diff --git a/INSTALL.md b/INSTALL.md
index aa9dd5f84..5596afc72 100644
--- a/INSTALL.md
+++ b/INSTALL.md
@@ -11,7 +11,7 @@ virtualenv, never `--all-extras`. `train` and `infer` do not pull a rollout engi
 | Engine extra | PyTorch | CUDA |
 |---|---|---|
 | `vllm` (vLLM + vLLM-Omni) | `2.13.0+cu130` | 13.0 |
-| `sglang` | `2.11.0+cu130` | 13.0 |
+| `sglang` | `2.13.0+cu130` | 13.0 |
 
 SGLang's wheel needs glibc >= 2.34. Put NVIDIA's CUDA 13 forward-compat
 libraries on `LD_LIBRARY_PATH` before launch; the launchers do not do this.
@@ -20,7 +20,7 @@ libraries on `LD_LIBRARY_PATH` before launch; the launchers do not do this.
 
 ```bash
 uv venv --python 3.12 --seed .venv && source .venv/bin/activate
-uv pip install -e ".[vllm,train,infer]" --prerelease=allow
+uv pip install -e ".[vllm,train,infer]"
 ```
 
 ## sglang
@@ -47,7 +47,7 @@ Limit build parallelism with `MAX_JOBS` if host RAM is constrained, then install
 the SGLang extra:
 
 ```bash
-uv pip install -e ".[sglang,train,infer]" --prerelease=allow
+uv pip install -e ".[sglang,train,infer]"
 ```
 
 ## Extras
@@ -58,7 +58,7 @@ uv pip install -e ".[sglang,train,infer]" --prerelease=allow
 | `sglang` | `sglang[diffusion]`, `checkpoint-engine`, `flash-attn-4`, `flash-linear-attention[conv1d]`, torch +cu130 stack, PyAV | SGLang-based AR/VLM and diffusion recipes |
 | `fastvideo` | FastVideo pinned to an upstream Git commit | WAN 2.1 / 2.2 rollout; [the extra does not currently resolve](#fastvideo-installation-blocker) |
 | `train` | `wandb`, `aiohttp`, `math-verify` | Training runs and local math-answer scoring |
-| `cosmos3` | `diffusers>=0.39` | [Cosmos3 SFT](unirl/models/cosmos3/README.md); apply the [version constraint](#cosmos3-version-prerequisite) |
+| `cosmos3` | `diffusers>=0.39` | [Cosmos3 SFT](unirl/models/cosmos3/README.md); uv's `diffusers==0.40.0` override already satisfies this extra |
 | `infer` | `accelerate`, `timm` | HunyuanImage3, Janus-Pro, and similar models |
 | `eval` | `torchvision`, `paddlepaddle`, `paddleocr`, `python-Levenshtein` | OCR-based reward components |
 | `veomni` | `veomni` | Recipes using the [VeOmni training backend](unirl/train/backend/veomni/) |
@@ -76,9 +76,9 @@ engine extras; install it when you need OCR rewards.
 For development tools (lint and tests):
 
 ```bash
-uv pip install -e ".[vllm,train,infer,eval,dev]" --prerelease=allow
+uv pip install -e ".[vllm,train,infer,eval,dev]"
 # or, for the sglang engine:
-uv pip install -e ".[sglang,train,infer,eval,dev]" --prerelease=allow
+uv pip install -e ".[sglang,train,infer,eval,dev]"
 ```
 
 Prefer these extras over the legacy [`requirements.txt`](requirements.txt) and
@@ -89,24 +89,19 @@ Prefer these extras over the legacy [`requirements.txt`](requirements.txt) and
 The `fastvideo` extra pins
 [hao-ai-lab/FastVideo@2095477](https://github.com/hao-ai-lab/FastVideo/blob/2095477eac7e289c7a7ab13acb367ca60687c304/pyproject.toml),
 which requires `transformers==4.57.3` and `wandb>=0.21.0`. The transformers pin
-conflicts with UniRL's `transformers>=5.6,<5.7`, so `.[fastvideo]` does not
+conflicts with UniRL's `transformers>=5.12,<5.13`, so `.[fastvideo]` does not
 resolve — a separate venv does not help, because UniRL's base deps still apply.
 Adding `train` also conflicts on `wandb`. Use `$FASTVIDEO_PATH` as in the
 [FastVideo engine README](unirl/rollout/engine/fastvideo/README.md) until the extra
 is solvable.
 
-### Cosmos3 version prerequisite
+### Cosmos3
 
-`cosmos3` asks for `diffusers>=0.39`, but uv's override `diffusers>=0.38.0`
-[replaces](https://docs.astral.sh/uv/concepts/resolution/#dependency-overrides)
-that floor instead of intersecting with it. Include `cosmos3` in the extras and
-pass `--constraint` when installing:
+`cosmos3` asks for `diffusers>=0.39`. The uv override pins `diffusers==0.40.0`,
+which already satisfies that floor, so install it as a normal extra:
 
 ```bash
-COSMOS3_CONSTRAINTS="$(mktemp)"
-printf '%s\n' 'diffusers>=0.39' > "$COSMOS3_CONSTRAINTS"
-uv pip install -e ".[vllm,train,infer,cosmos3]" --prerelease=allow \
-    --constraint "$COSMOS3_CONSTRAINTS"
+uv pip install -e ".[vllm,train,infer,cosmos3]"
 ```
 
 ## Environment
diff --git a/datasets/ucf101/README.md b/datasets/ucf101/README.md
index 324eb9105..07be49266 100644
--- a/datasets/ucf101/README.md
+++ b/datasets/ucf101/README.md
@@ -18,7 +18,7 @@ training environment can be installed with:
 ```bash
 uv venv --python 3.12 --seed .venv
 source .venv/bin/activate
-uv pip install -e ".[vllm,train,infer]" --prerelease=allow
+uv pip install -e ".[vllm,train,infer]"
 ```
 
 Both supported engine extras (`vllm` and `sglang`) install PyAV for raw video
diff --git a/examples/ar/qwen3_omni_audio_dcase_gspo_lora_vllm_omni_1x4.yaml b/examples/ar/qwen3_omni_audio_dcase_gspo_lora_vllm_omni_1x4.yaml
index d87a40663..05896b3e4 100644
--- a/examples/ar/qwen3_omni_audio_dcase_gspo_lora_vllm_omni_1x4.yaml
+++ b/examples/ar/qwen3_omni_audio_dcase_gspo_lora_vllm_omni_1x4.yaml
@@ -114,7 +114,7 @@ rollout:
     _target_: unirl.rollout.engine.vllm_omni.config.VLLMOmniEngineConfig
     model_path: ${oc.env:QWEN3_OMNI_PATH,/path/to/Qwen3-Omni-30B-A3B-Instruct}
     modality: qwen3_omni_thinker
-    stage_yaml_override: qwen3_omni_thinker_only_rl_1x4.yaml
+    deploy_config_override: qwen3_omni_thinker_only_rl_1x4.yaml
     enable_sleep_mode: true
     max_prompt_length: 16384
     video_fps: 2.0
diff --git a/examples/ar/qwen3_omni_audio_video_gspo_lora_vllm_omni_1x4.yaml b/examples/ar/qwen3_omni_audio_video_gspo_lora_vllm_omni_1x4.yaml
index 9d80b2584..a2ece8e31 100644
--- a/examples/ar/qwen3_omni_audio_video_gspo_lora_vllm_omni_1x4.yaml
+++ b/examples/ar/qwen3_omni_audio_video_gspo_lora_vllm_omni_1x4.yaml
@@ -124,7 +124,7 @@ rollout:
     model_path: ${oc.env:QWEN3_OMNI_PATH,/path/to/Qwen3-Omni-30B-A3B-Instruct}
     modality: qwen3_omni_thinker
     # Audio-in-video must serialize packed forwards to avoid elevated rollout/replay K3.
-    stage_yaml_override: qwen3_omni_thinker_only_rl_audio_video_1x4.yaml
+    deploy_config_override: qwen3_omni_thinker_only_rl_audio_video_1x4.yaml
     enable_sleep_mode: true
     max_prompt_length: 16384
     video_fps: 2.0
diff --git a/examples/ar/qwen3_omni_audio_video_gspo_lora_vllm_omni_1x8.yaml b/examples/ar/qwen3_omni_audio_video_gspo_lora_vllm_omni_1x8.yaml
index 8d3d0d99a..f42dc3b34 100644
--- a/examples/ar/qwen3_omni_audio_video_gspo_lora_vllm_omni_1x8.yaml
+++ b/examples/ar/qwen3_omni_audio_video_gspo_lora_vllm_omni_1x8.yaml
@@ -114,7 +114,7 @@ rollout:
     model_path: ${oc.env:QWEN3_OMNI_PATH,/path/to/Qwen3-Omni-30B-A3B-Instruct}
     modality: qwen3_omni_thinker
     # Audio-in-video must serialize packed forwards to avoid elevated rollout/replay K3.
-    stage_yaml_override: qwen3_omni_thinker_only_rl_audio_video_1x4.yaml
+    deploy_config_override: qwen3_omni_thinker_only_rl_audio_video_1x4.yaml
     enable_sleep_mode: true
     max_prompt_length: 16384
     video_fps: 2.0
diff --git a/examples/ar/qwen3_omni_image_video_r1_gspo_lora_vllm_omni_1x4.yaml b/examples/ar/qwen3_omni_image_video_r1_gspo_lora_vllm_omni_1x4.yaml
index 1131021c6..944ef2741 100644
--- a/examples/ar/qwen3_omni_image_video_r1_gspo_lora_vllm_omni_1x4.yaml
+++ b/examples/ar/qwen3_omni_image_video_r1_gspo_lora_vllm_omni_1x4.yaml
@@ -109,7 +109,7 @@ rollout:
     _target_: unirl.rollout.engine.vllm_omni.config.VLLMOmniEngineConfig
     model_path: ${oc.env:QWEN3_OMNI_PATH,/path/to/Qwen3-Omni-30B-A3B-Instruct}
     modality: qwen3_omni_thinker
-    stage_yaml_override: qwen3_omni_thinker_only_rl_1x4.yaml
+    deploy_config_override: qwen3_omni_thinker_only_rl_1x4.yaml
     enable_sleep_mode: true
     max_prompt_length: 16384
     image_max_pixels: ${bundle.config.image_max_pixels}
diff --git a/examples/ar/qwen3_omni_video_r1_gspo_lora_vllm_omni_1x4.yaml b/examples/ar/qwen3_omni_video_r1_gspo_lora_vllm_omni_1x4.yaml
index 17e123817..96391067d 100644
--- a/examples/ar/qwen3_omni_video_r1_gspo_lora_vllm_omni_1x4.yaml
+++ b/examples/ar/qwen3_omni_video_r1_gspo_lora_vllm_omni_1x4.yaml
@@ -122,7 +122,7 @@ rollout:
     _target_: unirl.rollout.engine.vllm_omni.config.VLLMOmniEngineConfig
     model_path: ${oc.env:QWEN3_OMNI_PATH,/path/to/Qwen3-Omni-30B-A3B-Instruct}
     modality: qwen3_omni_thinker
-    stage_yaml_override: qwen3_omni_thinker_only_rl_1x4.yaml   # TP=4 (audio tower 20 heads)
+    deploy_config_override: qwen3_omni_thinker_only_rl_1x4.yaml   # TP=4 (audio tower 20 heads)
     enable_sleep_mode: true
     omni_extra:
       stage_init_timeout: 1200
diff --git a/examples/ar/qwen3_omni_video_r1_gspo_lora_vllm_omni_1x8.yaml b/examples/ar/qwen3_omni_video_r1_gspo_lora_vllm_omni_1x8.yaml
index b924219fd..69da1bb14 100644
--- a/examples/ar/qwen3_omni_video_r1_gspo_lora_vllm_omni_1x8.yaml
+++ b/examples/ar/qwen3_omni_video_r1_gspo_lora_vllm_omni_1x8.yaml
@@ -109,7 +109,7 @@ rollout:
     _target_: unirl.rollout.engine.vllm_omni.config.VLLMOmniEngineConfig
     model_path: ${oc.env:QWEN3_OMNI_PATH,/path/to/Qwen3-Omni-30B-A3B-Instruct}
     modality: qwen3_omni_thinker
-    stage_yaml_override: qwen3_omni_thinker_only_rl_1x4.yaml   # TP=4 (audio tower 20 heads)
+    deploy_config_override: qwen3_omni_thinker_only_rl_1x4.yaml   # TP=4 (audio tower 20 heads)
     enable_sleep_mode: true
     omni_extra:
       stage_init_timeout: 1200
diff --git a/examples/diffusion/bagel/bagel_it2i_managed_editscore.yaml b/examples/diffusion/bagel/bagel_it2i_managed_editscore.yaml
index 984718439..c7d56e45f 100644
--- a/examples/diffusion/bagel/bagel_it2i_managed_editscore.yaml
+++ b/examples/diffusion/bagel/bagel_it2i_managed_editscore.yaml
@@ -137,7 +137,7 @@ backend:
     total_steps: 100000
   lora_cfg:
     _target_: unirl.train.configs.LoraConfig
-    rank: 64                           # must be <= the stage YAML's max_lora_rank (64)
+    rank: 64                           # must be <= the deploy config's max_lora_rank (64)
     alpha: 128
     dropout: 0.0
     bias: none
@@ -162,7 +162,7 @@ rollout:
     # Required; same checkpoint the bundle loads.
     model_path: ${oc.env:BAGEL_PATH,ByteDance-Seed/BAGEL-7B-MoT}
     # BAGEL single-stage editing modality (registers BagelIt2iAdapter + boots
-    # stage_configs/bagel_t2i_rl.yaml — one YAML serves both image-out modalities —
+    # deploy_configs/bagel_t2i_rl.yaml — one YAML serves both image-out modalities —
     # with RLBagelPipeline).
     modality: bagel_it2i
     # Required for colocate: the trainer calls sleep()/wake_up() around the train
diff --git a/examples/diffusion/bagel/bagel_it2i_vllmomni.yaml b/examples/diffusion/bagel/bagel_it2i_vllmomni.yaml
index f39eae450..bce22115e 100644
--- a/examples/diffusion/bagel/bagel_it2i_vllmomni.yaml
+++ b/examples/diffusion/bagel/bagel_it2i_vllmomni.yaml
@@ -134,7 +134,7 @@ backend:
     total_steps: 100000
   lora_cfg:
     _target_: unirl.train.configs.LoraConfig
-    rank: 64                           # must be <= the stage YAML's max_lora_rank (64)
+    rank: 64                           # must be <= the deploy config's max_lora_rank (64)
     alpha: 128
     dropout: 0.0
     bias: none
@@ -159,7 +159,7 @@ rollout:
     # Required; same checkpoint the bundle loads.
     model_path: ${oc.env:BAGEL_PATH,ByteDance-Seed/BAGEL-7B-MoT}
     # BAGEL single-stage editing modality (registers BagelIt2iAdapter + boots
-    # stage_configs/bagel_t2i_rl.yaml — one YAML serves both image-out modalities —
+    # deploy_configs/bagel_t2i_rl.yaml — one YAML serves both image-out modalities —
     # with RLBagelPipeline).
     modality: bagel_it2i
     # Required for colocate: the trainer calls sleep()/wake_up() around the train
diff --git a/examples/diffusion/bagel/bagel_vllmomni.yaml b/examples/diffusion/bagel/bagel_vllmomni.yaml
index 50cfcbaa0..364f9195a 100644
--- a/examples/diffusion/bagel/bagel_vllmomni.yaml
+++ b/examples/diffusion/bagel/bagel_vllmomni.yaml
@@ -119,7 +119,7 @@ rollout:
     # Required; same checkpoint the bundle loads.
     model_path: ${oc.env:BAGEL_PATH,/root/hf_model/BAGEL-7B-MoT}
     # BAGEL single-stage T2I diffusion modality (registers BagelT2iAdapter +
-    # boots stage_configs/bagel_t2i_rl.yaml with RLBagelPipeline).
+    # boots deploy_configs/bagel_t2i_rl.yaml with RLBagelPipeline).
     modality: bagel_t2i
     # Required for colocate: trainer calls sleep()/wake_up() around the train
     # phase, a no-op unless CuMemAllocator is enabled at vLLM-Omni init time.
diff --git a/examples/diffusion/bagel/bagel_vllmomni_async.yaml b/examples/diffusion/bagel/bagel_vllmomni_async.yaml
index 0e1f34495..9f6e85de5 100644
--- a/examples/diffusion/bagel/bagel_vllmomni_async.yaml
+++ b/examples/diffusion/bagel/bagel_vllmomni_async.yaml
@@ -108,7 +108,7 @@ rollout:
     # Required; same checkpoint the bundle loads.
     model_path: ${oc.env:BAGEL_PATH,/root/hf_model/BAGEL-7B-MoT}
     # BAGEL single-stage T2I diffusion modality (registers BagelT2iAdapter +
-    # boots stage_configs/bagel_t2i_rl.yaml with RLBagelPipeline).
+    # boots deploy_configs/bagel_t2i_rl.yaml with RLBagelPipeline).
     modality: bagel_t2i
     # Separate slabs don't time-share GPUs, so sleep/wake is unnecessary.
     enable_sleep_mode: false
diff --git a/examples/diffusion/hunyuan_video15/hunyuan_video15_t2v_vllmomni_colocate.yaml b/examples/diffusion/hunyuan_video15/hunyuan_video15_t2v_vllmomni_colocate.yaml
index 8a839f6a5..5c7a16aac 100644
--- a/examples/diffusion/hunyuan_video15/hunyuan_video15_t2v_vllmomni_colocate.yaml
+++ b/examples/diffusion/hunyuan_video15/hunyuan_video15_t2v_vllmomni_colocate.yaml
@@ -20,7 +20,7 @@
 #     sync (sd3/qwen v2 use it). Pushes the trained LoRA adapter into the
 #     co-located sibling engine in-process; the engine runs base + adapter,
 #     which is mathematically the merged model the separate recipe pushed.
-#     hv15's stage config sets enable_lora/max_lora_rank=64 so the adapter
+#     hv15's deploy config sets enable_lora/max_lora_rank=64 so the adapter
 #     actually applies. This avoids the CUDA-IPC path's SGLang dependency,
 #     which the vllm-omni-only venv (two-venv image) does not provide.
 #
diff --git a/examples/diffusion/sd3/sd3_vllmomni_lora_separate.yaml b/examples/diffusion/sd3/sd3_vllmomni_lora_separate.yaml
index 782b140ad..36a5e5521 100644
--- a/examples/diffusion/sd3/sd3_vllmomni_lora_separate.yaml
+++ b/examples/diffusion/sd3/sd3_vllmomni_lora_separate.yaml
@@ -6,7 +6,7 @@
 #   - LoRA training (lora_cfg kept), pushed as the bare ADAPTER each sync
 #     (RemoteLoraWeightSync ships lora_A/lora_B to the engine's set_lora_from_tensors).
 #   - No NCCL group / merged-model broadcast: rank 0 ships the few-MB adapter over a
-#     plain Ray RPC to each rollout Worker. The vLLM-Omni sd35_t2i stage config
+#     plain Ray RPC to each rollout Worker. The vLLM-Omni sd35_t2i deploy config
 #     enables LoRA (enable_lora: true, max_lora_rank: 32), so the engine accepts it.
 
