[Docs] Clarify TP1 DLO mmap support scope

Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>
Merge branch 'main' into codex-dlo-host-weight-plan
2026-08-16 10:54:24 +08:00 · 2026-08-16 00:13:33 +08:00 · 2026-08-15 23:57:03 +08:00 · 2026-08-15 23:40:24 +08:00 · 2026-08-15 23:32:27 +08:00 · 2026-08-15 23:32:27 +08:00
308 changed files with 3230 additions and 33224 deletions
@@ -159,7 +159,19 @@ steps:
          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
          - export VLLM_USE_DEEP_GEMM="0"
          - export VLLM_MOE_USE_DEEP_GEMM="0"
-          - pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy
+          - pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_voxcpm2_tts_expansion.py
+        mirror_hardwares: l4_1
+
+      - label: ":full_moon: TTS · Accuracy Test · VoxCPM2"
+        timeout_in_minutes: 30
+        commands:
+          - |
+            timeout 30m bash -c "
+              export VLLM_LOGGING_LEVEL=DEBUG
+              export VLLM_WORKER_MULTIPROC_METHOD=spawn
+              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+              pytest -s -v tests/e2e/online_serving/test_voxcpm2_tts_expansion.py -m 'full_model and cuda and tts' --run-level 'full_model'
+            "
        mirror_hardwares: l4_1

      - label: ":full_moon: TTS · Perf Test"
@@ -176,6 +188,35 @@ steps:
            exit $$EXIT
        mirror_hardwares: h100_2

+      - label: ":full_moon: TTS · VoxCPM2 · Perf Test"
+        key: nightly-tts-performance-voxcpm2
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_voxcpm2.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        mirror_hardwares: h100_1
+
+      - label: ":full_moon: TTS · Higgs-Audio-V3 · Perf Test"
+        key: nightly-tts-performance-higgs-audio-v3
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - export VLLM_USE_DEEP_GEMM="0"
+          - export VLLM_MOE_USE_DEEP_GEMM="0"
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_higgs_audio_v3.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        mirror_hardwares: h100_1
+
  - group: ":card_index_dividers: Tiny Model Tests [Multi-GPU]"
    key: nightly-tiny-model-test-group
    depends_on: upload-nightly-pipeline
@@ -205,7 +246,8 @@ steps:
      build.pull_request.labels includes "nightly-test" ||
      build.pull_request.labels includes "diffusion-x2iat-test"
    steps:
-      # Use explicit file shards so single-card cases do not reserve 2+ GPUs.
+      # Use explicit file shards so single-card cases do not reserve 2+ GPUs,
+      # and Bagel feature coverage is not duplicated via a broad tests/e2e sweep.

      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Single-GPU"
        timeout_in_minutes: 120
@@ -213,6 +255,15 @@ steps:
          - >-
            pytest -sv
            tests/e2e/online_serving/test_qwen_image_expansion.py
+            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
+            tests/e2e/online_serving/test_qwen_image_layered_expansion.py
+            tests/e2e/online_serving/test_boogu_image.py
+            tests/e2e/online_serving/test_boogu_image_edit.py
+            tests/e2e/online_serving/test_boogu_image_expansion.py
+            tests/e2e/online_serving/test_boogu_image_edit_expansion.py
+            tests/e2e/online_serving/test_bagel_expansion.py
+            tests/e2e/online_serving/test_krea2_expansion.py
+            tests/e2e/offline_inference/test_krea2_expansion.py
            -m "full_model and diffusion and H100 and not distributed_cuda"
            --run-level "full_model"
        mirror_hardwares: h100_1
@@ -223,14 +274,47 @@ steps:
          - >-
            pytest -sv
            tests/e2e/online_serving/test_qwen_image_expansion.py
+            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
+            tests/e2e/offline_inference/test_krea2_expansion.py
            -m "full_model and diffusion and H100 and distributed_cuda"
            --run-level "full_model"
        mirror_hardwares: h100_2

+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Layered/LongCat/FLUX"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_qwen_image_layered_expansion.py
+            -m "full_model and diffusion and H100 and distributed_cuda"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Bagel"
+        timeout_in_minutes: 120
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_bagel_expansion.py
+            -m "full_model and diffusion and H100 and distributed_cuda"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: ":full_moon: Diffusion X2I(&A&T) · GLM-Image Function Test with H100"
+        timeout_in_minutes: 120
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_glm_image_expansion.py
+            tests/e2e/offline_inference/test_glm_image_autoround_w4a16_expansion.py
+            -m "full_model and diffusion and H100"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4"
        timeout_in_minutes: 120
        commands:
-          - pytest -sv tests/e2e/ -k "not test_wan and not hunyuan" -m "full_model and diffusion and L4" --run-level "full_model" --ignore=tests/e2e/accuracy
+          - pytest -sv tests/e2e/ -k "not test_wan and not test_bagel_expansion and not hunyuan" -m "full_model and diffusion and L4" --run-level "full_model" --ignore=tests/e2e/accuracy
        mirror_hardwares: l4_4

      - label: ":full_moon: Diffusion X2I(&A&T) · Doc Test"
@@ -243,7 +327,7 @@ steps:
        timeout_in_minutes: 180
        commands:
          - export VLLM_HTTP_TIMEOUT_KEEP_ALIVE=120
-          - pytest -s -v tests/e2e/accuracy/test_qwen_image.py --run-level full_model
+          - pytest -s -v tests/e2e/accuracy/test_qwen_image*.py --run-level full_model
          - pytest -s -v tests/e2e/accuracy/test_diffusers_backend_similarity.py -k '2i_matches_diffusers' -m full_model --run-level full_model
        mirror_hardwares: h100_1

@@ -271,6 +355,70 @@ steps:
            exit $$EXIT
        mirror_hardwares: h100_4

+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Edit"
+        key: nightly-diffusion-x2iat-performance-qwen-image-edit
+        timeout_in_minutes: 180
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+          - export CACHE_DIT_VERSION=1.5.0
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_vllm_omni.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+        mirror_hardwares: h100_4
+
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Edit-2511"
+        key: nightly-diffusion-x2iat-performance-qwen-image-edit-2511
+        timeout_in_minutes: 180
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+          - export CACHE_DIT_VERSION=1.5.0
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_2511_vllm_omni.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+        mirror_hardwares: h100_4
+
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Layered"
+        key: nightly-diffusion-x2iat-performance-qwen-image-layered
+        timeout_in_minutes: 120
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+          - export CACHE_DIT_VERSION=1.5.0
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_layered_vllm_omni.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+        mirror_hardwares: h100_1
+
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · BAGEL"
+        key: nightly-diffusion-x2iat-performance-bagel
+        timeout_in_minutes: 180
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+          - export CACHE_DIT_VERSION=1.5.0
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_bagel_vllm_omni.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+        mirror_hardwares: h100_1
+
  - group: ":card_index_dividers: Diffusion X2V Model Test"
    key: nightly-diffusion-x2v-group
    depends_on: upload-nightly-pipeline
@@ -292,6 +440,12 @@ steps:
          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda" --run-level "full_model"
        mirror_hardwares: h100_2

+      - label: ":full_moon: Diffusion X2V · LingBot Video MoE Function Test"
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_lingbot_video_moe.py -k "video_generation" -m "full_model and cuda" --run-level "full_model"
+        mirror_hardwares: h100_1
+
      - label: ":full_moon: Diffusion X2V · Doc Test"
        timeout_in_minutes: 60
        commands:
@@ -302,8 +456,15 @@ steps:
        timeout_in_minutes: 90
        commands:
          - pytest -s -v tests/e2e/offline_inference/test_wan22_autoround_w4a16_expansion.py -m "full_model and cuda" --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_wan_2_1_vace_expansion.py tests/e2e/online_serving/test_ltx.py -m "full_model and cuda" --run-level "full_model"
        mirror_hardwares: h100_2

+      - label: ":full_moon: Diffusion X2V · LTX-2.5 Online Test"
+        timeout_in_minutes: 60
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_ltx25.py -m "full_model and cuda" --run-level "full_model"
+        mirror_hardwares: h100_1
+
      - label: ":full_moon: Diffusion X2V · Wan2.2 Accuracy Test"
        timeout_in_minutes: 180
        commands:
@@ -322,8 +483,23 @@ steps:
        commands:
          - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_t2v/test_hunyuanvideo15_t2v_video_similarity.py -m full_model --run-level full_model
          - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_i2v/test_hunyuanvideo15_i2v_video_similarity.py -m full_model --run-level full_model
+          - pytest -s -v tests/e2e/accuracy/test_video_streaming_output_similarity.py -m full_model --run-level full_model
        mirror_hardwares: h100_2

+      - label: ":full_moon: Diffusion X2V · LTX Accuracy Test"
+        timeout_in_minutes: 180
+        commands:
+          - pytest -s -v 'tests/e2e/accuracy/ltx/test_ltx_official_similarity.py::test_ltx_matches_official[ltx2_3_i2v]' -m full_model --run-level full_model
+          - pytest -s -v 'tests/e2e/accuracy/ltx/test_ltx_official_similarity.py::test_ltx_matches_official[ltx23_two_stage_layer_fused_i2v]' -m full_model --run-level full_model
+        mirror_hardwares: h100_1
+
+      - label: ":full_moon: Diffusion X2V · LTX-2.5 Official Parity Test"
+        timeout_in_minutes: 180
+        commands:
+          - pytest -s -v tests/e2e/accuracy/ltx/test_ltx25_official_similarity.py::test_ltx25_distilled_two_stage_matches_official -m full_model --run-level full_model
+          - pytest -s -v tests/e2e/accuracy/ltx/test_ltx25_official_similarity.py::test_ltx25_full_matches_official -m full_model --run-level full_model
+        mirror_hardwares: h100_1
+
      - label: ":full_moon: Diffusion X2V · Perf Test"
        key: nightly-diffusion-x2v-performance
        timeout_in_minutes: 180
@@ -399,6 +575,16 @@ steps:
          - pytest -sv tests/diffusion/quantization -m 'full_model and cuda and L4' --run-level "full_model"
        mirror_hardwares: l4_1

+  - label: ":bar_chart: Testcase Statistics"
+    key: nightly-testcase-statistics
+    timeout_in_minutes: 120
+    depends_on: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1"
+    commands:
+      - python tools/nightly/buildkite_testcase_statistics.py -o tests/dfx/perf/results/buildkite_testcase_statistics.html
+      - buildkite-agent artifact upload "tests/dfx/perf/results/*.html"
+    mirror_hardwares: h100_1
+
  # NOTE: ":email: Nightly Collection & Email" is deprecated and will be removed soon.
  # Do not add new depends_on entries or extend this step.
  - label: ":email: Nightly Collection & Email"
@@ -409,20 +595,32 @@ steps:
      - nightly-omni-performance-vllm-text
      - nightly-omni-performance-multi-replicas
      - nightly-tts-performance
+      - nightly-tts-performance-higgs-audio-v3
      - nightly-diffusion-x2iat-performance-qwen-image
+      - nightly-diffusion-x2iat-performance-qwen-image-edit
+      - nightly-diffusion-x2iat-performance-qwen-image-edit-2511
+      - nightly-diffusion-x2iat-performance-qwen-image-layered
+      - nightly-diffusion-x2iat-performance-bagel
      - nightly-diffusion-x2v-performance
+      - nightly-testcase-statistics
    if: build.env("NIGHTLY") == "1" && build.env("EMAIL_DISTRIBUTION") == "1"
    commands:
      - pip install openpyxl
      - export DEFAULT_INPUT_DIR=tests/dfx/perf/results
      - export DEFAULT_OUTPUT_DIR=tests/dfx/perf/results
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance-higgs-audio-v3
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-no-async-chunk
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-async-chunk
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-vllm-text
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-multi-replicas
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-edit
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-edit-2511
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-layered
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-bagel
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.html" . --step nightly-testcase-statistics
      - python tools/nightly/generate_nightly_perf_excel.py
      - python tools/nightly/generate_nightly_perf_html.py
      - python tools/nightly/send_nightly_email.py --report-file "tests/dfx/perf/results/*.xlsx, tests/dfx/perf/results/*.html"
@@ -30,6 +30,14 @@ steps:
      - pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m "slow"
    mirror_hardwares: h100_4

+  - label: "Reliability Test - voxcpm2"
+    timeout_in_minutes: 180
+    depends_on: upload-weekly-pipeline
+    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
+    commands:
+      - pytest -s -v tests/dfx/reliability/test_reliability_voxcpm2.py -m "slow"
+    mirror_hardwares: l4_1
+
  - group: "Reliability Test - Invalid parameters Test"
    depends_on: upload-weekly-pipeline
    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
@@ -141,4 +149,4 @@ steps:
            pytest -sv
            tests/e2e/ -m "slow and diffusion and H100 and distributed_cuda"
            --run-level "full_model"
-        mirror_hardwares: h100_4
+        mirror_hardwares: h100_2
@@ -69,12 +69,6 @@ steps:
          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
          HF_TOKEN: "${HF_TOKEN}"
        commands:
-          # Pin the Seed-TTS scorer to the second card: _get_eval_device() otherwise
-          # auto-picks npu:0, which the omni server under test already fills with its
-          # KV cache, so Whisper OOMs on `.to(device)` and every clip reports
-          # "ASR / WER failed".
-          - export SEED_TTS_WER_EVAL=1
-          - export SEED_TTS_EVAL_DEVICE=npu:1
          - |
            set +e
            pytest -s -v tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py -m "full_model and npu and A3" --run-level full_model
@@ -102,6 +96,33 @@ steps:
            EXIT=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
            exit $$EXIT
+  - group: ":card_index_dividers: Diffusion X2I(&A&T) Model Test"
+    key: nightly-diffusion-x2iat-group
+    depends_on: upload-nightly-pipeline
+    if: >-
+      build.env("NIGHTLY") == "1" ||
+      build.pull_request.labels includes "nightly-test" ||
+      build.pull_request.labels includes "diffusion-x2iat-test"
+    steps:
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Edit-2511"
+        key: nightly-diffusion-x2iat-performance-qwen-image-edit-2511
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_8
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+          DIFFUSION_ATTENTION_BACKEND: TORCH_SDPA
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - export CACHE_DIT_VERSION=1.5.0
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py -m npu --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_2511_vllm_omni.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
  - group: ":card_index_dividers: Diffusion I2I Model Test"
    key: nightly-HunyuanImage3-x2i-test-group
    depends_on: upload-nightly-pipeline
@@ -1,15 +1,4 @@
 steps:
-  - label: "Platforms Test"
-    key: ready-platforms-test
-    depends_on: upload-ready-pipeline
-    mirror_hardwares: a3_npu_2
-    commands:
-      - >-
-        pytest -s -v
-        tests/platforms/npu/test_minicpmo_code2wav_npugraph.py
-        -m 'core_model and npu and A3 and omni'
-        --run-level 'core_model'
-
  - group: ":card_index_dividers: Omni Model Test"
    key: ready-omni-group
    depends_on: upload-ready-pipeline
@@ -66,7 +66,7 @@ repos:
      entry: python tools/pre_commit/check_tts_adapter.py
      language: python
      pass_filenames: false
-      files: ^(vllm_omni/entrypoints/openai/serving_speech\.py|tools/pre_commit/check_tts_adapter\.py)$
+      files: ^(vllm_omni/entrypoints/openai/(serving_speech\.py|tts_adapters/__init__\.py)|tools/pre_commit/check_tts_adapter\.py)$
      # The budget-is-now-slack reminder is printed on a PASSING run, and
      # pre-commit hides output from passing hooks unless the hook is verbose.
      verbose: true
@@ -43,7 +43,6 @@ nav:
    - Overview: features/README.md
    - Runtime and Stage Execution:
      - Execution Modes and Streaming: user_guide/diffusion/execution_modes.md
-      - Prefill-Decode Disaggregation (experimental): features/pd_disaggregation.md
      - Sleep Mode: features/sleep_mode.md
    - Quantization:
      - Overview: user_guide/quantization/overview.md
@@ -117,7 +116,6 @@ nav:
        - Async Diffusion Output: design/feature/async_diffusion_output.md
        - Async Omni Output Materialization: design/feature/omni_async_output_materialization.md
        - Automatic Prefix Caching: design/feature/prefix_caching.md
-        - Realtime AR-Diffusion: design/feature/realtime_ar_diffusion.md
      - Communication:
        - OmniConnector Implementations:
          - Mooncake Store Connector: design/feature/omni_connectors/mooncake_store_connector.md
@@ -14,6 +14,10 @@ For introduction, please check [Introduction for stage config](./stage_configs.m

 - **[GPU Memory Calculation and Configuration](./gpu_memory_utilization.md)** - Guide on how to calculate memory requirements and set up `gpu_memory_utilization` for optimal performance

+## Multi-Stage Recipes
+
+- **[Prefill-Decode Disaggregation](./pd_disaggregation.md)** - How to derive a PD-aware Qwen3-Omni stage config from the default config without introducing another bundled YAML
+
 ## Optimization Features

 - **[Diffusion Features Overview](../user_guide/diffusion_features.md)** - Complete overview of all diffusion model features and supported models
@@ -0,0 +1,171 @@
+# Prefill-Decode (PD) Disaggregation
+
+PD disaggregation splits the Qwen3-Omni thinker into separate prefill and decode
+stages so prompt processing and token generation can run on different workers.
+
+This is documented as a stage-config recipe instead of a bundled YAML because the
+deployment-specific values usually change per environment:
+
+- GPU placement
+- `tensor_parallel_size`
+- connector backend and connector ports
+- connector IPs or bootstrap addresses
+
+Start from the [default Qwen3-Omni stage config](gh-file:vllm_omni/deploy/qwen3_omni_moe.yaml)
+and copy it to your own file, for example `qwen3_omni_pd.yaml`. Then apply the
+changes below.
+
+## Requirements
+
+- 3+ GPUs for a basic layout: prefill, decode, and talker+code2wav
+- A KV connector supported by vLLM, such as `MooncakeConnector`
+- Matching `tensor_parallel_size` on the prefill and decode thinker stages
+
+## 1. Split the thinker into prefill and decode stages
+
+Replace the original thinker stage with two stages:
+
+```yaml
+stage_args:
+  - stage_id: 0
+    stage_type: llm
+    is_prefill_only: true
+    runtime:
+      devices: "0"
+    engine_args:
+      max_num_seqs: 16
+      model_stage: thinker
+      model_arch: Qwen3OmniMoeForConditionalGeneration
+      worker_type: ar
+      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
+      gpu_memory_utilization: 0.9
+      enforce_eager: true
+      trust_remote_code: true
+      engine_output_type: latent
+      distributed_executor_backend: "mp"
+      enable_prefix_caching: false
+      max_num_batched_tokens: 32768
+      hf_config_name: thinker_config
+      tensor_parallel_size: 1
+      kv_transfer_config:
+        kv_connector: "MooncakeConnector"
+        kv_role: "kv_producer"
+        kv_rank: 0
+        kv_parallel_size: 2
+        kv_connector_extra_config:
+          mooncake_bootstrap_port: 25201
+    final_output: false
+    is_comprehension: true
+    default_sampling_params:
+      temperature: 0.4
+      top_p: 0.9
+      top_k: 1
+      max_tokens: 2048
+      seed: 42
+      detokenize: True
+      repetition_penalty: 1.05
+
+  - stage_id: 1
+    stage_type: llm
+    is_decode_only: true
+    runtime:
+      devices: "1"
+    engine_args:
+      max_num_seqs: 64
+      model_stage: thinker
+      model_arch: Qwen3OmniMoeForConditionalGeneration
+      worker_type: ar
+      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
+      gpu_memory_utilization: 0.9
+      enforce_eager: true
+      trust_remote_code: true
+      engine_output_type: latent
+      distributed_executor_backend: "mp"
+      enable_prefix_caching: false
+      max_num_batched_tokens: 32768
+      hf_config_name: thinker_config
+      tensor_parallel_size: 1
+      kv_transfer_config:
+        kv_connector: "MooncakeConnector"
+        kv_role: "kv_consumer"
+        kv_rank: 1
+        kv_parallel_size: 2
+        kv_connector_extra_config:
+          mooncake_bootstrap_port: 25202
+    engine_input_source: [0]
+    final_output: true
+    final_output_type: text
+    is_comprehension: true
+    default_sampling_params:
+      temperature: 0.4
+      top_p: 0.9
+      top_k: 1
+      max_tokens: 2048
+      seed: 42
+      detokenize: True
+      repetition_penalty: 1.05
+```
+
+Notes:
+
+- `is_prefill_only: true` marks the thinker stage that only saves KV.
+- `is_decode_only: true` marks the thinker stage that resumes from remote KV.
+- `kv_transfer_config` is required on both stages.
+- The orchestrator forces the prefill stage to run with `max_tokens=1`, so the
+  prefill side only processes the prompt and exports KV.
+
+## 2. Shift the downstream stages by one index
+
+After inserting the extra thinker stage, renumber the remaining stages:
+
+```yaml
+  - stage_id: 2
+    runtime:
+      devices: "2"
+    engine_input_source: [1]
+    sync_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker_token_only
+
+  - stage_id: 3
+    runtime:
+      devices: "2"
+    engine_args:
+      max_num_seqs: 1
+    engine_input_source: [2]
+    sync_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav_token_only
+```
+
+Compared with the default Qwen3-Omni config:
+
+- the talker becomes stage `2` instead of stage `1`
+- the code2wav stage becomes stage `3` instead of stage `2`
+- the talker now reads from decode stage `1`
+
+## 3. Add runtime edges for the four-stage pipeline
+
+```yaml
+runtime:
+  enabled: true
+  edges:
+    - from: 0
+      to: 1
+    - from: 1
+      to: 2
+    - from: 2
+      to: 3
+```
+
+## 4. Launch with your custom config
+
+```bash
+vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091 \
+    --deploy-config /path/to/qwen3_omni_pd.yaml
+```
+
+## Operational Notes
+
+- `MooncakeConnector` does not support heterogeneous TP sizes across the PD
+  pair. Keep prefill and decode at the same `tensor_parallel_size`.
+- If the thinker requires TP=2, both thinker stages must use TP=2 and be given
+  separate GPU sets, for example `"0,1"` for prefill and `"2,3"` for decode.
+- Choose connector ports and addresses that match your deployment. The values
+  shown above are examples only.
@@ -38,12 +38,11 @@ Each entry under `stages:` accepts any `StageDeployConfig` field directly (no ne
 |-------|------|----------|---------|-------------|
 | `stage_id` | int | required | — | Stage identity; matched against `PipelineConfig.stages[*].stage_id`. |
 | `max_num_seqs` | int | optional | `64` | Max concurrent sequences per stage. |
-| `gpu_memory_utilization` | float | optional | `0.9` | Per-stage total memory target; used for automatic KV-cache sizing. |
-| `kv_cache_memory_bytes` | int \| null | optional | `null` | Explicit per-rank KV-cache budget in bytes; overrides automatic sizing. |
+| `gpu_memory_utilization` | float | optional | `0.9` | Per-stage memory budget. |
 | `tensor_parallel_size` | int | optional | `1` | TP degree for this stage. |
 | `enforce_eager` | bool | optional | `false` | Disable CUDA graphs. |
-| `max_num_batched_tokens` | int | optional | `32768` | Per-stage prefill/token budget; also contributes to the native maximum in-flight token limit. |
-| `max_model_len` | int \| null | optional | `null` | Per-sequence context or KV length; `-1` enables native cache-capacity auto-fitting, while values above the HF default auto-set `VLLM_ALLOW_LONG_MAX_MODEL_LEN=1`. |
+| `max_num_batched_tokens` | int | optional | `32768` | Prefill budget. |
+| `max_model_len` | int \| null | optional | `null` | Per-stage context length (auto-sets `VLLM_ALLOW_LONG_MAX_MODEL_LEN=1` when larger than HF default). |
 | `async_scheduling` | bool \| null | optional | `null` | Per-stage async scheduling toggle. |
 | `devices` | str | optional | `"0"` | `CUDA_VISIBLE_DEVICES`-style device list. |
 | `output_connectors` | dict \| null | optional | `null` | Keyed by `to_stage_<n>`; values are names registered under top-level `connectors:`. |
@@ -184,6 +184,9 @@ Failed jobs: 1/2
    # Reliability: HunyuanImage DiT (H100 × 4)
    pytest -s -v dfx/reliability/test_reliability_hunyuan_image.py -m slow

+    # Reliability: VoxCPM2 (L4)
+    pytest -s -v dfx/reliability/test_reliability_voxcpm2.py -m slow
+
    ```

    The latest L5 CI jobs (reliability + invalid-parameter weekly steps) are in [test-weekly.yml](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/cuda/test-weekly.yml).
@@ -562,6 +562,7 @@ L5 level testing focuses on the performance of model services under ***long-runn
    -   `test_reliability_qwen3_omni.py` — Qwen3-Omni chat / multimodal reliability (GPU OOM, process kill, recovery, error contract under `--async-chunk` vs default).
    -   `test_reliability_wan22.py` — Wan2.2 I2V video API reliability (`/v1/videos` under OOM and process kill, recovery).
    -   `test_reliability_hunyuan_image.py` — HunyuanImage-3.0-Instruct DiT-only reliability (`/v1/images/generations` under OOM and process kill; deploy `hunyuan_image3_dit.yaml`, H100 × 4).
+    -   `test_reliability_voxcpm2.py` — VoxCPM2 TTS reliability (`/v1/audio/speech` under OOM and process kill; deploy `voxcpm2.yaml`, L4 × 1).
    -   `helpers.py` — Shared primitives used by current suites: raw HTTP probes for `/v1/chat/completions` and `/health`, OpenAI-style error parsing, GPU OOM sidecar (`inject_gpu_oom` / `stop_gpu_oom_hogs`), and `pgrep`-based process-kill injector construction (`make_process_kill_fault_injector`).
    -   `conftest.py` — `fault_injector` and `omni_server_after_fault` / `omni_server_after_fault_function` fixtures to run a callable **after** the server is ready.
    -   `README.md` — Short local run commands for this directory.
@@ -661,12 +662,13 @@ Reliability tests are **short fault-injection** integration runs (L5 **(b)** in
 | `test_reliability_qwen3_omni.py` | Qwen3-Omni: OOM vs **text vs text+audio** error contract, large multimodal chat under OOM, concurrent pressure, **SIGKILL** on `VLLM::Worker`, `/health` → 503 + fast-fail + concurrent chat; optional OOM recovery scenario (may be skipped while tracked in issues). |
 | `test_reliability_wan22.py` | Wan2.2 I2V: large `/v1/videos` under OOM, **SIGKILL** on `multiprocessing.spawn` chain, health / fast-fail / concurrent video requests; optional recovery test (may be skipped). |
 | `test_reliability_hunyuan_image.py` | HunyuanImage DiT-only: large `/v1/images/generations` under OOM, **SIGKILL** on `vLLM-Omni::` workers and serve/tree targets, health / fast-fail / concurrent image requests; some OOM/recovery cases may be skipped while tracked in issues. |
+| `test_reliability_voxcpm2.py` | VoxCPM2: `/v1/audio/speech` under OOM (error contract), **SIGKILL** on `VLLM::` workers and serve/tree targets, health / fast-fail / concurrent speech requests; some OOM cases may be skipped while tracked in issues. |
 | `README.md` | Minimal run / collect examples. |

 ##### Parametrization and markers

 - Each test module defines a **`RELIABILITY_SCENARIOS`** list (`test_name`, `server_params`: model, `stage_config_name` or diffusion `server_args`, etc.). **`create_reliability_omni_server_params()`** in `tests/dfx/conftest.py` resolves stage paths (including XPU substitutions where applicable) and builds **`OmniServerParams`** lists consumed by **`@pytest.mark.parametrize(..., indirect=True)`** on `omni_server` or `omni_server_function`.
- Cases are tagged **`@pytest.mark.slow`** for weekly / selective CI. GPU-heavy suites use **`@hardware_test(res={"cuda": "H100"}, num_cards=...)`** (Qwen3-Omni **2**× H100; Wan2.2 **1**× H100; HunyuanImage DiT **4**× H100).
+- Cases are tagged **`@pytest.mark.slow`** for weekly / selective CI. GPU-heavy suites use **`@hardware_test(res={"cuda": "H100"}, num_cards=...)`** or **`@hardware_test(res={"cuda": "L4"}, num_cards=1)`** (Qwen3-Omni **2**× H100; Wan2.2 **1**× H100; HunyuanImage DiT **4**× H100; VoxCPM2 **1**× L4).
 - **Process-kill** tests use **`@pytest.mark.skipif(os.name == "nt", ...)`** because injection uses POSIX **`pgrep` / `kill`**.

 ##### CI trigger
@@ -678,11 +680,13 @@ Weekly Buildkite (`.buildkite/cuda/test-weekly.yml`) runs one step per model sui
 | Reliability Test - qwen3-omni | `test_reliability_qwen3_omni.py` | H100 × 2 (`mithril-h100-pool`) |
 | Reliability Test - wan22 | `test_reliability_wan22.py` | H100 × 2 (`mithril-h100-pool`) |
 | Reliability Test - hunyuan-image | `test_reliability_hunyuan_image.py` | H100 × 4 (`mithril-h100-pool`) |
+| Reliability Test - voxcpm2 | `test_reliability_voxcpm2.py` | L4 × 1 (`gpu_1_queue`) |

 ```bash
 pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m "slow"
 pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m "slow"
 pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m "slow"
+pytest -s -v tests/dfx/reliability/test_reliability_voxcpm2.py -m "slow"
 ```

 ##### Local commands
@@ -692,6 +696,7 @@ pytest --collect-only tests/dfx/reliability
 pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m slow
 pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m slow
 pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m slow
+pytest -s -v tests/dfx/reliability/test_reliability_voxcpm2.py -m slow
 ```

 ##### Adding a new model suite
@@ -704,4 +709,4 @@ pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m slow
 </details>

 -   -   ***Stability***: `pytest -s -v tests/dfx/stability/scripts/test_stability_qwen3_omni.py` or `pytest -s -v tests/dfx/stability/scripts/test_stability_wan22.py` (or add `test_stability_<model>.py` alongside a matching JSON config)
-    -   ***Reliability***: `pytest -s -v tests/dfx/reliability/test_reliability_<model>.py -m slow` (current suites: `qwen3_omni`, `wan22`, `hunyuan_image`; add `test_reliability_<suite>.py` and a matching step in `.buildkite/cuda/test-weekly.yml` for new models)
+    -   ***Reliability***: `pytest -s -v tests/dfx/reliability/test_reliability_<model>.py -m slow` (current suites: `qwen3_omni`, `wan22`, `hunyuan_image`, `voxcpm2`; add `test_reliability_<suite>.py` and a matching step in `.buildkite/cuda/test-weekly.yml` for new models)
@@ -4,9 +4,6 @@
 > `experimental/ar_diffusion` contract. It follows the ownership and lifecycle
 > direction in [#5137](https://github.com/vllm-project/vllm-omni/issues/5137),
 > but it is not yet part of the stable module-design hierarchy.
->
-> The end-to-end request, identity, failure, and memory contracts are documented
-> in [Realtime AR-Diffusion sessions](feature/realtime_ar_diffusion.md).

