Explicit connector mapping for AR -> DiT
     output_connectors:
       to_stage_1: ar_to_dit_nixl
-    final_output: false
-    is_comprehension: true
     default_sampling_params:
       temperature: 0.9
       top_p: 0.75
@@ -57,59 +72,19 @@ stage_args:
   # Note: Devices and num_gpus controlled via DIT_GPUS and DIT_TP env vars.
   # oc.decode converts the env var string to proper Python type (int).
   - stage_id: 1
-    stage_type: diffusion
-    runtime:
-      process: true
-      devices: ${oc.env:DIT_GPUS,0}
-      requires_multimodal_data: true
-    engine_args:
-      model_stage: dit
-      max_num_seqs: 1
-      model_arch: GlmImagePipeline
-      num_gpus: ${oc.decode:${oc.env:DIT_TP,1}}
-      enforce_eager: true
-      trust_remote_code: true
-      distributed_executor_backend: "mp"
-    engine_input_source: [0]
-    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.glm_image.ar2diffusion
+    devices: ${oc.env:DIT_GPUS,0}
+    num_gpus: ${oc.decode:${oc.env:DIT_TP,1}}
+    max_num_seqs: 1
+    enforce_eager: true
     # Explicit connector mapping for AR -> DiT
     input_connectors:
       from_stage_0: ar_to_dit_nixl
     # Output connector for DIT -> Router (multi-node: replaces SHM)
     output_connectors:
       to_stage_router: dit_to_router
-    final_output: true
-    final_output_type: image
     default_sampling_params:
       seed: 42
       num_inference_steps: 50
       guidance_scale: 1.5
       height: 1024
       width: 1024
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 1
-  edges:
-    - from: 0
-      to: 1
-      window_size: -1
-  # Connector definitions referenced by input_connectors/output_connectors in stage_args.
-  connectors:
-    ar_to_dit_nixl:
-      name: NixlConnector
-      extra:
-        # Set transport backend(s) as needed for your environment, e.g. ["UCX"], ["LIBFABRIC"].
-        # If omitted, connector defaults apply.
-        # backends: ["UCX"]
-        # Optional tuning knobs; adjust/remove based on your NIXL deployment.
-        enforce_handshake_compat: false
-    dit_to_router:
-      # Final DIT stage -> Router connector.
-      # Enables multi-node: router and DIT on different machines (no SHM needed).
-      # Remove to fall back to single-node SHM.
-      name: NixlConnector
-      extra:
-        enforce_handshake_compat: false