 num_devices: 8
@@ -87,7 +87,7 @@ rollout:
   config:
     _target_: unirl.rollout.engine.vllm_omni.config.VLLMOmniEngineConfig
     model_path: ${oc.env:PRETRAINED_MODEL,stabilityai/stable-diffusion-3.5-medium}
-    # sd35_t2i stage config enables LoRA (enable_lora: true, max_lora_rank: 32),
+    # sd35_t2i deploy config enables LoRA (enable_lora: true, max_lora_rank: 32),
     # so the engine accepts the adapter pushed by RemoteLoraWeightSync.
     modality: sd3_t2i
     # Separate slabs do not time-share GPUs, so sleep/wake is unnecessary.
diff --git a/examples/unified_model/hi3_vllmomni.yaml b/examples/unified_model/hi3_vllmomni.yaml
index bc3398763..be6666cf4 100644
--- a/examples/unified_model/hi3_vllmomni.yaml
+++ b/examples/unified_model/hi3_vllmomni.yaml
@@ -118,7 +118,7 @@ backend:
 # texts per prompt; the DiT engine (modality dit_recaption, single diffusion
 # stage, GPUs 4-7) renders M distinct-noise images per recaption. UnifiedModelTrainer
 # wires both with remote() in the shared placement; each engine clears
-# CUDA_VISIBLE_DEVICES for its multi-GPU HI3 modality and its stage YAML's
+# CUDA_VISIBLE_DEVICES for its multi-GPU HI3 modality and its deploy config's
 # runtime.devices pins the physical cards (a real partition, NOT anchor+pop).
 # Both run enable_sleep_mode so the trainer can sleep/wake them around the
 # colocate train phase. The two engines share ONE backbone/LoRA via sync.
diff --git a/examples/unified_model/hi3_vllmomni_veomni_ep.yaml b/examples/unified_model/hi3_vllmomni_veomni_ep.yaml
index fcd710c79..f88a9c001 100644
--- a/examples/unified_model/hi3_vllmomni_veomni_ep.yaml
+++ b/examples/unified_model/hi3_vllmomni_veomni_ep.yaml
@@ -124,7 +124,7 @@ backend:
 # texts per prompt; the DiT engine (modality dit_recaption, single diffusion
 # stage, GPUs 4-7) renders M distinct-noise images per recaption. UnifiedModelTrainer
 # wires both with remote() in the shared placement; each engine clears
-# CUDA_VISIBLE_DEVICES for its multi-GPU HI3 modality and its stage YAML's
+# CUDA_VISIBLE_DEVICES for its multi-GPU HI3 modality and its deploy config's
 # runtime.devices pins the physical cards (a real partition, NOT anchor+pop).
 # Both run enable_sleep_mode so the trainer can sleep/wake them around the
 # colocate train phase. The two engines share ONE backbone/LoRA via sync.
diff --git a/pyproject.toml b/pyproject.toml
index 41d7769d2..f94eae7cc 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -41,10 +41,9 @@ dependencies = [
   "tensordict>=0.5",
 ]
 
-# Parent/inline engines load this. Multi-stage StageDiffusionProc does not —
-# VLLMOmniHijack.hijack() reinstalls the same hook in spawn children.
+# vllm-omni 0.28 loads this in parent and spawned stage processes.
 [project.entry-points."vllm_omni.general_plugins"]
-unirl_capture_flush = "unirl.rollout.engine.vllm_omni.plugin:register_capture_flush"
+unirl_runtime = "unirl.rollout.engine.vllm_omni.plugin:register_unirl_runtime"
 
 [project.optional-dependencies]
 # sglang pinned to the upstream release the UniRL sglang patch package
@@ -83,10 +82,12 @@ sglang = [
 ]
 # CUDA 13: vllm >=0.26 wants a bare `torch==2.13.0` whose wheel is cu130, and a
 # +cu129 pin would resolve anyway, linking a CUDA-12 torch under a CUDA-13 vllm.
-# Prerelease, so installs need --prerelease=allow, plus sglang's compat layer.
 vllm = [
-  "vllm==0.27.0 ; sys_platform == 'linux'",
-  "vllm-omni==0.27.0rc1 ; sys_platform == 'linux'",
+  "vllm==0.28.0 ; sys_platform == 'linux'",
+  "vllm-omni==0.28.0 ; sys_platform == 'linux'",
+  # vllm-omni 0.28 needs >=5.10.1,<5.15 (5.10.0 is yanked, 5.15 has a known
+  # construction regression); exact because there is no uv.lock to hold it.
+  "transformers==5.12.1 ; sys_platform == 'linux'",
   "torch==2.13.0+cu130 ; sys_platform == 'linux'",
   "torchvision==0.28.0+cu130 ; sys_platform == 'linux'",
   # torchaudio is not required directly by the vLLM backend, so this extra
@@ -169,17 +170,14 @@ index-strategy = "unsafe-best-match"
 # transformers 5.12.1 accepts tokenizers<=0.23.0, but tokenizers 0.23.0(rc)
 # dropped the `cls` kwarg from RobertaProcessing.__new__, so deserializing a
 # saved CLIP tokenizer.json raises TypeError at rollout init (the SD3 text
-# encoders, e2e-confirmed on H20). The sglang install's --prerelease=allow flag
-# (see INSTALL.md) otherwise selects 0.23.0rc*; this override outranks it AND
-# guards against 0.23.0 shipping stable. tokenizers 0.22.2 verified on both
-# engine venvs.
+# encoders, e2e-confirmed on H20). Keep both engine stacks on the verified
+# tokenizers 0.22 family.
 override-dependencies = [
   "kernels>=0.14.1,<0.15",
   "tokenizers>=0.22,<0.23",
-  # sglang[diffusion] hard-pins diffusers==0.37.0 (still pinned on sglang main),
-  # colliding with the base diffusers>=0.38.0 floor (#91); override to the floor
-  # so ".[sglang,...]" stays solvable (same tactic as kernels/tokenizers above).
-  "diffusers>=0.38.0",
+  # Override sglang[diffusion]'s stale 0.37 pin with the exact version required
+  # by vllm-omni 0.28.0. Both mutually exclusive engine stacks use this version.
+  "diffusers==0.40.0",
 ]
 environments = [
   "sys_platform == 'linux' and platform_machine == 'x86_64'",
diff --git a/unirl/rollout/README.md b/unirl/rollout/README.md
index 3577177dd..9fb8ba3fe 100644
--- a/unirl/rollout/README.md
+++ b/unirl/rollout/README.md
@@ -90,7 +90,7 @@ Engine dirs use two layouts. The compact engines (`trainside`, `fastvideo`,
 management), `utils/`, `weight_sync.py`, and a
 runtime-patch dir for the pinned upstream (`sglang_diffusion/_patches/`,
 `vllm_omni/patches/`). `vllm_omni` additionally carries worker-subprocess code
-(`pipelines/`, `worker/`) and stage boot configs (`stage_configs/`).
+(`pipelines/`, `worker/`) and deployment configs (`deploy_configs/`).
 
 Model onboarding is per-engine, and the adapter file is usually **not** the whole
 change surface:
@@ -107,7 +107,7 @@ change surface:
   model needs a new upstream patch.
 - **`vllm_omni`:** add an `adapters/<family>.py` binder (keyed by modality),
   register it, import it in `adapters/__init__.py`, and add the appropriate boot
-  YAML under `stage_configs/`. DiT families additionally need a worker-side
+  YAML under `deploy_configs/`. DiT families additionally need a worker-side
   `pipelines/<model>/pipeline.py`; if the AR/DiT worker needs new behavior, add a
   `worker/` extension or `patches/compat_<model>.py`.
 
diff --git a/unirl/rollout/engine/vllm/runtime.py b/unirl/rollout/engine/vllm/runtime.py
index 3a4726d37..05e965242 100644
--- a/unirl/rollout/engine/vllm/runtime.py
+++ b/unirl/rollout/engine/vllm/runtime.py
@@ -8,7 +8,7 @@
 from typing import Any, Dict, List
 
 _PROTOCOL_VERSION = 1
-_SUPPORTED_VLLM_VERSION = "0.27.0"
+_SUPPORTED_VLLM_VERSION = "0.28.0"
 
 
 class _ProtocolError(RuntimeError):
diff --git a/unirl/rollout/engine/vllm_omni/adapters/bagel.py b/unirl/rollout/engine/vllm_omni/adapters/bagel.py
index d88221941..6a4e8b95b 100644
--- a/unirl/rollout/engine/vllm_omni/adapters/bagel.py
+++ b/unirl/rollout/engine/vllm_omni/adapters/bagel.py
@@ -245,7 +245,7 @@ def build_conditions(self, sample: Sample, per_request: List[List[OmniRawResult]
 class BagelAdapter(ModelAdapter):
     """Bind BAGEL t2i and it2i to one single-stage DiT worker."""
 
-    stage_yaml = "bagel_t2i_rl.yaml"
+    deploy_config = "bagel_t2i_rl.yaml"
     omni_mode = "text-to-image"
     needs_driver_tokenizer = False
     image_input: bool = False  # Whether the modality requires an edit-source image.
diff --git a/unirl/rollout/engine/vllm_omni/adapters/base.py b/unirl/rollout/engine/vllm_omni/adapters/base.py
index a76a15eae..26ffc0ccd 100644
--- a/unirl/rollout/engine/vllm_omni/adapters/base.py
+++ b/unirl/rollout/engine/vllm_omni/adapters/base.py
@@ -45,7 +45,7 @@ class ModelAdapter(ABC):
 
     modality: str = ""
 
-    stage_yaml: str = ""
+    deploy_config: str = ""
     omni_mode: Optional[str] = None
     needs_sigmas: bool = True
     needs_driver_tokenizer: bool = True
@@ -75,9 +75,9 @@ def resolve_sde_label(strategy: Any) -> Optional[str]:
 
     def boot_kwargs(self) -> Dict[str, Any]:
         """Model-specific boot intent beyond the generic config spelling."""
-        require(bool(self.stage_yaml), f"{type(self).__name__} must set stage_yaml")
+        require(bool(self.deploy_config), f"{type(self).__name__} must set deploy_config")
         kwargs: Dict[str, Any] = {
-            "stage_yaml": self.stage_yaml,
+            "deploy_config": self.deploy_config,
             "needs_driver_tokenizer": bool(self.needs_driver_tokenizer),
             "clear_cuda_visible": bool(self.clear_cuda_visible),
         }
diff --git a/unirl/rollout/engine/vllm_omni/adapters/hi3.py b/unirl/rollout/engine/vllm_omni/adapters/hi3.py
index 5c9730c2f..6a497406b 100644
--- a/unirl/rollout/engine/vllm_omni/adapters/hi3.py
+++ b/unirl/rollout/engine/vllm_omni/adapters/hi3.py
@@ -84,7 +84,7 @@ def hi3_fused_conditions(diff_outputs: List[OmniRawResult], *, modality: str) ->
             "Check that RLHunyuanImage3Pipeline.prepare_inputs_for_generation "
             "hook ran in every DiT worker — the subclass swap may not have "
             "taken effect (verify custom_pipeline_args.pipeline_class in "
-            "the stage YAML)."
+            "the deploy config)."
         )
 
     sequence_lengths = [int(c["input_ids"].shape[-1]) for c in captures]
@@ -512,7 +512,7 @@ def build_conditions(self, sample: Sample, per_request: List[List[OmniRawResult]
 class Hi3T2iAdapter(ModelAdapter):
     """HI3 text → AR think → DiT image."""
 
-    stage_yaml = "hunyuan_image3_t2i_rl.yaml"
+    deploy_config = "hunyuan_image3_t2i_rl.yaml"
     omni_mode = "text-to-image"
     ar_lora_passthrough = True
     clear_cuda_visible = True
@@ -548,7 +548,7 @@ def build_response(self, sample: Sample, per_request: List[List[OmniRawResult]])
 class Hi3It2iAdapter(ModelAdapter):
     """HI3 image+text → AR recaption → DiT edited image."""
 
-    stage_yaml = "hunyuan_image3_it2i_rl.yaml"
+    deploy_config = "hunyuan_image3_it2i_rl.yaml"
     omni_mode = "text-to-image"
     ar_lora_passthrough = True
     clear_cuda_visible = True
@@ -582,9 +582,9 @@ def build_response(self, sample: Sample, per_request: List[List[OmniRawResult]])
 
 @register_adapter("hi3_i2t")
 class Hi3I2tAdapter(ModelAdapter):
-    """HI3 image+text → AR text (vendored comprehension YAML)."""
+    """HI3 image+text → AR text."""
 
-    stage_yaml = "hunyuan_image3_i2t.yaml"
+    deploy_config = "hunyuan_image3_i2t_rl.yaml"
     needs_sigmas = False
     ar_lora_passthrough = True
     clear_cuda_visible = True
@@ -617,9 +617,9 @@ def build_response(self, sample: Sample, per_request: List[List[OmniRawResult]])
 
 @register_adapter("hi3_t2t")
 class Hi3T2tAdapter(ModelAdapter):
-    """HI3 text → AR text (vendored comprehension YAML)."""
+    """HI3 text → AR text."""
 
-    stage_yaml = "hunyuan_image3_t2t.yaml"
+    deploy_config = "hunyuan_image3_ar_rl.yaml"
     needs_sigmas = False
     ar_lora_passthrough = True
     clear_cuda_visible = True
@@ -652,7 +652,7 @@ def build_response(self, sample: Sample, per_request: List[List[OmniRawResult]])
 class Hi3ArRecaptionAdapter(ModelAdapter):
     """Two-engine trainer's AR think/recaption producer."""
 
-    stage_yaml = "hunyuan_image3_ar_recaption_rl.yaml"
+    deploy_config = "hunyuan_image3_ar_recaption_rl.yaml"
     needs_sigmas = False
     ar_lora_passthrough = True
     clear_cuda_visible = True
@@ -681,7 +681,7 @@ def build_response(self, sample: Sample, per_request: List[List[OmniRawResult]])
 class Hi3DitRecaptionAdapter(ModelAdapter):
     """Standalone HI3 DiT — the two-engine trainer's image half."""
 
-    stage_yaml = "hunyuan_image3_dit_recaption_rl.yaml"
+    deploy_config = "hunyuan_image3_dit_recaption_rl.yaml"
     omni_mode = "text-to-image"
     clear_cuda_visible = True
     # HI3 two-engine stages are TP>1 — wake-time LoRA re-push must use the byte-copy transport.
diff --git a/unirl/rollout/engine/vllm_omni/adapters/hv15.py b/unirl/rollout/engine/vllm_omni/adapters/hv15.py
index 079d61867..61118c216 100644
--- a/unirl/rollout/engine/vllm_omni/adapters/hv15.py
+++ b/unirl/rollout/engine/vllm_omni/adapters/hv15.py
@@ -54,8 +54,8 @@ class Hv15VideoOutputAdapter(DitOutputAdapter):
         "on the output envelope's unirl metadata (or it lacked the dual-stream "
         "text_mllm/text_glyph embeds). Check that "
         "RLHunyuanVideo15Pipeline's encode_prompt hook ran in every DiT "
-        "worker — verify custom_pipeline_args.pipeline_class in the stage "
-        "YAML."
+        "worker — verify custom_pipeline_args.pipeline_class in the deploy "
+        "config."
     )
 
     def build_decoded(self, sample: Sample, per_request: List[List[OmniRawResult]]) -> Any:
@@ -116,7 +116,7 @@ def _cat_field(field_name: str) -> Optional[torch.Tensor]:
 class Hv15T2vAdapter(ModelAdapter):
     """HunyuanVideo-1.5 text → video (single diffusion stage, TP=1)."""
 
-    stage_yaml = "hunyuan_video15_t2v_rl.yaml"
+    deploy_config = "hunyuan_video15_t2v_rl.yaml"
     needs_driver_tokenizer = False
 
     def __init__(self, config: Any, model_config: Any, *, strategy: Any = None, tokenize_fn: Any = None) -> None:
diff --git a/unirl/rollout/engine/vllm_omni/adapters/qwen3_omni.py b/unirl/rollout/engine/vllm_omni/adapters/qwen3_omni.py
index 1df55989c..2ad3cddfb 100644
--- a/unirl/rollout/engine/vllm_omni/adapters/qwen3_omni.py
+++ b/unirl/rollout/engine/vllm_omni/adapters/qwen3_omni.py
@@ -457,8 +457,7 @@ def _status(cls, per_request: List[List[OmniRawResult]]) -> torch.Tensor:
         values: List[int] = []
         for group in per_request:
             output = cls._stage0(group)
-            request_output = getattr(output, "request_output", None)
-            completions = getattr(request_output, "outputs", None) or []
+            completions = getattr(output, "outputs", None) or []
             finish_reason = getattr(completions[0], "finish_reason", None) if completions else None
             values.append(int(mapping.get(str(finish_reason), SegmentStatus.PENDING)))
         return torch.tensor(values, dtype=torch.long)
@@ -495,7 +494,7 @@ def build(self, sample: Sample, per_request: List[List[OmniRawResult]]) -> Sampl
 class Qwen3OmniThinkerAdapter(ModelAdapter):
     """Qwen3-Omni Thinker — text/video → AR text (single stage, TP>1, LoRA)."""
 
-    stage_yaml = "qwen3_omni_thinker_only_rl_1x4.yaml"
+    deploy_config = "qwen3_omni_thinker_only_rl_1x4.yaml"
     omni_mode = None
     needs_sigmas = False
     needs_driver_tokenizer = False
diff --git a/unirl/rollout/engine/vllm_omni/adapters/qwen_image.py b/unirl/rollout/engine/vllm_omni/adapters/qwen_image.py
index 0d2e0abe3..cfe6b4f60 100644
--- a/unirl/rollout/engine/vllm_omni/adapters/qwen_image.py
+++ b/unirl/rollout/engine/vllm_omni/adapters/qwen_image.py
@@ -97,7 +97,7 @@ class QwenImageOutputAdapter(DitOutputAdapter):
         "the output envelope's unirl metadata. Check that RLQwenImagePipeline's "
         "encode_prompt tap ran in every DiT worker — the subclass swap may "
         "not have taken effect (verify custom_pipeline_args.pipeline_class "
-        "in the stage YAML)."
+        "in the deploy config)."
     )
 
     def build_conditions(self, sample: Sample, per_request: List[List[OmniRawResult]]) -> Dict[str, Any]:
@@ -138,7 +138,7 @@ def build_conditions(self, sample: Sample, per_request: List[List[OmniRawResult]
 class QwenImageT2iAdapter(ModelAdapter):
     """Qwen-Image text → image (single diffusion stage, TP=1)."""
 