 `ARDiffusionEngine` is an opt-in diffusion backend for pipelines that keep
 autoregressive attention KV across requests. The engine only selects
@@ -54,13 +51,14 @@ time, and its inherited batch/step entry points also fail explicitly. Supporting
 those paths requires a batch-aware state-binding contract rather than silently
 bypassing `bind_ar_diffusion_state()`.

-The implementation may retain multiple sessions per model instance. It selects
-the largest feasible capacity up to the pipeline's `session_capacity`, based on
-self-attention pages, cross-attention allocations, scratch, and declared
-model-owned state. One session is admitted whenever it fits actual free device
-memory; `gpu_memory_fraction` controls additional residency. Requests beyond
-that capacity evict the least-recently-used session through the same lifecycle
-hook used by explicit close and failure cleanup.
+The current implementation retains one session per model instance. A request
+for a different `session_id` closes the resident session through the same LRU
+lifecycle used by explicit close and failure cleanup before allocating the new
+one. Self-attention and cross-attention memory are therefore sized for one
+resident session even if the pipeline declares a larger `session_capacity`.
+This prioritizes rollout throughput and prevents block-pool exhaustion before a
+count-based eviction can run. Multi-session residency can be added later with a
+memory-backed capacity calculation or allocation-pressure eviction.

 Scratch storage is also capability-driven. `frames_per_block` reserves space for
 an uncommitted current video block, while `max_scratch_tokens_per_branch`
@@ -10,30 +10,6 @@ For user-facing commands, see the
 [distributed layerwise offloading guide](../../../user_guide/diffusion/offloader/distributed_layerwise_offload.md)
 and the [Cosmos3 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md).

-## Feature compatibility
-
-Host-storage optimization and runtime compatibility are separate decisions.
-When direct checkpoint mmap is unavailable, DLO can still use tensors produced
-by the ordinary loader. "Compatibility path" below means that fallback is
-implemented but has less end-to-end coverage than the primary path.
-
-Legend: ✅ supported, ⚠️ compatibility path or limited validation, ❌ unsupported.
-
-| Feature | DLO + AllGather | DLO without AllGather |
-|---|---|---|
-| **DP** | ✅ Primary path; host weights are sharded across the DP group. | ✅ Each DP rank streams complete rank-local blocks. |
-| **SP** | ✅ When DP=1, DLO uses the SP group for weight sharding. | ✅ SP remains active without a DLO weight collective. |
-| **TP > 1** | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. |
-| **HSDP** | ❌ Rejected to avoid double-sharding parameters. | ⚠️ Limited end-to-end coverage. |
-| **Per-tensor online FP8 linears** | ✅ Ordinary loader finalizes weights and scales before DLO sharding. | ✅ Ordinary loader retains complete rank-local tensors. |
-| **Other online quantization methods** | ❌ Rejected until runtime packing and scale layouts are validated. | ⚠️ Allowed through the ordinary loader; validation is method-specific. |
-| **Model-level or standard layerwise CPU offload** | ❌ Disabled because DLO takes priority. | ❌ Disabled because DLO takes priority. |
-| **Resident leading layers** | ❌ Rejected. | ✅ Requires eligible resident paths in the model's `OffloadPlan`. |
-
-See [Parallelism compatibility](#parallelism-compatibility) and
-[Request and loading constraints](#request-and-loading-constraints) for the
-detailed contracts and validation boundaries.
-
 ## Status

 DLO is implemented for multi-device diffusion execution. The default
@@ -43,291 +19,21 @@ DLO weight collective.

 Host storage is selected separately from the transfer protocol. The loader can
 produce a direct-checkpoint mmap plan for a proven-compatible runtime layout;
-otherwise it uses the ordinary loader. In no-AllGather mode, the Phase B host
-weight cache can explicitly normalize supported final ordinary-loader DiT
-tensors into an immutable node-local mmap entry. Ordinary pinned weights remain
-the default after checkpoint mmap fallback when the cache is not selected; cache
-publication and validation add startup and disk lifecycle costs. Replicas can
-share either proven-compatible checkpoint pages or, when selected, equivalent
-final runtime layouts. Under the existing
-pinned-memory policy, the backend attempts to make the complete final mapping
-directly H2D-copyable when the platform provides registration, avoiding the
-recurrent host packing otherwise required by mmap. An optional positive limit
-adds a per-worker safety ceiling. CUDA is the first implementation.
+otherwise it uses the ordinary loader. Consequently, no-AllGather replicas on
+the same node can share immutable checkpoint pages when direct mmap is
+selected, while the ordinary-loader fallback still keeps a private runtime
+copy per process.

 The Phase A shared-mmap support boundary is TP1. TP greater than one is an
 ordinary-loader compatibility path: DLO can consume the resulting TP-local
-tensors. Direct checkpoint mmap still does not cover TP, but the Phase B cache
-can share each final TP coordinate among equivalent processes.
+tensors, but those configurations do not use checkpoint mmap and must not be
+used to claim shared-mmap host-memory savings.

-The compatibility matrix below defines the current implementation boundary.
-Hardware and performance qualification is tracked separately from this
-long-lived design contract.
+The compatibility matrix below describes the current implementation. The
+unit-level guards are covered, but not every parallelism combination has a
+full model-and-hardware end-to-end test.

-## Architecture
-
-DLO is divided into four planes: configuration, loader-owned host storage,
-backend setup, and the per-block inference hot path. Keeping these decisions
-separate is intentional. In particular, an mmap cache is a host-storage choice;
-it does not imply a DLO AllGather, and disabling DLO AllGather does not decide
-whether the host source is private memory, checkpoint mmap, or host weight cache
-mmap. The shared control plane is described once below; the AllGather and
-no-AllGather startup and hot paths are then shown separately.
-
-### Component ownership
-
-The map includes optional mode-specific components, but the ownership rules
-and loader-to-backend boundary are shared.
-
-```mermaid
-flowchart TB
-    CLI["CLI / EngineArgs"] --> Config["OmniDiffusionConfig<br/>validation and operator policy"]
-    Config --> Runner["DiffusionModelRunner<br/>startup coordinator"]
-
-    subgraph LoaderPlane["Loader-owned host-storage plane"]
-        Loader["DiffusersPipelineLoader"]
-        Checkpoint["checkpoint mmap preflight"]
-        Ordinary["ordinary load + callbacks + calibration"]
-        HostWeightCache["host_weight_cache<br/>final-layout publication / join"]
-        Contract["HostWeightPlan<br/>backing + bindings + transforms"]
-        Loader --> Checkpoint
-        Loader -->|"registered final layout requested"| Ordinary
-        Checkpoint -->|"compatible"| Contract
-        Checkpoint -->|"fallback"| Ordinary
-        Ordinary -->|"host weight cache enabled"| HostWeightCache
-        HostWeightCache -->|"enabled and compatible"| Contract
-    end
-
-    Runner -->|"load_model"| Loader
-    Loader -->|"pipeline + optional HostWeightPlan"| Runner
-
-    Config --> Factory["get_offload_backend<br/>OffloadConfig normalization"]
-    Runner -->|"device + optional plan"| Factory
-    Factory --> Backend["DistributedLayerwiseOffloadBackend<br/>resource and transfer owner"]
-    Topology["existing DP / SP process groups"] --> Backend
-    Discovery["ModuleDiscovery + optional OffloadPlan"] --> Backend
-    Contract -.->|"passed once by the runner"| Backend
-
-    Backend -.->|"host weight cache + pinned-memory policy"| Registration["host_registration<br/>platform-neutral lifecycle"]
-    Registration --> CUDA["CUDA registration backend"]
-    Backend --> Hooks["DistributedLayerwiseOffloadHook<br/>one circular hook per streamed block"]
-    Forward["pipeline.forward"] --> Hooks
-    Hooks --> DataPath["two device slots<br/>H2D only or H2D + AllGather"]
-
-    Shutdown["DiffusionWorker.shutdown"] -->|"disable before process-group teardown"| Backend
-```
-
-| Component | Owns | Must not own |
-| --- | --- | --- |
-| CLI, engine arguments, and `OmniDiffusionConfig` | User policy, defaults, validation, and propagation to every worker | Checkpoint inspection, cache publication, or buffer allocation |
-| `OffloadConfig.from_od_config()` | Offload-strategy selection, effective DLO group size, AllGather policy, resident-layer count, pinned-memory policy, and optional registration ceiling | Host-weight compatibility decisions |
-| `DiffusersPipelineLoader` | Ordinary loading lifecycle and the decision to return an optional `HostWeightPlan` | DLO process groups or block-stream scheduling |
-| Checkpoint mmap planner | Proving that checkpoint tensors can represent the required runtime layout before ordinary DiT materialization is skipped | Host weight cache identity or transfer policy |
-| `host_weight_cache` | Normalizing final ordinary-loader tensors into an immutable node-local entry and returning a validated plan | Model mutation, host registration, H2D, or collectives |
-| `HostWeightPlan` | The one-way loader-to-offloader contract: backing kind, tensor bindings, source coverage, and any bounded transforms | Cache construction or runtime scheduling |
-| `ModuleDiscovery` and model `OffloadPlan` | Discovering DiT/components and declaring block, resident, and on-demand paths | Loader or cache behavior embedded in model pipelines |
-| `DistributedLayerwiseOffloadBackend` | Consuming the exact plan, selecting the existing group, mapping/repointing storage, allocating shared buffers and streams, optional host registration, installing hooks, and cleanup | Re-running loader compatibility checks or orchestrating replicas |
-| `DistributedLayerwiseOffloadHook` | The block hot path: prefetch, optional DLO AllGather, readiness events, parameter repointing, and replacement with offload placeholders | Cache keys, manifest validation, or request routing |
-| `host_registration` and platform backend | All-or-nothing registration lifecycle for existing immutable mappings | Choosing which weights are cached or changing the two-slot device-buffer contract |
-| `DiffusionWorker` | Calling backend teardown before destroying the distributed environment | Managing backend-owned mappings or registrations directly |
-
-The configuration fields are routed to the component that can enforce their
-contract. The cache root and lock timeout are consumed by the loader.
-`dlo_use_host_weight_cache` explicitly selects final-layout cache publication
-instead of direct-checkpoint mmap. It defaults to false so the ordinary
-pinned-loader path remains available without requiring an opt-out.
-With `pin_cpu_memory=True`, selecting the host weight cache deliberately chooses
-the transform-complete layout even when direct checkpoint mmap is available,
-because only that final layout can be registered without deferred transforms.
-The backend consumes `dlo_host_weight_cache_pin_limit_gib` as an optional
-registration ceiling; zero applies no additional ceiling and does not select
-the cache by itself.
-`dlo_use_allgather` controls loader plan routing because the host weight cache
-is no-AllGather-only, and it selects the backend transfer protocol. This is
-coordination through typed configuration and `HostWeightPlan`, not through
-model-specific global state.
-
-### Runtime host-layout identity (v1)
-
-Host weight cache v1 identifies the loader-produced host representation, not
-the topology used later to move weights to devices. These are separate
-contracts:
-
- `RuntimeHostLayout` describes tensor ownership and representation after the
-  ordinary loader has applied its parallel-aware transforms. This is the
-  layout stored in the host weight cache.
- `DLOTransferLayout` describes whether each worker copies its local weights
-  directly or participates in the existing DLO AllGather protocol. It is not
-  part of the host weight cache in v1; AllGather continues to use its existing
-  process-owned host shards.
-
-The v1 physical-sharing invariant is one cache entry per TP coordinate,
-shared by equivalent DP and SP consumers on the same node. The identity
-therefore includes TP world size and TP rank, because different TP coordinates
-own different tensors. It excludes DP world size/rank and SP rank, because
-those dimensions do not change the final tensor ownership supported by v1.
-SP world size and implementation details remain conservative compatibility
-guards until equivalent layouts across those configurations are proven.
-
-PP, HSDP/DTensor, expert parallelism, and CFG parallelism are rejected rather
-than assigned speculative cache identities. Supporting one of these layouts
-requires first defining its loader-produced ownership, representation, and
-sharing domain; it does not follow automatically from supporting its runtime
-collectives. Consequently, this phase implements a TP-coordinate runtime host
-layout, not a general parallel-layout cache.
-
-### Shared startup and lifecycle
-
-The shared control path selects host backing and transfers its ownership to
-the backend before either transfer mode installs hooks:
-
-```mermaid
-sequenceDiagram
-    participant C as OmniDiffusionConfig
-    participant R as DiffusionModelRunner
-    participant L as DiffusersPipelineLoader
-    participant B as DLO backend
-    participant W as DiffusionWorker
-
-    C->>R: validated offload policy
-    R->>L: load_model(config)
-    L->>L: select and validate host backing
-    L-->>R: pipeline + optional HostWeightPlan
-    R->>B: enable(config, device, optional plan)
-    B->>B: discover modules and consume host backing
-    B-->>R: mode-specific backend is ready
-    W->>B: shutdown calls disable()
-    B->>B: synchronize, unregister, remove hooks, close mappings
-```
-
-The loader returns a plan only after its contract is valid: checkpoint plans
-have preflight-proven source coverage and metadata, while host weight cache
-plans also have full manifest, file, and mapped-content validation. The runner
-then transfers ownership exactly once with
-`model_loader.take_host_weight_plan()`. If no plan is returned, the backend
-requires materialized ordinary-loader tensors. If a plan caused ordinary DiT
-materialization to be skipped, failure to create or enable its consumer is a
-hard startup error rather than a fallback to uninitialized parameters.
-
-At inference time, the model pipeline remains the caller: normal block forward
-calls activate the installed hooks. Both modes use one copy stream and two
-shared device slots sized for the largest streamed block. The hooks do not
-call the loader or cache.
-
-### AllGather startup and hot path
-
-```mermaid
-sequenceDiagram
-    participant B as DLO backend
-    participant H as block hooks
-    participant G as existing DP or SP group
-
-    B->>B: choose existing group and prepare persistent local shards
-    B->>B: allocate two device slots and a communication stream
-    B->>H: install hooks and prime the first block
-    loop each streamed block
-        H->>H: copy the next local shard to the alternate slot
-        H->>G: all_gather_into_tensor(next block)
-        G-->>H: complete next block is ready
-        H->>H: execute current block and rotate slots
-    end
-```
-
-AllGather uses the existing DP group when DP is greater than one, otherwise
-the existing SP group when SP is greater than one. TP is never selected as
-DLO's weight-sharding group. Every participating rank must request blocks in
-the same order; fast P2P reduces communication cost but does not remove that
-lockstep contract. An effective group size of one issues no collective and
-uses rank-local transfer.
-
-### No-AllGather startup and hot path
-
-```mermaid
-sequenceDiagram
-    participant L as DiffusersPipelineLoader
-    participant C as host_weight_cache
-    participant B as DLO backend
-    participant P as platform registration
-    participant H as block hooks
-
-    alt host weight cache explicitly selected
-        L->>L: load, mutate, validate, and calibrate final tensors
-        L->>C: publish or join final-layout entry
-        C-->>L: validated host weight cache plan or fail-closed fallback
-        L-->>B: cache plan or materialized fallback via runner
-    else automatic checkpoint selection
-        L->>L: checkpoint mmap preflight
-        alt loader selects a compatible checkpoint plan
-            L-->>B: checkpoint plan via runner
-        else ordinary loading
-            L->>L: load, mutate, validate, and calibrate final tensors
-            L-->>B: materialized tensors via runner
-        end
-    end
-    B->>B: consume complete rank-local block sources
-    opt host weight cache plan and pinned-memory policy
-        B->>P: register the complete immutable mapping
-        P-->>B: direct-H2D source or bounded-staging fallback
-    end
-    B->>H: install hooks and prime the first block
-    loop each streamed block
-        alt registered host weight cache mapping
-            H->>H: copy complete next block directly to device
-        else pageable mmap mapping
-            H->>H: pack into a bounded staging slot, then copy to device
-        else ordinary runtime tensors
-            H->>H: copy complete next block to device
-        end
-        H->>H: execute current block and rotate slots
-    end
-```
-
-No-AllGather creates no DLO weight group and issues no DLO collective.
-Workers may therefore schedule requests independently. Ordinary TP and SP
-collectives remain inside model execution; host registration changes only
-the source of the complete-block H2D copy.
-
-### Host backing and transfer composition
-
-| Loader-selected host backing | Backend transfer mode | Setup result | Per-block hot path |
-| --- | --- | --- | --- |
-| Ordinary runtime tensors | AllGather group larger than one | Each rank retains a pinned local weight shard | shard H2D → DLO AllGather → device slot |
-| Checkpoint mmap | AllGather group larger than one | Each rank prepares its persistent shard, then releases the source mapping | shard H2D → DLO AllGather → device slot |
-| Ordinary runtime tensors | Rank-local / no-AllGather | Each rank retains the complete loader-produced local layout | complete-block H2D → device slot |
-| Checkpoint mmap, or host weight cache with pinned memory disabled/registration unavailable | Rank-local / no-AllGather | Immutable mapping plus two bounded host-staging slots | mmap views → staging slot → device slot |
-| Host weight cache mmap with pinned memory enabled and registration available | Rank-local / no-AllGather | Complete mapping is registered by the platform backend | registered mmap views → device slot |
-
-The three axes therefore remain independent:
-
-1. **Offload policy:** whether DLO is enabled and which layers are resident.
-2. **Host backing:** ordinary tensors, checkpoint mmap, or final host weight cache
-   mmap, selected by the loader.
-3. **Transfer protocol:** rank-local H2D or H2D plus DLO AllGather, selected by
-   the backend from validated configuration and existing topology.
-
-### Failure and lifecycle boundaries
-
- Invalid option combinations fail during configuration validation.
- Checkpoint mmap preflight failure falls back before ordinary loading is
-  skipped.
- Host weight cache incompatibility, timeout, or publication failure retains
-  the already-valid ordinary tensors.
- Host-registration failure rolls back and retains bounded staging. A rollback
-  failure aborts startup because unmapping a range still owned by the platform
-  would be unsafe.
- During shutdown, DLO synchronizes outstanding device work, unregisters host
-  mappings, removes hooks, and closes mmap handles only after safe unregister;
-  the worker destroys its process groups afterward.
-
-DLO deliberately does not own DP replica placement, independent-engine request
-routing, cross-engine admission control, or cross-node cache distribution.
-Those belong to an external router/orchestrator. It also does not replace TP or
-SP model collectives; it only adds an optional weight collective over an
-already-existing DP or SP group.
-
-## Shared design
-
-The following topology and loader contracts apply to both transfer modes.
+## Design

 ### DLO consumes the existing parallel topology

@@ -335,10 +41,15 @@ DLO does not create a new DP, TP, or SP topology. It reads the configured
 `DiffusionParallelConfig` and attaches offload hooks to the DiT blocks after the
 standard distributed groups have been initialized.

-AllGather mode selects one existing group for DLO weight sharding; no-AllGather
-mode creates no DLO group. Neither mode changes ordinary TP or SP model
-collectives. The mode-specific group and scheduling contracts are described
-below.
+The DLO weight-sharding group is selected as follows:
+
+1. Use the existing DP group when `data_parallel_size > 1`.
+2. When DP is one and SP is greater than one, use the SP group.
+3. Otherwise, run rank-locally without a DLO process group.
+
+TP is deliberately not used as DLO's AllGather group. HSDP has its own
+parameter-sharding lifecycle and is not allowed to be sharded a second time by
+DLO's AllGather path.

 ### The loader owns host-weight planning

@@ -370,24 +81,7 @@ flags such as `_supports_mmap_loading` or parameter attributes for mmap-only
 transforms. Model-specific direct-layout knowledge, when required, lives in a
 checkpoint adapter beside the ordinary loader.

-## AllGather design
-
-### Group selection
-
-The DLO weight-sharding group is selected in this order:
-
-1. Use the existing DP group when `data_parallel_size > 1`.
-2. When DP is one and SP is greater than one, use the existing SP group.
-3. Otherwise, use an effective group size of one and issue no collective.
-
-TP is deliberately not used as DLO's AllGather group. HSDP has its own
-parameter-sharding lifecycle and cannot be sharded a second time by DLO.
-
-### Sharded transfer and double buffering
-
-The loader may provide ordinary runtime tensors or a compatible checkpoint
-plan. During backend setup, either source is normalized into persistent local
-shards. The final-layout host weight cache is not selected in AllGather mode.
+### AllGather path

 With the default `dlo_use_allgather=True`, each rank stores approximately
 `1 / group_size` of each streamable block in pinned host memory. The next
@@ -395,15 +89,6 @@ block's shard is copied to a device buffer and reconstructed with
 `all_gather_into_tensor` on a communication stream while the current block is
 executing.