-    stage_yaml = "qwen_image_t2i_rl.yaml"
+    deploy_config = "qwen_image_t2i_rl.yaml"
     omni_mode = "text-to-image"
     needs_driver_tokenizer = False
 
diff --git a/unirl/rollout/engine/vllm_omni/adapters/sd3.py b/unirl/rollout/engine/vllm_omni/adapters/sd3.py
index c4e06e935..44f264838 100644
--- a/unirl/rollout/engine/vllm_omni/adapters/sd3.py
+++ b/unirl/rollout/engine/vllm_omni/adapters/sd3.py
@@ -60,7 +60,7 @@ def build_conditions(self, sample: Sample, per_request: List[List[OmniRawResult]
                 "RLStableDiffusion3Pipeline._install_encode_prompt_hook ran "
                 "in every DiT worker — the subclass swap may not have taken "
                 "effect (verify custom_pipeline_args.pipeline_class in the "
-                "stage YAML)."
+                "deploy config)."
             )
 
         embeds = torch.cat([c["prompt_embeds"] for c in captures], dim=0)
@@ -93,7 +93,7 @@ def build_conditions(self, sample: Sample, per_request: List[List[OmniRawResult]
 class Sd3T2iAdapter(ModelAdapter):
     """SD3.5-medium text → image (single diffusion stage, TP=1)."""
 
-    stage_yaml = "sd35_t2i_rl.yaml"
+    deploy_config = "sd35_t2i_rl.yaml"
     omni_mode = "text-to-image"
     needs_driver_tokenizer = False
 
diff --git a/unirl/rollout/engine/vllm_omni/backends/base.py b/unirl/rollout/engine/vllm_omni/backends/base.py
index c7f724ad8..0b0270829 100644
--- a/unirl/rollout/engine/vllm_omni/backends/base.py
+++ b/unirl/rollout/engine/vllm_omni/backends/base.py
@@ -11,7 +11,6 @@
     Optional,
     Protocol,
     Sequence,
-    runtime_checkable,
 )
 
 if TYPE_CHECKING:
@@ -61,7 +60,7 @@ class OmniRawResult(Protocol):
     request_id: str
     stage_id: Optional[int]
     final_output_type: Optional[str]
-    request_output: Optional[Any]
+    outputs: Sequence[Any]
     prompt_token_ids: Optional[Sequence[int]]
     images: Optional[Sequence[Any]]
     trajectory_latents: Optional["torch.Tensor"]
@@ -70,7 +69,6 @@ class OmniRawResult(Protocol):
     multimodal_output: Optional[dict]
 
 
-@runtime_checkable
 class Backend(Protocol):
     """The seam every ``vllm_omni`` collaborator reaches the runtime through."""
 
diff --git a/unirl/rollout/engine/vllm_omni/backends/native.py b/unirl/rollout/engine/vllm_omni/backends/native.py
index 5acdd47ac..3f179e2f5 100644
--- a/unirl/rollout/engine/vllm_omni/backends/native.py
+++ b/unirl/rollout/engine/vllm_omni/backends/native.py
@@ -54,45 +54,15 @@ def _import_omni_runtime() -> Dict[str, Any]:
     }
 
 
-def _resolve_stage_yaml(name: str) -> str:
-    """Return the absolute path of the local stage-config YAML asset."""
+def _resolve_deploy_config(name: str) -> str:
+    """Return the absolute path of the local deploy-config YAML asset."""
     here = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
-    path = os.path.join(here, "stage_configs", name)
+    path = os.path.join(here, "deploy_configs", name)
     if not os.path.exists(path):
-        raise FileNotFoundError(f"_resolve_stage_yaml: YAML not found at {path}")
+        raise FileNotFoundError(f"_resolve_deploy_config: YAML not found at {path}")
     return path
 
 
-def _cfg_get(cfg: Any, key: str, default: Any = None) -> Any:
-    """Mapping/OmegaConf/attr-tolerant getter (``None`` coerces to default)."""
-    if cfg is None:
-        return default
-    getter = getattr(cfg, "get", None)
-    value = getter(key, default) if callable(getter) else getattr(cfg, key, default)
-    return default if value is None else value
-
-
-def _tp_from_stage_configs(stage_configs: Sequence[Any]) -> Dict[int, int]:
-    """Extract ``{stage_id: tensor_parallel_size}`` from the runtime's configs."""
-    tp_map: Dict[int, int] = {}
-    for entry in stage_configs:
-        sid = int(_cfg_get(entry, "stage_id", len(tp_map)))
-        ea = _cfg_get(entry, "engine_args", {})
-        tp = _cfg_get(ea, "tensor_parallel_size")
-        if tp is None:
-            tp = _cfg_get(_cfg_get(ea, "parallel_config", {}), "tensor_parallel_size")
-        tp_map[sid] = int(tp) if tp is not None else 1
-    return tp_map
-
-
-def _assemble_omni_kwargs(intent: Dict[str, Any]) -> Dict[str, Any]:
-    """Spell the boot intent into ``Omni`` ctor kwargs."""
-    omni_kwargs = dict(intent.get("omni_kwargs") or {})
-    if intent.get("enable_sleep_mode"):
-        omni_kwargs["enable_sleep_mode"] = True
-    return omni_kwargs
-
-
 @contextmanager
 def _master_port_env(port: Optional[int]):
     """Pin ``MASTER_PORT`` to the reserved engine port for the duration of ``Omni()``."""
@@ -120,12 +90,10 @@ def __init__(
         runtime: Dict[str, Any],
         *,
         tokenizer: Optional[Any],
-        tp_per_stage: Dict[int, int],
     ) -> None:
         self._omni: Optional[Any] = omni
         self._rt = runtime
         self._tokenizer = tokenizer
-        self._tp_per_stage = dict(tp_per_stage)
 
     @classmethod
     def boot(cls, intent: Dict[str, Any]) -> "VLLMOmniBackend":
@@ -141,9 +109,11 @@ def boot(cls, intent: Dict[str, Any]) -> "VLLMOmniBackend":
         # colocated train actor's state; spawned workers initialize their own.
         with _preserve_cudnn_sdp_state():
             from unirl.rollout.engine.vllm_omni.patches import install as install_patches
+            from unirl.rollout.engine.vllm_omni.plugin import register_unirl_runtime
 
             install_patches()
             rt = _import_omni_runtime()
+            register_unirl_runtime()
 
         if intent.get("clear_cuda_visible"):
             os.environ.pop("CUDA_VISIBLE_DEVICES", None)
@@ -159,8 +129,10 @@ def boot(cls, intent: Dict[str, Any]) -> "VLLMOmniBackend":
         except Exception:  # noqa: BLE001 - belt and braces; never block a boot
             pass
 
-        yaml_path = _resolve_stage_yaml(str(intent["stage_yaml"]))
-        omni_kwargs = _assemble_omni_kwargs(intent)
+        deploy_config_path = _resolve_deploy_config(str(intent["deploy_config"]))
+        omni_kwargs = dict(intent.get("omni_kwargs") or {})
+        if intent.get("enable_sleep_mode"):
+            omni_kwargs["enable_sleep_mode"] = True
         ports = intent.get("ports")
         boot_master_port = int(ports.master_port) if ports is not None else None
         logger.info(
@@ -168,7 +140,7 @@ def boot(cls, intent: Dict[str, Any]) -> "VLLMOmniBackend":
             pformat(
                 {
                     **intent,
-                    "stage_yaml_path": yaml_path,
+                    "deploy_config_path": deploy_config_path,
                     "assembled_omni_kwargs": omni_kwargs,
                     "boot_master_port": boot_master_port,
                 },
@@ -185,7 +157,7 @@ def boot(cls, intent: Dict[str, Any]) -> "VLLMOmniBackend":
                 with _master_port_env(boot_master_port):
                     omni = rt["Omni"](
                         model=str(intent["model_path"]),
-                        stage_configs_path=yaml_path,
+                        deploy_config=deploy_config_path,
                         **omni_kwargs,
                     )
             finally:
@@ -193,15 +165,16 @@ def boot(cls, intent: Dict[str, Any]) -> "VLLMOmniBackend":
                     fcntl.flock(lock_file, fcntl.LOCK_UN)
                     lock_file.close()
 
+            stage_configs = omni.engine.stage_configs
             try:
                 from omegaconf import OmegaConf
 
                 resolved_stage_configs = OmegaConf.to_container(
-                    OmegaConf.create(omni.stage_configs),
+                    OmegaConf.create(stage_configs),
                     resolve=True,
                 )
             except Exception:  # noqa: BLE001 - config logging must never block boot
-                resolved_stage_configs = omni.stage_configs
+                resolved_stage_configs = stage_configs
             logger.info(
                 "VLLM-Omni resolved runtime stage configs (after all overrides):\n%s",
                 pformat(resolved_stage_configs, sort_dicts=True),
@@ -215,15 +188,12 @@ def boot(cls, intent: Dict[str, Any]) -> "VLLMOmniBackend":
                 omni,
                 rt,
                 tokenizer=tokenizer,
-                tp_per_stage=_tp_from_stage_configs(omni.stage_configs),
             )
         except BaseException:
             logger.exception("VLLM-Omni boot failed; tearing down any engine processes")
             if omni is not None:
                 try:
-                    close = getattr(omni, "close", None)
-                    if callable(close):
-                        close()
+                    omni.close()
                 except Exception:
                     logger.exception("Failed to close the half-booted vLLM-Omni engine")
             terminate_descendants(os.getpid(), name_prefix=_ENGINE_PROC_PREFIX)
@@ -289,18 +259,80 @@ def tokenize_prompt(self, text: str, *, task: str, sys_type: str) -> List[int]:
             build_prompt_tokens,
         )
 
-        # vllm-omni 0.27 returns a PromptTokensResult here, not a bare id list.
+        # vLLM-Omni returns a PromptTokensResult here, not a bare id list.
         return list(build_prompt_tokens(text, self._tokenizer, task=task, sys_type=sys_type).token_ids)
 
     def num_stages(self) -> int:
         return int(self._require_omni().engine.num_stages)
 
     def tp_per_stage(self) -> Dict[int, int]:
-        return dict(self._tp_per_stage)
+        tp_map: Dict[int, int] = {}
+        for entry in self._require_omni().engine.stage_configs:
+            engine_args = entry.engine_args
+            tp = engine_args.get("tensor_parallel_size")
+            if tp is None:
+                parallel_config = engine_args.get("parallel_config")
+                if parallel_config is not None:
+                    tp = parallel_config.get("tensor_parallel_size")
+            tp_map[int(entry.stage_id)] = int(tp) if tp is not None else 1
+        return tp_map
 
     def _stage_ids(self) -> List[int]:
         return list(range(self.num_stages()))
 
+    def _stage_type(self, stage_id: int) -> str:
+        return self._require_omni().engine.get_stage_metadata(stage_id).stage_type
+
+    @staticmethod
+    def _require_rpc_success(
+        action: str,
+        stage_id: int,
+        results: object,
+        *,
+        allow_false: bool = False,
+    ) -> None:
+        """Fail closed when a stage or replica reports an unsupported/failed RPC."""
+
+        def validate(result: object) -> None:
+            if result is None:
+                return
+            if isinstance(result, (list, tuple)):
+                for item in result:
+                    validate(item)
+                return
+            if isinstance(result, bool):
+                if not result and not allow_false:
+                    raise RuntimeError(f"vllm-omni {action} returned False on stage {stage_id}")
+                return
+            if not isinstance(result, Mapping):
+                raise RuntimeError(f"vllm-omni {action} returned an unexpected result on stage {stage_id}: {result!r}")
+
+            status = result.get("status")
+            error = result.get("error")
+            if result.get("supported") is False or error or (status is not None and status != "SUCCESS"):
+                reason = result.get("reason") or result.get("error_msg") or error or status
+                raise RuntimeError(f"vllm-omni {action} failed on stage {stage_id}: {reason!r}")
+
+        validate(results)
+
+    def _collective_rpc(
+        self,
+        stage_id: int,
+        method: str,
+        *,
+        args: tuple = (),
+        kwargs: Optional[dict] = None,
+        allow_false: bool = False,
+    ) -> List[Any]:
+        results = self._require_omni().engine.collective_rpc(
+            method=method,
+            args=args,
+            kwargs=kwargs,
+            stage_ids=[stage_id],
+        )
+        self._require_rpc_success(method, stage_id, results, allow_false=allow_false)
+        return results
+
     @staticmethod
     def _require_ack_success(action: str, stage_id: int, task_id: str, acks: object) -> None:
         # Worker handlers catch their own exceptions and answer
@@ -312,9 +344,6 @@ def _require_ack_success(action: str, stage_id: int, task_id: str, acks: object)
         def ack_field(ack: object, name: str, default: object = None) -> object:
             return ack.get(name, default) if isinstance(ack, Mapping) else getattr(ack, name, default)
 
-        def ack_error(ack: object) -> object:
-            return ack_field(ack, "error_msg", ack_field(ack, "error"))
-
         def validate_result(result: object) -> None:
             nonlocal success_count
             if result is None:
@@ -330,10 +359,11 @@ def validate_result(result: object) -> None:
             if status is None:
                 raise RuntimeError(f"vllm-omni {action} returned no ACK status for stage {stage_id}: result={result!r}")
             if status != "SUCCESS":
+                error = ack_field(result, "error_msg", ack_field(result, "error"))
                 raise RuntimeError(
                     f"vllm-omni {action} failed on stage {stage_id}: worker rank "
                     f"{ack_field(result, 'rank', '?')} answered status={status!r} "
-                    f"error={ack_error(result)!r}"
+                    f"error={error!r}"
                 )
             ack_stage_id = ack_field(result, "stage_id")
             ack_task_id = ack_field(result, "task_id")
@@ -351,11 +381,14 @@ def validate_result(result: object) -> None:
             raise RuntimeError(f"vllm-omni {action} returned no successful ACK for stage {stage_id}")
 
     def sleep_task(self) -> None:
-        """Fan ``handle_sleep_task`` to every stage's workers (level 1)."""
+        """Sleep AR stages through EngineCore and diffusion stages through worker RPC."""
         import uuid
 
         omni = self._require_omni()
         for sid in self._stage_ids():
+            if self._stage_type(sid) != "diffusion":
+                self._collective_rpc(sid, "sleep", args=(1, "abort"))
+                continue
             task_id = str(uuid.uuid4())
             acks = omni.engine.collective_rpc(
                 method="handle_sleep_task",
@@ -365,13 +398,17 @@ def sleep_task(self) -> None:
             self._require_ack_success("sleep", int(sid), task_id, acks)
 
     def wake_task(self) -> None:
-        """Fan ``handle_wake_task`` to every stage's workers + sync CUDA."""
+        """Wake AR stages through EngineCore and diffusion stages through worker RPC."""
         import uuid
 
         import torch
 
         omni = self._require_omni()
         for sid in self._stage_ids():
+            if self._stage_type(sid) != "diffusion":
+                # ``None`` wakes every tag, including UniRL's MoE workspace tag.
+                self._collective_rpc(sid, "wake_up", kwargs={"tags": None})
+                continue
             task_id = str(uuid.uuid4())
             acks = omni.engine.collective_rpc(
                 method="handle_wake_task",
@@ -383,14 +420,19 @@ def wake_task(self) -> None:
             torch.cuda.synchronize()
 
     def ping(self) -> bool:
-        return self._omni is not None
+        omni = self._omni
+        if omni is None:
+            return False
+        try:
+            return bool(omni.is_running)
+        except Exception:
+            logger.exception("vLLM-Omni health check failed")
+            return False
 
     def shutdown(self) -> None:
         if self._omni is not None:
             try:
-                close = getattr(self._omni, "close", None)
-                if callable(close):
-                    close()
+                self._omni.close()
             finally:
                 self._omni = None
 
@@ -407,7 +449,6 @@ def update_from_ipc(
         replica_rank: Optional[int],
     ) -> None:
         """Fan a bucketed CUDA-IPC state-dict update out to per-stage workers."""
-        omni = self._require_omni()
         kwargs = {
             "peft_config": peft_config,
             "base_sync_done": base_sync_done,
@@ -415,11 +456,10 @@ def update_from_ipc(
             "replica_rank": replica_rank,
         }
         for sid in self._stage_ids():
-            omni.engine.collective_rpc(
-                method="update_weights_from_ipc",
-                args=(),
+            self._collective_rpc(
+                sid,
+                "update_weights_from_ipc",
                 kwargs={**kwargs, "stage_id": int(sid)},
-                stage_ids=[int(sid)],
             )
 
     def init_weights_group(
@@ -432,7 +472,6 @@ def init_weights_group(
         group_name: str,
         backend: str,
     ) -> None:
-        omni = self._require_omni()
         kwargs = {
             "master_address": str(master_address),
             "master_port": int(master_port),
@@ -442,11 +481,10 @@ def init_weights_group(
             "backend": str(backend),
         }
         for sid in self._stage_ids():
-            omni.engine.collective_rpc(
-                method="init_weights_update_group",
-                args=(),
+            self._collective_rpc(
+                sid,
+                "init_weights_update_group",
                 kwargs=kwargs,
-                stage_ids=[int(sid)],
             )
 
     def update_from_distributed(
@@ -459,7 +497,6 @@ def update_from_distributed(
         target_modules: Optional[List[str]],
         flush_cache: bool,
     ) -> None:
-        omni = self._require_omni()
         kwargs = {
             "names": list(names),
             "dtypes": list(dtypes),
@@ -469,22 +506,20 @@ def update_from_distributed(
             "flush_cache": bool(flush_cache),
         }
         for sid in self._stage_ids():
-            omni.engine.collective_rpc(
-                method="update_weights_from_distributed",
-                args=(),
+            self._collective_rpc(
+                sid,
+                "update_weights_from_distributed",
                 kwargs=kwargs,
-                stage_ids=[int(sid)],
             )
 
     def destroy_weights_group(self, *, group_name: str) -> None:
         if self._omni is None:
             return
         for sid in self._stage_ids():
-            self._omni.engine.collective_rpc(
-                method="destroy_weights_update_group",
-                args=(),
+            self._collective_rpc(
+                sid,
+                "destroy_weights_update_group",
                 kwargs={"group_name": str(group_name)},
-                stage_ids=[int(sid)],
             )
 
     def update_from_tensor(
@@ -496,7 +531,6 @@ def update_from_tensor(
         flush_cache: bool,
     ) -> None:
         """Fan a SGLang-shape tensor payload to per-stage workers."""
-        omni = self._require_omni()
         kwargs = {
             "serialized_named_tensors": list(serialized_named_tensors),
             "target_modules": list(target_modules) if target_modules else None,
@@ -504,11 +538,10 @@ def update_from_tensor(
             "flush_cache": bool(flush_cache),
         }
         for sid in self._stage_ids():
-            omni.engine.collective_rpc(
-                method="update_weights_from_tensor",
-                args=(),
+            self._collective_rpc(
+                sid,
+                "update_weights_from_tensor",
                 kwargs=kwargs,
-                stage_ids=[int(sid)],
             )
 
     def set_lora_handle(
@@ -526,9 +559,9 @@ def set_lora_handle(
             DIFFRL_LORA_NAME,
             DIFFRL_LORA_PATH,
         )
+        from unirl.utils.peft_merge import adapt_lora_for_vllm
 
-        omni = self._require_omni()
-        lora_tensors = self._wrap_peft_envelope(lora_tensors)
+        lora_tensors = adapt_lora_for_vllm(lora_tensors)
         self._remove_existing_lora(int(DIFFRL_LORA_INT_ID))
 
         from unirl.distributed.weight_sync.transfer.sgl_compat import (
@@ -540,8 +573,9 @@ def set_lora_handle(
                 name: t.detach().clone() if isinstance(t, torch.Tensor) else t for name, t in lora_tensors.items()
             }
             serialized = MultiprocessingSerializer.serialize(cloned, output_str=True)
-            omni.engine.collective_rpc(
-                method="set_lora_from_tensor_dict",
+            self._collective_rpc(
+                sid,
+                "set_lora_from_tensor_dict",
                 args=(
                     str(adapter_name) or DIFFRL_LORA_NAME,
                     int(DIFFRL_LORA_INT_ID),
@@ -549,7 +583,6 @@ def set_lora_handle(
                     dict(peft_config or {}),
                     serialized,
                 ),
-                stage_ids=[int(sid)],
             )
 
     def set_lora_copy(
@@ -570,9 +603,9 @@ def set_lora_copy(
             DIFFRL_LORA_NAME,
             DIFFRL_LORA_PATH,
         )
+        from unirl.utils.peft_merge import adapt_lora_for_vllm
 
-        omni = self._require_omni()
-        lora_tensors = self._wrap_peft_envelope(lora_tensors)
+        lora_tensors = adapt_lora_for_vllm(lora_tensors)
         self._remove_existing_lora(int(DIFFRL_LORA_INT_ID))
 
         cpu_tensors = {
@@ -583,8 +616,9 @@ def set_lora_copy(
         serialized = base64.b64encode(buf.getvalue()).decode("ascii")
 
         for sid in self._stage_ids():
-            omni.engine.collective_rpc(
-                method="set_lora_from_tensor_dict_copy",
+            self._collective_rpc(
+                sid,
+                "set_lora_from_tensor_dict_copy",
                 args=(
                     str(adapter_name) or DIFFRL_LORA_NAME,
                     int(DIFFRL_LORA_INT_ID),
@@ -592,54 +626,38 @@ def set_lora_copy(
                     dict(peft_config or {}),
                     serialized,
                 ),
-                stage_ids=[int(sid)],
             )
 
-    @staticmethod
-    def _wrap_peft_envelope(lora_tensors: Dict[str, Any]) -> Dict[str, Any]:
-        """Wrap canonical wire keys in the PEFT envelope vllm-omni expects."""
-        from unirl.utils.peft_merge import adapt_lora_for_vllm
-
-        first_key = next(iter(lora_tensors), "")
-        if lora_tensors and not first_key.startswith("base_model.model."):
-            return adapt_lora_for_vllm(lora_tensors)
-        return lora_tensors
-
     def _remove_existing_lora(self, adapter_id: int) -> None:
         """Drop the existing adapter on every stage before re-adding."""
-        omni = self._require_omni()
         for sid in self._stage_ids():
-            try:
-                omni.engine.collective_rpc(
-                    method="remove_lora",
-                    args=(int(adapter_id),),
-                    stage_ids=[int(sid)],
-                )
-            except Exception:
-                pass
+            self._collective_rpc(
+                sid,
+                "remove_lora",
+                args=(int(adapter_id),),
+                allow_false=True,
+            )
 
     def param_checksums(self, *, names: List[str]) -> dict:
         """Fan ``_diffrl_loaded_param_checksums`` across stages and ranks."""
-        omni = self._require_omni()
         out: dict = {}
         for sid in self._stage_ids():
-            results = omni.engine.collective_rpc(
-                method="_diffrl_loaded_param_checksums",
+            results = self._collective_rpc(
+                sid,
+                "_diffrl_loaded_param_checksums",
                 args=(list(names),),
-                stage_ids=[int(sid)],
             )
             out[int(sid)] = results[0] if isinstance(results, list) and results else results
         return out
 
     def lora_checksums(self, *, adapter_id: int, names: Optional[List[str]]) -> dict:
         """Fan ``_diffrl_loaded_lora_checksums`` across stages and ranks."""
-        omni = self._require_omni()
         out: dict = {}
         for sid in self._stage_ids():
-            results = omni.engine.collective_rpc(
-                method="_diffrl_loaded_lora_checksums",
+            results = self._collective_rpc(
+                sid,
+                "_diffrl_loaded_lora_checksums",
                 args=(int(adapter_id), list(names) if names else None),
-                stage_ids=[int(sid)],
             )
             out[int(sid)] = results[0] if isinstance(results, list) and results else results
         return out
@@ -649,15 +667,19 @@ def _group_by_request(flat_outputs: Sequence[Any], n: int) -> List[List[Any]]:
     """Group ``Omni.generate``'s flat output list into per-request lists."""
     grouped: List[List[Any]] = [[] for _ in range(n)]
     for out in flat_outputs:
-        rid = getattr(out, "request_id", "") or ""
-        if "_" in rid:
-            idx_part = rid.split("_", 1)[0]
-            try:
-                idx = int(idx_part)
-            except ValueError:
-                continue
-            if 0 <= idx < n:
-                grouped[idx].append(out)
+        try:
+            rid = out.request_id
+            idx_part, suffix = rid.split("_", 1)
+            if not suffix:
+                raise ValueError("empty request-id suffix")
+            idx = int(idx_part)
+        except (AttributeError, TypeError, ValueError) as exc:
+            raise RuntimeError(
+                f"vllm-omni returned an invalid request_id: {getattr(out, 'request_id', None)!r}"
+            ) from exc
+        if not 0 <= idx < n:
+            raise RuntimeError(f"vllm-omni request_id index {idx} is outside the prompt batch of size {n}")
+        grouped[idx].append(out)
     return grouped
 
 
diff --git a/unirl/rollout/engine/vllm_omni/config.py b/unirl/rollout/engine/vllm_omni/config.py
index ed4520524..7bd98562b 100644
--- a/unirl/rollout/engine/vllm_omni/config.py
+++ b/unirl/rollout/engine/vllm_omni/config.py
@@ -31,7 +31,7 @@ def make_engine(self, **deps: Any):
 
     enable_sleep_mode: bool = True
 
-    stage_yaml_override: Optional[str] = None
+    deploy_config_override: Optional[str] = None
 
     omni_extra: Dict[str, Any] = field(default_factory=dict)
 
@@ -71,8 +71,8 @@ def server_intent(
         }
         intent.update(extra)
 
-        if self.stage_yaml_override:
-            intent["stage_yaml"] = str(self.stage_yaml_override)
+        if self.deploy_config_override:
+            intent["deploy_config"] = str(self.deploy_config_override)
 
         omni_kwargs: Dict[str, Any] = dict(
             stage_init_timeout=1200,
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/bagel_t2i_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/bagel_t2i_rl.yaml
new file mode 100644
index 000000000..2ef9de384
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/bagel_t2i_rl.yaml
@@ -0,0 +1,35 @@
+# BAGEL-7B-MoT single-stage deployment shared by t2i and it2i rollout.
+pipeline: bagel_single_stage
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "0"
+    max_num_seqs: 1
+    enforce_eager: true
+    trust_remote_code: true
+    enable_prefix_caching: false
+    vae_use_slicing: false
+    vae_use_tiling: false
+    cache_backend: null
+    cache_config: null
+    enable_cache_dit_summary: false
+    enable_lora: true
+    max_lora_rank: 64
+    parallel_config:
+      pipeline_parallel_size: 1
+      data_parallel_size: 1
+      tensor_parallel_size: 1
+      enable_expert_parallel: false
+      sequence_parallel_size: 1
+      ulysses_degree: 1
+      ring_degree: 1
+      cfg_parallel_size: 1
+      vae_patch_parallel_size: 1
+      use_hsdp: false
+      hsdp_shard_size: -1
+      hsdp_replicate_size: 1
+    custom_pipeline_args:
+      pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.bagel.pipeline.RLBagelPipeline
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_ar_recaption_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_ar_recaption_rl.yaml
new file mode 100644
index 000000000..0b1b48337
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_ar_recaption_rl.yaml
@@ -0,0 +1,31 @@
+# HunyuanImage-3 terminal AR recaption deployment for the two-engine trainer.
+pipeline: unirl_hi3_ar_recaption
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "0,1,2,3"
+    max_num_seqs: 1
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.ar_extension.HI3ARWeightSyncExtension
+    enable_lora: true
+    fully_sharded_loras: true
+    max_model_len: 8192
+    gpu_memory_utilization: 0.7
+    enforce_eager: true
+    trust_remote_code: true
+    enable_prefix_caching: false
+    max_num_batched_tokens: 8192
+    tensor_parallel_size: 4
+    pipeline_parallel_size: 1
+    hf_overrides:
+      rope_parameters:
+        mrope_section: [0, 32, 32]
+        rope_type: default
+    default_sampling_params:
+      temperature: 0.7
+      top_p: 0.95
+      top_k: 1024
+      max_tokens: 4096
+      detokenize: true
+      repetition_penalty: 1.1
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_ar_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_ar_rl.yaml
new file mode 100644
index 000000000..4954d65e4
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_ar_rl.yaml
@@ -0,0 +1,29 @@
+# HunyuanImage-3 text-only AR deployment for the t2t adapter.
+pipeline: unirl_hi3_ar_text
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "0,1,2,3"
+    max_num_seqs: 1
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.ar_extension.HI3ARWeightSyncExtension
+    enable_lora: true
+    gpu_memory_utilization: 0.95
+    enforce_eager: true
+    trust_remote_code: true
+    enable_prefix_caching: false
+    max_num_batched_tokens: 32768
+    tensor_parallel_size: 4
+    pipeline_parallel_size: 1
+    hf_overrides:
+      rope_parameters:
+        mrope_section: [0, 32, 32]
+        rope_type: default
+    default_sampling_params:
+      temperature: 0.0
+      top_p: 0.95
+      top_k: 1024
+      max_tokens: 2048
+      stop_token_ids: [127957, 128024, 128026]
+      detokenize: true
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_dit_recaption_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_dit_recaption_rl.yaml
new file mode 100644
index 000000000..442163bbc
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_dit_recaption_rl.yaml
@@ -0,0 +1,25 @@
+# HunyuanImage-3 standalone DiT deployment for the two-engine trainer.
+pipeline: hunyuan_image3_dit
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "4,5,6,7"
+    max_num_seqs: 1
+    gpu_memory_utilization: 0.7
+    enforce_eager: true
+    trust_remote_code: true
+    vae_use_slicing: false
+    vae_use_tiling: false
+    cache_backend: null
+    cache_config: null
+    enable_cache_dit_summary: false
+    omni_kv_config:
+      need_recv_cache: true
+    parallel_config:
+      tensor_parallel_size: 4
+      enable_expert_parallel: true
+    custom_pipeline_args:
+      pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.hi3.pipeline.RLHunyuanImage3Pipeline
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_i2t_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_i2t_rl.yaml
new file mode 100644
index 000000000..dd99ba799
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_i2t_rl.yaml
@@ -0,0 +1,29 @@
+# HunyuanImage-3 multimodal AR deployment for the i2t adapter.
+pipeline: unirl_hi3_ar_multimodal_text
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "0,1,2,3"
+    max_num_seqs: 1
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.ar_extension.HI3ARWeightSyncExtension
+    enable_lora: true
+    gpu_memory_utilization: 0.95
+    enforce_eager: true
+    trust_remote_code: true
+    enable_prefix_caching: false
+    max_num_batched_tokens: 32768
+    tensor_parallel_size: 4
+    pipeline_parallel_size: 1
+    hf_overrides:
+      rope_parameters:
+        mrope_section: [0, 32, 32]
+        rope_type: default
+    default_sampling_params:
+      temperature: 0.0
+      top_p: 0.95
+      top_k: 1024
+      max_tokens: 2048
+      stop_token_ids: [127957, 128024, 128026]
+      detokenize: true
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_it2i_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_it2i_rl.yaml
new file mode 100644
index 000000000..17fdb9d9c
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_it2i_rl.yaml
@@ -0,0 +1,48 @@
+# HunyuanImage-3 image-to-image deployment with UniRL worker extensions.
+pipeline: hunyuan_image_3_moe
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "0,1,2,3"
+    max_num_seqs: 1
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.ar_extension.HI3ARWeightSyncExtension
+    enable_lora: true
+    gpu_memory_utilization: 0.95
+    enforce_eager: true
+    trust_remote_code: true
+    enable_prefix_caching: false
+    max_num_batched_tokens: 32768
+    tensor_parallel_size: 4
+    pipeline_parallel_size: 1
+    hf_overrides:
+      rope_parameters:
+        mrope_section: [0, 32, 32]
+        rope_type: default
+    default_sampling_params:
+      temperature: 0.6
+      top_p: 0.95
+      top_k: 1024
+      max_tokens: 4096
+      stop_token_ids: [127957]
+      detokenize: true
+
+  - stage_id: 1
+    devices: "4,5,6,7"
+    max_num_seqs: 1
+    enforce_eager: true
+    trust_remote_code: true
+    parallel_config:
+      tensor_parallel_size: 4
+      enable_expert_parallel: true
+    custom_pipeline_args:
+      pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.hi3.pipeline.RLHunyuanImage3Pipeline
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
+    default_sampling_params:
+      num_inference_steps: 50
+      guidance_scale: 2.5
+
+edges:
+  - from: 0
+    to: 1
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_t2i_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_t2i_rl.yaml
new file mode 100644
index 000000000..20839b06f
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_image3_t2i_rl.yaml
@@ -0,0 +1,66 @@
+# HunyuanImage-3 text-to-image deployment with UniRL worker extensions.
+pipeline: hunyuan_image_3_moe
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "0,1,2,3"
+    max_num_seqs: 1
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.ar_extension.HI3ARWeightSyncExtension
+    enable_lora: true
+    gpu_memory_utilization: 0.9
+    enforce_eager: true
+    trust_remote_code: true
+    enable_prefix_caching: false
+    max_num_batched_tokens: 32768
+    tensor_parallel_size: 4
+    pipeline_parallel_size: 1
+    hf_overrides:
+      rope_parameters:
+        mrope_section: [0, 32, 32]
+        rope_type: default
+    omni_kv_config:
+      need_send_cache: true
+    default_sampling_params:
+      temperature: 0.0
+      top_p: 1.0
+      top_k: -1
+      max_tokens: 2048
+      seed: 42
+      detokenize: true
+      repetition_penalty: 1.1
+
+  - stage_id: 1
+    devices: "4,5,6,7"
+    max_num_seqs: 1
+    enforce_eager: true
+    vae_use_slicing: false
+    vae_use_tiling: false
+    cache_backend: null
+    cache_config: null
+    enable_cache_dit_summary: false
+    omni_kv_config:
+      need_recv_cache: true
+    parallel_config:
+      pipeline_parallel_size: 1
+      data_parallel_size: 1
+      tensor_parallel_size: 4
+      enable_expert_parallel: false
+      sequence_parallel_size: 1
+      ulysses_degree: 1
+      ring_degree: 1
+      cfg_parallel_size: 1
+      vae_patch_parallel_size: 1
+      use_hsdp: false
+      hsdp_shard_size: -1
+      hsdp_replicate_size: 1
+    custom_pipeline_args:
+      pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.hi3.pipeline.RLHunyuanImage3Pipeline
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
+
+edges:
+  - from: 0
+    to: 1
+    window_size: -1
+    max_inflight: 1
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_video15_t2v_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_video15_t2v_rl.yaml
new file mode 100644
index 000000000..4e73f3dc0
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/hunyuan_video15_t2v_rl.yaml
@@ -0,0 +1,34 @@
+# HunyuanVideo-1.5 single-stage text-to-video deployment for RL rollout.
+pipeline: hunyuan_video_15
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "0"
+    max_num_seqs: 1
+    enforce_eager: true
+    vae_use_slicing: false
+    vae_use_tiling: false
+    cache_backend: null
+    cache_config: null
+    enable_cache_dit_summary: false
+    diffusion_attention_backend: TORCH_SDPA
+    enable_lora: true
+    max_lora_rank: 64
+    parallel_config:
+      pipeline_parallel_size: 1
+      data_parallel_size: 1
+      tensor_parallel_size: 1
+      enable_expert_parallel: false
+      sequence_parallel_size: 1
+      ulysses_degree: 1
+      ring_degree: 1
+      cfg_parallel_size: 1
+      vae_patch_parallel_size: 1
+      use_hsdp: false
+      hsdp_shard_size: -1
+      hsdp_replicate_size: 1
+    custom_pipeline_args:
+      pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.hv15.pipeline.RLHunyuanVideo15Pipeline
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/qwen3_omni_thinker_only_rl_1x4.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/qwen3_omni_thinker_only_rl_1x4.yaml
new file mode 100644
index 000000000..984d20fa1
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/qwen3_omni_thinker_only_rl_1x4.yaml
@@ -0,0 +1,30 @@
+# Qwen3-Omni Thinker-only deployment, TP=4.
+pipeline: unirl_qwen3_omni_thinker
+async_chunk: false
+distributed_executor_backend: mp
+dtype: bfloat16
+trust_remote_code: true
+enable_prefix_caching: false
+enable_chunked_prefill: true
+
+stages:
+  - stage_id: 0
+    devices: "0,1,2,3"
+    tensor_parallel_size: 4
+    gpu_memory_utilization: 0.35
+    max_num_seqs: 64
+    max_num_batched_tokens: 32768
+    limit_mm_per_prompt:
+      audio: 1
+      video: 1
+      image: 1
+    mm_processor_cache_gb: 0
+    load_format: safetensors
+    enforce_eager: false
+    enable_lora: true
+    max_lora_rank: 64
+    max_loras: 1
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.qwen3_omni_ar_extension.Qwen3OmniARWeightSyncExtension
+    enable_sleep_mode: true
+    logprobs_mode: processed_logprobs
+    disable_log_stats: true
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/qwen3_omni_thinker_only_rl_audio_video_1x4.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/qwen3_omni_thinker_only_rl_audio_video_1x4.yaml
new file mode 100644
index 000000000..75b73c381
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/qwen3_omni_thinker_only_rl_audio_video_1x4.yaml
@@ -0,0 +1,30 @@
+# Qwen3-Omni Thinker-only audio-in-video deployment, TP=4.
+pipeline: unirl_qwen3_omni_thinker
+async_chunk: false
+distributed_executor_backend: mp
+dtype: bfloat16
+trust_remote_code: true
+enable_prefix_caching: false
+enable_chunked_prefill: true
+
+stages:
+  - stage_id: 0
+    devices: "0,1,2,3"
+    tensor_parallel_size: 4
+    gpu_memory_utilization: 0.35
+    max_num_seqs: 1
+    max_num_batched_tokens: 32768
+    limit_mm_per_prompt:
+      audio: 1
+      video: 1
+      image: 0
+    mm_processor_cache_gb: 0
+    load_format: safetensors
+    enforce_eager: false
+    enable_lora: true
+    max_lora_rank: 64
+    max_loras: 1
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.qwen3_omni_ar_extension.Qwen3OmniARWeightSyncExtension
+    enable_sleep_mode: true
+    logprobs_mode: processed_logprobs
+    disable_log_stats: true
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/qwen_image_t2i_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/qwen_image_t2i_rl.yaml
new file mode 100644
index 000000000..fb089da30
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/qwen_image_t2i_rl.yaml
@@ -0,0 +1,33 @@
+# Single-stage Qwen-Image deployment for RL rollout.
+pipeline: unirl_rl_image_diffusion
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "0"
+    max_num_seqs: 1
+    enforce_eager: true
+    vae_use_slicing: false
+    vae_use_tiling: false
+    cache_backend: null
+    cache_config: null
+    enable_cache_dit_summary: false
+    enable_lora: true
+    max_lora_rank: 64
+    parallel_config:
+      pipeline_parallel_size: 1
+      data_parallel_size: 1
+      tensor_parallel_size: 1
+      enable_expert_parallel: false
+      sequence_parallel_size: 1
+      ulysses_degree: 1
+      ring_degree: 1
+      cfg_parallel_size: 1
+      vae_patch_parallel_size: 1
+      use_hsdp: false
+      hsdp_shard_size: -1
+      hsdp_replicate_size: 1
+    custom_pipeline_args:
+      pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.qwen_image.pipeline.RLQwenImagePipeline
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
diff --git a/unirl/rollout/engine/vllm_omni/deploy_configs/sd35_t2i_rl.yaml b/unirl/rollout/engine/vllm_omni/deploy_configs/sd35_t2i_rl.yaml
new file mode 100644
index 000000000..168f76c8b
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/deploy_configs/sd35_t2i_rl.yaml
@@ -0,0 +1,33 @@
+# Single-stage SD3.5-medium deployment for RL rollout.
+pipeline: unirl_rl_image_diffusion
+async_chunk: false
+distributed_executor_backend: mp
+
+stages:
+  - stage_id: 0
+    devices: "0"
+    max_num_seqs: 1
+    enforce_eager: true
+    vae_use_slicing: false
+    vae_use_tiling: false
+    cache_backend: null
+    cache_config: null
+    enable_cache_dit_summary: false
+    enable_lora: true
+    max_lora_rank: 32
+    parallel_config:
+      pipeline_parallel_size: 1
+      data_parallel_size: 1
+      tensor_parallel_size: 1
+      enable_expert_parallel: false
+      sequence_parallel_size: 1
+      ulysses_degree: 1
+      ring_degree: 1
+      cfg_parallel_size: 1
+      vae_patch_parallel_size: 1
+      use_hsdp: false
+      hsdp_shard_size: -1
+      hsdp_replicate_size: 1
+    custom_pipeline_args:
+      pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.sd3.pipeline.RLStableDiffusion3Pipeline
+    worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
diff --git a/unirl/rollout/engine/vllm_omni/engine.py b/unirl/rollout/engine/vllm_omni/engine.py
index 194b38139..0033499b4 100644
--- a/unirl/rollout/engine/vllm_omni/engine.py
+++ b/unirl/rollout/engine/vllm_omni/engine.py
@@ -129,7 +129,7 @@ def _mark_consistently_offloaded(self) -> None:
 