-![AllGather DLO sharded host weights and lockstep reconstruction](../../figures/dlo/dlo_allgather_sharded_host.svg)
-
-The figure uses DP2 with TP2 to expose both parallel dimensions. DLO creates
-one AllGather group across the DP replicas for each TP coordinate; TP0 workers
-reconstruct TP0 blocks together and TP1 workers reconstruct TP1 blocks
-together. Ordinary TP collectives remain inside each DP replica. Unlike the
-no-AllGather host weight cache path, these persistent pinned host shards are
-process-owned, and every participating rank must follow the same block order.
-
 ```text
 Compute:    [Block N]             [Block N+1]          [Block N+2]
 H2D:                      [shard N+1]           [shard N+2]
@@ -429,29 +114,11 @@ When DP is greater than one, the engine can process one request per DP rank in
 the same denoising wave. Because AllGather is a collective, all participating
 requests must take the same execution path at every denoising step.

-Fast NVLink/NVSwitch changes collective cost, not this execution contract. For
-independently scheduled replicas—even on one NVSwitch baseboard—no-AllGather
-plus shared mmap storage avoids coupling their request schedules and failure
-domains.
-
-## No-AllGather design
-
-### Rank-local transfer
+### Rank-local path without DLO AllGather

 With `--dlo-no-use-allgather`, DLO forces its internal offload shard size to
 one and streams complete blocks using H2D copies only. The host backing may be
-a loader-approved checkpoint mapping, a normalized host weight cache mapping, or
-ordinary runtime tensors.
-
-![No-AllGather DLO shared host weights and independent runtime orchestration](../../figures/dlo/dlo_no_allgather_shared_host.svg)
-
-The figure uses TP2 to make per-coordinate physical sharing explicit. Each
-engine contains two rank-local workers and devices. The cache keeps one
-immutable entry per TP coordinate, so equivalent TP0 workers across engines
-share the TP0 entry and equivalent TP1 workers share the separate TP1 entry.
-An external router or orchestrator chooses each engine's requests; DLO
-coordinates only the worker-local mapping, hooks, transfer streams, events,
-and device slots.
+either a loader-approved checkpoint mapping or ordinary runtime tensors.

 For direct mmap, each process retains immutable safetensors views and uses two
 bounded pinned host staging slots. Processes on the same node that map the same
@@ -460,17 +127,10 @@ the persistent private full-model copy per pure-DP process, but each process
 still packs and transfers every complete block. Sharing is node-local; each
 node has its own page cache.

-For a host weight cache mapping, the default pinned-memory policy attempts to
-replace those host slots with direct copies from the shared final tensor views.
-Disabling pinned memory or a failed registration retains staging behavior.
-
-When direct mmap preflight fails or the host weight cache is selected, the
-regular model loader remains responsible for preparing each rank's weights,
-including TP-local tensors or HSDP-managed parameters. Ordinary pinned tensors
-remain the default without cache selection. When the host weight cache is
-explicitly selected,
-supported CPU layouts enter the cache; unsupported layouts keep one private
-runtime copy per process.
+When direct mmap preflight fails, the regular model loader remains responsible
+for preparing each rank's weights, including TP-local tensors or HSDP-managed
+parameters. In that fallback, each pure-DP process keeps a private full runtime
+copy.

 This mode means:

@@ -481,109 +141,18 @@ This mode means:
 - TP/HSDP/SP collectives, if configured, are not disabled by this flag; only
  DLO's additional weight AllGather is disabled.
 - Pure DP deployments share one checkpoint-backed copy per node when direct
-  mmap is selected. The host weight cache can share equivalent ordinary-loader
-  outputs; its fallback keeps one private runtime copy per rank.
+  mmap is selected; the ordinary-loader fallback keeps one private runtime
+  copy per rank.
 - The scheduler does not require a synchronized DP request wave for DLO.

-### Post-load host weight cache
-
-The Phase B cache preserves the same ownership boundary. DLO neither derives
-cache keys nor interprets model loaders. No-AllGather DLO first attempts a
-compatible direct-checkpoint mmap plan. When that plan is unavailable, every
-rank completes the ordinary loader, custom callbacks, post-load processing,
-strict validation, and calibration. When cache use or registration is
-explicitly selected, the loader:
-
-1. discovers final CPU DiT parameters and persistent buffers;
-2. rejects unsupported non-contiguous, aliased, quantized, distributed, or
-   ambiguous layouts before changing the model;
-3. hashes tensor names, metadata, and final bytes;
-4. builds or joins one content-addressed entry under a strict POSIX lock; and
-5. returns a `HostWeightPlan(backing_kind="host_weight_cache")` only after file,
-   manifest, and remapped-tensor content validation succeeds.
-
-The identity excludes DP size/rank and SP rank. It includes TP world size/rank
-and conservative SP implementation/world-size guards. DP replicas and
-equivalent SP ranks can therefore share; different TP coordinates cannot.
-Final tensor content is authoritative, so an unmodeled loader/config change
-that changes bytes creates a different entry instead of risking silent
-sharing.
-
-Publication writes sharded safetensors and a JSON manifest into a same-
-filesystem temporary directory, flushes them, and atomically renames the
-directory. The manifest includes per-file SHA-256 digests and the final tensor
-content digest. Kernel `flock` ownership distinguishes a live writer from a
-stale lock file: process death releases the lock automatically. Stale temporary
-entries are removed only by the next lock owner. A bounded wait or any cache
-error retains the already-valid ordinary tensors.
-
-DLO maps this plan without moving the DiT to `meta` and without rerunning any
-post-load hook. It changes tensor storage while preserving `Parameter` objects,
-then runs the existing `gc.collect()`/`malloc_trim()` cleanup before installing
-streaming hooks. The source mappings remain open through DLO disable. With
-pinned memory disabled, or when registration is unavailable, they feed the
-existing bounded two-slot host-staging path. Every streamed block is first
-packed from pageable mmap storage into a staging slot and is then copied to the
-device. This removes model-sized private pinned storage, but adds a recurrent
-host-to-host copy and can amplify TP synchronization waits when several ranks
-compete for host bandwidth.
-
-This version is deliberately steady-state-oriented. It requires roughly one
-local-disk model copy per runtime identity, has no eviction, and still performs
-ordinary loading plus full content-hash passes in every rank. Cache publication
-or validation failure falls back to the already-valid ordinary host tensors.
-Skip-load-on-hit is a separate follow-up after loader side effects can be
-modeled safely.
-
-### Registered host weight cache transfer
-
-Selecting the host weight cache with `pin_cpu_memory=True` opts a worker into
-registration of the complete mapping when its platform supports it. It also
-changes loader selection: no-AllGather uses the ordinary loader and final-layout
-host weight cache even when a TP1 direct-checkpoint plan is available. A
-checkpoint binding can carry a deferred adapter such as
-grouped-QKV reordering. Registering that raw source would not remove the
-adapter's recurrent CPU allocation/copy, whereas the host weight cache contains
-the transform-complete final tensors.
-
-After mapping and validating every final tensor, DLO groups source storages by
-backing file, page-aligns their address ranges, and coalesces overlapping or
-adjacent ranges within that mapping. When a positive per-worker ceiling is
-configured, the complete byte count is checked against it before any platform
-mutation. The CUDA backend
-first requires `cudaDevAttrHostRegisterReadOnlySupported`, then registers every
-range with the read-only host-registration flag and verifies that PyTorch
-recognizes every mapped tensor as pinned. A device without that capability
-retains bounded staging; vLLM-Omni does not replace immutable mappings with a
-writable or copy-on-write CUDA-specific layout. Other registration failures
-roll back any ranges already registered and preserve the existing two-slot
-staging path. Partial direct transfer is deliberately unsupported. A rollback
-failure aborts worker initialization so a still-registered range is retained
-until process/context teardown rather than being unsafely unmapped.
-
-On success, the hooks copy each immutable tensor view directly into its offset
-in the existing flattened rotating device buffer on the copy stream. Parameter
-repointing, two-block HBM residency, and H2D/compute overlap are unchanged. The
-direct path trades recurrent host packing for finer-grained H2D operations. A
-packed cache schema should be considered only if that additional format and
-publication complexity becomes necessary.
-
-Registration is per process and platform context; physical file-cache pages
-remain shared and are not copied into one private host allocation per worker.
-The operator must nevertheless budget page-locked memory and satisfy the
-platform and OS registration limits for every worker. During shutdown, DLO
-first synchronizes outstanding device work, unregisters all ranges, and then
-closes the safetensors mappings. Worker teardown invokes this cleanup before
-destroying the distributed environment and platform context.
-
 ## Parallelism compatibility

 | Parallelism | DLO + AllGather | DLO without AllGather |
 |---|---|---|
-| **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. Compatible TP1 replicas can share checkpoint pages; the optional host weight cache shares equivalent final layouts while excluding DP rank from its identity. CUDA workers may register the complete final mapping for direct H2D. |
-| **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active without a DLO weight collective. Host weight cache v1 excludes SP rank and includes conservative SP implementation/world-size guards. Registered transfer does not remove ordinary SP activation/attention collectives. |
-| **TP > 1** | Outside the Phase A direct-mmap scope. The loader preserves TP-local layouts and DLO may apply DP/SP host sharding to ordinary runtime tensors. | The ordinary TP-aware loader produces rank-local tensors. Host weight cache v1 creates one entry per TP coordinate, shareable by equivalent DP/SP processes. Registering those final mappings avoids the recurrent staging copy but does not remove ordinary TP collectives. |
-| **HSDP** | Rejected. HSDP has already sharded parameters, so DLO AllGather would double-shard them. | Accepted only with ordinary-loader rank-local tensors. HSDP owns parameter sharding and its own gathers; host weight cache v1 rejects HSDP/DTensor layouts. |
+| **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. Compatible TP1 replicas can share checkpoint pages on each node; fallback runtime tensors remain private. |
+| **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active, but DLO keeps standard-loader rank-local weights and adds no SP weight collective. |
+| **TP > 1** | Outside the Phase A shared-mmap support scope. The loader falls back before mutation, preserves TP-local layouts, and DLO may apply DP/SP host sharding to those ordinary runtime tensors. | Outside the Phase A shared-mmap support scope. The ordinary TP-aware loader produces rank-local tensors, which DLO streams without an additional weight collective; DP replicas retain private runtime storage. |
+| **HSDP** | Rejected. HSDP has already sharded parameters, so DLO AllGather would double-shard them. | Accepted by configuration. HSDP owns parameter sharding and its own gathers; DLO only stages rank-local parameters. End-to-end coverage is limited. |

 ### Combined dimensions

@@ -599,9 +168,7 @@ destroying the distributed environment and platform context.
 - **HSDP + DP or TP:** rejected independently by the diffusion parallel
  configuration.

-## Mode constraints
-
-### AllGather constraints
+## Request and loading constraints

 AllGather DP multi-concurrency requires:

@@ -609,54 +176,104 @@ AllGather DP multi-concurrency requires:
 - the same `num_inference_steps` for all requests in a wave; and
 - identical request arguments that affect the collective execution path.

-Per-tensor online FP8 linears can use DLO AllGather after the ordinary loader
-finalizes their runtime weights and scales. Other online methods must use
-`--dlo-no-use-allgather` or disable online quantization until their runtime
-layouts are validated.
-
-### No-AllGather constraints
-
 The no-AllGather path does not impose these DLO-specific synchronized-wave
-requirements. Host weight cache v1 is no-AllGather only and rejects all
-quantization configurations, HSDP/DTensor, expert parallelism, CFG
-parallelism, PP, non-contiguous tensors, and tied/shared storage.
-
-### Shared loading constraint
+requirements.

 Direct checkpoint mmap can back either transfer path. It is currently limited
 to proven TP1, non-HSDP, non-online-quantized layouts. Other layouts use the
-ordinary loader.
+ordinary loader. Online quantization remains incompatible with DLO AllGather;
+use `--dlo-no-use-allgather` or disable online quantization.

-## Validation status
+A normalized runtime mmap cache, built through the ordinary loader, is the
+proposed general mechanism for sharing transformed TP or quantized layouts.
+That cache and its publication/lifecycle protocol are intentionally outside
+this phase; see [RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195).

-Historical test counts and hardware measurements are intentionally not kept
-in this design document because they can drift as the implementation changes.
-Revalidation must run against the final commit and publish commit-pinned
-artifacts before making hardware or performance claims. Track that work in
-[issue #6231](https://github.com/vllm-project/vllm-omni/issues/6231).
+## Validation coverage
+
+Current source-level validation includes:
+
+- HSDP + DLO + AllGather rejection;
+- HSDP + DLO without AllGather acceptance at configuration level;
+- loader preflight fallback for TP, HSDP, online quantization, unknown custom
+  loaders, missing keys, and shape/dtype mismatches;
+- exact loader-to-backend plan transfer and ordinary-loader fallback;
+- rank-local mmap source retention, bounded two-slot staging, and adapter
+  transforms without parameter-side flags;
+- resident-layer requests requiring no-AllGather;
+- DP request-wave validation for denoising-step compatibility;
+- sharding, double-buffer, AllGather-size, and heterogeneous-block regression
+  tests.
+
+### B300 parallel-topology smoke matrix
+
+A four-GPU B300 smoke test covered MiniMax-H3 FL2VA with the same prompt, seed,
+CUDNN attention backend, 256x256 output, two denoising steps, and
+`dlo_resident_layers=0`. The TP2 rows used DiT DP2xTP2 with the text encoder and
+VAEs at TP1. They validate the ordinary-loader fallback only, not direct mmap
+or shared-mmap host-memory savings.
+
+| Configuration | Result | Warm E2E | Peak device memory | Host PSS |
+|---|---:|---:|---:|---:|
+| DP4xTP1 AllGather | Passed, 4 concurrent requests | 2.87 s / 4 requests | 13.84 GiB | 211.99 GiB |
+| DP4xTP1 no-AllGather | Passed, 1 request | 15.02 s | 13.23 GiB | 187.77 GiB |
+| DP2xTP2 AllGather | Passed, 2 concurrent requests | 4.16 s / 2 requests | 12.50 GiB | 211.97 GiB |
+| DP2xTP2 no-AllGather | Passed, 1 request | 3.51 s | 11.88 GiB | 314.01 GiB |
+
+Within each topology, the AllGather and no-AllGather video and audio outputs
+were byte-identical. All four runs completed without an `ERROR` or traceback
+and released their device allocations. For DP4xTP1, no-AllGather direct mmap
+reduced total PSS by 24.22 GiB (11.4%) and `Private_Dirty` from 211.33 to
+125.32 GiB (40.7%) relative to AllGather. For DP2xTP2, preflight selected the
+ordinary loader as designed; no-AllGather PSS was 314.01 GiB, about 48% above
+AllGather, because DP replicas did not share checkpoint-backed runtime
+weights. This is a functional and memory smoke test, not a production-quality
+performance or output-quality benchmark.
+
+### Host-memory measurement
+
+A two-worker MiniMax-H3 FL2VA measurement on one L20X node compared the
+ordinary-loader fallback with direct mmap. Both runs used
+DP=2, TP=1, no DLO AllGather, BF16 weights, two denoising steps, and a
+256x256 four-second request. The ordinary-loader workers were sampled after
+initialization. The mmap workers were sampled after one completed request, so
+the checkpoint working set had been faulted into the page cache; this is the
+more conservative point for mmap.
+
+The values below come from `/proc/<worker>/smaps_rollup` and include the whole
+worker, not only the DiT. The stable rank-to-rank difference comes from other
+pipeline components, so each worker should be compared with the same worker in
+the other storage mode.
+
+| Worker | Ordinary RSS | mmap RSS | Ordinary PSS | mmap PSS | PSS reduction |
+|---|---:|---:|---:|---:|---:|
+| DP worker 0 | 168.27 GiB | 132.76 GiB | 167.84 GiB | 101.43 GiB | 66.40 GiB |
+| DP worker 1 | 116.19 GiB | 79.97 GiB | 115.73 GiB | 48.64 GiB | 67.09 GiB |
+| **Two-worker total** | — | — | **283.56 GiB** | **150.08 GiB** | **133.48 GiB (47.1%)** |
+
+The direct-mmap workers each reported 62.45 GiB `Shared_Clean` but only
+31.20 GiB `Pss_File`, which is the proportional charge expected when the same
+resident checkpoint pages are mapped by two workers. `Private_Dirty` also fell
+from 167.53 to 70.24 GiB for worker 0 and from 115.40 to 17.44 GiB for worker
+1, a reduction of about 97–98 GiB per worker. RSS understates this benefit
+because it counts a shared physical page in every process that maps it; summed
+PSS is the appropriate node-memory comparison.
+
+The highest-value missing coverage is broader end-to-end numerical and
+lifecycle comparison against ordinary layerwise offload for DP+SP,
+HSDP+SP+no-AllGather, and TP greater than one across additional models and
+target CUDA/NCCL or CANN/HCCL hardware. That broader TP coverage does not
+change the Phase A direct-mmap TP1 support boundary.

 ## Recommendations

-### AllGather
-
- Use **DP + DLO AllGather** when ranks already execute the same block path in
-  lockstep and belong to the same fast P2P domain. It shards persistent host
-  weights across that group.
+- Use **DP + DLO AllGather** for the supported throughput and host-memory
+  scaling path.
 - Use **SP + DLO AllGather** for long-sequence workloads when DP concurrency is
  not the goal.
-
-### No-AllGather
-
 - Use **no-AllGather** when independent replica execution is required. TP1
-  direct-mmap deployments can share checkpoint pages per node. Enable the
-  host weight cache only when independent replicas have repeated final layouts
-  and benchmark the target CPU/memory topology, especially for matching
-  coordinates across multiple TP engines. With the default pinned-memory
-  policy, supported CUDA workers prefer registered direct H2D. Configure a
-  positive ceiling when an explicit page-locking bound is required, or disable
-  pinned memory when pageable staging is operationally preferable.
-
-### HSDP
-
- Do not combine HSDP with DLO AllGather or host weight cache v1. HSDP with
-  no-AllGather remains an ordinary-loader rank-local path.
+  direct-mmap deployments can share checkpoint pages per node; other layouts
+  retain the ordinary loader's private host memory behavior and are outside
+  the Phase A shared-mmap support scope.
+- Prefer **HSDP alone** for production HSDP deployments until the combined
+  HSDP + DLO no-AllGather path has broader end-to-end coverage.
@@ -190,15 +190,12 @@ configured parallelism strategy:
 - `WeightsMapper` transformations must be applied before matching ignored-layer
  lists or component prefixes.

-Distributed layerwise offload has an additional boundary: direct checkpoint
-mmap cannot prepare online-quantized tensors because it bypasses the ordinary
-loader that creates their runtime weights and scales. Per-tensor online FP8
-linears can still use DLO AllGather through the ordinary loader: each rank first
-finalizes the FP8 tensors, then DLO shards those runtime tensors for
-reconstruction during prefetch. Other online methods remain limited to
-`--dlo-no-use-allgather` until their packing and scale layouts are validated.
-See the [DLO design document](offloader/distributed_layerwise_offload.md) for
-the full compatibility matrix.
+Distributed layerwise offload has an additional boundary: its sharded mmap
+AllGather path is not compatible with online quantization because the path
+expects already materialized checkpoint weights. Use the rank-local
+`--dlo-no-use-allgather` path or a compatible pre-quantized checkpoint when
+combining these features. See the [DLO design document](offloader/distributed_layerwise_offload.md)
+for the full compatibility matrix.

 ## Adding a quantization backend

@@ -1,213 +0,0 @@
-# Realtime AR-Diffusion sessions
-
-> **Status:** experimental. This document describes the internal execution
-> contract implemented under `vllm_omni.experimental.ar_diffusion`. It does not
-> define a public HTTP or WebSocket schema.
-
-## Scope
-
-Realtime AR-Diffusion serves world models that generate one autoregressive
-latent block per request while retaining model state across requests. The
-generic runtime owns identity, event ordering, session lifecycle, routing, and
-KV allocation. A model adapter owns model-specific controls and conditioning.
-
-LingBot World v2 is the first single-KV-branch integration. DreamZero uses the
-same runner capability with two logical CFG branches.
-
-## Identities
-
-Four identifiers have distinct lifetimes:
-
-| Identifier | Lifetime | Owner | Contract |
-| --- | --- | --- | --- |
-| `session_id` | One persistent world | Session manager | Selects runner KV and model-owned state. |
-| `event_id` | One control/prompt update | Transport adapter | Unique and monotonically increasing within a session, including across reset. |
-| `chunk_index` | One committed AR block position | Session | Contiguous from zero; reset starts again at zero. |
-| `request_id` | One chunk execution | Session | Correlates the tick with its returned metadata. |
-
-`AsyncOmni` keeps its normal UUID-suffixed engine routing ID. The tick
-`request_id` remains inside the immutable AR-Diffusion snapshot and the
-standard output metadata, while caller-visible `OmniRequestOutput.request_id`
-continues to use AsyncOmni's existing external-ID contract. Protocol identity
-therefore does not require a private entrypoint override.
-
-## Request lifecycle
-
-1. `ARDiffusionSession.accept_event()` validates and queues an immutable
-   `ARDiffusionSessionEvent`. Prompt and opaque control tracks may be updated in
-   the same event.
-2. `ARDiffusionSession.next_chunk()` serializes chunk execution and calls
-   `_build_tick_locked()`. The resulting `ARDiffusionTickRequest` is a snapshot
-   of the current prompt, reduced controls, ordered event IDs, and chunk index.
-3. `ARDiffusionOmniTickConsumer.execute_tick()` copies the stage sampling
-   parameters, stores the typed tick in
-   `sampling_params.extra_args["ar_diffusion_tick"]`, constructs the normal Omni
-   prompt through its `prompt_provider`, and invokes `AsyncOmni.generate()`.
-4. `AsyncOmni` allocates its normal unique engine routing ID and submits one
-   `OmniDiffusionRequest` through the orchestrator.
-5. `ARDiffusionModelRunner._request_session()` parses the typed tick.
-   `_get_or_create_session()` obtains runner-owned paged KV, and
-   `bind_ar_diffusion_state()` exposes it to the pipeline only for the duration
-   of `execute_model()`.
-6. The model pipeline validates the tick and generates exactly one latent
-   block. LingBot interprets camera controls in
-   `LingBotCameraControlReducer` and
-   `LingBotWorldCausalDMDPipeline.forward()`; the generic runtime does not parse
-   camera or action schemas.
-7. The pipeline returns one `DiffusionOutput` with the generated payload and an
-   `ARDiffusionChunkMetadata` mapping under `metadata["ar_diffusion"]`.
-   `AsyncOmni` exposes it as
-   `OmniRequestOutput.multimodal_output["metadata"]["ar_diffusion"]`.
-8. The consumer parses the standard envelope. The session commits its reducer
-   state, prompt, controls, event IDs, and next chunk index only when returned
-   metadata exactly equals the submitted tick snapshot.
-
-The metadata envelope is:
-
-```json
-{
-  "metadata": {
-    "ar_diffusion": {
-      "session_id": "world-7",
-      "request_id": "ar-world-7-3-...",
-      "chunk_index": 3,
-      "applied_event_ids": [10, 11]
-    }
-  }
-}
-```
-
-## Transaction and failure semantics
-
-Accepted events are not acknowledged as applied until the model output and
-metadata are validated. Reducer `prepare()` is speculative; reducer `commit()`
-and the session snapshot form one logical commit. A model, metadata, or reducer
-failure leaves the chunk index and queued events uncommitted and transitions the
-session to `FAILED`.
-
-After a tick failure, worker state is closed because a forward may have
-partially modified or already released KV. The same chunk is not retried in
-place. A caller must explicitly reset the session, starting again at chunk
-zero, or close it.
-
-Explicit close follows this state machine:
-
-```text
-ACTIVE or FAILED
-       |
-       v
-    CLOSING ---- worker cleanup succeeds ----> CLOSED
-       |
-       +------- worker cleanup fails --------> CLEANUP_FAILED
-                                                   |
-                                                   +-- close retry --> CLOSING
-```
-
-`CLEANUP_FAILED` is a local tombstone. The session manager retains it, rejects
-creation of the same `session_id`, and rejects reset. Only an idempotent close
-retry that succeeds removes the session. Transport disconnect uses the same
-path.
-
-## KV ownership and capacity
-
-`ARDiffusionKVCacheSpec` is the pipeline-owned immutable geometry contract. It
-declares:
-
- TP-local self-attention geometry, AR block size, sink, and recent window;
- logical KV branches and their worker-local slots;
- fixed-length cross-attention caches;
- scratch tokens required by an uncommitted forward;
- persistent model-owned CUDA bytes per resident session; and
- the requested resident-session capacity.
-
-The runner computes:
-
-```text
-required(capacity) =
-    scratch
-  + managed_self_attention_pages(capacity)
-  + capacity * cross_attention_bytes_per_session
-  + capacity * model_owned_state_bytes_per_session
-```
-
-`gpu_memory_fraction` is a soft expansion budget:
-
-```text
-configured_budget = available_device_bytes * gpu_memory_fraction
-effective_budget = max(configured_budget, required(1))
-```
-
-If `required(1)` exceeds actual available device memory, initialization fails.
-Otherwise the runner always admits one viable session and uses the configured
-fraction to select additional resident sessions up to the model-declared cap.
-All model-owned reservations are deducted before the paged self-attention pool
-is allocated.
-
-LingBot reports its persistent full-horizon image-condition tensor. DreamZero
-reports the measured 603 MiB per-session Wan VAE causal-convolution state.
-
-## Concurrency and routing limits
-
-The current runner executes one request at a time:
-
- `max_num_seqs` must be one;
- request batching and diffusion step execution are rejected;
- one AR block is generated per engine request; and
- an AR-Diffusion stage must have exactly one replica.
-
-The single-replica restriction is intentional. Multi-replica support requires
-session-affine routing so every tick, reset, and close reaches the worker that
-owns the session. Random or round-robin routing is not correct.
-
-Multiple resident sessions may share one runner. Capacity selection accounts
-for all persistent pools, and the runner performs LRU eviction through the same
-model lifecycle hook used by explicit close.
-
-## Model boundary
-
-The generic runtime transports controls as opaque
-`ARDiffusionControlInput(track, schema, data)` values. Model adapters validate
-and reduce those values. This keeps target/velocity, key state, pose,
-intrinsics, and other model-specific conventions out of the session protocol.
-
-For LingBot World v2:
-
- `lingbot_world/actions.py` parses key-state and trajectory controls;
- `lingbot_world/camera.py` loads or constructs camera trajectories and Plücker embeddings;
- `lingbot_world/transformer.py` implements checkpoint-compatible causal attention; and
- `lingbot_world/pipeline.py` constructs conditioning, owns small non-KV session state, and
-  produces one block plus the standard metadata envelope.
-
-## Non-goals
-
-This contract does not currently provide:
-
- a public HTTP or WebSocket request schema (the structured-interaction
-  frontend is tracked separately in
-  [#5527](https://github.com/vllm-project/vllm-omni/pull/5527));
- camera/action semantics shared by every world model;
- session migration or replication across workers;
- retry of an ambiguous partially executed chunk;
- cross-stage KV transfer; or
- stateful streaming VAE decode.
-
-Those features can be layered on top without adding LingBot-specific identity
-or lifecycle fields to the generic runtime.
-
-## Required regression coverage
-
-CPU contract tests cover:
-
- separation of tick correlation IDs from internal engine routing IDs;
- event ordering, metadata equality, and reducer fail-closed commit;
- close/disconnect failure tombstones and cleanup retry;
- capacity one under the default fraction for shipped LingBot geometry;
- capacity expansion under a tuned fraction;
- DreamZero model-owned state accounting;
- runner session reuse, reset, close, LRU eviction, and forward cleanup; and
- LingBot model registration, imports, camera/action reduction, and one-block
-  output metadata.
-
-GPU validation must additionally exercise real model loading, action input,
-prompt switching, rolling-window replacement, reset, close, finite output
-latents, and exact metadata/request identity.
@@ -1,165 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" width="1280" height="850" viewBox="0 0 1280 850" role="img" aria-labelledby="title desc">
-  <title id="title">AllGather DLO sharded host weights and lockstep device reconstruction</title>
-  <desc id="desc">An external orchestrator submits one synchronized wave to a DP2 by TP2 diffusion engine. Each DP replica contains TP0 and TP1 workers with private pinned host shards and two rotating GPU weight slots. DLO AllGather groups connect matching TP coordinates across DP replicas, reconstructing full rank-local blocks on each device in lockstep. Ordinary TP groups remain within each DP replica.</desc>
-  <defs>
-    <marker id="request-arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="7" markerHeight="7" orient="auto">
-      <path d="M 0 0 L 10 5 L 0 10 z" fill="#7c3aed"/>
-    </marker>
-    <marker id="h2d-arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="7" markerHeight="7" orient="auto">
-      <path d="M 0 0 L 10 5 L 0 10 z" fill="#d97706"/>
-    </marker>
-    <marker id="collective-arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse">
-      <path d="M 0 0 L 10 5 L 0 10 z" fill="#7c3aed"/>
-    </marker>
-    <style>
-      .canvas { fill: #ffffff; stroke: #e2e8f0; }
-      .node { fill: #f8fafc; stroke: #cbd5e1; }
-      .engine { fill: #ffffff; stroke: #cbd5e1; }
-      .replica { fill: #f8fafc; stroke: #cbd5e1; }
-      .router { fill: #f5f3ff; stroke: #c4b5fd; }
-      .scope { fill: #f8fafc; stroke: #cbd5e1; }
-      .host { fill: #fff7ed; stroke: #fdba74; }
-      .shard { fill: #d97706; }
-      .gpu { fill: #f0fdf4; stroke: #86efac; }
-      .slot-current { fill: #16a34a; }
-      .slot-next { fill: #bbf7d0; stroke: #22c55e; }
-      .lockstep { fill: #f5f3ff; stroke: #c4b5fd; }
-      .title { font: 700 25px system-ui, -apple-system, Segoe UI, sans-serif; fill: #0f172a; }
-      .subtitle { font: 500 13px system-ui, -apple-system, Segoe UI, sans-serif; fill: #64748b; }
-      .section { font: 700 15px system-ui, -apple-system, Segoe UI, sans-serif; fill: #1e293b; }
-      .section-sub { font: 500 11.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #64748b; }
-      .eyebrow { font: 700 10px system-ui, -apple-system, Segoe UI, sans-serif; letter-spacing: .8px; fill: #64748b; }
-      .box-title { font: 700 12px system-ui, -apple-system, Segoe UI, sans-serif; fill: #1e293b; }
-      .box-text { font: 500 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #475569; }
-      .shard-title { font: 700 13px system-ui, -apple-system, Segoe UI, sans-serif; fill: #ffffff; }
-      .shard-text { font: 500 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #ffffff; }
-      .slot-title { font: 700 10px system-ui, -apple-system, Segoe UI, sans-serif; fill: #ffffff; }
-      .slot-text { font: 500 9px system-ui, -apple-system, Segoe UI, sans-serif; fill: #ffffff; }
-      .slot-next-title { font: 700 10px system-ui, -apple-system, Segoe UI, sans-serif; fill: #166534; }
-      .slot-next-text { font: 500 9px system-ui, -apple-system, Segoe UI, sans-serif; fill: #166534; }
-      .request-label { font: 600 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #6d28d9; }
-      .h2d-label { font: 600 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #b45309; }
-      .collective-label { font: 700 10px system-ui, -apple-system, Segoe UI, sans-serif; fill: #6d28d9; }
-      .lock-title { font: 700 12px system-ui, -apple-system, Segoe UI, sans-serif; fill: #6d28d9; }
-      .lock-text { font: 500 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #5b21b6; }
-      .request-path { fill: none; stroke: #7c3aed; stroke-width: 2; marker-end: url(#request-arrow); }
-      .h2d-path { fill: none; stroke: #d97706; stroke-width: 2; marker-end: url(#h2d-arrow); }
-      .collective-path { fill: none; stroke: #7c3aed; stroke-width: 2.5; marker-start: url(#collective-arrow); marker-end: url(#collective-arrow); }
-    </style>
-  </defs>
-
-  <rect x="12" y="12" width="1256" height="826" rx="14" class="canvas"/>
-
-  <text id="title-text" x="40" y="50" class="title">AllGather DLO: sharded host weights, lockstep reconstruction</text>
-  <text x="40" y="74" class="subtitle">DP2 &#215; TP2 shown: DLO groups span DP replicas per TP coordinate; ordinary TP stays within each replica</text>
-
-  <!-- The external control plane submits one synchronized DP wave. -->
-  <rect x="390" y="94" width="500" height="68" rx="10" class="router"/>
-  <text x="640" y="121" text-anchor="middle" class="section">External router / orchestrator</text>
-  <text x="640" y="144" text-anchor="middle" class="section-sub">submits one synchronized DP wave &#183; outside vLLM-Omni DLO</text>
-
-  <rect x="30" y="184" width="1220" height="634" rx="14" class="node"/>
-  <text x="52" y="211" class="eyebrow">ONE NODE &#183; ONE SYNCHRONIZED ENGINE, PROCESS-OWNED HOST SHARDS</text>
-
-  <path d="M 640 162 L 640 232" class="request-path"/>
-  <text x="654" y="202" class="request-label">request wave</text>
-
-  <!-- One engine contains two DP replicas that execute in lockstep. -->
-  <rect x="45" y="232" width="1190" height="262" rx="12" class="engine"/>
-  <text x="68" y="261" class="section">One synchronized diffusion engine &#183; DP2 &#215; TP2</text>
-  <text x="68" y="281" class="section-sub">two DP requests follow the same block order; DLO reuses existing DP groups at each TP coordinate</text>
-
-  <!-- DP replica 0. -->
-  <rect x="65" y="296" width="515" height="186" rx="10" class="replica"/>
-  <text x="82" y="318" class="box-title">DP replica 0 &#183; request 0</text>
-  <text x="563" y="318" text-anchor="end" class="section-sub">ordinary TP group</text>
-
-  <rect x="82" y="332" width="473" height="64" rx="8" class="gpu"/>
-  <text x="98" y="351" class="eyebrow">TP0 &#183; GPU 0</text>
-  <text x="98" y="376" class="box-text">DLO backend &#183; local shard H2D</text>
-  <rect x="250" y="340" width="132" height="44" rx="6" class="slot-current"/>
-  <text x="316" y="357" text-anchor="middle" class="slot-title">SLOT 0 &#183; CURRENT</text>
-  <text x="316" y="374" text-anchor="middle" class="slot-text">full TP-local N &#183; compute</text>
-  <rect x="397" y="340" width="142" height="44" rx="6" class="slot-next"/>
-  <text x="468" y="357" text-anchor="middle" class="slot-next-title">SLOT 1 &#183; PREFETCH</text>
-  <text x="468" y="374" text-anchor="middle" class="slot-next-text">full TP-local N+1 after AG</text>
-
-  <rect x="82" y="410" width="473" height="64" rx="8" class="gpu"/>
-  <text x="98" y="429" class="eyebrow">TP1 &#183; GPU 1</text>
-  <text x="98" y="454" class="box-text">DLO backend &#183; local shard H2D</text>
-  <rect x="250" y="418" width="132" height="44" rx="6" class="slot-current"/>
-  <text x="316" y="435" text-anchor="middle" class="slot-title">SLOT 0 &#183; CURRENT</text>
-  <text x="316" y="452" text-anchor="middle" class="slot-text">full TP-local N &#183; compute</text>
-  <rect x="397" y="418" width="142" height="44" rx="6" class="slot-next"/>
-  <text x="468" y="435" text-anchor="middle" class="slot-next-title">SLOT 1 &#183; PREFETCH</text>
-  <text x="468" y="452" text-anchor="middle" class="slot-next-text">full TP-local N+1 after AG</text>
-
-  <!-- DP replica 1. -->
-  <rect x="700" y="296" width="515" height="186" rx="10" class="replica"/>
-  <text x="717" y="318" class="box-title">DP replica 1 &#183; request 1</text>
-  <text x="1198" y="318" text-anchor="end" class="section-sub">ordinary TP group</text>
-
-  <rect x="725" y="332" width="473" height="64" rx="8" class="gpu"/>
-  <text x="741" y="351" class="eyebrow">TP0 &#183; GPU 2</text>
-  <text x="741" y="376" class="box-text">DLO backend &#183; local shard H2D</text>
-  <rect x="893" y="340" width="132" height="44" rx="6" class="slot-current"/>
-  <text x="959" y="357" text-anchor="middle" class="slot-title">SLOT 0 &#183; CURRENT</text>
-  <text x="959" y="374" text-anchor="middle" class="slot-text">full TP-local N &#183; compute</text>
-  <rect x="1040" y="340" width="142" height="44" rx="6" class="slot-next"/>
-  <text x="1111" y="357" text-anchor="middle" class="slot-next-title">SLOT 1 &#183; PREFETCH</text>
-  <text x="1111" y="374" text-anchor="middle" class="slot-next-text">full TP-local N+1 after AG</text>
-
-  <rect x="725" y="410" width="473" height="64" rx="8" class="gpu"/>
-  <text x="741" y="429" class="eyebrow">TP1 &#183; GPU 3</text>
-  <text x="741" y="454" class="box-text">DLO backend &#183; local shard H2D</text>
-  <rect x="893" y="418" width="132" height="44" rx="6" class="slot-current"/>
-  <text x="959" y="435" text-anchor="middle" class="slot-title">SLOT 0 &#183; CURRENT</text>
-  <text x="959" y="452" text-anchor="middle" class="slot-text">full TP-local N &#183; compute</text>
-  <rect x="1040" y="418" width="142" height="44" rx="6" class="slot-next"/>
-  <text x="1111" y="435" text-anchor="middle" class="slot-next-title">SLOT 1 &#183; PREFETCH</text>
-  <text x="1111" y="452" text-anchor="middle" class="slot-next-text">full TP-local N+1 after AG</text>
-
-  <!-- DLO collectives connect matching TP coordinates across DP replicas. -->
-  <text x="640" y="350" text-anchor="middle" class="collective-label">DLO AG &#183; TP0 GROUP</text>
-  <path d="M 555 364 L 725 364" class="collective-path"/>