     @distributed(dispatch_mode=Dispatch.BROADCAST)
     def sleep(self) -> None:
-        """Fan ``handle_sleep_task`` to every stage's workers (level 1)."""
+        """Sleep every stage at level 1 (AR via EngineCore, diffusion via worker task)."""
         if self._is_offloaded and not self._transition_failed:
             return
         try:
@@ -145,7 +145,7 @@ def sleep(self) -> None:
 
     @distributed(dispatch_mode=Dispatch.BROADCAST)
     def wake_up(self) -> None:
-        """Fan ``handle_wake_task`` to every stage's workers + restore LoRA."""
+        """Wake every stage (AR via EngineCore, diffusion via worker task) + restore LoRA."""
         if self._transition_failed:
             # Recover an unknown partial stage state to one known boundary
             # before attempting another wake. If this retry fails, retain the
diff --git a/unirl/rollout/engine/vllm_omni/patches/README.md b/unirl/rollout/engine/vllm_omni/patches/README.md
index e318c649f..2a8f3184e 100644
--- a/unirl/rollout/engine/vllm_omni/patches/README.md
+++ b/unirl/rollout/engine/vllm_omni/patches/README.md
@@ -39,7 +39,7 @@ All in `runtime.py` unless noted.
 | Patch | Why | DELETE-WHEN |
 | --- | --- | --- |
 | `wrap_mp_process_for_children` | Re-installs the bundle inside every spawn child. **Must run first** | the rest of the bundle is empty |
-| `register_capture_flush` (`plugin.py`) | Flushes `_unirl_captures` into formatter `metadata["unirl"]` after postprocess. The `vllm_omni.general_plugins` entry point is not loaded in `StageDiffusionProc` (HI3's 2-stage path), so hijack installs it in spawn children | vllm-omni loads general plugins in every stage process including `StageDiffusionProc` |
+| `register_capture_flush` (`plugin.py`) | Flushes `_unirl_captures` into formatter `metadata["unirl"]` after postprocess. Stable 0.28 loads the general plugin in every stage process, so no child-process monkeypatch is needed | vllm-omni exposes the captured RL metadata natively |
 | `patch_dit_lora_loader` / `patch_ar_lora_loader` | Stock `DiffusionLoRAManager._load_adapter` loads only from a file path; RL pushes freshly-trained adapter tensors without a disk round-trip (`OmniTensorLoRARequest`). Lifted verbatim from verl-omni | vllm-omni's LoRA managers accept tensor-bag requests natively |
 | `patch_dit_hi3_lora_weights` | Resolves HI3 DiT `transformer.layers.*` wrappers against PEFT `model.layers.*` keys and converts GQA-interleaved fused-QKV LoRA-B rows to vLLM's packed `[q, k, v]` slices | [vllm-omni #6411](https://github.com/vllm-project/vllm-omni/issues/6411) is fixed and the pinned release includes it |
 | `patch_fp32_skip` | Punica kernels hard-assert dtype; HI3's MoE router gate is fp32, so non-fp16/bf16 layers must be skipped for LoRA wrapping | vllm's `from_layer` skips unsupported dtypes itself |
@@ -50,8 +50,8 @@ All in `runtime.py` unless noted.
 | `patch_sigmas_passthrough` | HI3's DiT `scheduler.set_timesteps` never receives `sampling_params.sigmas` | upstream forwards `sigmas` itself |
 | `install_fate_sharing` | `PR_SET_PDEATHSIG` is bound by Linux to the **specific creating thread**, so arming it for children of short-lived init threads kills healthy workers; and a worker inside a CUDA/NCCL call never observes vLLM's `death_pipe` EOF | vllm's own child-reaping is thread-safe |
 | `compat_tokenizer` (module) | HI3's `__init__` looks up `<img_ratio_36>` and computes `ratio_36 + 1`; the Base checkpoint ships ratio tokens 0-32 only → `TypeError: … 'NoneType' and 'int'`. Both the slow **and** fast tokenizer classes must be patched, not the shared base. The module import *is* the install trigger (it is the `HI3ARWorkerExtension` qualname target). Upstream ≥ v0.20.0 raises a clean `ValueError` instead — a better error, but the Base ckpt still needs this 0-fallback to work | Base-ckpt support is dropped (Instruct ships the tokens) |
-| `compat_hi3_lora` (module) | vllm 0.20 expects a flat `list[tuple[str, str, int, str]]` from `get_expert_mapping`; HI3 returns a 2-tuple, so `process_packed_modules_mapping` trips `ValueError: too many values to unpack` at boot under `enable_lora` | vllm handles the 2-tuple / HI3 returns the flat list |
-| `patch_moe_workspace_pool` (`compat_moe_workspace`) | vLLM's reusable MoE `WorkspaceManager` scratch buffer is allocated outside the `weights`/`kv_cache` CuMem pools, so worker sleep leaves it resident through the colocated train phase. Gives each DBO ubatch an independently replaceable CuMem pool while tagging all allocations `moe_workspace`; worker sleep then discards the scratch pages with the other non-weight pools and UniRL's untagged wake remaps them. Self-skips when the allocator is absent (sleep mode off → `CuMemAllocator.instance is None`) | vLLM allocates the MoE workspace inside a sleep-managed pool itself |
+| `compat_hi3_lora` (module) | HI3's AR model returns `(expert_mapping, replacements)`, while vLLM 0.28's LoRA manager still iterates the result as a flat four-field mapping and fails at boot | vLLM consumes the mapping-plus-replacements tuple natively |
+| `patch_moe_workspace_pool` (`compat_moe_workspace`) | vLLM's reusable MoE `WorkspaceManager` scratch buffer is allocated outside the `weights`/`kv_cache` CuMem pools, so worker sleep leaves it resident through the colocated train phase. Gives each DBO×lane workspace an independently replaceable CuMem pool while tagging all allocations `moe_workspace`; worker sleep then discards the scratch pages and UniRL's untagged wake remaps them. Self-skips when the allocator is absent | vLLM allocates the MoE workspace inside a sleep-managed pool itself |
 
 ## HI3 DiT LoRA namespace and layout
 
@@ -80,6 +80,6 @@ known-invalid adapter.
   submodules, loaded lazily — `import unirl.rollout.engine.vllm_omni.patches` must
   not pull vllm.
 - **Every patch needs a DELETE-WHEN row.** Without one it is permanent by default.
-- **0.27 honors `engine_args.diffusion_attention_backend`, not `attention_backend`.**
-  The old AR field is silently dropped by `OmniDiffusionConfig.from_kwargs`;
-  HV1.5 pins the diffusion backend in its stage YAML.
+- **0.28 honors `diffusion_attention_backend`, not `attention_backend`.**
+  The AR field is silently dropped by `OmniDiffusionConfig.from_kwargs`;
+  HV1.5 pins the diffusion backend in its deploy config.
diff --git a/unirl/rollout/engine/vllm_omni/patches/compat_hi3_lora.py b/unirl/rollout/engine/vllm_omni/patches/compat_hi3_lora.py
index e735a6335..8120a4c8f 100644
--- a/unirl/rollout/engine/vllm_omni/patches/compat_hi3_lora.py
+++ b/unirl/rollout/engine/vllm_omni/patches/compat_hi3_lora.py
@@ -1,4 +1,4 @@
-"""Runtime patch making ``get_moe_expert_mapping`` tolerate HI3's 2-tuple ``get_expert_mapping`` shape."""
+"""Make vLLM's MoE LoRA mapping accept HI3's mapping-plus-replacements tuple."""
 
 from __future__ import annotations
 
@@ -6,7 +6,7 @@
 
 
 def install() -> None:
-    """Patch ``get_moe_expert_mapping`` everywhere it's used, unwrapping HI3's 2-tuple to vllm 0.20's flat list."""
+    """Unwrap HI3's ``(expert_mapping, replacements)`` result for vLLM LoRA."""
     global _INSTALLED
     if _INSTALLED:
         return
diff --git a/unirl/rollout/engine/vllm_omni/patches/compat_moe_workspace.py b/unirl/rollout/engine/vllm_omni/patches/compat_moe_workspace.py
index 3fa1d336d..2a58912d2 100644
--- a/unirl/rollout/engine/vllm_omni/patches/compat_moe_workspace.py
+++ b/unirl/rollout/engine/vllm_omni/patches/compat_moe_workspace.py
@@ -10,16 +10,16 @@
 _WORKSPACE_MANAGER_MARKER = "_diffrl_moe_workspace_pool_installed"
 
 
-def _ubatch_pool_key(ubatch_id: int) -> str:
-    return f"{MOE_WORKSPACE_TAG}:{ubatch_id}"
+def _workspace_pool_key(workspace_id: int) -> str:
+    return f"{MOE_WORKSPACE_TAG}:{workspace_id}"
 
 
 @contextmanager
-def _workspace_pool_context(allocator: Any, ubatch_id: int) -> Iterator[None]:
-    """Create one replaceable pool per ubatch while keeping one sleep tag."""
-    # The dictionary key is ubatch-specific so DBO slots do not replace one
-    # another, while AllocationData keeps one model-level workspace tag.
-    pool_key = _ubatch_pool_key(ubatch_id)
+def _workspace_pool_context(allocator: Any, workspace_id: int) -> Iterator[None]:
+    """Create one replaceable pool per workspace while keeping one sleep tag."""
+    # The dictionary key is per-workspace so DBO and lane slots do not replace
+    # one another, while AllocationData keeps one model-level workspace tag.
+    pool_key = _workspace_pool_key(workspace_id)
     previous_tag = allocator.current_tag
     with allocator.use_memory_pool(tag=pool_key):
         allocator.current_tag = MOE_WORKSPACE_TAG
@@ -30,9 +30,9 @@ def _workspace_pool_context(allocator: Any, ubatch_id: int) -> Iterator[None]:
             allocator.current_tag = previous_tag
 
 
-def _release_ubatch_pool(allocator: Any, ubatch_id: int) -> None:
+def _release_workspace_pool(allocator: Any, workspace_id: int) -> None:
     """Release a retained MemPool before its pluggable allocator."""
-    pool_state = allocator.allocator_and_pools.pop(_ubatch_pool_key(ubatch_id), None)
+    pool_state = allocator.allocator_and_pools.pop(_workspace_pool_key(workspace_id), None)
     if pool_state is None:
         return
     memory_pool, pluggable_allocator = pool_state
@@ -46,7 +46,7 @@ def _patch_workspace_manager_class(
     workspace_manager_class: type,
     *,
     allocator_provider: Callable[[], Any],
-    ubatch_provider: Callable[[], int],
+    workspace_id_provider: Callable[[Any], int | None],
 ) -> None:
     """Route ``_ensure_workspace_size`` growth through the dedicated CuMem tag."""
     if getattr(workspace_manager_class, _WORKSPACE_MANAGER_MARKER, False):
@@ -55,12 +55,14 @@ def _patch_workspace_manager_class(
     original_ensure_workspace_size = workspace_manager_class._ensure_workspace_size
 
     def ensure_workspace_size(manager: Any, required_bytes: int) -> Any:
-        ubatch_id = ubatch_provider()
         try:
-            current_workspace = manager._current_workspaces[ubatch_id]
+            workspace_id = workspace_id_provider(manager)
+            if workspace_id is None:
+                return original_ensure_workspace_size(manager, required_bytes)
+            current_workspace = manager._current_workspaces[workspace_id]
             current_size = manager._workspace_size_bytes(current_workspace)
             workspace_locked = manager.is_locked()
-        except AttributeError:
+        except (AttributeError, IndexError):
             return original_ensure_workspace_size(manager, required_bytes)
         if current_size >= required_bytes:
             return original_ensure_workspace_size(manager, required_bytes)
@@ -77,21 +79,21 @@ def ensure_workspace_size(manager: Any, required_bytes: int) -> Any:
         import torch
 
         # ``use_memory_pool`` replaces the retained MemPool for a key. Release
-        # this ubatch's old tensor before replacing its pool; reversing that
+        # this slot's old tensor before replacing its pool; reversing that
         # order aborts in MemPool::~MemPool on torch 2.11.
         replacing_workspace = current_workspace is not None
         if replacing_workspace:
             # vLLM 0.20's CuMem free callback can unmap immediately, so a
             # device-wide sync must drain every stream first.
             torch.accelerator.synchronize()
-            manager._current_workspaces[ubatch_id] = None
+            manager._current_workspaces[workspace_id] = None
             del current_workspace
             gc.collect()
-        _release_ubatch_pool(allocator, ubatch_id)
+        _release_workspace_pool(allocator, workspace_id)
         if replacing_workspace:
             torch.accelerator.empty_cache()
 
-        with _workspace_pool_context(allocator, ubatch_id):
+        with _workspace_pool_context(allocator, workspace_id):
             return original_ensure_workspace_size(manager, required_bytes)
 
     workspace_manager_class._ensure_workspace_size = ensure_workspace_size
@@ -103,16 +105,25 @@ def patch_moe_workspace_pool() -> None:
     try:
         from vllm.device_allocator.cumem import CuMemAllocator
         from vllm.v1.worker.ubatching import dbo_current_ubatch_id
-        from vllm.v1.worker.workspace import WorkspaceManager
+        from vllm.v1.worker.workspace import WorkspaceManager, _workspace_lane
 
         WorkspaceManager._ensure_workspace_size
     except (ImportError, AttributeError):
         return
 
+    def workspace_id_provider(manager: Any) -> int | None:
+        # Delegate invalid lanes before indexing. A lane can otherwise flatten
+        # onto the next ubatch's valid slot and release an unrelated workspace.
+        lane = _workspace_lane.get()
+        num_lanes = manager._num_lanes
+        if lane >= num_lanes:
+            return None
+        return dbo_current_ubatch_id() * num_lanes + lane
+
     _patch_workspace_manager_class(
         WorkspaceManager,
         allocator_provider=lambda: CuMemAllocator.instance,
-        ubatch_provider=dbo_current_ubatch_id,
+        workspace_id_provider=workspace_id_provider,
     )
 
 
diff --git a/unirl/rollout/engine/vllm_omni/patches/runtime.py b/unirl/rollout/engine/vllm_omni/patches/runtime.py
index e47a49acf..3631fe324 100644
--- a/unirl/rollout/engine/vllm_omni/patches/runtime.py
+++ b/unirl/rollout/engine/vllm_omni/patches/runtime.py
@@ -11,15 +11,9 @@
 
 import torch
 from msgspec import field
-
-try:
-    from vllm.lora.lora_model import LoRAModel
-except ImportError:
-    from vllm.lora.models import LoRAModel  # type: ignore[no-redef]
-
+from vllm.lora.lora_model import LoRAModel
 from vllm.lora.lora_weights import LoRALayerWeights, PackedLoRALayerWeights
 from vllm.lora.peft_helper import PEFTHelper
-from vllm.lora.utils import get_adapter_absolute_path
 from vllm_omni.diffusion.lora.manager import DiffusionLoRAManager, logger
 from vllm_omni.lora.request import LoRARequest as OmniLoRARequest
 