-  <text x="640" y="428" text-anchor="middle" class="collective-label">DLO AG &#183; TP1 GROUP</text>
-  <path d="M 555 442 L 725 442" class="collective-path"/>
-
-  <!-- All participants must reach the same collective in the same order. -->
-  <rect x="365" y="516" width="550" height="48" rx="8" class="lockstep"/>
-  <circle cx="395" cy="540" r="12" fill="#7c3aed"/>
-  <text x="395" y="545" text-anchor="middle" font-family="system-ui, sans-serif" font-size="15" font-weight="700" fill="#ffffff">&#8644;</text>
-  <text x="420" y="536" class="lock-title">DLO AllGather requires lockstep block order</text>
-  <text x="420" y="553" class="lock-text">every participant enters each block collective in the same sequence</text>
-
-  <!-- Each replica copies two process-owned host shards before reconstruction. -->
-  <path d="M 455 650 C 412 625, 340 610, 340 575 L 340 540 C 340 516, 318 500, 300 482" class="h2d-path"/>
-  <text x="276" y="618" text-anchor="middle" class="h2d-label">local shard H2D &#215;2</text>
-  <path d="M 825 650 C 868 625, 940 610, 940 575 L 940 540 C 940 516, 962 500, 980 482" class="h2d-path"/>
-  <text x="1004" y="618" text-anchor="middle" class="h2d-label">local shard H2D &#215;2</text>
-
-  <!-- Persistent host shards are private to workers, not mmap-shared. -->
-  <rect x="58" y="650" width="242" height="140" rx="10" class="scope"/>
-  <text x="78" y="676" class="eyebrow">GROUP TOPOLOGY</text>
-  <text x="78" y="702" class="box-title">DP2 &#215; TP2</text>
-  <text x="78" y="725" class="box-text">DLO groups: same TP coordinate</text>
-  <text x="78" y="746" class="box-text">across DP replicas</text>
-  <text x="78" y="767" class="box-text">TP stays the model-parallel group.</text>
-
-  <rect x="330" y="650" width="620" height="140" rx="10" class="host"/>
-  <text x="350" y="676" class="eyebrow">PROCESS-OWNED PINNED HOST SHARDS &#183; NOT HOST WEIGHT CACHE MMAP</text>
-  <rect x="370" y="692" width="260" height="60" rx="8" class="shard"/>
-  <text x="500" y="716" text-anchor="middle" class="shard-title">DP replica 0 &#183; private</text>
-  <text x="500" y="738" text-anchor="middle" class="shard-text">TP0 shard 0/2 &#183; TP1 shard 0/2</text>
-  <rect x="650" y="692" width="260" height="60" rx="8" class="shard"/>
-  <text x="780" y="716" text-anchor="middle" class="shard-title">DP replica 1 &#183; private</text>
-  <text x="780" y="738" text-anchor="middle" class="shard-text">TP0 shard 1/2 &#183; TP1 shard 1/2</text>
-  <text x="640" y="773" text-anchor="middle" class="box-text">Across each DLO group, shards total approximately one rank-local runtime block copy</text>
-
-  <rect x="980" y="650" width="242" height="140" rx="10" class="scope"/>
-  <text x="1000" y="676" class="eyebrow">HOT PATH &#183; PER WORKER</text>
-  <text x="1000" y="702" class="box-title">Prefetch next block</text>
-  <text x="1000" y="725" class="box-text">local pinned shard &#8594; GPU</text>
-  <text x="1000" y="746" class="box-text">DLO AllGather &#8594; full block</text>
-  <text x="1000" y="767" class="box-text">rotate slots; overlap compute</text>
-</svg>
@@ -1,202 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" width="1280" height="850" viewBox="0 0 1280 850" role="img" aria-labelledby="title desc">
-  <title id="title">No-AllGather DLO shared host weights and independent runtime orchestration</title>
-  <desc id="desc">An external orchestrator routes requests to two independently scheduled TP2 diffusion engines on one node. Each engine contains two rank-local workers and devices, with private loaders, mmap views, DLO runtime hooks, streams, events, and rotating GPU slots. Equivalent TP0 workers across engines map one immutable TP0 entry in the host weight cache, while equivalent TP1 workers map a separate TP1 entry, without a cross-engine DLO AllGather.</desc>
-  <defs>
-    <marker id="request-arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="7" markerHeight="7" orient="auto">
-      <path d="M 0 0 L 10 5 L 0 10 z" fill="#7c3aed"/>
-    </marker>
-    <marker id="mapping-arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="7" markerHeight="7" orient="auto">
-      <path d="M 0 0 L 10 5 L 0 10 z" fill="#d97706"/>
-    </marker>
-    <marker id="runtime-arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="7" markerHeight="7" orient="auto">
-      <path d="M 0 0 L 10 5 L 0 10 z" fill="#2563eb"/>
-    </marker>
-    <style>
-      .canvas { fill: #ffffff; stroke: #e2e8f0; }
-      .node { fill: #f8fafc; stroke: #cbd5e1; }
-      .engine { fill: #ffffff; stroke: #cbd5e1; }
-      .router { fill: #f5f3ff; stroke: #c4b5fd; }
-      .scope { fill: #f8fafc; stroke: #cbd5e1; }
-      .cache { fill: #fff7ed; stroke: #fdba74; }
-      .payload { fill: #d97706; }
-      .loader { fill: #f1f5f9; stroke: #cbd5e1; }
-      .mapping { fill: #fffbeb; stroke: #f59e0b; }
-      .hook { fill: #eff6ff; stroke: #60a5fa; }
-      .gpu { fill: #f0fdf4; stroke: #86efac; }
-      .device { fill: #ffffff; stroke: #86efac; }
-      .slot-current { fill: #16a34a; }
-      .slot-next { fill: #bbf7d0; stroke: #22c55e; }
-      .no-collective { fill: #fef2f2; stroke: #fca5a5; }
-      .title { font: 700 25px system-ui, -apple-system, Segoe UI, sans-serif; fill: #0f172a; }
-      .subtitle { font: 500 13px system-ui, -apple-system, Segoe UI, sans-serif; fill: #64748b; }
-      .section { font: 700 15px system-ui, -apple-system, Segoe UI, sans-serif; fill: #1e293b; }
-      .section-sub { font: 500 11.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #64748b; }
-      .eyebrow { font: 700 10px system-ui, -apple-system, Segoe UI, sans-serif; letter-spacing: .8px; fill: #64748b; }
-      .box-title { font: 700 12px system-ui, -apple-system, Segoe UI, sans-serif; fill: #1e293b; }
-      .box-text { font: 500 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #475569; }
-      .payload-title { font: 700 14px system-ui, -apple-system, Segoe UI, sans-serif; fill: #ffffff; }
-      .payload-text { font: 500 11px system-ui, -apple-system, Segoe UI, sans-serif; fill: #ffffff; }
-      .device-title { font: 700 9px system-ui, -apple-system, Segoe UI, sans-serif; fill: #166534; }
-      .slot-text { font: 700 8px system-ui, -apple-system, Segoe UI, sans-serif; fill: #ffffff; }
-      .slot-next-text { font: 700 8px system-ui, -apple-system, Segoe UI, sans-serif; fill: #166534; }
-      .arrow-label { font: 600 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #475569; }
-      .request-label { font: 600 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #6d28d9; }
-      .mapping-label { font: 600 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #b45309; }
-      .no-title { font: 700 12px system-ui, -apple-system, Segoe UI, sans-serif; fill: #b91c1c; }
-      .no-text { font: 500 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #7f1d1d; }
-      .request-path { fill: none; stroke: #7c3aed; stroke-width: 2; marker-end: url(#request-arrow); }
-      .mapping-path { fill: none; stroke: #d97706; stroke-width: 2; marker-end: url(#mapping-arrow); }
-      .runtime-path { fill: none; stroke: #2563eb; stroke-width: 2; marker-end: url(#runtime-arrow); }
-      .startup-path { fill: none; stroke: #94a3b8; stroke-width: 1.5; stroke-dasharray: 4 3; marker-end: url(#runtime-arrow); }
-    </style>
-  </defs>
-
-  <rect x="12" y="12" width="1256" height="826" rx="14" class="canvas"/>
-
-  <text id="title-text" x="40" y="50" class="title">No-AllGather DLO: TP-coordinate host sharing, independent engines</text>
-  <text x="40" y="74" class="subtitle">TP2 shown: each engine owns two devices; equivalent workers share one host payload per TP coordinate</text>
-
-  <!-- The control plane remains outside DLO. -->
-  <rect x="390" y="94" width="500" height="68" rx="10" class="router"/>
-  <text x="640" y="121" text-anchor="middle" class="section">External router / orchestrator</text>
-  <text x="640" y="144" text-anchor="middle" class="section-sub">independent request placement &#183; outside vLLM-Omni DLO</text>
-
-  <!-- One node contains independent engine runtimes over one shared physical backing. -->
-  <rect x="30" y="184" width="1220" height="634" rx="14" class="node"/>
-  <text x="52" y="211" class="eyebrow">ONE NODE &#183; PROCESS-LOCAL ENGINES OVER SHARED PHYSICAL HOST MEMORY</text>
-
-  <!-- Independent request routing from the external orchestrator. -->
-  <path d="M 500 162 C 456 181, 400 204, 332 232" class="request-path"/>
-  <text x="416" y="194" class="request-label">request A</text>
-  <path d="M 780 162 C 824 181, 880 204, 948 232" class="request-path"/>
-  <text x="864" y="194" class="request-label">request B</text>
-
-  <!-- Independent engine A. -->
-  <rect x="45" y="232" width="575" height="250" rx="12" class="engine"/>
-  <text x="68" y="261" class="section">Independent engine A</text>
-  <text x="68" y="281" class="section-sub">TP2 &#183; two rank-local workers &#183; own request schedule and DLO state</text>
-
-  <rect x="68" y="310" width="98" height="130" rx="8" class="loader"/>
-  <text x="117" y="335" text-anchor="middle" class="eyebrow">STARTUP &#215;2</text>
-  <text x="117" y="359" text-anchor="middle" class="box-title">rank-local</text>
-  <text x="117" y="377" text-anchor="middle" class="box-title">loaders</text>
-  <text x="117" y="402" text-anchor="middle" class="box-text">final transforms</text>
-  <text x="117" y="421" text-anchor="middle" class="box-text">publish / join</text>
-
-  <rect x="180" y="310" width="112" height="130" rx="8" class="mapping"/>
-  <text x="236" y="335" text-anchor="middle" class="eyebrow">HOST VIEWS &#215;2</text>
-  <text x="236" y="359" text-anchor="middle" class="box-title">TP0 mmap</text>
-  <text x="236" y="377" text-anchor="middle" class="box-title">TP1 mmap</text>
-  <text x="236" y="402" text-anchor="middle" class="box-text">process-local</text>
-  <text x="236" y="421" text-anchor="middle" class="box-text">optional registration</text>
-
-  <rect x="306" y="310" width="120" height="130" rx="8" class="hook"/>
-  <text x="366" y="335" text-anchor="middle" class="eyebrow">RUNTIME &#215;2</text>
-  <text x="366" y="359" text-anchor="middle" class="box-title">DLO backends</text>
-  <text x="366" y="377" text-anchor="middle" class="box-title">+ block hooks</text>
-  <text x="366" y="402" text-anchor="middle" class="box-text">private streams</text>
-  <text x="366" y="421" text-anchor="middle" class="box-text">events + state</text>
-
-  <rect x="440" y="310" width="157" height="130" rx="8" class="gpu"/>
-  <text x="518" y="330" text-anchor="middle" class="eyebrow">TWO DEVICES</text>
-  <rect x="450" y="340" width="137" height="42" rx="6" class="device"/>
-  <text x="518" y="353" text-anchor="middle" class="device-title">TP0 &#183; GPU A0</text>
-  <rect x="457" y="359" width="58" height="16" rx="4" class="slot-current"/>
-  <text x="486" y="370" text-anchor="middle" class="slot-text">SLOT 0 NOW</text>
-  <rect x="521" y="359" width="58" height="16" rx="4" class="slot-next"/>
-  <text x="550" y="370" text-anchor="middle" class="slot-next-text">SLOT 1 NEXT</text>
-  <rect x="450" y="389" width="137" height="42" rx="6" class="device"/>
-  <text x="518" y="402" text-anchor="middle" class="device-title">TP1 &#183; GPU A1</text>
-  <rect x="457" y="408" width="58" height="16" rx="4" class="slot-current"/>
-  <text x="486" y="419" text-anchor="middle" class="slot-text">SLOT 0 NOW</text>
-  <rect x="521" y="408" width="58" height="16" rx="4" class="slot-next"/>
-  <text x="550" y="419" text-anchor="middle" class="slot-next-text">SLOT 1 NEXT</text>
-
-  <path d="M 166 375 L 180 375" class="startup-path"/>
-  <path d="M 292 375 L 306 375" class="runtime-path"/>
-  <path d="M 426 375 L 440 375" class="runtime-path"/>
-
-  <!-- Independent engine B. -->
-  <rect x="660" y="232" width="575" height="250" rx="12" class="engine"/>
-  <text x="683" y="261" class="section">Independent engine B</text>
-  <text x="683" y="281" class="section-sub">TP2 &#183; two rank-local workers &#183; own request schedule and DLO state</text>
-
-  <rect x="683" y="310" width="98" height="130" rx="8" class="loader"/>
-  <text x="732" y="335" text-anchor="middle" class="eyebrow">STARTUP &#215;2</text>
-  <text x="732" y="359" text-anchor="middle" class="box-title">rank-local</text>
-  <text x="732" y="377" text-anchor="middle" class="box-title">loaders</text>
-  <text x="732" y="402" text-anchor="middle" class="box-text">final transforms</text>
-  <text x="732" y="421" text-anchor="middle" class="box-text">publish / join</text>
-
-  <rect x="795" y="310" width="112" height="130" rx="8" class="mapping"/>
-  <text x="851" y="335" text-anchor="middle" class="eyebrow">HOST VIEWS &#215;2</text>
-  <text x="851" y="359" text-anchor="middle" class="box-title">TP0 mmap</text>
-  <text x="851" y="377" text-anchor="middle" class="box-title">TP1 mmap</text>
-  <text x="851" y="402" text-anchor="middle" class="box-text">process-local</text>
-  <text x="851" y="421" text-anchor="middle" class="box-text">optional registration</text>
-
-  <rect x="921" y="310" width="120" height="130" rx="8" class="hook"/>
-  <text x="981" y="335" text-anchor="middle" class="eyebrow">RUNTIME &#215;2</text>
-  <text x="981" y="359" text-anchor="middle" class="box-title">DLO backends</text>
-  <text x="981" y="377" text-anchor="middle" class="box-title">+ block hooks</text>
-  <text x="981" y="402" text-anchor="middle" class="box-text">private streams</text>
-  <text x="981" y="421" text-anchor="middle" class="box-text">events + state</text>
-
-  <rect x="1055" y="310" width="157" height="130" rx="8" class="gpu"/>
-  <text x="1133" y="330" text-anchor="middle" class="eyebrow">TWO DEVICES</text>
-  <rect x="1065" y="340" width="137" height="42" rx="6" class="device"/>
-  <text x="1133" y="353" text-anchor="middle" class="device-title">TP0 &#183; GPU B0</text>
-  <rect x="1072" y="359" width="58" height="16" rx="4" class="slot-current"/>
-  <text x="1101" y="370" text-anchor="middle" class="slot-text">SLOT 0 NOW</text>
-  <rect x="1136" y="359" width="58" height="16" rx="4" class="slot-next"/>
-  <text x="1165" y="370" text-anchor="middle" class="slot-next-text">SLOT 1 NEXT</text>
-  <rect x="1065" y="389" width="137" height="42" rx="6" class="device"/>
-  <text x="1133" y="402" text-anchor="middle" class="device-title">TP1 &#183; GPU B1</text>
-  <rect x="1072" y="408" width="58" height="16" rx="4" class="slot-current"/>
-  <text x="1101" y="419" text-anchor="middle" class="slot-text">SLOT 0 NOW</text>
-  <rect x="1136" y="408" width="58" height="16" rx="4" class="slot-next"/>
-  <text x="1165" y="419" text-anchor="middle" class="slot-next-text">SLOT 1 NEXT</text>
-
-  <path d="M 781 375 L 795 375" class="startup-path"/>
-  <path d="M 907 375 L 921 375" class="runtime-path"/>
-  <path d="M 1041 375 L 1055 375" class="runtime-path"/>
-
-  <!-- The runtimes are deliberately not coupled by a DLO collective. -->
-  <rect x="365" y="516" width="550" height="44" rx="8" class="no-collective"/>
-  <circle cx="395" cy="538" r="12" fill="#dc2626"/>
-  <path d="M 389 532 L 401 544 M 401 532 L 389 544" stroke="#ffffff" stroke-width="2.2"/>
-  <text x="420" y="534" class="no-title">No cross-engine DLO AllGather</text>
-  <text x="420" y="551" class="no-text">shared host pages do not couple request order, progress, or failure domains</text>
-  <text x="640" y="592" text-anchor="middle" class="section-sub">Ordinary TP / SP model collectives, when configured, remain inside each engine</text>
-
-  <!-- Each engine maps both TP-coordinate payloads from the shared host cache. -->
-  <path d="M 455 650 C 412 625, 340 610, 340 570 L 340 525 C 340 478, 278 454, 236 440" class="mapping-path"/>
-  <text x="286" y="617" text-anchor="middle" class="mapping-label">TP0 + TP1 mappings</text>
-  <path d="M 825 650 C 868 625, 940 610, 940 570 L 940 525 C 940 478, 893 454, 851 440" class="mapping-path"/>
-  <text x="994" y="617" text-anchor="middle" class="mapping-label">TP0 + TP1 mappings</text>
-
-  <!-- One node-local physical backing, visually beneath its consumers. -->
-  <rect x="58" y="650" width="242" height="140" rx="10" class="scope"/>
-  <text x="78" y="676" class="eyebrow">SHARING SCOPE</text>
-  <text x="78" y="702" class="box-title">TP2: two coordinate entries</text>
-  <text x="78" y="725" class="box-text">one physical copy per coordinate</text>
-  <text x="78" y="746" class="box-text">A/TP0 + B/TP0 &#8594; TP0 entry</text>
-  <text x="78" y="767" class="box-text">A/TP1 + B/TP1 &#8594; TP1 entry</text>
-
-  <rect x="330" y="650" width="620" height="140" rx="10" class="cache"/>
-  <text x="350" y="676" class="eyebrow">NODE-LOCAL IMMUTABLE HOST WEIGHT CACHE &#183; ONE COPY PER TP COORDINATE</text>
-  <rect x="370" y="692" width="260" height="60" rx="8" class="payload"/>
-  <text x="500" y="717" text-anchor="middle" class="payload-title">TP0 physical payload</text>
-  <text x="500" y="738" text-anchor="middle" class="payload-text">shared by A/TP0 + B/TP0</text>
-  <rect x="650" y="692" width="260" height="60" rx="8" class="payload"/>
-  <text x="780" y="717" text-anchor="middle" class="payload-title">TP1 physical payload</text>
-  <text x="780" y="738" text-anchor="middle" class="payload-text">shared by A/TP1 + B/TP1</text>
-  <text x="640" y="773" text-anchor="middle" class="box-text">Identity includes final layout + TP coordinate; engine and DP replica are excluded</text>
-
-  <rect x="980" y="650" width="242" height="140" rx="10" class="scope"/>
-  <text x="1000" y="676" class="eyebrow">WHAT STAYS PRIVATE</text>
-  <text x="1000" y="702" class="box-title">Per worker / process</text>
-  <text x="1000" y="725" class="box-text">mmap object and validation state</text>
-  <text x="1000" y="746" class="box-text">optional host registration</text>
-  <text x="1000" y="767" class="box-text">hooks, streams, events, GPU slots</text>
-</svg>
@@ -22,7 +22,6 @@ implementation contract; it is not, by itself, a general support claim.
 - [Async Diffusion Output](feature/async_diffusion_output.md)
 - [Async Omni Output Materialization](feature/omni_async_output_materialization.md)
 - [Automatic Prefix Caching in Omni Models](feature/prefix_caching.md)
- [Realtime AR-Diffusion Sessions](feature/realtime_ar_diffusion.md)

 ### Communication

@@ -18,8 +18,8 @@ validation_paths:
  - tests/config/**
 upstream_refs:
  - vllm.config
-last_reviewed: 2026-08-11
-last_verified_commit: e356708da4cd39992405686dacfc32b18fcccc7f
+last_reviewed: 2026-08-07
+last_verified_commit: 3d7fc3b9ba3cac88d579d4dc35b78b0b641675fc
 ---

 # vLLM-Omni configuration
@@ -39,66 +39,11 @@ The PR review proposal to capture environment-derived values in configuration
 objects at initialization is an open design question, not a current
 invariant.

-## Current structured-stage reuse behavior
-
-The current RFC #4021 implementation uses upstream vLLM config classes as the
-schema source for AR and generation stages:
-
- `OmniStageLoadConfig` inherits `vllm.config.LoadConfig`.
- `OmniStageCacheConfig` inherits `vllm.config.CacheConfig`.
- `OmniStageSchedulerConfig` inherits `vllm.config.SchedulerConfig`.
- `OmniStageParallelConfig` inherits `vllm.config.ParallelConfig`.
-
-This inheritance preserves the existing load, cache, scheduler, and parallel
-concern boundaries while reducing duplicate downstream declarations. It does
-not, by itself, make every inherited field an effective vLLM-Omni runtime
-option. Construction and runtime consumption are separate surfaces:
-
-| Surface | Current behavior |
-| --- | --- |
-| Structured schema | Upstream dataclass fields are inherited. Their defaults, default factories, and applicable Pydantic validation can participate when the structured object is constructed. |
-| Omni input ownership | Pipeline/deploy/CLI construction continues to accept only fields with an existing structured owner. Direct construction of an inherited sub-config can expose a wider upstream schema. |
-| Engine projection | For AR and generation stages, reusable fields are discovered by intersecting each upstream config dataclass with upstream `EngineArgs`. Only constructor-explicit values are emitted; inherited defaults remain deferred to terminal materialization. |
-| Terminal materialization | The engine-owning process constructs the final upstream `VllmConfig` and performs model-, platform-, rank-, port-, and backend-dependent initialization. |
-
-The generated projection maps include known upstream naming differences such
-as `cache_dtype` to `kv_cache_dtype`, `policy` to `scheduling_policy`, and
-`data_parallel_master_ip` to `data_parallel_address`. A field added to both an
-upstream concern config and `EngineArgs` therefore enters the AR/generation
-projection without requiring a second downstream allowlist. Explicit inherited
-fields that have no projection raise an error rather than being silently
-dropped.
-
-Ownership exclusions remain deliberate. Stage topology owns `scheduler_cls`,
-cache owns `disable_hybrid_kv_cache_manager`, and Omni runtime owns
-`distributed_executor_backend` and `worker_cls`; vLLM's private API-process
-fields are terminal internals. These exclusions prevent one input from being
-constructed or projected through two config concerns. Effective-engine-argument
-tests cover both the dynamic mapping and this exclusion boundary.
-
-`CompilationConfig` and `ProfilerConfig` are direct structured-stage fields:
-`stage.compilation_config` and `stage.profiler_config`. They accept mapping/YAML
-inputs and complete upstream value objects. Mapping inputs are validated and
-materialized as their concrete upstream types when the structured config is
-constructed, so downstream consumers see one resolved representation.
-Prebuilt upstream objects retain their type across the structured projection
-boundary.
-
-Quantization remains on its existing Omni-owned transport contract. This
-reuse step does not adopt upstream `QuantizationConfigArgs`, because the
-current engine-specific quantization split needs to be resolved separately.
-
-Diffusion keeps its existing engine-field ownership and projection surface.
-Shared Python inheritance must not be treated as evidence that an LLM-only
-parallel input is effective for a diffusion stage.
-
 ## Safe-change guide while deferred

 Changes should trace every affected producer to its runtime consumer and test
 the structured, legacy, CLI, and deployment paths that are actually supported.
-For inherited vLLM fields, tests should separately cover constructor behavior
-and effective `EngineArgs` projection. Do not infer a stable contract from
-this placeholder page.
+Do not infer a stable contract from this placeholder page.

 ## Promotion gate

@@ -20,19 +20,13 @@ an implementation contract has a user-facing workflow.
 | Choose serial, batched, step-wise, or streaming diffusion execution | [Execution Modes and Streaming](../user_guide/diffusion/execution_modes.md) | [Diffusion Continuous Batching](../design/feature/diffusion_continuous_batching.md), [Async Diffusion Output](../design/feature/async_diffusion_output.md) |
 | Reclaim stage memory without restarting the server | [Sleep Mode](sleep_mode.md) | Runtime lifecycle behavior is documented in the user guide |

-The [Prefill-Decode Disaggregation design reference](pd_disaggregation.md)
-describes the experimental Qwen3-Omni runtime topology and its current
-integration boundaries. It is not yet a supported launch recipe.
-
 Some runtime designs are deliberately not promoted as standalone User Guide
 features yet:

- Generic, OmniConnector-based
-  [Disaggregated Inference](../design/feature/disaggregated_inference.md) is not
-  yet a standalone User Guide feature. Its
+- [Disaggregated Inference](../design/feature/disaggregated_inference.md) and
  [OmniConnector implementations](../design/index.md#communication) describe
-  transport contracts, while practical connector configuration remains under
-  [Pipeline and deploy configurations](../configuration/stage_configs.md).
+  topology and transport contracts. Practical deployment configuration remains
+  under [Pipeline and deploy configurations](../configuration/stage_configs.md).
 - [Async Chunk](../design/feature/async_chunk.md) and
  [Async Omni Output Materialization](../design/feature/omni_async_output_materialization.md)
  are model- and pipeline-dependent. Use the selected model's deploy
@@ -1,140 +0,0 @@
-# Prefill-Decode (PD) Disaggregation (experimental)
-
-!!! warning "Experimental design reference"
-    PD disaggregation is only partially integrated on the current `main` branch.
-    The runtime contains PD detection and routing scaffolding, but there is no
-    supported deploy overlay or end-to-end validated launch recipe yet. Treat this
-    page as a description of the current design, not as production guidance.
-
-Prefill-decode disaggregation splits the Qwen3-Omni Thinker into two logical
-stages: a prefill worker that processes the prompt and produces KV cache, and a
-decode worker that imports that cache and generates tokens. The Talker and
-Code2Wav stages remain downstream; PD does not disaggregate the entire
-Qwen3-Omni pipeline.
-
-## Current design
-
-### Runtime topology
-
-```mermaid
-flowchart LR
-    client["Client request"] --> orchestrator["Omni Orchestrator"]
-    orchestrator -->|"original prompt<br/>max_tokens = 1"| prefill["Thinker prefill<br/>KV producer"]
-    prefill -->|"KV blocks"| connector[("vLLM KV connector")]
-    prefill -->|"KV routing metadata"| orchestrator
-    orchestrator -->|"original prompt<br/>remote KV parameters"| decode["Thinker decode<br/>KV consumer"]
-    connector -->|"remote KV load"| decode
-    decode -->|"generated text"| client
-```
-
-The orchestrator owns the request lifecycle, while the vLLM KV connector owns
-the bulk KV transfer. The original prompt is submitted to both Thinker stages:
-the decode worker still needs the request tokens and multimodal features to
-construct its request, while the imported KV cache prevents it from repeating
-the completed prefill computation.
-
-For the full Qwen3-Omni pipeline, splitting the Thinker adds one logical stage:
-
-| Stage | Role | Output or handoff |
-| --- | --- | --- |
-| 0 | Thinker prefill | Saves prompt KV and prompt-side multimodal output |
-| 1 | Thinker decode | Loads remote KV, generates text, and produces decode-side conditioning |
-| 2 | Talker | Converts combined Thinker conditioning into codec codes |
-| 3 | Code2Wav | Converts codec codes into audio |
-
-Text remains a final output of the decode stage, and audio remains a final
-output of Code2Wav.
-
-### Request lifecycle
-
-1. At startup, PD detection looks for one stage marked `is_prefill_only` and
-   one downstream stage marked `is_decode_only`. The decode stage must list the
-   prefill stage as an input source.
-2. The orchestrator clones the Thinker sampling parameters for prefill, forces
-   `max_tokens=1`, clears stop conditions, and sets producer-side
-   `kv_transfer_params`. This makes prefill finish by length after exporting KV.
-3. When prefill finishes, the orchestrator captures its KV routing metadata and
-   Qwen multimodal output under the original request ID.
-4. The orchestrator submits the original prompt to decode with consumer-side
-   `kv_transfer_params`. Those parameters identify the transfer, prefill engine,
-   bootstrap endpoint, and, when required by the connector, remote request.
-5. The decode engine imports the KV cache and continues generation. The normal
-   Omni stage path then carries Thinker conditioning through Talker and
-   Code2Wav.
-
-### Control and data paths
-
-PD uses two different connector layers. They solve different handoffs and are
-not interchangeable.
-
-| Path | Payload | Owner |
-| --- | --- | --- |
-| PD control path | Request ID, original prompt, `transfer_id`, bootstrap address, engine ID, and connector-specific request metadata | Omni Orchestrator |
-| PD data path | Thinker attention KV blocks | vLLM KV connector |
-| Downstream model path | Thinker hidden states and embeddings, Talker codec codes, and audio tensors | vLLM-Omni stage connectors and processors |
-
-### Detection and validation
-
-The current runtime accepts one logical prefill/decode pair per pipeline. Each
-stage may use tensor parallelism, but both sides of the pair must satisfy these
-invariants:
-
- both stages define `kv_transfer_config`
- prefill uses `kv_producer` or `kv_both`
- decode uses `kv_consumer` or `kv_both`
- both stages use the same KV connector
- `kv_buffer_device` and `kv_buffer_size` agree when set on both stages
- both stages use the same `tensor_parallel_size`
-
-These checks happen before request processing so an inconsistent PD topology
-fails during engine initialization.
-
-## What is implemented today
-
-| Area | Current state |
-| --- | --- |
-| Pair detection and topology validation | Implemented in the shared PD helper |
-| Prefill sampling preparation | Implemented for sync and async entrypoints |
-| Prefill-to-decode routing | Implemented in the current Orchestrator |
-| Deploy-based configuration | Not yet migrated; the supported `stages:` deploy schema cannot currently expand Qwen3-Omni into the four-stage PD topology |
-| Qwen3 Thinker-to-Talker handoff | Bridge state and embedding-merge helpers exist, but the merge is not wired into the live stage processor |
-| Connector compatibility | Incomplete; the Orchestrator currently requires `remote_request_id`, while the pinned Mooncake connector completes prefill without returning that metadata |
-| Validation coverage | PD entrypoint tests and the PD deploy overlay in online end-to-end tests are temporarily disabled during migration |
-| Performance | No non-regression or throughput/latency result is established for the current path |
-
-Because these pieces are still converging, the previous legacy `stage_args` YAML
-and `vllm serve` command have been removed. The serve CLI now accepts the
-deploy-based `stages:` schema, so retaining the legacy recipe would imply a
-launch path that current `main` rejects.
-
-## Configuration contract under development
-
-The intended deploy integration starts from the
-[default Qwen3-Omni deploy config](gh-file:vllm_omni/deploy/qwen3_omni_moe.yaml)
-and expands its frozen three-stage pipeline into the four stages shown above.
-A complete integration must:
-
- split the Thinker definition without duplicating the Talker or Code2Wav model
-  contracts
- remap downstream stage IDs, input sources, and stage connectors
- inject producer and consumer `kv_transfer_config` values into the two Thinker
-  stages
- allocate separate device sets for prefill and decode
- preserve both final outputs: decode text and Code2Wav audio
- provide connector-specific routing metadata without hard-coding one connector
-  contract
-
-The required number of GPUs depends on tensor-parallel sizes and whether Talker
-and Code2Wav share a device. A basic TP=1 layout uses one device for prefill, one
-for decode, and one for the two downstream audio stages.
-
-## Implementation map
-
- PD detection, validation, and sampling-parameter preparation:
-  [`PDDisaggregationMixin`](gh-file:vllm_omni/entrypoints/pd_utils.py)
- startup detection and connector metadata:
-  [`AsyncOmniEngine`](gh-file:vllm_omni/engine/async_omni_engine.py)
- request-scoped prefill-to-decode routing:
-  [`Orchestrator`](gh-file:vllm_omni/engine/orchestrator.py)
- Qwen3 downstream conditioning helpers:
-  [Qwen3-Omni stage input processor](gh-file:vllm_omni/model_executor/stage_input_processors/qwen3_omni.py)
@@ -37,7 +37,6 @@ th {
 | `Krea2Pipeline` | Krea 2 (Raw + Turbo) | `krea/Krea-2-Raw`, `krea/Krea-2-Turbo` | ✅︎ | | | | — |
 | `WanPipeline` | Wan2.1-T2V, Wan2.2-T2V, Wan2.2-TI2V | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`, `Wan-AI/Wan2.1-T2V-14B-Diffusers`, `Wan-AI/Wan2.2-T2V-A14B-Diffusers`, `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `WanImageToVideoPipeline` | Wan2.2-I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `LingBotWorldCausalDMDPipeline` | LingBot-World 2.0 v1 (experimental) | `robbyant/lingbot-world-v2-14b-causal-fast-diffusers` | Experimental | | | | — |
 | `Cosmos3OmniDiffusersPipeline` | Cosmos3 T2I, T2V, I2V, V2V, T2V with sound, action policy | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Wan22S2VPipeline` | Wan2.2-S2V | `Wan-AI/Wan2.2-S2V-14B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Wan22VACEPipeline` | Wan2.1-VACE | `Wan-AI/Wan2.1-VACE-1.3B-diffusers`, `Wan-AI/Wan2.1-VACE-14B-diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
@@ -53,7 +52,6 @@ th {
 | `LingBotVideoPipeline` | LingBot-Video dense and MoE T2I, T2V, TI2V | `robbyant/lingbot-video-dense-1.3b`, `robbyant/lingbot-video-moe-30b-a3b` | ✅︎ | | | | — |
 | `MiniMaxH3Pipeline` | MiniMax H3 T2VA, first/last-keyframe FL2VA, and mixed-reference Ref2VA | `MiniMaxAI/MiniMax-H3` | ✅︎ | ✅︎ | | | [Published](https://recipes.vllm.ai/MiniMaxAI/MiniMax-H3) |
 | `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `Pi0Pipeline` | π0 (Pi-Zero) VLA | `lerobot/pi0_base` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/lerobot/Pi0.md) |
 | `HeliosPipeline`, `HeliosPyramidPipeline` | Helios | `BestWishYsh/Helios-Base`, `BestWishYsh/Helios-Mid`, `BestWishYsh/Helios-Distilled` | ✅︎ | ✅︎ | ✅︎ | | — |
 | `MagiHumanPipeline` | MagiHuman | `SII-GAIR/daVinci-MagiHuman-Base-1080p` | ✅︎ | ✅︎ | | | — |
 | `OvisImagePipeline` | Ovis-Image | `OvisAI/Ovis-Image` | ✅︎ | ✅︎ | | ✅︎ | — |
@@ -78,7 +76,6 @@ th {
 | `SoulXSingerPipeline` | SoulX-Singer (SVS) | `Soul-AILab/SoulX-Singer` | ✅︎ | | | | — |
 | `SoulXSingerSVCPipeline` | SoulX-Singer-SVC | `Soul-AILab/SoulX-Singer` (`model-svc.pt`) | ✅︎ | | | | — |
 | `AudioXPipeline` | AudioX | `zhangj1an/AudioX` | ✅︎ | ✅︎ | | | — |
-| `MiniMaxMusic3ForConditionalGeneration` | MiniMax Music 3 (text-to-music) | `MiniMaxAI/MiniMax-Music3` | ✅︎ | | | | — |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-CustomVoice | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` | ✅︎ | | | | [Published](https://recipes.vllm.ai/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice) |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-VoiceDesign | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-Base | `Qwen/Qwen3-TTS-12Hz-1.7B-Base` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
@@ -120,10 +117,6 @@ support metadata until their recipe is audited.

 ✅︎ indicates the model is supported on that backend. Empty cells mean not listed as supported on that backend.

-See the [LingBot-World 2.0 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Robbyant/LingBot-World-2.0.md)
-for offline and experimental realtime usage, supported controls, validation,
-and current limitations.
-
 <sup>H3</sup> MiniMax H3 AMD GPU support is validated on gfx942 (MI300X) and
 gfx950 (MI350) in BF16, using the AITER `FLASH_ATTN` diffusion attention backend.
 Other AMD SKUs (e.g. MI325X) are not listed until their
@@ -141,20 +141,7 @@ The response shape depends on the streaming parameters:

 **Non-streaming (default).** With `stream=false` and no `stream_format`, returns the
 complete clip as binary audio data with an appropriate `Content-Type` header (e.g.
-`audio/wav`). Because the raw-bytes body has no JSON carrier, successful
-non-streaming responses from non-diffusion speech servers report usage through
-response headers:
-
-| Header | Description |
-| --- | --- |
-| `x-vllm-omni-input-tokens` | Total input tokens (`text_tokens` + `audio_tokens`). |
-| `x-vllm-omni-output-tokens` | Generated codec/audio tokens. |
-| `x-vllm-omni-total-tokens` | `input_tokens` + `output_tokens`. |
-| `x-vllm-omni-input-text-tokens` | Tokens from the synthesized text (`input` plus `instructions`). |
-| `x-vllm-omni-input-audio-tokens` | Reference-audio codec frames, non-zero only for in-context voice cloning. |
-
-Diffusion-mode speech servers route through a separate response path and do not
-emit these headers.
+`audio/wav`). The raw-bytes body has no JSON carrier, so no `usage` is reported.

 **Raw audio stream** (`stream_format="audio"`). Streams raw audio bytes (PCM or
 WAV) as they are decoded.
@@ -167,14 +167,6 @@ transfer options such as `control_path`, `control_guidance`,
 `num_conditional_frames`, `show_control_condition`, and `show_input`; see the
 Cosmos3 recipe for complete examples.

-HTTP redirects for `image_reference.image_url` follow vLLM's
-`VLLM_MEDIA_URL_ALLOW_REDIRECTS` setting. Before starting the server, set it to
-`1` (the default) to allow redirects or `0` to reject them. A redirect target
-can differ from the original host, and vLLM-Omni does not yet fully implement
-upstream vLLM's media URL allowlist protection. Deployments should therefore
-treat remote media URLs as untrusted and choose this setting as part of their
-URL access policy.
-
 ### Speech-to-Video

 For models that support audio-driven generation (e.g., Wan2.2-S2V), pass both
@@ -1,127 +1,153 @@
 # Distributed Layerwise Offloading

-Distributed layerwise offloading (DLO) streams DiT blocks through two device
-buffers instead of keeping the complete DiT in HBM. It supports sharded host
-weights with AllGather or independently streamed rank-local weights. The
-no-AllGather path can optionally share node-local host weights through mmap.
+Distributed layerwise offloading (DLO) extends block streaming to multi-device
+deployments. With AllGather enabled, each rank stores roughly `1 / dp_size` of
+the host weights and reconstructs each layer at runtime. Without AllGather,
+each rank streams a complete block independently. Compatible TP1 deployments
+can share checkpoint-backed host pages among processes on the same node;
+otherwise DLO streams the ordinary loader's rank-local tensors.

-See the [shared DLO component architecture](../../../design/feature/offloader/distributed_layerwise_offload.md#architecture)
-for ownership and lifecycle coordination. The mode-specific internals are
-documented separately in the
-[AllGather design](../../../design/feature/offloader/distributed_layerwise_offload.md#allgather-design)
-and
-[no-AllGather design](../../../design/feature/offloader/distributed_layerwise_offload.md#no-allgather-design).
+See the [DLO feature design](../../../design/feature/offloader/distributed_layerwise_offload.md)
+for the implementation contract and compatibility matrix.

-## Choose a mode
+## Execution model

-| Mode | Use when | Host weights | Runtime synchronization |
-| --- | --- | --- | --- |
-| DLO AllGather (default) | DP ranks execute the same block path in lockstep | About `1 / dp_size` per rank | DLO weight AllGather |
-| no-AllGather | Ranks or engines must schedule independently | Complete rank-local layout | No DLO weight collective |
-| no-AllGather + host weight cache | Equivalent independent workers share one node | Shared final mmap layout per TP coordinate | No DLO weight collective |
-| host weight cache + host registration | The platform supports registration and recurrent staging is too expensive | Shared registered mmap layout | No DLO weight collective |
+DLO overlaps three operations with a fixed two-block device buffer:

-AllGather is normally the best choice for synchronized DP. Host weight cache mode
-targets independently scheduled replicas, especially repeated TP engines on
-one node.
+```text
+Compute stream:  [Layer N]          [Layer N+1]        [Layer N+2]
+H2D stream:      [H2D shard N+1]    [H2D shard N+2]
+AllGather:       [AG N+1]           [AG N+2]
+Slots:           slot 0: Layer N    slot 1: Layer N+1
+```
+
+AllGather communicates only request-independent weight shards, so data-
+parallel ranks may process different requests concurrently.

 ## Usage