@@ -137,28 +131,23 @@ def start(self):
 
 def patch_dit_lora_loader() -> None:
     """Patch ``DiffusionLoRAManager._load_adapter`` (DiT stage) to support in-memory tensors."""
+    original = DiffusionLoRAManager._load_adapter
+    if getattr(original, "_diffrl_tensor_lora_loader", False):
+        return
 
-    def hijack__load_adapter(self, lora_request: OmniTensorLoRARequest) -> tuple[LoRAModel, PEFTHelper]:
+    def hijack__load_adapter(
+        self,
+        lora_request: OmniTensorLoRARequest,
+        _orig=original,
+    ) -> tuple[LoRAModel, PEFTHelper]:
+        if not isinstance(lora_request, OmniTensorLoRARequest):
+            return _orig(self, lora_request)
         if not self._expected_lora_modules:
             raise ValueError("No supported LoRA modules found in the diffusion pipeline.")
 
         logger.debug("Supported LoRA modules: %s", self._expected_lora_modules)
 
-        lora_tensors = None
-
-        if isinstance(lora_request, OmniTensorLoRARequest):
-            peft_config = lora_request.peft_config
-            lora_tensors = lora_request.lora_tensors
-            peft_helper = PEFTHelper.from_dict(peft_config)
-        else:
-            lora_path = get_adapter_absolute_path(lora_request.lora_path)
-            logger.debug("Resolved LoRA path: %s", lora_path)
-
-            peft_helper = PEFTHelper.from_local_dir(
-                lora_path,
-                max_position_embeddings=None,
-                tensorizer_config_dict=lora_request.tensorizer_config_dict,
-            )
+        peft_helper = PEFTHelper.from_dict(lora_request.peft_config or {})
 
         logger.info(
             "Loaded PEFT config: r=%d, lora_alpha=%d, target_modules=%s",
@@ -167,28 +156,15 @@ def hijack__load_adapter(self, lora_request: OmniTensorLoRARequest) -> tuple[LoR
             peft_helper.target_modules,
         )
 
-        if isinstance(lora_request, OmniTensorLoRARequest):
-            lora_model = LoRAModel.from_lora_tensors(
-                tensors=lora_tensors,
-                peft_helper=peft_helper,
-                lora_model_id=lora_request.lora_int_id,
-                device="cpu",
-                dtype=self.dtype,
-                model_vocab_size=None,
-                weights_mapper=None,
-            )
-        else:
-            lora_model = LoRAModel.from_local_checkpoint(
-                lora_path,
-                expected_lora_modules=self._expected_lora_modules,
-                peft_helper=peft_helper,
-                lora_model_id=lora_request.lora_int_id,
-                device="cpu",
-                dtype=self.dtype,
-                model_vocab_size=None,
-                tensorizer_config_dict=lora_request.tensorizer_config_dict,
-                weights_mapper=None,
-            )
+        lora_model = LoRAModel.from_lora_tensors(
+            tensors=lora_request.lora_tensors or {},
+            peft_helper=peft_helper,
+            lora_model_id=lora_request.lora_int_id,
+            device="cpu",
+            dtype=self.dtype,
+            model_vocab_size=None,
+            weights_mapper=None,
+        )
 
         logger.info(
             "Loaded LoRA model: id=%d, num_modules=%d, modules=%s",
@@ -202,6 +178,7 @@ def hijack__load_adapter(self, lora_request: OmniTensorLoRARequest) -> tuple[LoR
 
         return lora_model, peft_helper
 
+    hijack__load_adapter._diffrl_tensor_lora_loader = True  # type: ignore[attr-defined]
     setattr(DiffusionLoRAManager, "_load_adapter", hijack__load_adapter)
 
 
@@ -317,6 +294,9 @@ def hijack_ar__load_adapter(self, lora_request, _orig=_orig_ar_load_adapter) ->
 
         model = self._adapter_manager.model
         hf_to_vllm_mapper = getattr(model, "hf_to_vllm_mapper", None)
+        if hf_to_vllm_mapper is not None:
+            hf_to_vllm_mapper = hf_to_vllm_mapper.get_unstacked_mapper()
+        lora_skip_prefixes = getattr(model, "lora_skip_prefixes", None)
         lora = self._lora_model_cls.from_lora_tensors(
             tensors=lora_request.lora_tensors or {},
             peft_helper=peft_helper,
@@ -325,6 +305,7 @@ def hijack_ar__load_adapter(self, lora_request, _orig=_orig_ar_load_adapter) ->
             dtype=self.lora_config.lora_dtype,
             model_vocab_size=self.vocab_size,
             weights_mapper=hf_to_vllm_mapper,
+            skip_prefixes=lora_skip_prefixes,
         )
         return lora
 
@@ -401,7 +382,6 @@ def _patched_from_layer(
 
     for _modname in (
         "vllm.lora.lora_model",
-        "vllm.lora.models",
         "vllm.lora.model_manager",
         "vllm.lora.worker_manager",
     ):
@@ -678,11 +658,6 @@ class VLLMOmniHijack:
     def hijack() -> None:
         wrap_mp_process_for_children()
 
-        # StageDiffusionProc never loads vllm_omni.general_plugins, so spawn children get the flush only via wrap_mp.
-        from unirl.rollout.engine.vllm_omni.plugin import register_capture_flush
-
-        register_capture_flush()
-
         patch_dit_lora_loader()
         patch_dit_hi3_lora_weights()
         patch_ar_lora_loader()
diff --git a/unirl/rollout/engine/vllm_omni/pipeline_configs.py b/unirl/rollout/engine/vllm_omni/pipeline_configs.py
new file mode 100644
index 000000000..e1f986a24
--- /dev/null
+++ b/unirl/rollout/engine/vllm_omni/pipeline_configs.py
@@ -0,0 +1,116 @@
+"""Out-of-tree vLLM-Omni pipeline topology registrations."""
+
+from __future__ import annotations
+
+from vllm_omni.config.pipeline_registry import OMNI_PIPELINES, register_pipeline
+from vllm_omni.config.stage_config import (
+    PipelineConfig,
+    StageExecutionType,
+    StagePipelineConfig,
+)
+
+UNIRL_RL_IMAGE_DIFFUSION = PipelineConfig(
+    model_type="unirl_rl_image_diffusion",
+    stages=(
+        StagePipelineConfig(
+            stage_id=0,
+            model_stage="diffusion",
+            execution_type=StageExecutionType.DIFFUSION,
+            final_output=True,
+            final_output_type="image",
+        ),
+    ),
+)
+
+_HI3_MODEL_ARCH = "HunyuanImage3ForCausalMM"
+
+
+def _hi3_ar_pipeline(
+    model_type: str,
+    *,
+    requires_multimodal_data: bool,
+    engine_output_type: str,
+) -> PipelineConfig:
+    return PipelineConfig(
+        model_type=model_type,
+        model_arch=_HI3_MODEL_ARCH,
+        stages=(
+            StagePipelineConfig(
+                stage_id=0,
+                model_stage="AR",
+                execution_type=StageExecutionType.LLM_AR,
+                final_output=True,
+                final_output_type="text",
+                owns_tokenizer=True,
+                requires_multimodal_data=requires_multimodal_data,
+                model_arch=_HI3_MODEL_ARCH,
+                engine_output_type=engine_output_type,
+            ),
+        ),
+    )
+
+
+UNIRL_HI3_AR_TEXT = _hi3_ar_pipeline(
+    "unirl_hi3_ar_text",
+    requires_multimodal_data=False,
+    engine_output_type="text",
+)
+UNIRL_HI3_AR_MULTIMODAL_TEXT = _hi3_ar_pipeline(
+    "unirl_hi3_ar_multimodal_text",
+    requires_multimodal_data=True,
+    engine_output_type="text",
+)
+UNIRL_HI3_AR_RECAPTION = _hi3_ar_pipeline(
+    "unirl_hi3_ar_recaption",
+    requires_multimodal_data=False,
+    engine_output_type="latent",
+)
+
+UNIRL_QWEN3_OMNI_THINKER = PipelineConfig(
+    model_type="unirl_qwen3_omni_thinker",
+    model_arch="Qwen3OmniMoeThinkerForConditionalGeneration",
+    stages=(
+        StagePipelineConfig(
+            stage_id=0,
+            model_stage="thinker",
+            execution_type=StageExecutionType.LLM_AR,
+            final_output=True,
+            final_output_type="text",
+            owns_tokenizer=True,
+            requires_multimodal_data=True,
+            hf_config_name="thinker_config",
+            model_arch="Qwen3OmniMoeThinkerForConditionalGeneration",
+            engine_output_type="text",
+            sampling_constraints={"detokenize": True},
+        ),
+    ),
+)
+
+
+def register_unirl_pipeline_configs() -> None:
+    """Register UniRL's custom topologies once, rejecting key collisions."""
+    for pipeline in (
+        UNIRL_RL_IMAGE_DIFFUSION,
+        UNIRL_HI3_AR_TEXT,
+        UNIRL_HI3_AR_MULTIMODAL_TEXT,
+        UNIRL_HI3_AR_RECAPTION,
+        UNIRL_QWEN3_OMNI_THINKER,
+    ):
+        existing = OMNI_PIPELINES.get(pipeline.model_type)
+        if existing is None:
+            register_pipeline(pipeline)
+        elif existing != pipeline:
+            raise RuntimeError(
+                f"vLLM-Omni pipeline key {pipeline.model_type!r} is already "
+                f"registered to an incompatible pipeline: {existing!r}"
+            )
+
+
+__all__ = [
+    "UNIRL_HI3_AR_MULTIMODAL_TEXT",
+    "UNIRL_HI3_AR_RECAPTION",
+    "UNIRL_HI3_AR_TEXT",
+    "UNIRL_QWEN3_OMNI_THINKER",
+    "UNIRL_RL_IMAGE_DIFFUSION",
+    "register_unirl_pipeline_configs",
+]
diff --git a/unirl/rollout/engine/vllm_omni/plugin.py b/unirl/rollout/engine/vllm_omni/plugin.py
index 93b47b49a..d4539cafe 100644
--- a/unirl/rollout/engine/vllm_omni/plugin.py
+++ b/unirl/rollout/engine/vllm_omni/plugin.py
@@ -9,6 +9,14 @@
 _PATCH_FLAG = "_unirl_capture_flush"
 
 
+def register_unirl_runtime() -> None:
+    """Register UniRL topology and output hooks in this vllm-omni process."""
+    from unirl.rollout.engine.vllm_omni.pipeline_configs import register_unirl_pipeline_configs
+
+    register_unirl_pipeline_configs()
+    register_capture_flush()
+
+
 def register_capture_flush() -> None:
     """Flush pipeline captures into formatter metadata after postprocess sees raw media."""
     from vllm_omni.diffusion import diffusion_engine, output_formatter
@@ -33,3 +41,6 @@ def patched(*args, **kwargs):
     # diffusion_engine took a from-import, so it holds its own binding.
     diffusion_engine.format_diffusion_outputs = patched
     logger.info("unirl: diffusion formatter flushes capture metadata after postprocess")
+
+
+__all__ = ["register_capture_flush", "register_unirl_runtime"]
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/bagel_t2i_rl.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/bagel_t2i_rl.yaml
deleted file mode 100644
index 38f82e89f..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/bagel_t2i_rl.yaml
+++ /dev/null
@@ -1,103 +0,0 @@
-# Single-stage BAGEL-7B-MoT image-out config for RL rollout.
-#
-# BAGEL is the single-stage topology where the DiT worker owns a full
-# Qwen2-MoT LLM + ViT + VAE + tokenizer and handles text2img end-to-end (no
-# separate AR/Thinker prelude — think mode is off for the GRPO path).
-#
-# Shared by BOTH image-out modalities — ``bagel_t2i`` and ``bagel_it2i``
-# (editing). The custom worker pipeline owns the ViT and editing prefill; the
-# source image travels on the request (``multi_modal_data["image"]``), not in this
-# config, so the two modalities need no stage-level difference.
-#
-# Diff vs upstream's bagel_single_stage deploy yaml:
-#   - custom_pipeline_args installs our RLBagelPipeline subclass. Unlike SD3/Qwen
-#     (which swap a diffusers scheduler), RLBagelPipeline sets self.scheduler to
-#     the BAGEL-specific BagelFlowSDEScheduler that upstream generate_image
-#     already threads through its hand-rolled denoise loop — giving per-step SDE
-#     transitions + dense trajectory capture (latents, reconstructed sigmas,
-#     sparse log_probs) whose math matches the trainside FlowSDEStrategy. It also
-#     taps bagel.prepare_vae_latent to inject the driver-authored x_T.
-#   - worker_extension_cls installs DiTWeightSyncExtension so the worker exposes
-#     update_weights_from_ipc (bucketed CUDA-IPC) and the NCCL update_weights
-#     paths the trainer's LoRA sync needs.
-#
-# Contract notes:
-#   - The adapter translates canonical guidance_scale to BAGEL's
-#     extra_args.cfg_text_scale; guidance_scale is not forwarded as a top-level
-#     worker argument. It also always sends cfg_img_scale because upstream
-#     BAGEL defaults a missing cfg_text_scale to 4.0 (CFG ON); trainside cfg=1
-#     must remain a single-forward path.
-#   - num_inference_steps is sent as trainside_steps + 1: BAGEL builds
-#     linspace(1, 0, num_timesteps) and loops num_timesteps-1 steps. BAGEL builds
-#     its sigma schedule internally (it ignores sampling_params.sigmas), with a
-#     hardwired timestep_shift=3.0 == the trainside shift, so the worker schedule
-#     equals the diffusion Part's pinned sigmas; response sigma-echo verification asserts it.
-#   - trust_remote_code: BAGEL ships custom configs (llm_config.json / vit_config.json).
-#
-# NB: no ``pipeline:`` key — the legacy stage_configs loader reads only
-# ``stage_args``, and ``custom_pipeline_args.pipeline_class`` bypasses model-type
-# detection entirely (the worker imports RLBagelPipeline by qualname, which loads
-# the BAGEL checkpoint itself), exactly like the SD3 / Qwen-Image RL configs.
-#
-# Sized for BAGEL-7B-MoT on TP=1 (navit bs=1; one sample per generate call).
-# Single-GPU rollout matches the trainer's full-forward math more closely than
-# tensor-parallel rollout and avoids the systematic bf16 drift seen in replay
-# log-prob parity checks.
-
-async_chunk: false
-
-stage_args:
-  - stage_id: 0
-    stage_type: diffusion
-    runtime:
-      process: true
-      devices: "0"
-    engine_args:
-      model_stage: dit
-      max_num_seqs: 1
-      enforce_eager: true
-      trust_remote_code: true
-      enable_prefix_caching: false
-      distributed_executor_backend: "mp"
-      vae_use_slicing: false
-      vae_use_tiling: false
-      cache_backend: null
-      cache_config: null
-      enable_cache_dit_summary: false
-      # Required for vLLM to wrap DiT Linear modules in LoRA-aware variants.
-      # Without ``enable_lora``, ``add_lora`` still registers the adapter (so
-      # checksum probes pass) but no Linear consults it and the forward silently
-      # runs base weights. ``max_lora_rank`` must be >= the recipe's LoRA rank
-      # (64 in the BAGEL LoRA recipes); vLLM's default of 16 rejects larger ranks.
-      enable_lora: true
-      max_lora_rank: 64
-      parallel_config:
-        pipeline_parallel_size: 1
-        data_parallel_size: 1
-        tensor_parallel_size: 1
-        enable_expert_parallel: false
-        sequence_parallel_size: 1
-        ulysses_degree: 1
-        ring_degree: 1
-        cfg_parallel_size: 1
-        vae_patch_parallel_size: 1
-        use_hsdp: false
-        hsdp_shard_size: -1
-        hsdp_replicate_size: 1
-      # unirl-specific: install RLBagelPipeline so SDE-mode rollouts get the
-      # trajectory-capturing scheduler + driver-noise injection.
-      custom_pipeline_args:
-        pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.bagel.pipeline.RLBagelPipeline
-      # unirl-specific: install the weight-sync extension so the DiT worker
-      # exposes update_weights_from_ipc (bucketed CUDA-IPC) and later
-      # init_weight_transfer_engine / update_weights (NCCL). Coexists with the
-      # custom pipeline above; vllm-omni composes them via multiple inheritance.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
-    final_output: true
-    final_output_type: image
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 1
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_ar_recaption_rl.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_ar_recaption_rl.yaml
deleted file mode 100644
index 1bac171e7..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_ar_recaption_rl.yaml
+++ /dev/null
@@ -1,77 +0,0 @@
-# AR-ONLY think_recaption stage config (two-engine v2 trainer).
-#
-# A single, TERMINAL AR (LLM) stage that, in is_comprehension:false mode,
-# generates <think>...</think><recaption>...</recaption><answer><boi><img_size>...
-# and stops at EOS. The full AR text is the recaption the DiT engine consumes as
-# cot_text (extra["ar_generated_text"]); per-token logp (logprobs=1) feeds
-# ar_capture.extract_ar_segment for ARGRPO replay.
-#
-# Differs from upstream i2t/t2t (which are is_comprehension:true comprehension
-# answers, NOT think/recaption) — see hi3_v2_IMPL_spec.md plan-correction #1.
-#
-# Terminal AR-stage specifics:
-#   - final_output_type stays text (terminal; no DiT downstream).
-#   - omni_kv_config.need_send_cache DROPPED (no DiT stage to receive KV).
-#
-# Devices 0-3 (AR engine half of the 8-GPU colocate pool); the DiT engine
-# (hunyuan_image3_dit_recaption_rl.yaml) takes 4-7.
-
-stage_args:
-  - stage_id: 0
-    stage_type: llm
-    runtime:
-      process: true
-      devices: "0,1,2,3"
-    engine_args:
-      model_stage: AR
-      max_num_seqs: 1
-      model_arch: HunyuanImage3ForCausalMM
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      # Side-effect import installs the Base-ckpt tokenizer fix AND composes
-      # the bucketed CUDA-IPC weight-sync mixin (LoRA sync target). See
-      # t2i_rl.yaml for the full rationale.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.ar_extension.HI3ARWeightSyncExtension
-      enable_lora: true
-      # Routes MoE LoRA off vLLM 0.27's fused one-shot kernel, which faults here.
-      fully_sharded_loras: true
-      # Colocate-memory trim: the AR stage only needs prompt + max_new_tokens
-      # (~4096), not the model's full 22800 ctx. Capping max_model_len shrinks
-      # the KV-cache reservation (~21 GiB → a few GiB) and gpu_memory_util 0.9→0.7
-      # lowers the resident/slept footprint — both free room on the shared train
-      # cards (engine sleep keeps a few GiB on GPU 0-3, which the FSDP backward
-      # competes with). max_num_batched_tokens 32768→8192 also clears the
-      # "exceeds max_num_seqs*max_model_len" boot warning.
-      max_model_len: 8192
-      gpu_memory_utilization: 0.7
-      enforce_eager: true
-      trust_remote_code: true
-      engine_output_type: latent
-      enable_prefix_caching: false
-      max_num_batched_tokens: 8192
-      tensor_parallel_size: 4
-      pipeline_parallel_size: 1
-      distributed_executor_backend: "mp"
-      hf_overrides:
-        rope_parameters:
-          mrope_section: [0, 32, 32]
-          rope_type: default
-    # is_comprehension:false → AR generates think/recaption (not a plain
-    # comprehension answer). No omni_kv_config: this stage is terminal.
-    is_comprehension: false
-    final_output: true
-    final_output_type: text
-    # No default seed on purpose: a fixed seed + temperature would make the N
-    # temperature-sampled recaptions of a prompt identical (→ AR GRPO advantage
-    # 0). Omitting it lets vLLM draw a fresh seed per request so the N recaptions
-    # actually differ.
-    default_sampling_params:
-      temperature: 0.7
-      top_p: 0.95
-      top_k: 1024
-      max_tokens: 4096
-      detokenize: True
-      repetition_penalty: 1.1
-
-runtime:
-  enabled: true
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_dit_recaption_rl.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_dit_recaption_rl.yaml
deleted file mode 100644
index 0e0312905..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_dit_recaption_rl.yaml
+++ /dev/null
@@ -1,68 +0,0 @@
-# DiT-ONLY stage config (two-engine v2 trainer) — critical gotcha.
-#
-# A single standalone diffusion stage that consumes an EXTERNAL recaption as
-# cot_text (injected per-prompt via extra["ar_generated_text"] by request.py's
-# dit_recaption branch) and renders M images at distinct noise. No AR stage,
-# no KV transfer — the DiT pipeline self-prefills the text context from
-# prompt + cot_text (RLHunyuanImage3Pipeline.forward → upstream forward →
-# prepare_model_inputs(mode="gen_image")).
-#
-# Base = upstream hunyuan_image3_t2i.yaml, the ONLY HI3 DiT config verified
-# standalone (single diffusion stage, 4x H20): enable_expert_parallel:true,
-# need_recv_cache:true (no upstream edge feeds it, so it self-prefills — same
-# as the verified standalone). Plus our RL bits from t2i_rl.yaml's DiT stage:
-#   - custom_pipeline_args → RLHunyuanImage3Pipeline (SDE scheduler swap +
-#     dense trajectory capture + fused_mm_capture for DiffusionGRPO replay).
-#   - worker_extension_cls → DiTWeightSyncExtension (LoRA receive target).
-#   - vae_use_slicing/tiling:false, cache_*:null (trajectory-capture parity).
-#
-# Devices 4-7 (DiT engine half of the 8-GPU colocate pool); AR engine
-# (hunyuan_image3_ar_recaption_rl.yaml) takes 0-3.
-
-stage_args:
-  - stage_id: 0
-    stage_type: diffusion
-    runtime:
-      devices: "4,5,6,7"
-    engine_args:
-      max_num_seqs: 1
-      model_stage: dit
-      # Colocate-memory trim: lower the DiT engine's resident/slept footprint on
-      # GPU 4-7 so the FSDP backward has more room (mirrors the AR stage trim).
-      gpu_memory_utilization: 0.7
-      enforce_eager: true
-      trust_remote_code: true
-      distributed_executor_backend: "mp"
-      vae_use_slicing: false
-      vae_use_tiling: false
-      cache_backend: null
-      cache_config: null
-      enable_cache_dit_summary: false
-      # No upstream sender; the recv side is set up but no transfer arrives,
-      # so the pipeline prefills its own text context (matches the verified
-      # standalone upstream t2i.yaml, which is also need_recv_cache:true with
-      # no edge feeding it).
-      omni_kv_config:
-        need_recv_cache: true
-      parallel_config:
-        tensor_parallel_size: 4
-        enable_expert_parallel: true
-      # RL pipeline subclass: SDE scheduler + per-step trajectory + fused_mm
-      # capture (DiffusionGRPO replay needs conditions["fused"]).
-      custom_pipeline_args:
-        pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.hi3.pipeline.RLHunyuanImage3Pipeline
-      # Weight-sync extension: DiT worker exposes update_weights_from_ipc /
-      # the LoRA receive path so set_lora_from_tensors lands the adapter.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
-    final_output: true
-    final_output_type: image
-    is_comprehension: false
-    # No default seed on purpose. Per-image distinct seeds are set by the engine
-    # (one generate() per prompt, seed = seed_from_sample_id) — NOT here and NOT
-    # via the request's sampling params. A fixed default (e.g. seed:42) here, OR
-    # relying on vllm-omni's random-seed fallback, makes every per-recaption
-    # request share ONE seed → byte-identical images → diffusion advantage 0.
-    # See engine.seed_from_sample_id / request._to_omni_dit_recaption.
-
-runtime:
-  enabled: true
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_i2t.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_i2t.yaml
deleted file mode 100644
index ad0be96ef..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_i2t.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-# Vendored at its last upstream revision, since vllm-omni deleted stage_configs/ in #3172.
-# HunyuanImage-3.0 image-to-text: one AR stage reads image + prompt and emits text.
-
-stage_args:
-  - stage_id: 0
-    stage_type: llm
-    runtime:
-      process: true
-      devices: "0,1,2,3"
-      requires_multimodal_data: true
-    engine_args:
-      model_stage: AR
-      max_num_seqs: 1
-      model_arch: HunyuanImage3ForCausalMM
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.95
-      enforce_eager: true
-      trust_remote_code: true
-      enable_prefix_caching: false
-      max_num_batched_tokens: 32768
-      tensor_parallel_size: 4
-      pipeline_parallel_size: 1
-      hf_overrides:
-        rope_parameters:
-          mrope_section: [0, 32, 32]
-          rope_type: default
-    is_comprehension: true
-    final_output: true
-    final_output_type: text
-    default_sampling_params:
-      temperature: 0.0
-      top_p: 0.95
-      top_k: 1024
-      max_tokens: 2048
-      stop_token_ids: [127957, 128024, 128026]  # <|endoftext|>, </think>, </answer>
-      detokenize: True
-
-runtime:
-  enabled: true
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_it2i_rl.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_it2i_rl.yaml
deleted file mode 100644
index 0124650df..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_it2i_rl.yaml
+++ /dev/null
@@ -1,87 +0,0 @@
-# RL-aware copy of upstream hunyuan_image3_it2i.yaml.
-# Diff vs upstream:
-#   1. Stage 0 final_output: true (was false). Surfaces the AR text + per-token
-#      logp our ar_capture.py reads to fill the AR generation Part's segment. The KV transfer
-#      to Stage 1 is independent of this flag, so the AR→DiT bridge is
-#      unaffected.
-#   2. Stage 1 engine_args.custom_pipeline_args injected to install our
-#      RLHunyuanImage3Pipeline subclass (scheduler swap + trajectory drain
-#      for SDE-mode rollouts).
-
-stage_args:
-  - stage_id: 0
-    stage_type: llm
-    runtime:
-      process: true
-      devices: "0,1,2,3"
-      requires_multimodal_data: true
-    engine_args:
-      max_num_seqs: 1
-      model_stage: AR
-      model_arch: HunyuanImage3ForCausalMM
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      # See t2i YAML for rationale; same tokenizer patch + weight-sync mixin.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.ar_extension.HI3ARWeightSyncExtension
-      # Enable LoRA on the AR stage; mirrors the t2i YAML.
-      enable_lora: true
-      gpu_memory_utilization: 0.95
-      enforce_eager: true
-      trust_remote_code: true
-      engine_output_type: latent
-      enable_prefix_caching: false
-      max_num_batched_tokens: 32768
-      tensor_parallel_size: 4
-      pipeline_parallel_size: 1
-      hf_overrides:
-        rope_parameters:
-          mrope_section: [0, 32, 32]
-          rope_type: default
-    is_comprehension: false
-    # unirl-specific: was `false` upstream — flipped so AR tokens +
-    # per-token logp surface to the orchestrator's caller for RL.
-    final_output: true
-    final_output_type: text
-    default_sampling_params:
-      temperature: 0.6
-      top_p: 0.95
-      top_k: 1024
-      max_tokens: 4096
-      stop_token_ids: [127957]
-      detokenize: true
-
-  - stage_id: 1
-    stage_type: diffusion
-    runtime:
-      process: true
-      devices: "4,5,6,7"
-      requires_multimodal_data: true
-    engine_args:
-      max_num_seqs: 1
-      model_stage: dit
-      model_arch: HunyuanImage3ForCausalMM
-      enforce_eager: true
-      trust_remote_code: true
-      distributed_executor_backend: "mp"
-      parallel_config:
-        tensor_parallel_size: 4
-        enable_expert_parallel: true
-      # unirl-specific: install RLHunyuanImage3Pipeline.
-      custom_pipeline_args:
-        pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.hi3.pipeline.RLHunyuanImage3Pipeline
-      # unirl-specific: install the weight-sync extension on the DiT
-      # worker (mirrors t2i YAML — see that file for rationale).
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
-    engine_input_source: [0]
-    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.hunyuan_image3.ar2diffusion
-    final_output: true
-    final_output_type: image
-    default_sampling_params:
-      num_inference_steps: 50
-      guidance_scale: 2.5
-
-runtime:
-  enabled: true
-  edges:
-    - from: 0
-      to: 1
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_t2i_rl.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_t2i_rl.yaml
deleted file mode 100644
index 44a5a3642..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_t2i_rl.yaml
+++ /dev/null
@@ -1,126 +0,0 @@
-# RL-aware copy of upstream hunyuan_image3_moe.yaml.
-# Diff vs upstream:
-#   - Stage 1 engine_args.custom_pipeline_args injected to install our
-#     RLHunyuanImage3Pipeline subclass (scheduler swap + trajectory drain
-#     for SDE-mode rollouts).
-# Stage 0 (AR) is unchanged: upstream already ships final_output: true,
-# which surfaces the AR text + per-token logp our ar_capture.py reads.
-#
-# Verified on 8x L40S-48G GPU (4 for AR + 4 for DiT) per upstream docs.
-
-stage_args:
-  - stage_id: 0
-    stage_type: llm
-    runtime:
-      process: true
-      devices: "0,1,2,3"
-    engine_args:
-      model_stage: AR
-      max_num_seqs: 1
-      model_arch: HunyuanImage3ForCausalMM
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      # Side-effect import: ``HI3ARWeightSyncExtension`` subclasses
-      # ``HI3ARWorkerExtension`` (in ``compat/tokenizer``), so resolving
-      # the qualname imports that module and triggers ``install()`` —
-      # patching ``convert_tokens_to_ids`` -> 0-on-None for the Base
-      # HunyuanImage-3 ckpt (its tokenizer is missing ``<img_ratio_33+>``;
-      # vllm-omni's HI3 model does ratio_36 + 1 unconditionally and would
-      # otherwise TypeError). No-op on Instruct since the lookup returns
-      # a real id there. Composing the weight-sync mixin on top adds
-      # ``update_weights_from_ipc`` (bucketed CUDA-IPC) without losing
-      # the tokenizer fix.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.ar_extension.HI3ARWeightSyncExtension
-      # unirl-specific: enable LoRA on the AR stage so
-      # ``set_lora_from_tensors`` can install in-memory adapters via
-      # ``add_lora`` (the LoRA tensor-bag transport). Defaults from
-      # ``vllm.config.lora.LoRAConfig`` (max_loras=1, max_lora_rank=16)
-      # are sufficient for our single-adapter trainer-side LoRA flow.
-      enable_lora: true
-      gpu_memory_utilization: 0.9
-      enforce_eager: true
-      trust_remote_code: true
-      engine_output_type: latent
-      enable_prefix_caching: false
-      max_num_batched_tokens: 32768
-      tensor_parallel_size: 4
-      pipeline_parallel_size: 1
-      # When running inside a Ray actor (e.g. RolloutActor), vllm's auto-
-      # detection picks the ray executor backend, but the StageEngineCoreProc
-      # subprocess loses the placement-group handle (KeyError: 'bundles').
-      # Force mp executor so each stage spawns workers via multiprocessing,
-      # matching the legacy in-process smoke and the DiT stage below.
-      distributed_executor_backend: "mp"
-      hf_overrides:
-        rope_parameters:
-          mrope_section: [0, 32, 32]
-          rope_type: default
-      omni_kv_config:
-        need_send_cache: true
-        kv_transfer_criteria:
-          type: prefill_finished
-    is_comprehension: true
-    final_output: true
-    final_output_type: text
-    default_sampling_params:
-      temperature: 0.0
-      top_p: 1.0
-      top_k: -1
-      max_tokens: 2048
-      seed: 42
-      detokenize: True
-      repetition_penalty: 1.1
-
-  - stage_id: 1
-    stage_type: diffusion
-    runtime:
-      process: true
-      devices: "4,5,6,7"
-    engine_args:
-      model_stage: diffusion
-      max_num_seqs: 1
-      enforce_eager: true
-      distributed_executor_backend: "mp"
-      vae_use_slicing: false
-      vae_use_tiling: false
-      cache_backend: null
-      cache_config: null
-      enable_cache_dit_summary: false
-      omni_kv_config:
-        need_recv_cache: true
-      parallel_config:
-        pipeline_parallel_size: 1
-        data_parallel_size: 1
-        tensor_parallel_size: 4
-        enable_expert_parallel: false
-        sequence_parallel_size: 1
-        ulysses_degree: 1
-        ring_degree: 1
-        cfg_parallel_size: 1
-        vae_patch_parallel_size: 1
-        use_hsdp: false
-        hsdp_shard_size: -1
-        hsdp_replicate_size: 1
-      # unirl-specific: install RLHunyuanImage3Pipeline so SDE-mode
-      # rollouts get a swappable scheduler + per-step trajectory capture.
-      custom_pipeline_args:
-        pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.hi3.pipeline.RLHunyuanImage3Pipeline
-      # unirl-specific: install the weight-sync extension so the DiT
-      # worker exposes ``update_weights_from_ipc`` (bucketed CUDA-IPC) and
-      # later ``init_weight_transfer_engine`` / ``update_weights`` (NCCL).
-      # Must coexist with the custom pipeline above; vllm-omni's worker
-      # composes these via multiple inheritance at instantiation time.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
-    engine_input_source: [0]
-    final_output: true
-    final_output_type: image
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 1
-  edges:
-    - from: 0
-      to: 1
-      window_size: -1
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_t2t.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_t2t.yaml
deleted file mode 100644
index 9edae5c24..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_image3_t2t.yaml
+++ /dev/null
@@ -1,41 +0,0 @@
-# Vendored at its last upstream revision, since vllm-omni deleted stage_configs/ in #3172.
-# HunyuanImage-3.0 text-to-text: one AR stage reads a prompt and emits text.
-# Sampling params match the official generation_config.json.
-
-stage_args:
-  - stage_id: 0
-    stage_type: llm
-    runtime:
-      process: true
-      devices: "0,1,2,3"
-      requires_multimodal_data: false
-    engine_args:
-      model_stage: AR
-      max_num_seqs: 1
-      model_arch: HunyuanImage3ForCausalMM
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.95
-      enforce_eager: true
-      trust_remote_code: true
-      enable_prefix_caching: false
-      max_num_batched_tokens: 32768
-      tensor_parallel_size: 4
-      pipeline_parallel_size: 1
-      hf_overrides:
-        rope_parameters:
-          mrope_section: [0, 32, 32]
-          rope_type: default
-    is_comprehension: true
-    final_output: true
-    final_output_type: text
-    default_sampling_params:
-      temperature: 0.0
-      top_p: 0.95
-      top_k: 1024
-      max_tokens: 2048
-      stop_token_ids: [127957, 128024, 128026]  # <|endoftext|>, </think>, </answer>
-      detokenize: True
-
-runtime:
-  enabled: true
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_video15_t2v_rl.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_video15_t2v_rl.yaml
deleted file mode 100644
index a215511bb..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/hunyuan_video15_t2v_rl.yaml
+++ /dev/null
@@ -1,79 +0,0 @@
-# Single-stage HunyuanVideo-1.5 text-to-video config for RL rollout.
-#
-# Diff vs SD3.5 stage config:
-#   - Same single-stage diffusion architecture (no AR prelude — HV1.5 is pure
-#     generative text-to-video, not multi-modal like HunyuanImage-3).
-#   - custom_pipeline_args installs our RLHunyuanVideo15Pipeline subclass
-#     so SDE-mode rollouts get trajectory capture (latents, sigmas, log_probs)
-#     and dual-text-encoder condition extraction (text_mllm + text_glyph).
-#   - worker_extension_cls installs DiTWeightSyncExtension (same as SD3,
-#     reused for HV1.5 DiT backbone).
-#   - Dual-text-encoder condition fields (Qwen2.5-VL MLLM + ByT5 glyph)
-#     are captured by the pipeline subclass and routed via the output envelope
-#     for the response handler to unpack into HunyuanVideo15Conditions.
-#
-# Sized for HunyuanVideo-1.5 on TP=1. Single-GPU rollout matches the trainer's
-# full-forward math more closely than tensor-parallel rollout.
-
-stage_args:
-  - stage_id: 0
-    stage_type: diffusion
-    runtime:
-      process: true
-      devices: "0"
-    engine_args:
-      model_stage: diffusion
-      # HV1.5 is non-batching upstream, so the engine requires this.
-      max_num_seqs: 1
-      enforce_eager: true
-      distributed_executor_backend: "mp"
-      vae_use_slicing: false
-      vae_use_tiling: false
-      cache_backend: null
-      cache_config: null
-      enable_cache_dit_summary: false
-      # Match the trainer's PyTorch SDPA forward path.
-      # 0.27 reads ``diffusion_attention_backend`` (not the AR field ``attention_backend``).
-      diffusion_attention_backend: TORCH_SDPA
-      # Required for vLLM to wrap DiT Linear modules in LoRA-aware variants.
-      # Without ``enable_lora``, ``add_lora`` still registers the adapter in
-      # ``DiffusionLoRAManager._registered_adapters`` (so checksum probes
-      # pass), but no Linear consults it and the forward silently runs base
-      # weights. ``max_lora_rank`` must be >= the recipe's LoRAPolicy rank;
-      # the vllm.config.lora default of 16 silently rejects larger adapters
-      # at add time.
-      enable_lora: true
-      max_lora_rank: 64
-      parallel_config:
-        pipeline_parallel_size: 1
-        data_parallel_size: 1
-        tensor_parallel_size: 1
-        enable_expert_parallel: false
-        sequence_parallel_size: 1
-        ulysses_degree: 1
-        ring_degree: 1
-        cfg_parallel_size: 1
-        vae_patch_parallel_size: 1
-        use_hsdp: false
-        hsdp_shard_size: -1
-        hsdp_replicate_size: 1
-      # unirl-specific: install RLHunyuanVideo15Pipeline so SDE-mode
-      # rollouts get a swappable scheduler, per-step trajectory capture
-      # (latents, sigmas, log_probs), and dual-text-encoder condition
-      # extraction (text_mllm + text_glyph from the dual-encoder architecture).
-      custom_pipeline_args:
-        pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.hv15.pipeline.RLHunyuanVideo15Pipeline
-      # unirl-specific: install the weight-sync extension so the DiT
-      # worker exposes update_weights_from_ipc (bucketed CUDA-IPC) and
-      # later init_weight_transfer_engine / update_weights (NCCL).
-      # Must coexist with the custom pipeline above; vllm-omni's worker
-      # composes these via multiple inheritance at instantiation time.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
-    final_output: true
-    final_output_type: video
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 1
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/qwen3_omni_thinker_only_rl_1x4.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/qwen3_omni_thinker_only_rl_1x4.yaml
deleted file mode 100644
index a728f248b..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/qwen3_omni_thinker_only_rl_1x4.yaml
+++ /dev/null
@@ -1,51 +0,0 @@
-# Qwen3-Omni Thinker-only stage, TP=4.
-stage_args:
-  - stage_id: 0
-    runtime:
-      devices: "0,1,2,3"
-    engine_args:
-      model_stage: thinker
-      model_arch: Qwen3OmniMoeThinkerForConditionalGeneration
-      worker_type: ar
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      engine_output_type: text
-      hf_config_name: thinker_config
-
-      tensor_parallel_size: 4
-      distributed_executor_backend: "mp"
-
-      gpu_memory_utilization: 0.35
-      # Standalone image, audio, and video prompts are safe to batch.
-      # Audio-in-video uses the dedicated serialized stage config instead.
-      max_num_seqs: 64
-      max_num_batched_tokens: 32768
-      enable_chunked_prefill: true
-
-      # Audio fusion is enabled per request; setting use_audio_in_video globally
-      # breaks vLLM's video-only dummy profiling during startup.
-      limit_mm_per_prompt:
-        audio: 1
-        video: 1
-        image: 1
-
-      mm_processor_cache_gb: 0
-
-      dtype: bfloat16
-      load_format: safetensors
-      trust_remote_code: true
-      enable_prefix_caching: false
-      enforce_eager: false
-
-      enable_lora: true
-      max_lora_rank: 64
-      max_loras: 1
-
-      # Add the weight-sync receivers to the AR worker.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.qwen3_omni_ar_extension.Qwen3OmniARWeightSyncExtension
-
-      enable_sleep_mode: true
-      logprobs_mode: processed_logprobs
-      disable_log_stats: true
-    final_output: true
-    final_output_type: text
-    is_comprehension: true
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/qwen3_omni_thinker_only_rl_audio_video_1x4.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/qwen3_omni_thinker_only_rl_audio_video_1x4.yaml
deleted file mode 100644
index b79dacfac..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/qwen3_omni_thinker_only_rl_audio_video_1x4.yaml
+++ /dev/null
@@ -1,52 +0,0 @@
-# Qwen3-Omni Thinker-only audio-in-video stage, TP=4.
-stage_args:
-  - stage_id: 0
-    runtime:
-      devices: "0,1,2,3"
-    engine_args:
-      model_stage: thinker
-      model_arch: Qwen3OmniMoeThinkerForConditionalGeneration
-      worker_type: ar
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      engine_output_type: text
-      hf_config_name: thinker_config
-
-      tensor_parallel_size: 4
-      distributed_executor_backend: "mp"
-
-      gpu_memory_utilization: 0.35
-      # Audio-in-video decoder numerics become batch-sensitive when several
-      # long multimodal requests share one packed forward. Serialize requests
-      # to keep rollout/replay K3 near the bf16 numerical floor.
-      max_num_seqs: 1
-      max_num_batched_tokens: 32768
-      enable_chunked_prefill: true
-
-      # Audio fusion is enabled per request; setting use_audio_in_video globally
-      # breaks vLLM's video-only dummy profiling during startup.
-      limit_mm_per_prompt:
-        audio: 1
-        video: 1
-        image: 0
-
-      mm_processor_cache_gb: 0
-
-      dtype: bfloat16
-      load_format: safetensors
-      trust_remote_code: true
-      enable_prefix_caching: false
-      enforce_eager: false
-
-      enable_lora: true
-      max_lora_rank: 64
-      max_loras: 1
-
-      # Add the weight-sync receivers to the AR worker.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.qwen3_omni_ar_extension.Qwen3OmniARWeightSyncExtension
-
-      enable_sleep_mode: true
-      logprobs_mode: processed_logprobs
-      disable_log_stats: true
-    final_output: true
-    final_output_type: text
-    is_comprehension: true
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/qwen_image_t2i_rl.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/qwen_image_t2i_rl.yaml
deleted file mode 100644
index 5468c503f..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/qwen_image_t2i_rl.yaml
+++ /dev/null
@@ -1,85 +0,0 @@
-# Single-stage Qwen-Image text-to-image config for RL rollout.
-#
-# Diff vs the (non-existent) upstream Qwen-Image stage config (upstream boots
-# it through the default single-diffusion-stage factory):
-#   - custom_pipeline_args installs our RLQwenImagePipeline subclass so
-#     SDE-mode rollouts get a swappable scheduler + per-step trajectory
-#     capture (latents, sigmas, log_probs). The subclass also owns the
-#     packed-latent boundary: x_T injection packs the driver's spatial
-#     [B, C, H, W] noise; trajectory harvest unpacks back from the
-#     transformer's [B, S, C*4] patch layout.
-#   - worker_extension_cls installs DiTWeightSyncExtension so the worker
-#     exposes update_weights_from_ipc (bucketed CUDA-IPC) and the NCCL
-#     update_weights paths the trainer needs.
-#
-# Contract notes:
-#   - CFG is off by default (recipes run guidance_scale 1.0 = true_cfg 1.0);
-#     the driver-side adapter omits negative_prompt entirely when CFG is off
-#     so upstream's "" -> has_neg_prompt + true_cfg 4.0 default never fires.
-#   - height/width must be multiples of 16 (vae_scale_factor * packing);
-#     otherwise upstream normalizes them and the driver-pinned dynamic-shift
-#     sigma schedule desyncs — the response layer's sigma-echo verify then
-#     fails loudly.
-#
-# Sized for Qwen-Image (20B MMDiT + co-resident Qwen2.5-VL text encoder) on
-# TP=1. Single-GPU rollout matches the trainer's full-forward math more
-# closely than tensor-parallel rollout and avoids the systematic bf16 drift
-# seen in replay log-prob parity checks.
-
-stage_args:
-  - stage_id: 0
-    stage_type: diffusion
-    runtime:
-      process: true
-      devices: "0"
-    engine_args:
-      model_stage: diffusion
-      # The pipeline sets supports_request_batch=False, which requires this.
-      max_num_seqs: 1
-      enforce_eager: true
-      distributed_executor_backend: "mp"
-      vae_use_slicing: false
-      vae_use_tiling: false
-      cache_backend: null
-      cache_config: null
-      enable_cache_dit_summary: false
-      # Required for vLLM to wrap DiT Linear modules in LoRA-aware variants.
-      # Without ``enable_lora``, ``add_lora`` still registers the adapter in
-      # ``DiffusionLoRAManager._registered_adapters`` (so checksum probes
-      # pass), but no Linear consults it and the forward silently runs base
-      # weights. ``max_lora_rank`` must be >= the recipe's LoRAPolicy rank
-      # (64 in qwen_image_dancegrpo); the vllm.config.lora default of 16
-      # silently rejects larger adapters at add time.
-      enable_lora: true
-      max_lora_rank: 64
-      parallel_config:
-        pipeline_parallel_size: 1
-        data_parallel_size: 1
-        tensor_parallel_size: 1
-        enable_expert_parallel: false
-        sequence_parallel_size: 1
-        ulysses_degree: 1
-        ring_degree: 1
-        cfg_parallel_size: 1
-        vae_patch_parallel_size: 1
-        use_hsdp: false
-        hsdp_shard_size: -1
-        hsdp_replicate_size: 1
-      # unirl-specific: install RLQwenImagePipeline so SDE-mode rollouts get
-      # a swappable scheduler + per-step trajectory capture.
-      custom_pipeline_args:
-        pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.qwen_image.pipeline.RLQwenImagePipeline
-      # unirl-specific: install the weight-sync extension so the DiT
-      # worker exposes update_weights_from_ipc (bucketed CUDA-IPC) and
-      # later init_weight_transfer_engine / update_weights (NCCL).
-      # Must coexist with the custom pipeline above; vllm-omni's worker
-      # composes these via multiple inheritance at instantiation time.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
-    final_output: true
-    final_output_type: image
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 1
diff --git a/unirl/rollout/engine/vllm_omni/stage_configs/sd35_t2i_rl.yaml b/unirl/rollout/engine/vllm_omni/stage_configs/sd35_t2i_rl.yaml
deleted file mode 100644
index 2c932a1b3..000000000
--- a/unirl/rollout/engine/vllm_omni/stage_configs/sd35_t2i_rl.yaml
+++ /dev/null
@@ -1,73 +0,0 @@
-# Single-stage SD3.5-medium text-to-image config for RL rollout.
-#
-# Diff vs the (non-existent) upstream SD3 stage config:
-#   - Single diffusion stage (no AR prelude — SD3.5 has no chain-of-thought
-#     stage like HunyuanImage-3).
-#   - custom_pipeline_args installs our RLStableDiffusion3Pipeline subclass
-#     so SDE-mode rollouts get a swappable scheduler + per-step trajectory
-#     capture (latents, sigmas, log_probs).
-#   - worker_extension_cls installs DiTWeightSyncExtension so the worker
-#     exposes update_weights_from_ipc (bucketed CUDA-IPC) and the NCCL
-#     update_weights paths the trainer needs.
-#
-# Sized for SD3.5-medium on TP=1. Single-GPU rollout matches the trainer's
-# full-forward math more closely than tensor-parallel rollout and avoids the
-# systematic bf16 drift seen in replay log-prob parity checks.
-
-stage_args:
-  - stage_id: 0
-    stage_type: diffusion
-    runtime:
-      process: true
-      devices: "0"
-    engine_args:
-      model_stage: diffusion
-      # The pipeline sets supports_request_batch=False, which requires this.
-      max_num_seqs: 1
-      enforce_eager: true
-      distributed_executor_backend: "mp"
-      vae_use_slicing: false
-      vae_use_tiling: false
-      cache_backend: null
-      cache_config: null
-      enable_cache_dit_summary: false
-      # Required for vLLM to wrap DiT Linear modules in LoRA-aware variants.
-      # Without ``enable_lora``, ``add_lora`` still registers the adapter in
-      # ``DiffusionLoRAManager._registered_adapters`` (so checksum probes
-      # pass), but no Linear consults it and the forward silently runs base
-      # weights. ``max_lora_rank`` must be >= the recipe's LoRAPolicy rank
-      # (32 in flowgrpo_fast_sd3); the vllm.config.lora default of 16
-      # silently rejects larger adapters at add time.
-      enable_lora: true
-      max_lora_rank: 32
-      parallel_config:
-        pipeline_parallel_size: 1
-        data_parallel_size: 1
-        tensor_parallel_size: 1
-        enable_expert_parallel: false
-        sequence_parallel_size: 1
-        ulysses_degree: 1
-        ring_degree: 1
-        cfg_parallel_size: 1
-        vae_patch_parallel_size: 1
-        use_hsdp: false
-        hsdp_shard_size: -1
-        hsdp_replicate_size: 1
-      # unirl-specific: install RLStableDiffusion3Pipeline so SDE-mode
-      # rollouts get a swappable scheduler + per-step trajectory capture.
-      custom_pipeline_args:
-        pipeline_class: unirl.rollout.engine.vllm_omni.pipelines.sd3.pipeline.RLStableDiffusion3Pipeline
-      # unirl-specific: install the weight-sync extension so the DiT
-      # worker exposes update_weights_from_ipc (bucketed CUDA-IPC) and
-      # later init_weight_transfer_engine / update_weights (NCCL).
-      # Must coexist with the custom pipeline above; vllm-omni's worker
-      # composes these via multiple inheritance at instantiation time.
-      worker_extension_cls: unirl.rollout.engine.vllm_omni.worker.dit_extension.DiTWeightSyncExtension
-    final_output: true
-    final_output_type: image
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 1
diff --git a/unirl/rollout/engine/vllm_omni/utils/tracks.py b/unirl/rollout/engine/vllm_omni/utils/tracks.py
index 7ef35b4c9..19e6de3bb 100644
--- a/unirl/rollout/engine/vllm_omni/utils/tracks.py
+++ b/unirl/rollout/engine/vllm_omni/utils/tracks.py
@@ -189,11 +189,9 @@ def decoded_text_from_ar(per_request: Sequence[Sequence[Any]]) -> Texts:
         ar = pick_stage_output(outputs, final_output_type="text", stage_id=0)
         text_str = ""
         if ar is not None:
-            ro = getattr(ar, "request_output", None)
-            if ro is not None:
-                completions = getattr(ro, "outputs", None) or []
-                if completions:
-                    text_str = getattr(completions[0], "text", "") or ""
+            completions = getattr(ar, "outputs", None) or []
+            if completions:
+                text_str = getattr(completions[0], "text", "") or ""
         texts.append(Text(text=text_str))
     return Texts.from_list(texts)
 