 ```bash
-# Sharded host weights with DLO AllGather
+# Four ranks with sharded host weights and AllGather
 vllm serve /path/to/model --omni \
  --enable-distributed-layerwise-offload \
  --data-parallel-size 4

-# Independently streamed rank-local weights
+# Standard-loader rank-local weights, without DLO AllGather
 vllm serve /path/to/model --omni \
  --enable-distributed-layerwise-offload \
+  --data-parallel-size 4 \
  --dlo-no-use-allgather

-# Share final host weights; prefer registered direct H2D under the pinned-memory policy
+# Sequence parallel deployment
 vllm serve /path/to/model --omni \
  --enable-distributed-layerwise-offload \
-  --dlo-no-use-allgather \
-  --dlo-use-host-weight-cache
-
-# Apply an 80 GiB per-worker safety ceiling to the same registration attempt
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --tensor-parallel-size 2 \
-  --dlo-no-use-allgather \
-  --dlo-use-host-weight-cache \
-  --dlo-host-weight-cache-pin-limit-gib 80
+  --usp 4
 ```

-## Relevant options
+```python
+from vllm_omni import Omni
+
+omni = Omni(
+    model="/path/to/model",
+    enable_distributed_layerwise_offload=True,
+    dlo_use_allgather=True,
+)
+```
+
+## Flags

 | Flag | Meaning | Default |
 | --- | --- | --- |
 | `--enable-distributed-layerwise-offload` | Enable DLO | `false` |
-| `--dlo-no-use-allgather` | Stream complete rank-local blocks independently | `false` |
-| `--dlo-use-host-weight-cache` | Use the final-layout cache and prefer registration when pinned memory is enabled | `false` |
-| `--dlo-host-weight-cache-pin-limit-gib GIB` | Optional per-worker registration ceiling; zero means no additional ceiling | `0` |
-| `--dlo-resident-layers N` | Keep eligible leading DiT blocks resident in HBM | `0` |
+| `--data-parallel-size N` | DP ranks and AllGather weight-sharding group | `1` |
+| `--dlo-use-allgather` | Shard host weights and reconstruct with AllGather | `true` |
+| `--dlo-no-use-allgather` | Stream complete rank-local blocks without a DLO weight collective | `false` |
+| `--dlo-resident-layers N` | Keep N leading main-DiT blocks on device; requires no-AllGather and model-declared resident paths | `0` |

-No-AllGather DLO automatically uses a compatible checkpoint mmap plan. When
-checkpoint mmap is unavailable, it preserves ordinary pinned loader weights by
-default. `--dlo-use-host-weight-cache` explicitly selects the transform-complete
-final-layout cache. Under the default `pin_cpu_memory=True` policy, DLO attempts
-to register the complete immutable mapping for direct H2D. Setting
-`pin_cpu_memory=False` through structured or programmatic configuration retains
-the staged mmap path instead. A positive pin limit is only an optional safety
-ceiling and does not select the cache. The cache uses
-`~/.cache/vllm-omni/dlo-host-weights`. Programmatic configuration may override
-`dlo_host_weight_cache_dir` and the writer lock timeout; these advanced storage
-controls are intentionally not separate CLI flags.
+## Host-weight loading

-## Operational notes
+The diffusion loader chooses host storage before DLO is enabled. It first
+attempts to build a complete, validated direct-checkpoint mmap plan. If names,
+coverage, shape, dtype, topology, or loader-callback compatibility cannot be
+proven, it runs the ordinary model loader instead. DLO consumes that result and
+does not make a second checkpoint-compatibility decision.

- All workers that should share must see the same local, disk-backed cache
-  directory. Do not use tmpfs or a cross-node filesystem for this version.
- Cache entries are immutable and validated before use. A cache or registration
-  failure keeps the ordinary loader weights or falls back to bounded staging.
- The bounded-staging fallback is a memory-first mode: recurrent pageable-mmap to
-  pinned-slot copies can reduce throughput, especially across concurrent TP
-  engines. Use it intentionally by disabling pinned host memory, or when the
-  active platform cannot register the immutable mapping.
- When a positive registration ceiling is configured, it must cover the
-  complete page-aligned mapping reported in the worker log. Registration is
-  all-or-nothing. Zero applies no additional ceiling.
- CUDA is the first registration backend and requires read-only host
-  registration support for immutable cache mappings. Platforms or devices
-  without that capability continue to use bounded staging; pinned-memory policy
-  requests direct H2D but does not guarantee it.
- Registration is process-local but does not duplicate the underlying file
-  pages. Each worker must still satisfy its platform and OS page-locking limits.
- Shutdown unregisters host ranges before closing their mmap handles.
- Host weight cache v1 has no automatic eviction. Stop all users of an entry
-  before deleting it.
+The shared-mmap optimization in this phase is supported only with TP1. TP
+greater than one falls back before model mutation to ordinary TP-aware loading.
+DLO may still consume those TP-local tensors, but this is a compatibility path:
+it does not share checkpoint-backed runtime weights across DP replicas and
+provides no shared-mmap host-memory guarantee.

-## Scheduling constraints
+The mmap plan skips only dedicated DiT weight sources. Other component sources,
+such as a text encoder loaded through the shared diffusion loader, continue to
+use their ordinary component loader. A checkpoint source that mixes DiT and
+non-DiT weights falls back completely rather than leaving an unplanned
+component uninitialized.

-AllGather ranks must request blocks in the same collective order. Concurrent DP
-requests may use different prompts, but they must follow the same denoising and
-block-execution path and set the same explicit `num_inference_steps`.
+With direct checkpoint mmap, the loader:

-No-AllGather workers do not have this DLO lockstep requirement. Ordinary TP or
-SP model collectives still synchronize ranks within each engine.
+1. saves non-persistent buffers such as RoPE frequencies;
+2. moves the normally created transformer to the meta device;
+3. loads checkpoint tensors as mmap views backed by the shared OS page cache;
+4. applies any loader-owned bounded layout adapters while packing blocks;
+5. restores saved non-persistent buffers; and
+6. preserves `post_load_weights()` and `validate_loaded_weights()` lifecycle
+   hooks.
+
+For AllGather with a group larger than one, each process copies only its
+persistent shard and then releases the source mapping. For no-AllGather, each
+process keeps the mapping open and packs complete blocks through two bounded
+pinned staging slots. Processes mapping the same files on one node share the
+immutable pages; no-AllGather still performs a complete-block H2D copy in each
+process.
+
+When the effective DLO group size is one, `dlo_use_allgather=True` does not
+perform a collective and uses the same rank-local transfer behavior.
+
+## Declarative topology
+
+Models may declare an `OffloadPlan` instead of embedding offload logic:
+
+```python
+from vllm_omni.diffusion.offloader import OffloadPlan
+
+
+class MyPipeline(nn.Module):
+    _dit_modules = ["transformer"]
+    _offload_plan = OffloadPlan(
+        block_attrs={"transformer": ("blocks",)},
+        offload_submodules={"context_encoder": "layers"},
+    )
+```
+
+When no plan exists, discovery falls back to
+`_layerwise_offload_blocks_attrs` and then heuristic attribute lookup.
+
+## Data-parallel concurrency
+
+With `data_parallel_size > 1` and AllGather enabled, the scheduler can process
+up to `dp_size` requests per denoising step. Every concurrent request must set
+the same explicit `num_inference_steps`; `None` is rejected because every rank
+must enter each collective.

 ## Limitations

- Direct checkpoint mmap currently requires TP1. The host weight cache can share
-  final ordinary-loader layouts between matching TP coordinates at TP1 or TP>1.
- Per-tensor online FP8 linears use the ordinary loader and can run with either
-  DLO transfer path. Other online quantization methods require no-AllGather
-  until their runtime layouts are validated.
- Host weight cache v1 rejects quantized, non-contiguous, aliased/tied,
-  device-only, HSDP, expert-parallel, CFG-parallel, and PP layouts.
- HSDP plus DLO AllGather is unsupported, and host weight cache v1 rejects HSDP.
- The host weight cache changes steady-state host backing, not startup loading:
-  each worker still performs ordinary loading and full content validation.
- Skip-load-on-hit, cache eviction, and cross-node sharing remain follow-up
-  work in
-  [RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195).
+- Direct checkpoint mmap currently requires TP1. TP greater than one is
+  outside the Phase A shared-mmap support scope and falls back before model
+  mutation to the ordinary TP-aware loader. DLO can stream that runtime layout,
+  but it provides no shared-mmap host-memory benefit or guarantee.
+- HSDP plus AllGather is rejected to avoid double sharding. HSDP without
+  AllGather has limited end-to-end validation.
+- Online quantization uses the ordinary loader with no-AllGather. It remains
+  incompatible with DLO AllGather.
+- Resident leading layers require `--dlo-no-use-allgather` and a model
+  `OffloadPlan` that declares eligible `resident_dit_paths`.
+- DP concurrency requires an explicit, identical inference-step count.
+
+Sharing transformed TP or quantized runtime layouts through a normalized mmap
+cache is a follow-up design in
+[RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195), not part of
+the direct-checkpoint path.

 See the [Cosmos3 DistOffload recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md)
 for an end-to-end example.
@@ -65,7 +65,7 @@ More failure modes (for example, network jitter and network interruption) will b

 ## Source of Conclusions

-The behaviors and conclusions above are summarized from current fault injection validation results on **`Qwen3-Omni`**, **`Wan2.2`**, and **`HunyuanImage-3.0-Instruct`** (DiT-only, `/v1/images/generations`). Automated coverage lives under `tests/dfx/reliability/` (`test_reliability_qwen3_omni.py`, `test_reliability_wan22.py`, `test_reliability_hunyuan_image.py`) and runs weekly via `.buildkite/cuda/test-weekly.yml`.
+The behaviors and conclusions above are summarized from current fault injection validation results on **`Qwen3-Omni`**, **`Wan2.2`**, **`HunyuanImage-3.0-Instruct`** (DiT-only, `/v1/images/generations`), and **`VoxCPM2`** (`/v1/audio/speech`). Automated coverage lives under `tests/dfx/reliability/` (`test_reliability_qwen3_omni.py`, `test_reliability_wan22.py`, `test_reliability_hunyuan_image.py`, `test_reliability_voxcpm2.py`) and runs weekly via `.buildkite/cuda/test-weekly.yml`.

 ### Example: `SIGTERM` fault injection log (Qwen3-Omni)

@@ -179,24 +179,6 @@ python text_to_image.py --model <your-model> --quantization fp8 --ignored-layers
 vllm serve <your-model> --omni --quantization fp8
 ```

-### Distributed layerwise offload
-
-Per-tensor online FP8 linears can be combined with DLO's default AllGather path:
-
-```bash
-vllm serve <your-model> \
-  --omni \
-  --quantization fp8 \
-  --enable-distributed-layerwise-offload
-```
-
-Direct checkpoint mmap is not used for this combination. Each rank runs the
-ordinary loader to create the final FP8 weights and scales before DLO shards
-them, so startup temporarily materializes the complete FP8 model in host memory
-on every rank. Runtime DLO host residency is sharded normally. Use
-`--dlo-no-use-allgather` when the synchronized AllGather request-wave contract
-does not fit the workload.
-
 ## Parameters

 | Parameter | Type | Default | Description |
@@ -24,6 +24,7 @@ from vllm.assets.video import VideoAsset, video_to_ndarrays
 from vllm.multimodal.media.audio import load_audio
 from vllm.utils.argparse_utils import FlexibleArgumentParser

+from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.model_executor.stage_input_processors.aura_omni import (
    DEFAULT_QWEN3_TTS_REF_TEXT,
@@ -227,6 +228,7 @@ def parse_args():
    parser.add_argument("--init-timeout", type=int, default=2000, help="Overall initialization timeout.")
    parser.add_argument("--stage-init-timeout", type=int, default=2000, help="Per-stage initialization timeout.")
    parser.add_argument("--log-stats", action="store_true", default=False, help="Enable stats logging.")
+    nullify_stage_engine_defaults(parser)
    return parser.parse_args()