@@ -229,10 +227,7 @@ def _flatten_logprobs(logprobs: Any, fallback_len: int) -> Optional[torch.Tensor
 
 def _extract_completion(out: Any) -> Tuple[List[int], Optional[torch.Tensor]]:
     """Pull ``(token_ids, per_token_logp)`` out of a Stage-0 result."""
-    request_output = getattr(out, "request_output", None)
-    if request_output is None:
-        return [], None
-    completions = getattr(request_output, "outputs", None) or []
+    completions = getattr(out, "outputs", None) or []
     if not completions:
         return [], None
     completion = completions[0]
diff --git a/unirl/rollout/engine/vllm_omni/worker/ipc_receive_mixin.py b/unirl/rollout/engine/vllm_omni/worker/ipc_receive_mixin.py
index 4d10273b7..08ff0904f 100644
--- a/unirl/rollout/engine/vllm_omni/worker/ipc_receive_mixin.py
+++ b/unirl/rollout/engine/vllm_omni/worker/ipc_receive_mixin.py
@@ -226,20 +226,43 @@ def set_lora_from_tensor_dict_copy(
         )
         return self.add_lora(request)
 
+    def _diffrl_parameter_source(self):
+        """Return the loaded module behind AR or diffusion runner wrappers."""
+        queue = [self]
+        seen: set[int] = set()
+        while queue:
+            obj = queue.pop(0)
+            if obj is None or id(obj) in seen:
+                continue
+            seen.add(id(obj))
+            named_parameters = getattr(obj, "named_parameters", None)
+            if callable(named_parameters):
+                try:
+                    next(iter(named_parameters()))
+                except StopIteration:
+                    pass
+                else:
+                    return obj
+            queue.extend(
+                getattr(obj, attr, None)
+                for attr in (
+                    "worker",
+                    "model_runner",
+                    "runner",
+                    "pipeline",
+                    "transformer",
+                    "model",
+                    "bagel",
+                )
+            )
+        return None
+
     def _diffrl_describe_params(
         self,
         names: Optional[list] = None,
     ) -> dict:
         """Return ``{name: (shape_tuple, dtype_str)}`` for the worker's loaded model."""
-        runner = getattr(self, "model_runner", None)
-        if runner is None:
-            return {}
-        param_source = None
-        for attr in ("pipeline", "model"):
-            obj = getattr(runner, attr, None)
-            if obj is not None and hasattr(obj, "named_parameters"):
-                param_source = obj
-                break
+        param_source = self._diffrl_parameter_source()
         if param_source is None:
             return {}
 
@@ -258,15 +281,7 @@ def _diffrl_param_checksums(
         """Return ``{name: short_sha256_hex}`` for the worker's loaded model."""
         import hashlib
 
-        runner = getattr(self, "model_runner", None)
-        if runner is None:
-            return {}
-        param_source = None
-        for attr in ("pipeline", "model"):
-            obj = getattr(runner, attr, None)
-            if obj is not None and hasattr(obj, "named_parameters"):
-                param_source = obj
-                break
+        param_source = self._diffrl_parameter_source()
         if param_source is None:
             return {}
 
@@ -298,15 +313,7 @@ def _diffrl_loaded_param_checksums(
             fingerprint_tensor,
         )
 
-        runner = getattr(self, "model_runner", None)
-        if runner is None:
-            return {}
-        param_source = None
-        for attr in ("pipeline", "model"):
-            obj = getattr(runner, attr, None)
-            if obj is not None and hasattr(obj, "named_parameters"):
-                param_source = obj
-                break
+        param_source = self._diffrl_parameter_source()
         if param_source is None:
             return {}
 
diff --git a/unirl/utils/peft_merge.py b/unirl/utils/peft_merge.py
index 6aef50419..c7e273ec4 100644
--- a/unirl/utils/peft_merge.py
+++ b/unirl/utils/peft_merge.py
@@ -346,7 +346,12 @@ def extract_lora_tensors(
 
 
 def adapt_lora_for_vllm(tensors: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
-    """Wrap canonical-format LoRA keys in the vllm-omni PEFT envelope."""
+    """Idempotently wrap canonical LoRA keys in the vllm-omni PEFT envelope."""
+    enveloped = [name.startswith(_PEFT_PREFIX) for name in tensors]
+    if all(enveloped):
+        return tensors
+    if any(enveloped):
+        raise ValueError("LoRA tensors mix canonical and PEFT-enveloped keys")
     return {f"{_PEFT_PREFIX}{k}": v for k, v in tensors.items()}