@@ -1,288 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Generate a LingBot-World v2 video from an image and camera trajectory.
-
-The official checkpoint is licensed separately under CC BY-NC-SA and is
-restricted to non-commercial use. This vLLM-Omni integration remains
-Apache-2.0 licensed.
-
-Example:
-    python examples/offline_inference/diffusion/lingbot_world_v2.py \
-        --prompt "The camera moves slowly forward through the scene." \
-        --image /path/to/first_frame.png \
-        --action-dir /path/to/actions/forward \
-        --output lingbot_world_v2.mp4
-
-``--action-dir`` must contain ``poses.npy`` and ``intrinsics.npy``. The
-example configures its parent as the trusted action root and sends the
-contained directory through ``extra_args["action_path"]``.
-"""
-
-from __future__ import annotations
-
-import argparse
-import math
-from collections.abc import Sequence
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-
-_MODEL = "robbyant/lingbot-world-v2-14b-causal-fast-diffusers"
-_NUM_INFERENCE_STEPS = 4
-_MAX_SEQUENCE_LENGTH = 512
-_MAX_PIXEL_AREA = 480 * 832
-_MAX_RAW_FRAMES = 117
-_MAX_SOURCE_CAMERA_FRAMES = 4096
-_TEMPORAL_COMPRESSION = 4
-_LATENT_FRAMES_PER_BLOCK = 3
-_SPATIAL_ALIGNMENT = 16
-
-
-@dataclass(frozen=True)
-class LingBotPaths:
-    image: Path
-    action_dir: Path
-    action_root: Path
-    action_relative: Path
-    camera_frames: int
-    output: Path
-
-
-def _validate_camera_arrays(action_dir: Path) -> int:
-    """Validate the bounded camera contract before starting the engine."""
-
-    try:
-        poses = np.load(action_dir / "poses.npy", mmap_mode="r", allow_pickle=False)
-        intrinsics = np.load(action_dir / "intrinsics.npy", mmap_mode="r", allow_pickle=False)
-    except (OSError, TypeError, ValueError):
-        raise ValueError("Camera files must be valid numeric poses.npy and intrinsics.npy arrays.") from None
-    if poses.ndim != 3 or poses.shape[1:] != (4, 4):
-        raise ValueError("poses.npy must have shape [frames, 4, 4].")
-    if intrinsics.ndim != 2 or intrinsics.shape[1:] != (4,):
-        raise ValueError("intrinsics.npy must have shape [frames, 4].")
-    if poses.shape[0] != intrinsics.shape[0]:
-        raise ValueError("poses.npy and intrinsics.npy must contain the same frame count.")
-    if not 1 <= poses.shape[0] <= _MAX_SOURCE_CAMERA_FRAMES:
-        raise ValueError("Camera arrays must contain between 1 and 4096 source frames.")
-    if poses.dtype.kind not in "fiu" or intrinsics.dtype.kind not in "fiu":
-        raise ValueError("Camera arrays must contain real numeric values.")
-    if not np.isfinite(poses).all() or not np.isfinite(intrinsics).all():
-        raise ValueError("Camera arrays must contain only finite values.")
-    return int(poses.shape[0])
-
-
-def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
-    parser = argparse.ArgumentParser(
-        description="Generate a LingBot-World v2 video from a first-frame image and camera trajectory."
-    )
-    parser.add_argument("--model", default=_MODEL, help="Official Hugging Face model ID or local checkpoint path.")
-    parser.add_argument("--prompt", required=True, help="Text description of the scene and desired motion.")
-    parser.add_argument("--image", required=True, help="Path to the first-frame image.")
-    parser.add_argument(
-        "--action-dir",
-        required=True,
-        help="Camera action directory containing poses.npy and intrinsics.npy.",
-    )
-    parser.add_argument("--height", type=int, default=480, help="Output height; must be divisible by 16.")
-    parser.add_argument("--width", type=int, default=832, help="Output width; must be divisible by 16.")
-    parser.add_argument(
-        "--num-frames",
-        type=int,
-        default=9,
-        help="Raw output frame count. Nine frames are one three-latent-frame causal block.",
-    )
-    parser.add_argument("--seed", type=int, default=42, help="Random seed.")
-    parser.add_argument(
-        "--tensor-parallel-size",
-        type=int,
-        default=1,
-        help="Number of GPUs used for tensor parallelism inside the DiT.",
-    )
-    parser.add_argument("--flow-shift", type=float, default=5.0, help="Positive FlowUniPC scheduler shift.")
-    parser.add_argument("--fps", type=int, default=16, help="Frames per second in the exported MP4.")
-    parser.add_argument("--output", default="lingbot_world_v2.mp4", help="Output MP4 path.")
-    parser.add_argument(
-        "--enforce-eager",
-        action="store_true",
-        help="Disable compilation and force eager execution.",
-    )
-    return parser.parse_args(argv)
-
-
-def resolve_cli_paths(args: argparse.Namespace) -> LingBotPaths:
-    """Canonicalize local inputs and construct a safe trusted-root mapping."""
-
-    image = Path(args.image).expanduser().resolve()
-    if not image.is_file():
-        raise ValueError("--image must point to an existing file.")
-
-    action_dir = Path(args.action_dir).expanduser().resolve()
-    if not action_dir.is_dir():
-        raise ValueError("--action-dir must point to an existing directory.")
-    if not all((action_dir / name).is_file() for name in ("poses.npy", "intrinsics.npy")):
-        raise ValueError("--action-dir must contain poses.npy and intrinsics.npy.")
-    camera_frames = _validate_camera_arrays(action_dir)
-
-    action_root = action_dir.parent.resolve()
-    try:
-        action_relative = action_dir.relative_to(action_root)
-    except ValueError:
-        raise ValueError("--action-dir must be contained by its canonical trusted action root.") from None
-    if not action_relative.parts:
-        raise ValueError("--action-dir must name a directory contained by a trusted action root.")
-
-    output = Path(args.output).expanduser().resolve()
-    return LingBotPaths(
-        image=image,
-        action_dir=action_dir,
-        action_root=action_root,
-        action_relative=action_relative,
-        camera_frames=camera_frames,
-        output=output,
-    )
-
-
-def _positive_finite(value: float, flag: str) -> float:
-    result = float(value)
-    if not math.isfinite(result) or result <= 0:
-        raise ValueError(f"{flag} must be a positive finite number.")
-    return result
-
-
-def build_omni_kwargs(
-    args: argparse.Namespace,
-    paths: LingBotPaths,
-    *,
-    parallel_config: Any,
-) -> dict[str, Any]:
-    """Build Omni configuration without overriding checkpoint class discovery."""
-
-    if args.tensor_parallel_size <= 0:
-        raise ValueError("--tensor-parallel-size must be a positive integer.")
-    flow_shift = _positive_finite(args.flow_shift, "--flow-shift")
-    model_path = Path(args.model).expanduser()
-    model = str(model_path.resolve()) if model_path.exists() else args.model
-    return {
-        "model": model,
-        "flow_shift": flow_shift,
-        "parallel_config": parallel_config,
-        "enforce_eager": args.enforce_eager,
-        "model_config": {"lingbot_action_root": str(paths.action_root)},
-    }
-
-
-def build_request(
-    args: argparse.Namespace,
-    paths: LingBotPaths,
-) -> tuple[dict[str, Any], dict[str, Any]]:
-    """Build the prompt and fixed v1 causal-DMD sampling contract."""
-
-    prompt_text = args.prompt.strip()
-    if not prompt_text:
-        raise ValueError("--prompt must contain non-whitespace text.")
-    if args.height <= 0 or args.width <= 0:
-        raise ValueError("--height and --width must be positive integers.")
-    if args.height % _SPATIAL_ALIGNMENT or args.width % _SPATIAL_ALIGNMENT:
-        raise ValueError("--height and --width must each be divisible by 16.")
-    if args.height * args.width > _MAX_PIXEL_AREA:
-        raise ValueError("--height * --width must not exceed 480 * 832 pixels.")
-    if args.num_frames <= 0:
-        raise ValueError("--num-frames must be a positive integer.")
-    if args.num_frames > _MAX_RAW_FRAMES:
-        raise ValueError("--num-frames must not exceed 117 raw frames.")
-    if args.num_frames > paths.camera_frames:
-        raise ValueError("Camera arrays must contain at least --num-frames frames.")
-    if (args.num_frames - 1) % _TEMPORAL_COMPRESSION:
-        raise ValueError("--num-frames must satisfy (num_frames - 1) divisible by 4.")
-    latent_frames = (args.num_frames - 1) // _TEMPORAL_COMPRESSION + 1
-    if latent_frames % _LATENT_FRAMES_PER_BLOCK:
-        raise ValueError("--num-frames must map to whole three-frame latent blocks.")
-    if args.fps <= 0:
-        raise ValueError("--fps must be a positive integer.")
-    flow_shift = _positive_finite(args.flow_shift, "--flow-shift")
-
-    prompt = {
-        "prompt": prompt_text,
-        "multi_modal_data": {"image": str(paths.image)},
-    }
-    sampling_kwargs = {
-        "height": args.height,
-        "width": args.width,
-        "num_frames": args.num_frames,
-        "num_inference_steps": _NUM_INFERENCE_STEPS,
-        "max_sequence_length": _MAX_SEQUENCE_LENGTH,
-        "seed": args.seed,
-        "fps": args.fps,
-        "extra_args": {
-            "action_path": paths.action_relative.as_posix(),
-            "flow_shift": flow_shift,
-        },
-    }
-    return prompt, sampling_kwargs
-
-
-def extract_video_array(outputs: Any) -> np.ndarray:
-    """Unwrap one standard Omni diffusion video into ``[frames, H, W, C]``."""
-
-    if not isinstance(outputs, list) or len(outputs) != 1:
-        raise ValueError("Expected exactly one Omni output for the single LingBot request.")
-    output = outputs[0]
-    request_output = getattr(output, "request_output", None)
-    if request_output is not None and getattr(request_output, "images", None):
-        output = request_output
-
-    images = getattr(output, "images", None)
-    if not images:
-        multimodal_output = getattr(output, "multimodal_output", None) or {}
-        images = multimodal_output.get("video")
-    if images is None:
-        raise ValueError("No video frames found in Omni output.")
-    if isinstance(images, list) and len(images) == 1:
-        images = images[0]
-
-    video: np.ndarray = np.asarray(images)
-    if video.ndim == 5:
-        if video.shape[0] != 1:
-            raise ValueError(f"Expected one generated video, got batch shape {video.shape}.")
-        video = video[0]
-    if video.ndim != 4 or video.shape[-1] not in (3, 4):
-        raise ValueError(f"Expected video shape [frames, height, width, channels], got {video.shape}.")
-    return video
-
-
-def main(argv: Sequence[str] | None = None) -> Path:
-    args = parse_args(argv)
-    paths = resolve_cli_paths(args)
-    prompt, sampling_kwargs = build_request(args, paths)
-    paths.output.parent.mkdir(parents=True, exist_ok=True)
-
-    # Keep vLLM imports below pure CLI/path validation so ``--help`` and the
-    # helper tests work on machines without the native vLLM runtime.
-    from diffusers.utils import export_to_video
-
-    from vllm_omni.diffusion.data import DiffusionParallelConfig
-    from vllm_omni.entrypoints.omni import Omni
-    from vllm_omni.inputs.data import OmniDiffusionSamplingParams
-
-    parallel_config = DiffusionParallelConfig(tensor_parallel_size=args.tensor_parallel_size)
-    omni_kwargs = build_omni_kwargs(args, paths, parallel_config=parallel_config)
-    omni = Omni(**omni_kwargs)
-    try:
-        outputs = omni.generate(
-            prompt,
-            OmniDiffusionSamplingParams(**sampling_kwargs),
-            use_tqdm=False,
-        )
-    finally:
-        omni.close()
-
-    frames = extract_video_array(outputs)
-    export_to_video(frames, str(paths.output), fps=args.fps)
-    print(f"Saved {frames.shape[0]} frames to {paths.output}")
-    return paths.output
-
-
-if __name__ == "__main__":
-    main()
@@ -1,239 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Run LingBot-World 2.0 as an in-process realtime AR-Diffusion session.
-
-This example exercises the internal session API. It is not a public HTTP or
-WebSocket protocol. Each JSONL event advances the world by one three-latent
-frame AR block and writes the latent plus its identity metadata. The current
-scoped implementation supports at most ten ticks per generation epoch because
-its image condition is bounded to 117 pixel frames; reset or create a session
-to start a new world.
-"""
-
-from __future__ import annotations
-
-import argparse
-import asyncio
-import json
-import math
-import time
-from collections.abc import Sequence
-from pathlib import Path
-from typing import Any
-
-_MODEL = "robbyant/lingbot-world-v2-14b-causal-fast-diffusers"
-_CAMERA_ACTION_SCHEMA = "lingbot.camera_actions.v1"
-_FRAMES_PER_BLOCK = 3
-# ((117 pixel frames - 1) / VAE temporal factor 4 + 1) / 3 latent frames.
-_MAX_REALTIME_TICKS = 10
-
-
-def _camera_event_data(frames: list[list[str]]) -> dict[str, Any]:
-    """Build the event-side script consumed by LingBotCameraControlReducer."""
-    return {"mode": "script", "frames": frames}
-
-
-def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Run an in-process realtime LingBot-World 2.0 session.")
-    parser.add_argument("--model", default=_MODEL, help="Hugging Face model ID or local checkpoint path.")
-    parser.add_argument("--image", required=True, help="Initial RGB image.")
-    parser.add_argument("--prompt", required=True, help="Initial scene prompt.")
-    parser.add_argument(
-        "--events",
-        required=True,
-        help="JSONL file with one prompt/action event per AR block; at most 10 events.",
-    )
-    parser.add_argument("--output-dir", required=True, help="Directory for chunk latents and metadata.")
-    parser.add_argument("--session-id", default="lingbot-world", help="Persistent world session identifier.")
-    parser.add_argument("--height", type=int, default=480)
-    parser.add_argument("--width", type=int, default=832)
-    parser.add_argument("--seed", type=int, default=42)
-    parser.add_argument("--gpu-memory-fraction", type=float, default=0.1)
-    parser.add_argument("--tensor-parallel-size", type=int, default=1)
-    parser.add_argument("--enforce-eager", action="store_true")
-    return parser.parse_args(argv)
-
-
-def _load_events(path: Path) -> list[dict[str, Any]]:
-    events: list[dict[str, Any]] = []
-    for line_number, raw_line in enumerate(path.read_text().splitlines(), start=1):
-        line = raw_line.strip()
-        if not line:
-            continue
-        try:
-            value = json.loads(line)
-        except json.JSONDecodeError as exc:
-            raise ValueError(f"events line {line_number} is not valid JSON: {exc.msg}.") from None
-        if not isinstance(value, dict):
-            raise ValueError(f"events line {line_number} must be a JSON object.")
-        event_id = value.get("event_id")
-        if isinstance(event_id, bool) or not isinstance(event_id, int) or event_id < 0:
-            raise ValueError(f"events line {line_number} requires a non-negative integer event_id.")
-        prompt = value.get("prompt")
-        if prompt is not None and (not isinstance(prompt, str) or not prompt.strip()):
-            raise ValueError(f"events line {line_number} prompt must be a non-empty string.")
-        frames = value.get("frames")
-        if frames is not None:
-            if not isinstance(frames, list) or len(frames) != _FRAMES_PER_BLOCK:
-                raise ValueError(f"events line {line_number} frames must contain exactly three lists.")
-            for frame in frames:
-                if not isinstance(frame, list) or any(not isinstance(action, str) for action in frame):
-                    raise ValueError(f"events line {line_number} frames must contain only action strings.")
-        if prompt is None and frames is None:
-            raise ValueError(f"events line {line_number} must update prompt and/or frames.")
-        events.append({"event_id": event_id, "prompt": prompt, "frames": frames})
-    if not events:
-        raise ValueError("events file must contain at least one event.")
-    if len(events) > _MAX_REALTIME_TICKS:
-        raise ValueError(
-            "events file must contain at most "
-            f"{_MAX_REALTIME_TICKS} events because the current LingBot realtime "
-            "image-condition horizon is 117 pixel frames."
-        )
-    event_ids = [event["event_id"] for event in events]
-    if event_ids != sorted(set(event_ids)):
-        raise ValueError("event_id values must be unique and strictly increasing.")
-    return events
-
-
-def _validate_args(args: argparse.Namespace) -> tuple[Path, Path, Path]:
-    image = Path(args.image).expanduser().resolve()
-    events = Path(args.events).expanduser().resolve()
-    output_dir = Path(args.output_dir).expanduser().resolve()
-    if not image.is_file():
-        raise ValueError("--image must point to an existing file.")
-    if not events.is_file():
-        raise ValueError("--events must point to an existing JSONL file.")
-    if not args.prompt.strip():
-        raise ValueError("--prompt must contain non-whitespace text.")
-    if args.height <= 0 or args.width <= 0 or args.height % 16 or args.width % 16:
-        raise ValueError("--height and --width must be positive multiples of 16.")
-    if args.tensor_parallel_size <= 0:
-        raise ValueError("--tensor-parallel-size must be positive.")
-    if not math.isfinite(args.gpu_memory_fraction) or not 0 < args.gpu_memory_fraction <= 1:
-        raise ValueError("--gpu-memory-fraction must be in (0, 1].")
-    return image, events, output_dir
-
-
-async def run(argv: Sequence[str] | None = None) -> Path:
-    args = parse_args(argv)
-    image, events_path, output_dir = _validate_args(args)
-    events = _load_events(events_path)
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    # Imports stay below pure input validation so --help and helper tests do not
-    # require a CUDA-enabled vLLM installation.
-    import torch
-
-    from vllm_omni.diffusion.data import DiffusionParallelConfig
-    from vllm_omni.diffusion.models.lingbot_world.actions import LingBotCameraControlReducer
-    from vllm_omni.entrypoints.async_omni import AsyncOmni
-    from vllm_omni.experimental.ar_diffusion.consumer import ARDiffusionOmniTickConsumer
-    from vllm_omni.experimental.ar_diffusion.session import (
-        ARDiffusionSessionEvent,
-        ARDiffusionSessionManager,
-        ARDiffusionWorkerLifecycle,
-    )
-    from vllm_omni.experimental.ar_diffusion.tick_protocol import ARDiffusionControlInput
-    from vllm_omni.inputs.data import OmniDiffusionSamplingParams
-
-    engine = AsyncOmni(
-        model=args.model,
-        engine_backend="vllm_omni.experimental.ar_diffusion.engine.ARDiffusionEngine",
-        enforce_eager=args.enforce_eager,
-        parallel_config=DiffusionParallelConfig(tensor_parallel_size=args.tensor_parallel_size),
-        max_num_seqs=1,
-        model_config={
-            "ar_diffusion_height": args.height,
-            "ar_diffusion_width": args.width,
-            "ar_diffusion_kv_config": {
-                "gpu_memory_fraction": args.gpu_memory_fraction,
-                "warmup_cudagraph": True,
-            },
-        },
-    )
-    sampling = OmniDiffusionSamplingParams(
-        height=args.height,
-        width=args.width,
-        num_frames=9,
-        num_inference_steps=4,
-        max_sequence_length=512,
-        seed=args.seed,
-        output_type="latent",
-        extra_args={"flow_shift": 5.0},
-    )
-    consumer = ARDiffusionOmniTickConsumer(
-        engine,
-        prompt_provider=lambda tick: {
-            "prompt": tick.prompt,
-            "multi_modal_data": {"image": str(image)},
-        },
-        sampling_params_list=[sampling],
-        diffusion_stage_id=0,
-    )
-    manager = ARDiffusionSessionManager(
-        tick_consumer=consumer,
-        lifecycle=ARDiffusionWorkerLifecycle(engine, stage_ids=[0], timeout=180.0),
-        max_pending_events=32,
-        control_reducer_factory=LingBotCameraControlReducer,
-    )
-    session = await manager.create_session(args.session_id)
-    measurements: list[dict[str, Any]] = []
-    try:
-        for chunk_index, event in enumerate(events):
-            controls = ()
-            if event["frames"] is not None:
-                controls = (
-                    ARDiffusionControlInput(
-                        track="camera",
-                        schema=_CAMERA_ACTION_SCHEMA,
-                        data=_camera_event_data(event["frames"]),
-                    ),
-                )
-            await session.accept_event(
-                ARDiffusionSessionEvent(
-                    event_id=event["event_id"],
-                    prompt=event["prompt"]
-                    if event["prompt"] is not None
-                    else (args.prompt if chunk_index == 0 else None),
-                    controls=controls,
-                )
-            )
-            started = time.perf_counter()
-            output = await session.next_chunk()
-            elapsed = time.perf_counter() - started
-            if len(output.images) != 1 or not isinstance(output.images[0], torch.Tensor):
-                raise RuntimeError("Expected one latent tensor from each realtime LingBot chunk.")
-            latent = output.images[0].detach().float().cpu()
-            metadata = consumer.chunk_metadata(output).to_dict()
-            torch.save(latent, output_dir / f"chunk_{chunk_index:03d}.pt")
-            (output_dir / f"chunk_{chunk_index:03d}.json").write_text(
-                json.dumps(metadata, indent=2, sort_keys=True) + "\n"
-            )
-            measurements.append(
-                {
-                    "chunk_index": chunk_index,
-                    "latency_seconds": elapsed,
-                    "shape": list(latent.shape),
-                    "finite": bool(torch.isfinite(latent).all()),
-                    "metadata": metadata,
-                }
-            )
-            print(json.dumps(measurements[-1], sort_keys=True), flush=True)
-    finally:
-        try:
-            await manager.close_session(args.session_id)
-        finally:
-            engine.shutdown()
-
-    summary_path = output_dir / "summary.json"
-    summary_path.write_text(json.dumps({"chunks": measurements}, indent=2, sort_keys=True) + "\n")
-    return summary_path
-
-
-def main(argv: Sequence[str] | None = None) -> Path:
-    return asyncio.run(run(argv))
-
-
-if __name__ == "__main__":
-    main()
@@ -82,6 +82,9 @@ def parse_args():
    # ``max_num_seqs=1`` …) are passed as flat kwargs to ``Omni`` below
    # and ``create_default_diffusion`` materializes the stage config.

+    from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
+
+    nullify_stage_engine_defaults(p)
    return p.parse_args()


@@ -67,7 +67,7 @@ def main() -> None:
    parser.add_argument("--output-dir", default="results/nemotron_voicechat")
    parser.add_argument("--system-prompt", default=DEFAULT_SYSTEM_PROMPT)
    parser.add_argument(
-        "--deploy-config",
+        "--stage-configs-path",
        default=None,
        help=(
            "Deploy yaml override. Pass vllm_omni/deploy/nemotron_labs_voicechat_streaming.yaml "
@@ -134,8 +134,8 @@ def main() -> None:
    ]

    omni_kwargs: dict = {}
-    if args.deploy_config:
-        omni_kwargs["deploy_config"] = args.deploy_config
+    if args.stage_configs_path:
+        omni_kwargs["stage_configs_path"] = args.stage_configs_path
    omni = Omni(model=str(ckpt), trust_remote_code=True, **omni_kwargs)
    outputs = omni.generate(inputs, sampling_params_list)

@@ -38,6 +38,17 @@ from vllm_omni.utils.tracking_parser import TrackingArgumentParser
 logger = logging.getLogger(__name__)

 DEFAULT_MODEL = "fishaudio/s2-pro"
+DEFAULT_DEPLOY_CONFIG = os.path.join(
+    os.path.dirname(__file__),
+    "..",
+    "..",
+    "..",
+    "..",
+    "vllm_omni",
+    "model_executor",
+    "stage_configs",
+    "fish_speech_s2_pro.yaml",
+)


 def build_prompt(
@@ -141,8 +152,7 @@ def _build_omni_kwargs(args, model_name: str) -> dict:
        log_stats=args.log_stats,
        stage_init_timeout=args.stage_init_timeout,
    )
-    if args.deploy_config:
-        omni_kwargs["deploy_config"] = args.deploy_config
+    omni_kwargs["deploy_config"] = args.deploy_config or DEFAULT_DEPLOY_CONFIG
    return omni_kwargs


@@ -32,10 +32,11 @@ from end2end import _estimate_prompt_len
 from vllm.utils.argparse_utils import FlexibleArgumentParser

 from vllm_omni import Omni
+from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 from vllm_omni.utils.forced_aligner import align, build_forced_aligner_config


-def _default_deploy_config() -> str:
+def _default_stage_config() -> str:
    repo_root = Path(__file__).resolve().parents[4]
    return str(repo_root / "vllm_omni" / "deploy" / "qwen3_tts.yaml")

@@ -108,7 +109,7 @@ def main(args: Any) -> None:
        "output_dir",
    ):
        omni_kwargs.pop(key, None)
-    omni_kwargs["deploy_config"] = args.deploy_config or _default_deploy_config()
+    omni_kwargs["stage_configs_path"] = args.stage_configs_path or _default_stage_config()
    omni_kwargs["log_stats"] = args.log_stats
    omni = Omni(**omni_kwargs)

@@ -155,7 +156,7 @@ def parse_args() -> Any:
        default=None,
        help="Optional YAML file for forced aligner settings (incl. gpu_memory_utilization)",
    )
-    parser.add_argument("--deploy-config", default=None, help="Qwen3-TTS deploy YAML")
+    parser.add_argument("--stage-configs-path", default=None, help="Qwen3-TTS deploy YAML")
    parser.add_argument("--text", default="Hello world.", help="Text to synthesize and align")
    parser.add_argument("--language", default="English", help="Qwen3-TTS language field")
    parser.add_argument("--speaker", default="Vivian", help="CustomVoice speaker name")
@@ -163,6 +164,7 @@ def parse_args() -> Any:
    parser.add_argument("--max-new-tokens", type=int, default=2048, help="TTS max_new_tokens")
    parser.add_argument("--output-dir", default="output_audio", help="Directory for WAV and JSON sidecar")
    parser.add_argument("--log-stats", action="store_true", default=False, help="Enable vLLM-Omni stats logging")
+    nullify_stage_engine_defaults(parser)
    return parser.parse_args()


@@ -1,104 +0,0 @@
-# π0 (Pi-Zero) VLA — OpenPI realtime serving
-
-[π0](https://www.physicalintelligence.company/blog/pi0) is a Vision-Language-Action
-model from Physical Intelligence: multi-camera images + a language instruction +
-robot proprioceptive state → a continuous action chunk via flow-matching denoising
-(it does **not** emit text tokens). This example serves π0 over the OpenPI realtime
-websocket protocol at `/v1/realtime/robot/openpi`.
-
-## Install extras
-
-The core `pip install -e .` does not include the OpenPI client used here:
-
- `openpi-client` (from the [openpi](https://github.com/Physical-Intelligence/openpi)
-  repo: `pip install -e packages/openpi-client`), `websockets`, `msgpack`, `msgpack-numpy`
-
-## Weights
-
-π0 uses the LeRobot `lerobot/pi0_base` checkpoint (HF, ~13 GB). Either let the server
-download it (`MODEL=lerobot/pi0_base`) or point at a local copy
-(`MODEL=/path/to/pi0_base`).
-
-## Run the server
-
-```bash
-vllm serve lerobot/pi0_base --omni --port 8000 \
-    --served-model-name pi0 \
-    --deploy-config vllm_omni/deploy/pi0.yaml \
-    --enforce-eager --disable-log-stats
-```
-
-The deploy config (`vllm_omni/deploy/pi0.yaml`) declares a single diffusion stage
-(`Pi0Pipeline`), float32, `max_num_seqs: 1`, and the `policy_server_config` handshake
-metadata (3 cameras, `joint_position`, action horizon 50, action dim 32).
-
-### Commonly adjusted `model_config` keys
-
-These live under `stages[0].model_config` in `vllm_omni/deploy/pi0.yaml`:
-
-| Key | Default | Meaning |
-|---|---|---|
-| `chunk_size` | `50` | Action-chunk length (timesteps) the model predicts per inference. |
-| `num_inference_steps` | `10` | Flow-matching Euler denoising steps. |
-| `max_action_dim` | `32` | Action dimensionality (state/action are padded to this). |
-| `max_state_dim` | `32` | Proprioceptive-state dimensionality (zero-padded to this). |
-| `image_resolution` | `[224, 224]` | Per-camera input size (square; SigLIP). |
-| `tokenizer_max_length` | `48` | Max PaliGemma prompt tokens. |
-| `max_cameras` | `3` | Camera slots the model attends to (real + `-1`-padded). |
-| `image_feature_keys` | 3 `observation.images.*` keys | Camera order the model attends to. |
-| `image_key_map` | `{}` | Map raw obs camera keys → `image_feature_keys` (empty = verbatim). |
-
-The `policy_server_config` block below them is the OpenPI handshake metadata
-advertised to the client; keep its `action_horizon` / `action_dim` /
-`image_resolution` in sync with the `model_config` values above. The e2e test
-`tests/e2e/online_serving/test_pi0_expansion.py::test_pi0_openpi_online` connects to a
-live server and asserts the advertised metadata matches these `pi0.yaml` values.
-
-## Run the client
-
-```bash
-python examples/online_serving/pi0/openpi_client.py --host 127.0.0.1 --port 8000 \
-    --prompt "pick up the red block and place it in the bin"
-```
-
-It connects, prints the server metadata, sends robot observations, and prints the
-returned `[action_horizon, action_dim] = [50, 32]` action chunks. Replace the blank
-cameras / zero state in `_make_dummy_obs` with real frames (HWC uint8) and
-proprioceptive state to drive a robot.
-
-### Observation format
-
-The client sends a flat dict per inference:
-
-```python
-{
-    "observation.images.base_0_rgb":       np.uint8[H, W, 3],
-    "observation.images.left_wrist_0_rgb": np.uint8[H, W, 3],
-    "observation.images.right_wrist_0_rgb":np.uint8[H, W, 3],
-    "state":   np.float32[state_dim],   # zero-padded to max_state_dim=32 server-side
-    "prompt":  "pick up the red block",
-    "session_id": "<uuid>",             # accepted but ignored (π0 is stateless)
-}
-```
-
-Camera keys must match the server's `image_feature_keys` (the checkpoint's
-`input_features` order). If your robot uses different camera names, set
-`model_config.image_key_map` in `pi0.yaml` to map raw obs keys → feature keys.
-
-## Correctness
-
-π0's flow-matching kernel is bit-for-bit matched to the LeRobot `PI0Policy` reference
-(`max|Δ| = 7.15e-07`, CPU/float32, fixed noise; see
-`tests/diffusion/models/pi0/test_pi0_parity.py::test_pi0_vllm_omni_vs_lerobot`) and is
-version-stable across transformers releases (the version-stability checks in
-`tests/diffusion/models/pi0/test_pi0_units.py`). An OpenPI websocket online-serving e2e
-lives in `tests/e2e/online_serving/test_pi0_expansion.py::test_pi0_openpi_online`.
-
-## Limitations
-
- **Normalization stats**: `lerobot/pi0_base` uses identity normalization, which is
-  fully supported (state passes through, actions are returned in the model's space).
-  Per-dataset `norm_stats` declared in a checkpoint's `config.json` are honored, but
-  stats stored only in LeRobot's `policy_preprocessor.json` companion safetensors are
-  not yet loaded — a fine-tuned checkpoint that relies on those would need that bridge
-  added before its actions are in real-world units.
@@ -1,119 +0,0 @@
-#!/usr/bin/env python3
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Minimal OpenPI client for the π0 (Pi-Zero) VLA policy server.
-
-Connects to a running π0 server (see README.md), sends a single robot
-observation (3 cameras + state + language instruction), and prints the returned
-continuous action chunk ``[action_horizon, action_dim]``.
-
-The server speaks the OpenPI websocket protocol:
-    connect -> server sends msgpack(policy_server_config metadata)
-    infer   -> client sends msgpack(obs), server replies msgpack(actions ndarray)
-
-Usage:
-    python openpi_client.py --host 127.0.0.1 --port 8000 \
-        --prompt "pick up the red block"
-"""
-
-from __future__ import annotations
-
-import argparse
-import logging
-import sys
-import uuid
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-
-try:
-    # Avoid shadowing the installed ``openpi_client`` package with this dir.
-    example_dir = str(Path(__file__).resolve().parent)
-    removed = sys.path and sys.path[0] == example_dir
-    if removed:
-        sys.path.pop(0)
-    try:
-        from openpi_client.websocket_client_policy import WebsocketClientPolicy
-    finally:
-        if removed:
-            sys.path.insert(0, example_dir)
-except ImportError as exc:  # pragma: no cover - runtime dependency guard
-    raise ImportError("π0 OpenPI example requires `openpi-client`.") from exc
-
-logger = logging.getLogger("pi0_openpi_client")
-
-DEFAULT_HOST = "127.0.0.1"
-DEFAULT_PORT = 8000
-DEFAULT_PATH = "/v1/realtime/robot/openpi"
-DEFAULT_PROMPT = "pick up the red block and place it in the bin"
-
-# π0 / pi0_base camera identities (must match the server's image_feature_keys).
-CAMERA_KEYS = (
-    "observation.images.base_0_rgb",
-    "observation.images.left_wrist_0_rgb",
-    "observation.images.right_wrist_0_rgb",
-)
-STATE_DIM = 32
-
-
-def _make_dummy_obs(prompt: str, session_id: str, image_size: int = 224) -> dict[str, Any]:
-    """Build a single π0 observation with blank cameras + zero state.
-
-    Replace the zeros with real camera frames (HWC uint8) and proprioceptive
-    state to drive an actual robot.
-    """
-    obs: dict[str, Any] = {cam: np.zeros((image_size, image_size, 3), dtype=np.uint8) for cam in CAMERA_KEYS}
-    obs["state"] = np.zeros(STATE_DIM, dtype=np.float32)
-    obs["prompt"] = prompt
-    obs["session_id"] = session_id
-    return obs
-
-
-def _ws_uri(host: str, port: int, path: str) -> str:
-    """Build the full websocket URI. WebsocketClientPolicy takes the whole URI as
-    ``host`` (it has no separate path arg)."""
-    if host.startswith(("ws://", "wss://")):
-        return host
-    return f"ws://{host}:{port}{path}"
-
-
-def _as_action_array(response: Any) -> np.ndarray:
-    if isinstance(response, dict) and response.get("type") == "error":
-        raise RuntimeError(f"Inference failed: {response.get('message', response)}")
-    return np.asarray(response, dtype=np.float32)
-
-
-def main() -> int:
-    parser = argparse.ArgumentParser(description="π0 OpenPI client")
-    parser.add_argument("--host", default=DEFAULT_HOST)
-    parser.add_argument("--port", type=int, default=DEFAULT_PORT)
-    parser.add_argument("--path", default=DEFAULT_PATH)
-    parser.add_argument("--prompt", default=DEFAULT_PROMPT)
-    parser.add_argument("--num-steps", type=int, default=2, help="Inference observations to send.")
-    args = parser.parse_args()
-    logging.basicConfig(level=logging.INFO, format="%(message)s")
-
-    policy = WebsocketClientPolicy(host=_ws_uri(args.host, args.port, args.path))
-    metadata = dict(policy.get_server_metadata())
-    logger.info("Server metadata: %s", metadata)
-
-    session_id = str(uuid.uuid4())
-    for step in range(args.num_steps):
-        obs = _make_dummy_obs(args.prompt, session_id)
-        actions = _as_action_array(policy.infer(obs))
-        logger.info(
-            "[step %d] actions shape=%s mean=%.4f std=%.4f",
-            step,
-            actions.shape,
-            float(actions.mean()),
-            float(actions.std()),
-        )
-        if not np.isfinite(actions).all():
-            raise RuntimeError("Server returned non-finite actions.")
-    logger.info("Done — π0 OpenPI round-trip OK.")
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())
@@ -79,7 +79,6 @@ plugins:
        user_guide/diffusion/step_execution.md: user_guide/diffusion/execution_modes.md
        user_guide/diffusion/cpu_offload_diffusion.md: user_guide/diffusion/cpu_offload.md
        user_guide/quantization/index.md: user_guide/quantization/overview.md
-        configuration/pd_disaggregation.md: features/pd_disaggregation.md
        design/feature/diffusion_request_level_batching.md: design/feature/diffusion_continuous_batching.md
        design/feature/diffusion_step_execution.md: design/feature/diffusion_continuous_batching.md
        design/feature/cpu_offloading.md: design/feature/offloader/README.md
@@ -45,7 +45,7 @@ MusicGen-style delay pattern).
 **Online serving:**

 ```bash
-vllm serve bosonai/higgs-audio-v3-tts-4b \
+vllm-omni serve bosonai/higgs-audio-v3-tts-4b \
    --host 0.0.0.0 --port 8095 \
    --trust-remote-code --omni
 ```
@@ -81,7 +81,7 @@ MammothModa2 generation parameters as a JSON object through `--extra-body`:
 ```bash
 python examples/offline_inference/text_to_image/text_to_image.py \
  --model ./MammothModa2-Preview \
-  --deploy-config vllm_omni/deploy/mammoth_moda2.yaml \
+  --stage-configs-path vllm_omni/deploy/mammoth_moda2.yaml \
  --prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
  --height 1024 \
  --width 1024 \
@@ -132,44 +132,6 @@ ls -lh mammoth_t2i.png
 python -c "from PIL import Image; print(Image.open('mammoth_t2i.png').size)"
 ```

-### 1x AMD MI300X, MammothModa2 Preview
-
-#### Environment
-
- OS: Linux 6.8.0-134-generic, x86_64
- Container: official ROCm image built from `docker/Dockerfile.rocm`
- Python: 3.12.13
- PyTorch: 2.11.0+gitd0c8b1f
- Driver / runtime: AMD 6.19.14.31400000 / ROCm 7.2.53211
- GPU: one AMD Instinct MI300X, `gfx942:sramecc+:xnack-`, 191.69 GiB visible HBM
- vLLM version: 0.27.0+rocm723
- vLLM Omni version or commit: `73e1368c7bb940efe1a025859c9d6c8eeeb2e3f0`
- Installed vLLM Omni package metadata: `0.27.0rc2.dev44+g55abdade9.rocm`
-
-#### Offline Commands
-
-The checked run used the committed stage split, with `gpu_memory_utilization` set to 0.5 for AR and 0.3 for DiT:
-
-```bash
-python3 examples/offline_inference/text_to_image/text_to_image.py \
-    --model bytedance-research/MammothModa2-Preview \
-    --deploy-config vllm_omni/deploy/mammoth_moda2.yaml \
-    --prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
-    --height 1024 \
-    --width 1024 \
-    --seed 42 \
-    --extra-body '{"text_guidance_scale": 4.0, "cfg_range": [0.0, 1.0], "num_inference_steps": 50}' \
-    --enable-diffusion-pipeline-profiler \
-    --log-stats \
-    --output mammoth_t2i.png
-```
-
-#### Verification
-
-The first request took 85.224 seconds. The AR stage generated 4,161 visual tokens in 72.996 seconds, and the DiT stage took 12.163 seconds. AR weight loading used 21.4 GiB and took 8.250 seconds. DiT weight loading used 5.49 GiB and took 1.824 seconds. The largest one second whole device memory sample was 106.57 GiB, including the AR KV cache reserved by the 0.5 memory setting.
-
-The output was a valid 1024 by 1024 RGB PNG.
-
 ## MammothModa2-Dev unified inference

 MammothModa2-Dev uses a Qwen3-VL AR backbone, while MammothModa2-Preview uses
@@ -416,10 +416,9 @@ For example, keep the first main block's attention projections in BF16 with:
 ```

 The structured option replaces `--quantization fp8`. Online FP8 can be used
-with H3 layerwise offload and with both DLO transfer paths. The default
-AllGather path uses the ordinary loader to finalize FP8 weights and scales
-before sharding them across ranks. `--dlo-no-use-allgather` instead retains
-complete rank-local tensors and avoids the synchronized request-wave contract.
+with H3 layerwise offload and with distributed layerwise offload's full-weight
+per-rank path (`--dlo-no-use-allgather`). The sharded DLO AllGather path is not
+supported for runtime-created FP8 weights.

 ## AMD ROCm (gfx942 / gfx950)

@@ -871,9 +870,8 @@ vllm serve "${MODEL_ROOT}/FL2VA" \
  H3 native `tile` mode only.
 - A U2 x Ring2 hybrid currently fails with an attention-mask length mismatch; use
  pure Ulysses.
- Online FP8 with DLO AllGather temporarily materializes the complete FP8 model
-  in host memory on every rank during startup before retaining only each rank's
-  shard. Size startup host memory for that transient peak.
+- Runtime-created FP8 weights do not support the sharded DLO AllGather path;
+  use `--dlo-no-use-allgather` when combining online FP8 with DLO.
 - TeaCache and Cache-DiT cannot be enabled on the same server.
 - Image+audio Ref2VA accepts exactly one image and one audio reference.
 - Video Ref2VA accepts one or more video files, but not an additional standalone
@@ -1,173 +0,0 @@
-# MiniMax Music 3
-
-> Text-to-music: lyrics plus a style caption in, a 32 kHz stereo song out
-
-## Summary
-
- Vendor: MiniMax
- Model: `MiniMaxAI/MiniMax-Music3`
- Task: Text-to-music generation (vocals and instrumental together)
- Mode: Online serving through the OpenAI-compatible `/v1/audio/speech` API
- Maintainer: @linyueqian
-
-## When to use this recipe
-
-Use it to serve `MiniMaxAI/MiniMax-Music3` for full-song generation up to six
-minutes. This is not a TTS model with a musical voice: there is no reference
-speaker, no `voice`, and no `temperature`. What you control is the lyrics, the
-caption, the seed, and the length.
-
-## References
-
- Model card: [MiniMaxAI/MiniMax-Music3](https://huggingface.co/MiniMaxAI/MiniMax-Music3)
- Demo samples: [MiniMax Music 3 demo](https://minimax-ai.github.io/music3-demo/)
- Deploy config: `vllm_omni/deploy/minimax_music3.yaml` (single GPU),
-  `vllm_omni/deploy/minimax_music3_2gpu.yaml` (split across two)
-
-## Architecture
-
-| Component | Spec |
-|---|---|
-| Backbone | Qwen3 decoder, 36 layers, hidden 4096, GQA 32/8, vocab 200k |
-| Depth decoder | 4 layers, hidden 4096, 16 heads; 7 residual codebooks of 1024 |
-| Frame | 8 codebooks; `c0` in the backbone vocabulary, `c1..c7` in the depth decoder |
-| AR guidance | Classifier-free, scale 1.5, `c0` masked to the conditioned branch's top 50 |
-| Acoustic stage | Flow-matching transformer (36 layers, dim 2048) plus a DAC-style decoder, 512x upsampling |
-| Solver | Euler, 30 steps, acoustic guidance scale 1.7 |
-| Window | 200 frames per acoustic chunk, 100-frame hop |
-| Frame rate | 25 frames per second |
-| Context length | 10,240 tokens |
-| Output | 32 kHz stereo WAV |
-
-Generation is two-stage. The backbone predicts one audio frame per decode step,
-each frame eight RVQ codebooks deep. Every 200 frames are handed to the
-flow-matching transformer, which solves a latent that the decoder turns into
-waveform.
-
-**Guidance doubles the rows.** Every request decodes twice: one row sees the
-real prompt, the other sees the same prompt with the caption-and-lyrics span
-replaced by `<|audio_cfg|>`. Both rows hold their own KV cache for the whole
-song, so a request costs twice the KV its length suggests. Size the pool for
-rows, not requests.
-
-## Hardware Support
-
-## GPU
-
-### 1x H200 141GB
-
-#### Environment
-
- OS: Linux
- Python: 3.12
- CUDA: 12.9
- vLLM version: 0.27.0
- vLLM-Omni version or commit: branch `feat/minimax-music3`
-
-#### Command
-
-```bash
-vllm serve MiniMaxAI/MiniMax-Music3 \
-    --host 0.0.0.0 --port 8000 \
-    --trust-remote-code --omni
-```
-
-The default deploy config `vllm_omni/deploy/minimax_music3.yaml` is discovered
-automatically from the checkpoint's `model_type=minimax_music3`. It colocates
-both stages on device 0.
-
-#### Verification
-
-```bash
-curl -X POST http://localhost:8000/v1/audio/speech \
-  -H "Content-Type: application/json" \
-  -d '{
-    "model": "MiniMaxAI/MiniMax-Music3",
-    "input": "[Verse]\nWalking down the empty street at midnight\nStreetlights flicker like a broken dream\n[Chorus]\nAnd I keep on walking\nTill the morning finds me",
-    "instructions": "A melancholic lo-fi hip-hop track at 85 BPM in F minor: mellow Rhodes piano riff, soft vinyl crackle, dusty boom-bap drums with a laid-back swing, warm upright bass. Intimate bedroom production, gentle tape saturation, no bright cymbals.",
-    "seed": 1,
-    "max_new_tokens": 750
-  }' \
-  --output song.wav
-```
-
-Expected: a 30 second 32 kHz stereo WAV. `max_new_tokens` counts audio frames
-at 25 per second, so 750 frames is at most 30 seconds. It is a cap, not a
-target: the model ends the song itself when it emits the audio-end token, and a
-response shorter than the cap is the model finishing rather than a truncation.
-
-#### Notes
-
- **Memory:** stage 0 takes `gpu_memory_utilization: 0.6` for the bf16 backbone
-  and its KV pool; stage 1 takes 0.2 for the transformer and decoder weights,
-  which sit outside any KV budget.
- **Key flags:** none required. Guidance, the solver step count and the window
-  geometry are fixed by the checkpoint and are not request parameters.
- **Known limitations:**
-  - The external API is non-streaming; `stream: true` is rejected.
-  - `max_model_len` is 10,240, so a maximum-length prompt (5,000 tokens) and a
-    maximum-length song (9,000 frames) cannot both fit. Long captions shorten
-    the maximum song. This is a property of the checkpoint.
-  - The acoustic stage runs in float32. bfloat16 is accepted but measurably
-    degrades the solver.
-
-### 2x H200 141GB
-
-Use `minimax_music3_2gpu.yaml` to put the acoustic stage on the second device:
-
-```bash
-vllm serve MiniMaxAI/MiniMax-Music3 \
-    --host 0.0.0.0 --port 8000 \
-    --trust-remote-code --omni \
-    --deploy-config minimax_music3_2gpu.yaml
-```
-
-Only the placement differs; both layouts run the acoustic stage in float32.
-
-## Writing the request
-
-Two fields carry the request, and both are required.
-
- `input` is the **lyrics**. Structure tags (`[Verse]`, `[Chorus]`, `[Bridge]`,
-  `[Outro]`) steer the arrangement and are part of the prompt contract.
-  **Put a tag on its own line.** Normalization keeps only the tags on a line
-  that starts with one and drops the rest of that line, so lyrics written next
-  to a tag are silently lost:
-
-  ```text
-  "[Verse]\nWalking down the street"   ->   [start] [verse] Walking down the street
-  "[Verse] Walking down the street"    ->   [start] [verse]
-  ```
-
- `instructions` is the **caption** describing genre, instrumentation, tempo,
-  mood and production. It is the strongest control available: vague captions
-  give generic arrangements. Long structured captions work well, since the
-  model was trained on descriptions covering global attributes, emotional
-  progression, vocal detail and arrangement.
-
-A request is deterministic in its seed: the same lyrics, caption, seed and
-length return the same audio. Changing only the seed gives another take of the
-same song.
-
-## Request parameters
-
-| Parameter | Type | Default | Description |
-|---|---|---|---|
-| `input` | string | required | Lyrics, non-empty. Tags on their own lines |
-| `instructions` | string | required | Caption, non-empty |
-| `seed` | int | `0` | Non-negative. Fixes the output for a given request |
-| `max_new_tokens` | int | `9000` | Cap on audio frames, 25 per second. Maximum 9,000 |
-| `response_format` | string | `"wav"` | Output container |
-
-## Parameters the model rejects
-
-The request is refused rather than silently ignored, so a mistake is visible
-immediately.
-
-| Parameter | Reason |
-|---|---|
-| `temperature`, `top_p`, `top_k`, `repetition_penalty` | Sampling is fixed: guidance at 1.5, then a seeded top-k 50 draw |
-| `voice` | There is no speaker to select; the vocal comes from the caption |
-| `ref_audio`, `ref_text`, `language`, `task_type` | No reference-audio conditioning and no language tag |
-| `speed` | Only `1.0`. Tempo belongs in the caption, e.g. "at 92 BPM" |
-| `stream: true` | The external API is non-streaming |
@@ -51,7 +51,7 @@ NemotronH (~3B active parameters). Pick the pipeline by task:
 `model_type` auto-resolves `audex_tts.yaml`):

 ```bash
-vllm serve nvidia/Nemotron-Labs-Audex-2B \
+vllm-omni serve nvidia/Nemotron-Labs-Audex-2B \
    --host 0.0.0.0 --port 8097 \
    --trust-remote-code --omni
 # other modes: add --stage-configs-path vllm_omni/deploy/audex_{tta,thinker_only,s2s}.yaml
@@ -109,7 +109,7 @@ The 30B REQUIRES its explicit deploy yaml (the shared HF `model_type`
 auto-resolves to the 2B-tuned config otherwise):

 ```bash
-vllm serve nvidia/Nemotron-Labs-Audex-30B-A3B \
+vllm-omni serve nvidia/Nemotron-Labs-Audex-30B-A3B \
    --host 0.0.0.0 --port 8097 \
    --trust-remote-code --omni \
    --stage-configs-path vllm_omni/deploy/audex_tts_30b.yaml
@@ -71,35 +71,6 @@ single-GPU entry point; `minicpmo_4_5_2gpu.yaml` is the recommended
 two-GPU profile. The removed fused two-stage implementation is not retained as
 a fallback because it would duplicate state machines and correctness paths.

-### Graph execution
-
-Graph boundaries follow each stage's state model:
-
-| Stage | Graph mode on Ascend | Eager boundary |
-| --- | --- | --- |
-| 0 Thinker | vLLM `PIECEWISE` | multimodal preprocessing and output routing |
-| 1 Talker | vLLM `PIECEWISE` | conditioning preprocessing, codec sampling, request-state updates |
-| 2 Code2Wav | inner exact-shape NPUGraph for the CFM DiT estimator | encoder, timestep embedding, HiFT/RNG, request parsing, stream-state commit |
-
-Stage 2 keeps `enforce_eager: true` for the generation runner because its
-Python request metadata and per-request cache dictionaries are not valid
-outer-graph inputs. The backend captures only the deterministic CFM DiT
-estimator, with an exact graph key for each tensor shape. Timestep embedding
-stays eager because the upstream implementation creates a CPU frequency tensor;
-HiFT stays eager because it generates random phase. The graph cache stores up
-to 32 entries by default, after which unseen shapes run eagerly.
-
-An ACL graph capture failure is fatal to that Stage-2 process because older
-torch-npu releases can leave allocator and RNG capture state invalid. Restart
-the service after a capture failure. To run without capture, set
-`code2wav_enable_npu_graph: false` under the Stage-2
-`platforms.npu.stages[].additional_config` block before startup. Tune the cache
-limit there with `code2wav_max_npu_graphs`. Graph mode also requires
-`ASCEND_LAUNCH_BLOCKING` to be unset or set to `0`.
-
-The inner NPUGraph is independent of the outer runner setting, so do not use a
-global `--enforce-eager` override when Stage 0/1 `PIECEWISE` replay is desired.
-
 ## GPU

 ### 1 x GPU (default — single command)
@@ -240,9 +211,9 @@ speech output (TTS)"** checkbox on / off.
  model processes still share one CUDA device.
 - `--trust-remote-code` is required — the HF repo ships a custom
  `MiniCPMO` config / model class.
- Stage 0 Thinker and Stage 1 Talker enable vLLM CUDA Graphs. Stage 2 keeps its
-  request orchestration eager while using inner CFM and HiFT CUDA Graphs. On
-  Ascend, the exact-shape CFM DiT estimator instead uses inner NPUGraph replay.
+- Stage 0 Thinker and Stage 1 Talker enable vLLM CUDA Graphs. Stage 2 remains
+  eager because its request-owned Flow/HiFT caches and variable chunk/cache
+  shapes are not yet exposed through a static exact-shape graph wrapper.
 - All default stages use `max_num_seqs: 4` to reduce cross-process GPU
  contention. Talker AR
  state and Code2Wav caches are request-owned; Code2Wav batches only
@@ -253,8 +224,8 @@ speech output (TTS)"** checkbox on / off.
 - `StageRequestStats.batch_size` is a request-scoped placeholder, not the
  scheduler's execution batch.
 - Single-GPU co-location trades throughput for hardware density: Stage 0/1
-  CUDA Graph replay and Stage 2 vocoder kernels compete across three CUDA
-  contexts. Use the 8x4090 config or a custom multi-GPU mapping for
+  CUDA Graph replay and eager Stage 2 vocoder kernels compete across three
+  CUDA contexts. Use the 8x4090 config or a custom multi-GPU mapping for
  throughput-sensitive serving.

 ### 8 x RTX 4090 24GB (consumer-GPU layout)
@@ -252,53 +252,3 @@ curl -X POST http://localhost:8091/v1/audio/speech \
    - stage_id: 1
      gpu_memory_utilization: 0.15
  ```
-
-### 1x AMD MI300X, 1.7B checkpoints
-
-#### Environment
-
- OS: Linux 6.8.0-134-generic, x86_64
- Container: official ROCm image built from `docker/Dockerfile.rocm`
- Python: 3.12.13
- PyTorch: 2.11.0+gitd0c8b1f
- Driver / runtime: AMD 6.19.14.31400000 / ROCm 7.2.53211
- GPU: one AMD Instinct MI300X, `gfx942:sramecc+:xnack-`, 191.69 GiB visible HBM
- vLLM version: 0.27.0+rocm723
- vLLM Omni version or commit: `3fecb6953ca8dc51210cc0421ef24552267a41ef`
- Installed vLLM Omni package metadata: `0.27.0rc2.dev44+g55abdade9.rocm`
- ONNX Runtime: onnxruntime-rocm 1.22.2.post3 with `ROCMExecutionProvider`
- transformers: 5.15.0
-
-#### Command
-
-The MI300X tests also worked without `enforce_eager`. To use that setup, remove `enforce_eager: true` from stage 1 under `platforms.rocm` in `vllm_omni/deploy/qwen3_tts.yaml`.
-
-Set this environment variable before you start Qwen3-TTS:
-
-```bash
-export MIOPEN_FIND_MODE=FAST
-```
-
-#### Verification
-
-Each row records one offline request with one prompt and batch size one. Request time is the logged `e2e_wall_time_ms`, and stage 1 time is the logged `e2e_stage_1_wall_time_ms`.
-
-| Checkpoint | Code2Wav configuration | `MIOPEN_FIND_MODE` | Request time | Stage 1 time | Output duration | Real time factor | Maximum sampled device memory |
-| --- | --- | --- | ---: | ---: | ---: | ---: | ---: |
-| CustomVoice | Eager | Unset | 1.683 s | 1.613 s | 5.68 s | 0.30 | 60.91 GiB |
-| CustomVoice | Graphs allowed | `FAST` | 1.326 s | 1.259 s | 5.68 s | 0.23 | 62.19 GiB |
-| VoiceDesign | Eager | Unset | 1.195 s | 1.180 s | 4.40 s | 0.27 | 60.91 GiB |
-| VoiceDesign | Graphs allowed | `FAST` | 1.127 s | 1.054 s | 4.40 s | 0.26 | 62.21 GiB |
-| Base | Eager | Unset | 13.330 s | 13.263 s | 4.32 s | 3.09 | 61.74 GiB |
-| Base | Graphs allowed | `FAST` | 14.314 s | 14.245 s | 4.80 s | 2.98 | 62.84 GiB |
-
-All six processes finished successfully, and each output passed the 24 kHz mono WAV checks.
-
-#### Notes
-
- The current ROCm config keeps Code2Wav in eager mode.
- The tested alternative allows graphs and sets [`MIOPEN_FIND_MODE=FAST`](https://rocm.docs.amd.com/projects/MIOpen/en/develop/reference/env_variables.html). MIOpen uses a saved kernel choice when available and uses its immediate fallback otherwise.
- CustomVoice used a graph for one of four logged Code2Wav batches. VoiceDesign used a graph for two of four batches. Base used no graphs for its four logged batches.
- The recorded request time with graphs allowed was lower for CustomVoice and VoiceDesign and higher for Base. The Base runs generated different amounts of audio, so their request times are not a direct comparison.
- Each setup was tested with one request, so the timing is not a performance benchmark.
- GPU memory was sampled once per second with `rocm-smi`.
@@ -49,7 +49,6 @@ recipes/
 | [`Helios/Helios.md`](./Helios/Helios.md) | Text-to-video, image-to-video, and video-to-video generation | 1x NVIDIA H20 |
 | [`inclusionAI/Ming-flash-omni-2.0.md`](./inclusionAI/Ming-flash-omni-2.0.md) | Online serving for multimodal chat + standalone TTS | 4x H100 / 1x H100 80GB |
 | [`inclusionAI/Ming-omni-tts.md`](./inclusionAI/Ming-omni-tts.md) | Offline + online dense Ming TTS/audio generation | 1x H100 80GB / 1x AMD MI300X (ROCm 7.2) |
-| [`k2-fsa/OmniVoice.md`](./k2-fsa/OmniVoice.md) | Offline and online multilingual text-to-speech | 1x AMD MI300X (ROCm 7.2) |
 | [`IndexTeam/IndexTTS-2.md`](./IndexTeam/IndexTTS-2.md) | Online serving for voice-cloned TTS with optional emotion control | 1x L4 24GB or larger CUDA GPU |
 | [`IndexTeam/IndexTTS-2_5.md`](./IndexTeam/IndexTTS-2_5.md) | Online + offline multilingual voice-cloned TTS with native speed and emotion control | 1x NVIDIA H20 96GB |
 | [`MiniMaxAI/MiniMax-H3.md`](./MiniMaxAI/MiniMax-H3.md) | T2VA, FL2VA, image+audio Ref2VA, and multi-video Ref2VA serving | 1x GPU with CPU offload / 4x B300 |
@@ -61,7 +60,6 @@ recipes/
 | [`meituan-longcat/LongCat-Video-Avatar-1.5.md`](./meituan-longcat/LongCat-Video-Avatar-1.5.md) | Audio-driven avatar video generation (AT2V / AI2V, single- and multi-speaker, AVC continuation) | 1x H100 80GB |
 | [`mistralai/Voxtral-TTS.md`](./mistralai/Voxtral-TTS.md) | Online serving for TTS | 1x RTX 4090 24GB |
 | [`Robbyant/LingBot-Video.md`](./Robbyant/LingBot-Video.md) | Native dense and MoE text-to-video serving | Dense: 1x L20X; MoE: 1x L20X (~67.7 GiB peak) |
-| [`Robbyant/LingBot-World-2.0.md`](./Robbyant/LingBot-World-2.0.md) | Offline and experimental realtime interactive world generation | 1x H200/B200 or 2x B200 for TP=2 |
 | [`cosmos3/Cosmos3-Nano.md`](./cosmos3/Cosmos3-Nano.md) | Text-to-image, text-to-video, image-to-video, video-to-video generation, text to video with sound, action policy | 1x H200 141GB / B300 |
 | [`cosmos3/Cosmos3-Edge.md`](./cosmos3/Cosmos3-Edge.md) | T2I / T2V / I2V + action (Physical-AI) generation on the Nemotron-based Edge transformer | 1x GPU (~8 GiB) |
 | [`cosmos3/Cosmos3-Super.md`](./cosmos3/Cosmos3-Super.md) | 64B T2I / T2V / I2V / V2V generation (+ optional audio) / Action policy | 8x H200/H100/A100 / 2x H200 / B300 |
@@ -74,9 +72,9 @@ recipes/
 | [`Qwen/Qwen-Image-2512.md`](./Qwen/Qwen-Image-2512.md) | Text-to-image serving with step-wise continuous batching replay and ModelOpt FP8 / mixed FP8/NVFP4 | 1x A800 80GB / 2x B200 |
 | [`Qwen/Qwen-Image-Edit.md`](./Qwen/Qwen-Image-Edit.md) | Text-guided single-image editing | 1x or 2x H200 141GB |
 | [`Qwen/Qwen3-Omni.md`](./Qwen/Qwen3-Omni.md) | Online serving for multimodal chat | 1x A100 80GB |
-| [`Qwen/Qwen3-TTS.md`](./Qwen/Qwen3-TTS.md) | Text-to-speech serving (CustomVoice / VoiceDesign / Base) | 1x H100/A100 80GB / 1x AMD MI300X |
-| [`SenseNova/SenseNova-U1.md`](./SenseNova/SenseNova-U1.md) | Unified image generation and understanding | 1x or 2x H200 / 1x AMD MI300X |
-| [`StabilityAI/Stable-Audio-Open.md`](./StabilityAI/Stable-Audio-Open.md) | Offline + online text-to-audio generation (Stable Audio Open) | 1x RTX 4090 24GB / 1x AMD MI300X |
+| [`Qwen/Qwen3-TTS.md`](./Qwen/Qwen3-TTS.md) | Text-to-speech serving (CustomVoice / VoiceDesign / Base) | 1x H100/A100 80GB |
+| [`SenseNova/SenseNova-U1.md`](./SenseNova/SenseNova-U1.md) | Unified image generation and understanding | 1x H200 (144GB) |
+| [`StabilityAI/Stable-Audio-Open.md`](./StabilityAI/Stable-Audio-Open.md) | Offline + online text-to-audio generation (Stable Audio Open) | 1x RTX 4090 24GB |
 | [`Tencent/Covo-Audio-Chat.md`](./Tencent/Covo-Audio-Chat.md) | Online serving for audio chat | 1x A100 80GB |
 | [`Tencent/HunyuanImage-3.0-Instruct.md`](./Tencent/HunyuanImage-3.0-Instruct.md) | DiT-only text-to-image serving and benchmark, including ModelOpt mixed FP8/NVFP4 | 4x H100/H800 80GB / 2x B200 |
 | [`Wan-AI/Wan2.2-T2V.md`](./Wan-AI/Wan2.2-T2V.md) | Text-to-video serving (Wan2.2 14B) with Skip-Softmax sparse attention | 1x datacenter Blackwell (B200/B300) |
@@ -1,122 +0,0 @@
-# LingBot-World 2.0
-
-> Offline and experimental realtime interactive world generation
-
-## Summary
-
- Vendor: Robbyant
- Model: `robbyant/lingbot-world-v2-14b-causal-fast-diffusers`
- Task: image-conditioned interactive world generation
- Modes: offline trajectory replay and in-process realtime AR-Diffusion ticks
- Hardware validated: NVIDIA H200 and B200
- Maintainer: Community
-
-The checkpoint is separately licensed under CC BY-NC-SA and restricted to
-non-commercial use. The vLLM-Omni integration code remains Apache-2.0.
-
-## Offline generation
-
-The offline path consumes one source image and an action directory containing
-`poses.npy` with shape `[frames, 4, 4]` and `intrinsics.npy` with shape
-`[frames, 4]`.
-
-```bash
-python examples/offline_inference/diffusion/lingbot_world_v2.py \
-  --prompt "The camera moves slowly forward through the scene." \
-  --image /path/to/first_frame.png \
-  --action-dir /path/to/actions/forward \
-  --num-frames 81 \
-  --output lingbot_world_v2.mp4
-```
-
-The current causal-fast checkpoint uses four DMD sampling steps. Raw frame
-counts must be `9 + 12k`, up to 117 frames. Camera files may contain a
-longer bounded source trajectory (the official example contains 269 frames);
-the runtime consumes only the prefix needed by the request.
-
-## Realtime in-process generation
-
-The realtime example keeps AR-Diffusion KV and model-owned state across
-requests. Each JSONL line describes the prompt and/or three latent-frame camera
-actions applied at the next chunk boundary:
-
-```json
-{"event_id":1,"prompt":"A road through a forest","frames":[["j"],[],[]]}
-{"event_id":2,"frames":[["w"],["w"],["w"]]}
-{"event_id":3,"prompt":"The road enters a snowy valley","frames":[[],[],[]]}
-```
-
-Run:
-
-```bash
-python examples/offline_inference/diffusion/lingbot_world_v2_realtime.py \
-  --image /path/to/first_frame.png \
-  --events /path/to/events.jsonl \
-  --output-dir /tmp/lingbot-realtime \
-  --gpu-memory-fraction 0.6
-```
-
-The example writes one latent tensor and one metadata JSON file per chunk. It
-exercises the same `ARDiffusionSessionManager -> ARDiffusionOmniTickConsumer
-> AsyncOmni -> ARDiffusionEngine` path used by a future HTTP or WebSocket
-transport.
-
-This PR intentionally does not define a public realtime HTTP/WebSocket schema.
-An online serving client should be added together with the public transport
-API rather than exposing LingBot-specific event fields from the generic model
-runtime. The structured camera-interaction frontend and transport work is
-tracked separately in
-[vllm-project/vllm-omni#5527](https://github.com/vllm-project/vllm-omni/pull/5527).
-
-## Realtime identity and controls
-
- `session_id` identifies the persistent world and its worker-owned state.
- `event_id` identifies a prompt/control update and remains monotonic across reset.
- `chunk_index` is contiguous from zero and restarts from zero after reset.
- `request_id` correlates one chunk snapshot with its output metadata.
- AsyncOmni uses a separate UUID-suffixed internal engine routing ID.
-
-The generic runtime transports controls as opaque snapshots. LingBot's adapter
-accepts:
-
- `lingbot.camera_actions.v1` for per-latent-frame key states such as `w`, `a`,
-  `s`, `d`, `i`, `j`, `k`, and `l`;
- `lingbot.camera_trajectory.v1` for explicit pose/intrinsics trajectories.
-
-## Validation
-
-Real-checkpoint validation uses 480x832 output, four DMD steps, and seed 42.
-The exercised matrix includes:
-
- TP=1 and TP=2 execution;
- two interleaved resident sessions;
- action input and prompt switching;
- seven contiguous chunks crossing the sink plus recent rolling window;
- direct versus paged replay;
- CUDA Graph execution;
- reset, close, failure cleanup, and exact metadata matching; and
- nine-frame VAE decode.
-
-Official `generate.py` accuracy and performance numbers, together with
-generation artifacts, are recorded in the PR validation section for the exact
-tested commit.
-
-## Current limitations
-
- Only the 14B causal-fast checkpoint is supported.
- The realtime control plane is internal; there is no public server transport yet.
- AR-Diffusion stages currently require one replica because session-affine
-  routing across replicas is not implemented.
- One AR block is generated per request and `max_num_seqs` must be one.
- Stateful streaming VAE decode is not implemented; the realtime example emits
-  latent chunks.
- SP/USP, pipeline/CFG parallelism, HSDP, VAE parallelism, quantization,
-  Cache-DiT, TeaCache, causal-pretrain, and the 1.3B checkpoint are not claimed.
- No AMD GPU, Ascend NPU, or Intel GPU support is claimed.
-
-## References
-
- Checkpoint: <https://huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast-diffusers>
- Official implementation: <https://github.com/robbyant/lingbot-world-v2>
- Offline example: [`examples/offline_inference/diffusion/lingbot_world_v2.py`](../../examples/offline_inference/diffusion/lingbot_world_v2.py)
- Realtime design: [`docs/design/feature/realtime_ar_diffusion.md`](../../docs/design/feature/realtime_ar_diffusion.md)
@@ -178,48 +178,6 @@ with the expected 1536×2720 resolution.
 - The LLM transformer uses `QKVParallelLinear` and `MergedColumnParallelLinear`
  for fused QKV and gate/up projections with TP support.

-### 1x AMD MI300X 192GB
-
-#### Environment
-
- OS: Linux 6.8.0-134-generic, x86_64
- Container: official ROCm image built from `docker/Dockerfile.rocm`
- Python: 3.12.13
- PyTorch: 2.11.0+gitd0c8b1f
- Driver / runtime: AMD 6.19.14.31400000 / ROCm 7.2.53211
- GPU: one AMD Instinct MI300X, `gfx942:sramecc+:xnack-`, 191.69 GiB visible HBM
- vLLM version: 0.27.0+rocm723
- vLLM Omni version or commit: `a704c8759c96e123c0d7c89b11f120b1c0f120cf`
- Installed vLLM Omni package metadata: `0.27.0rc2.dev44+g55abdade9.rocm`
-
-#### Command
-
-```bash
-python3 examples/offline_inference/text_to_image/text_to_image.py \
-    --model SenseNova/SenseNova-U1-8B-MoT \
-    --prompt "Close portrait of an elderly woman by a farmhouse window, textured skin, gentle smile, warm natural light, emotional documentary look." \
-    --width 1536 \
-    --height 2720 \
-    --seed 42 \
-    --num-inference-steps 50 \
-    --cfg-scale 4.0 \
-    --extra-body '{"think": true, "cfg_norm": "none", "timestep_shift": 3.0, "t_eps": 0.02}' \
-    --enable-diffusion-pipeline-profiler \
-    --log-stats \
-    --output sensenova_u1_mi300x.png
-```
-
-#### Verification
-
-The command completed and wrote a valid 1536 by 2720 RGB PNG.
-
-#### Notes
-
- Generation with 50 inference steps took 34.011 seconds.
- Model loading used 32.774 GiB and took 13.987 seconds.
- The internal profiler recorded 35.67 GB reserved and 35.10 GB allocated for the request.
- The highest one second whole device memory sample was 37.93 GiB.
-
 ## Online Serving

 SenseNova-U1 supports all four modalities via the OpenAI-compatible
@@ -78,7 +78,7 @@ python examples/offline_inference/text_to_audio/text_to_audio.py \
 Start the online serving endpoint:

 ```bash
-vllm serve /path/to/stable-audio-open-1.0 \
+vllm-omni serve /path/to/stable-audio-open-1.0 \
  --host 0.0.0.0 \
  --port 8091 \
  --gpu-memory-utilization 0.9 \
@@ -164,50 +164,7 @@ PY
  `torchaudio==2.11.0+cu130`.
 - The `NIXL is not available`, `GLOO_SOCKET_IFNAME`, and `torchsde` boundary
  warnings observed during validation did not prevent successful generation.
- The RTX 4090 entry was validated on one 24 GB GPU. The MI300X entry below covers one 192 GB GPU.
+- This recipe was validated on one RTX 4090 24GB GPU only. Other GPU counts,
+  ROCm, XPU, and NPU setups are not covered here.
 - Long generations, higher inference-step counts, and non-WAV response formats
  were not benchmarked in this recipe.
-
-#### 1x AMD MI300X 192GB
-
-##### Environment
-
- OS: Linux 6.8.0-134-generic, x86_64
- Container: official ROCm image built from `docker/Dockerfile.rocm`
- Python: 3.12.13
- PyTorch: 2.11.0+gitd0c8b1f
- Driver / runtime: AMD 6.19.14.31400000 / ROCm 7.2.53211
- GPU: one AMD Instinct MI300X, `gfx942:sramecc+:xnack-`, 191.69 GiB visible HBM
- vLLM version: 0.27.0+rocm723
- vLLM Omni version or commit: `73e1368c7bb940efe1a025859c9d6c8eeeb2e3f0`
- Installed vLLM Omni package metadata: `0.27.0rc2.dev44+g55abdade9.rocm`
-
-##### Commands
-
-```bash
-python3 examples/offline_inference/text_to_audio/text_to_audio.py \
-    --model stabilityai/stable-audio-open-1.0 \
-    --prompt "A gentle piano melody with soft room ambience" \
-    --negative-prompt "Low quality, distorted, noisy" \
-    --seed 42 \
-    --guidance-scale 7.0 \
-    --audio-length 10.0 \
-    --num-inference-steps 50 \
-    --cache-backend tea_cache \
-    --enable-diffusion-pipeline-profiler \
-    --output stable_audio_10s.wav
-```
-
-##### Verification
-
-The command completed and wrote a valid 44.1 kHz stereo WAV with 10.00 seconds of audio.
-
-##### Notes
-
- TeaCache ran with `rel_l1_thresh=0.2`.
- Model loading used 2.7891 GiB and took 3.706 seconds.
- Generation took 4.750 seconds, which gives a real time factor of 0.475 for the 10.00 second output.
- The internal profiler recorded 15.65 GB reserved and 9.68 GB allocated for the request.
- The highest one second whole device memory sample was 19.61 GiB.
- The output RMS was 0.0887, and its peak absolute amplitude was 0.5761.
- The full process took 384 seconds, including startup and compilation.
@@ -39,9 +39,6 @@ platform section, document one or more tested hardware configurations.

 #### Command

-Serve commands use `vllm serve <model> --omni`. The `--omni` flag is what
-selects the Omni pipeline, so it is required and cannot be dropped.
-
 ```bash
 # Add the exact command(s) here
 ```
@@ -189,10 +189,8 @@ wait

 - `--dlo-use-allgather` (default) requires all concurrent requests to have the same
  `num_inference_steps` (AllGather is a collective that needs all ranks synchronized)
- Online per-tensor FP8 linears with DLO AllGather use the ordinary loader
-  instead of direct checkpoint mmap, then shard the finalized FP8 weights and
-  scales. Size host memory for a transient complete FP8 model on every rank at
-  startup
+- Online quantization (FP8) is incompatible with the DLO+AllGather mmap path;
+  use `--dlo-no-use-allgather` or disable online quantization
 - Tensor Parallel > 1 is rejected in the DLO+AllGather mmap path. TP with
  `--dlo-no-use-allgather` retains the standard TP loader and should be treated
  as an experimental compatibility path
@@ -39,7 +39,7 @@ Other hardware is welcome as community validation lands.
 Launch the two-stage talker:

 ```bash
-vllm serve inclusionAI/Ming-omni-tts-0.5B \
+vllm-omni serve inclusionAI/Ming-omni-tts-0.5B \
    --deploy-config vllm_omni/deploy/ming_tts.yaml \
    --omni \
    --port 8091 \
@@ -170,35 +170,6 @@ python examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py \
 `--ref-audio` matches upstream `use_spk_emb=True`; do not add `--ref-text`
 for the dialect case.

-**Offline smoke result on ROCm 7.2:**
-
-A later offline smoke run used the same MI300X model with the following software:
-
- GPU: one AMD Instinct MI300X with 191.69 GiB visible HBM and gfx942
- Kernel: Linux 6.8.0-134-generic, x86_64
- Python: 3.12.13
- PyTorch: 2.11.0+gitd0c8b1f
- ROCm or HIP: 7.2.53211
- vLLM: 0.27.0+rocm723
- vLLM Omni checkout: `73e1368c7bb940efe1a025859c9d6c8eeeb2e3f0`
-
-```bash
-python3 examples/offline_inference/text_to_speech/ming_tts/end2end.py \
-    --model inclusionAI/Ming-omni-tts-0.5B \
-    --deploy-config vllm_omni/deploy/ming_tts.yaml \
-    --case basic \
-    --text "你好，这是 AMD MI300X ROCm 配方验证。" \
-    --output-dir ./ming_tts_output \
-    --output-name ming_omni_tts_mi300x.wav \
-    --enforce-eager \
-    --log-stats \
-    --metadata-json ./ming_tts_manifest.json
-```
-
-The checked run produced a valid 44.1 kHz mono WAV with 4.80 seconds of audio, RMS 0.1449, and peak absolute amplitude 0.8621. The first request took 20.406 seconds, which gives a real time factor of 4.25. The stage weight loads used 1.31 GiB and 1.47 GiB and took 1.756 seconds and 1.264 seconds. The largest one second whole device memory sample was 91.05 GiB, including the reserved KV cache. Both stages used eager mode, and the AR stage selected `TRITON_ATTN`.
-
-The run used the official ROCm image built from `docker/Dockerfile.rocm`.
-
 ## Notes

 - The official ROCm image includes the platform dependencies.
@@ -1,98 +0,0 @@
-# OmniVoice
-
-> Offline and online text to speech with automatic voice, voice design, and reference voice cloning
-
-## Summary
-
- Vendor: k2-fsa
- Model: `k2-fsa/OmniVoice`
- Task: Multilingual text to speech
- Mode: Offline inference for all modes and online serving for automatic voice
- Maintainer: Community
-
-## References
-
- [Hugging Face model](https://huggingface.co/k2-fsa/OmniVoice)
- [Offline text to speech examples](../../docs/user_guide/examples/offline_inference/text_to_speech.md#omnivoice)
- [Online text to speech examples](../../docs/user_guide/examples/online_serving/text_to_speech.md#omnivoice)
- [Deploy config](../../vllm_omni/deploy/omnivoice.yaml)
-
-## Requirements
-
-Download the checkpoint before the first run:
-
-```bash
-huggingface-cli download k2-fsa/OmniVoice
-```
-
-Automatic voice and voice design require `transformers>=4.57.0`. Reference voice cloning requires `transformers>=5.3.0`.
-
-The vLLM Omni deploy config runs the OmniVoice generator and decoder inside one float32 diffusion stage.
-
-## ROCm
-
-### 1x AMD MI300X
-
-#### Environment
-
- GPU: one AMD Instinct MI300X with 191.69 GiB visible HBM and gfx942
- Kernel: Linux 6.8.0-134-generic, x86_64
- Python: 3.12.13
- PyTorch: 2.11.0+gitd0c8b1f
- ROCm or HIP: 7.2.53211
- vLLM: 0.27.0+rocm723
- vLLM Omni checkout: `73e1368c7bb940efe1a025859c9d6c8eeeb2e3f0`
- transformers: 5.15.0
-
-The run used the official ROCm image built from `docker/Dockerfile.rocm`.
-
-#### Automatic voice
-
-Run the offline example from the repository root:
-
-```bash
-python3 examples/offline_inference/text_to_speech/omnivoice/end2end.py \
-    --model k2-fsa/OmniVoice \
-    --deploy-config vllm_omni/deploy/omnivoice.yaml \
-    --text "Hello, this is OmniVoice running on one AMD MI300X." \
-    --seed 42 \
-    --output omnivoice_mi300x.wav
-```
-
-The checked run produced a valid 24 kHz mono WAV with 3.60 seconds of audio, RMS 0.1194, and peak absolute amplitude 0.9129. The first request took 0.685 seconds, which gives a real time factor of 0.19. Model loading used 3.5739 GiB and took 9.564 seconds. The request used 4.56 GB reserved and 3.64 GB allocated according to the internal profiler, while the largest one second whole device sample was 8.33 GiB. The log states that the OmniVoice Triton kernels loaded.
-
-#### Reference voice cloning
-
-The offline interface also accepts a reference WAV and its transcript:
-
-```bash
-python3 examples/offline_inference/text_to_speech/omnivoice/end2end.py \
-    --model k2-fsa/OmniVoice \
-    --text "Hello, this is a cloned voice." \
-    --ref-audio ref.wav \
-    --ref-text "This is the reference transcription." \
-    --output cloned.wav
-```
-
-#### Voice design
-
-```bash
-python3 examples/offline_inference/text_to_speech/omnivoice/end2end.py \
-    --model k2-fsa/OmniVoice \
-    --text "Hello, this is a designed voice." \
-    --instruct "female, low pitch, British accent" \
-    --output designed.wav
-```
-
-## Online serving
-
-Online serving currently supports automatic voice:
-
-```bash
-vllm serve k2-fsa/OmniVoice --omni --port 8091 --trust-remote-code
-```
-
-```bash
-cd examples/online_serving/text_to_speech/omnivoice
-python speech_client.py --text "Hello, how are you?"
-```
@@ -1,108 +0,0 @@
-# Pi0 (π0, Pi-Zero) VLA
-
-> Vision-Language-Action policy serving over the OpenPI realtime websocket
-
-## Summary
-
- Vendor: Physical Intelligence (checkpoint redistributed by LeRobot)
- Model: `lerobot/pi0_base`
- Task: Vision-Language-Action — multi-camera images + a language instruction +
-  robot proprioceptive state → a continuous action chunk
-  `[action_horizon, action_dim] = [50, 32]` via flow-matching denoising. π0 does
-  **not** emit text tokens.
- Mode: Online serving via the OpenPI realtime robot API
-  (`/v1/realtime/robot/openpi` websocket)
- Maintainer: Community
-
-## When to use this recipe
-
-Use this recipe to deploy `lerobot/pi0_base` as a robot policy server: a robot
-client connects over the OpenPI websocket and, per control step, sends an
-observation (camera frames + state + instruction) and receives an action chunk.
-
-A single pipeline class (`Pi0Pipeline`) serves the policy. π0 is stateless across
-calls (no KV reuse), so the protocol's `session_id` / `reset` are accepted but
-ignored.
-
-## Weights
-
-π0 uses the LeRobot `lerobot/pi0_base` checkpoint (HF, ~13 GB). Either let the
-server download it (`MODEL=lerobot/pi0_base`) or point at a local copy.
-
-## Hardware Support
-
-### 1x H200 141GB (Online serving)
-
-#### Environment
-
- OS: Ubuntu 22.04+
- Python: 3.12+
- Driver / runtime: NVIDIA CUDA environment
- vLLM / vLLM-Omni: use the commit you are deploying from
- OpenPI client extras (for the example client): `openpi-client`
-  (`pip install -e packages/openpi-client` from the
-  [openpi](https://github.com/Physical-Intelligence/openpi) repo), `websockets`,
-  `msgpack`, `msgpack-numpy`
-
-#### Command
-
-```bash
-vllm serve lerobot/pi0_base \
-  --omni \
-  --host 0.0.0.0 --port 8000 \
-  --served-model-name pi0 \
-  --deploy-config vllm_omni/deploy/pi0.yaml \
-  --enforce-eager --disable-log-stats
-```
-
-The deploy config (`vllm_omni/deploy/pi0.yaml`) declares a single diffusion stage
-(`Pi0Pipeline`), `float32`, `max_num_seqs: 1`, and the `policy_server_config`
-handshake metadata (3 cameras, `joint_position`, action horizon 50, action dim
-32) that enables the OpenPI realtime serving layer.
-
-#### Verification
-
-Run the example client (sends a synthetic observation and prints the returned
-action chunk):
-
-```bash
-python examples/online_serving/pi0/openpi_client.py \
-  --host 127.0.0.1 --port 8000 \
-  --prompt "pick up the red block and place it in the bin"
-```
-
-It connects, prints the server metadata, sends robot observations, and prints the
-returned `[action_horizon, action_dim] = [50, 32]` action chunks. The observation
-is a flat dict per inference:
-
-```python
-{
-    "observation.images.base_0_rgb":        np.uint8[H, W, 3],
-    "observation.images.left_wrist_0_rgb":  np.uint8[H, W, 3],
-    "observation.images.right_wrist_0_rgb": np.uint8[H, W, 3],
-    "state":   np.float32[state_dim],   # zero-padded to max_state_dim=32 server-side
-    "prompt":  "pick up the red block",
-    "session_id": "<uuid>",             # accepted but ignored (π0 is stateless)
-}
-```
-
-Camera keys must match the server's `image_feature_keys` (the checkpoint's
-`input_features` order). If your robot uses different camera names, set
-`model_config.image_key_map` in `pi0.yaml` to map raw obs keys → feature keys.
-
-## Correctness
-
-π0's flow-matching kernel is bit-for-bit matched to the LeRobot `PI0Policy`
-reference (`max|Δ| = 7.15e-07`, CPU/float32, fixed noise; see
-`tests/diffusion/models/pi0/test_pi0_parity.py::test_pi0_vllm_omni_vs_lerobot`).
-
-## References
-
- Model blog: <https://www.physicalintelligence.company/blog/pi0>
- Checkpoint: <https://huggingface.co/lerobot/pi0_base>
- Pipeline: [`vllm_omni/diffusion/models/pi0/pipeline_pi0.py`](../../vllm_omni/diffusion/models/pi0/pipeline_pi0.py)
- Deploy config: [`vllm_omni/deploy/pi0.yaml`](../../vllm_omni/deploy/pi0.yaml)
- Example client: [`examples/online_serving/pi0/`](../../examples/online_serving/pi0/)
- Tests: CPU units + LeRobot parity in [`tests/diffusion/models/pi0/`](../../tests/diffusion/models/pi0/),
-  OpenPI websocket e2e in
-  [`tests/e2e/online_serving/test_pi0_expansion.py`](../../tests/e2e/online_serving/test_pi0_expansion.py)
@@ -44,20 +44,20 @@ conditioned on reference audio and its transcript.
 Start the server from the repository root:

 ```bash
-vllm serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091
+vllm-omni serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091
 ```

 Async chunking is enabled by default in the bundled deployment config. For
 the sync (non-streaming) path:

 ```bash
-vllm serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091 --no-async-chunk
+vllm-omni serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091 --no-async-chunk
 ```

 Use a custom deploy config for advanced cases:

 ```bash
-vllm serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091 \
+vllm-omni serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091 \
  --deploy-config /path/to/your_glm_tts_overrides.yaml
 ```

@@ -9,17 +9,13 @@ from pathlib import Path
 import pytest

 from vllm_omni.benchmarks.data_modules.daily_omni_dataset import (
-    DailyOmniDataset,
    daily_omni_local_qa_json,
    daily_omni_local_videos_dir,
-    ensure_daily_omni_hub_root,
    resolve_daily_omni_local_root,
 )

 pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.benchmark]

-_HUB_ID = "liarliar/Daily-Omni"
-

 def _write_qa_json(root: Path) -> Path:
    qa = root / "qa.json"
@@ -27,28 +23,6 @@ def _write_qa_json(root: Path) -> Path:
    return qa


-def _stage_offline_hub_cache(tmp_path: Path, monkeypatch: pytest.MonkeyPatch, *, with_qa: bool = True) -> Path:
-    """Build a hub cache holding ``liarliar/Daily-Omni`` and cut off all network access.
-
-    Reproduces the air-gapped NPU runner: the dataset is pre-staged with ``hf download``, so it
-    only exists in the *hub* cache -- never in the ``datasets`` cache that ``load_dataset`` reads.
-    """
-    constants = pytest.importorskip("huggingface_hub").constants
-
-    cache = tmp_path / "hub"
-    repo_dir = cache / "datasets--liarliar--Daily-Omni"
-    snapshot = repo_dir / "snapshots" / "deadbeef"
-    snapshot.mkdir(parents=True)
-    if with_qa:
-        _write_qa_json(snapshot)
-    (repo_dir / "refs").mkdir()
-    (repo_dir / "refs" / "main").write_text("deadbeef", encoding="utf-8")
-
-    monkeypatch.setattr(constants, "HF_HUB_CACHE", str(cache))
-    monkeypatch.setattr(constants, "HF_HUB_OFFLINE", True)
-    return snapshot
-
-
 def test_resolve_local_root_returns_none_for_hub_id() -> None:
    assert resolve_daily_omni_local_root("liarliar/Daily-Omni") is None
    assert resolve_daily_omni_local_root("") is None
@@ -116,37 +90,3 @@ def test_missing_qa_json_reports_none(tmp_path: Path) -> None:
    root = resolve_daily_omni_local_root(str(tmp_path))
    assert daily_omni_local_qa_json(root) is None
    assert daily_omni_local_videos_dir(root) is None
-
-
-def test_hub_id_resolves_to_offline_cache(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
-    snapshot = _stage_offline_hub_cache(tmp_path, monkeypatch)
-
-    root = ensure_daily_omni_hub_root(_HUB_ID)
-
-    assert root == snapshot.resolve()
-    assert daily_omni_local_qa_json(root) == (snapshot / "qa.json").resolve()
-
-
-def test_dataset_loads_hub_id_without_network(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
-    """The regression this file exists for: a Hub id must not require ``load_dataset``."""
-    snapshot = _stage_offline_hub_cache(tmp_path, monkeypatch)
-    # Any use of `datasets` here means we went back to the path that cannot read the hub cache.
-    monkeypatch.setattr("vllm_omni.benchmarks.data_modules.daily_omni_dataset.load_dataset", None)
-
-    dataset = DailyOmniDataset(dataset_path=_HUB_ID, random_seed=0)
-
-    assert dataset.qa_json_path == (snapshot / "qa.json").resolve()
-    assert [row["video_id"] for row in dataset.data] == ["vid0"]
-
-
-def test_uncached_hub_id_reports_both_attempts(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
-    _stage_offline_hub_cache(tmp_path, monkeypatch, with_qa=False)
-    monkeypatch.setattr("vllm_omni.benchmarks.data_modules.daily_omni_dataset.load_dataset", None)
-
-    with pytest.raises(RuntimeError) as excinfo:
-        DailyOmniDataset(dataset_path="someone-else/Daily-Omni", random_seed=0)
-
-    message = str(excinfo.value)
-    assert "someone-else/Daily-Omni" in message
-    assert "snapshot_download failed" in message
-    assert "--daily-omni-qa-json" in message
@@ -1102,13 +1102,6 @@ class TestResolveScheduler:


 class TestDeployConfigLoading:
-    def test_rejects_legacy_stage_args_schema(self, tmp_path):
-        deploy_path = tmp_path / "legacy.yaml"
-        deploy_path.write_text("stage_args:\n  - stage_id: 0\n", encoding="utf-8")
-
-        with pytest.raises(ValueError, match=r"stage_args.*PipelineConfig.*stages"):
-            load_deploy_config(deploy_path)
-
    def test_load_minicpmo_duplex_deploy_config(self):
        deploy_path = Path(get_deploy_config_path("minicpmo_4_5_duplex.yaml"))

@@ -1384,8 +1377,7 @@ stages:
        assert async_stages[1].custom_process_input_func is None

    def test_no_bundled_legacy_stage_config_yamls(self):
-        repo_root = Path(__file__).resolve().parents[2]
-        stage_config_dir = repo_root / "vllm_omni" / "model_executor" / "stage_configs"
+        stage_config_dir = Path(__file__).parent.parent / "vllm_omni" / "model_executor" / "stage_configs"
        assert not list(stage_config_dir.glob("*.yaml"))

    def test_merge_pipeline_deploy(self):
@@ -1,119 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
-"""Guard the NPU pytest circular import through DiffusionOutput."""
-
-from __future__ import annotations
-
-import ast
-import os
-import subprocess
-import sys
-from pathlib import Path
-
-import pytest
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-REPO_ROOT = Path(__file__).resolve().parents[2]
-
-
-def _top_level_import_modules(path: Path) -> set[str]:
-    tree = ast.parse(path.read_text(encoding="utf-8"))
-    modules: set[str] = set()
-    for node in tree.body:
-        if isinstance(node, ast.ImportFrom) and node.module:
-            modules.add(node.module)
-        elif isinstance(node, ast.Import):
-            modules.update(alias.name for alias in node.names)
-    return modules
-
-
-def _assert_isolated_import_succeeds(script: str) -> None:
-    env = os.environ.copy()
-    env.pop("CUDA_VISIBLE_DEVICES", None)
-    env.pop("HIP_VISIBLE_DEVICES", None)
-    result = subprocess.run(
-        [sys.executable, "-c", script],
-        cwd=REPO_ROOT,
-        env=env,
-        capture_output=True,
-        text=True,
-        timeout=120,
-        check=False,
-    )
-    assert result.returncode == 0, result.stderr
-
-
-def test_config_package_does_not_eagerly_import_pipeline_registry() -> None:
-    imported = _top_level_import_modules(REPO_ROOT / "vllm_omni" / "config" / "__init__.py")
-    assert "vllm_omni.config.config_factory" not in imported
-    assert "vllm_omni.config.pipeline_registry" not in imported
-
-
-def test_config_package_lazy_exports_still_resolve() -> None:
-    from vllm_omni.config import StageConfigFactory, register_pipeline
-    from vllm_omni.config.config_factory import StageConfigFactory as DirectStageConfigFactory
-    from vllm_omni.config.pipeline_registry import register_pipeline as direct_register_pipeline
-
-    assert StageConfigFactory is DirectStageConfigFactory
-    assert register_pipeline is direct_register_pipeline
-
-
-def test_npu_platform_defers_minimax_h3_encoder_patch() -> None:
-    src = (REPO_ROOT / "vllm_omni" / "platforms" / "npu" / "platform.py").read_text(encoding="utf-8")
-    init_start = src.index("def __init__(self) -> None:")
-    runtime_start = src.index("def init_diffusion_model_runner_runtime")
-    init_src = src[init_start:runtime_start]
-    runtime_src = src[runtime_start:]
-    assert "apply_minimax_h3_qwen3vl_patch" not in init_src
-    assert "apply_minimax_h3_qwen3vl_patch" in runtime_src
-    assert "apply_minimax_h3_qwen3vl_swiglu_patch" in runtime_src
-
-
-def test_lora_config_import_does_not_load_pipeline_registry() -> None:
-    """Import LoRAConfig and diffusion.data in a fresh interpreter.
-
-    The NPU failure was ``data.py`` importing LoRAConfig while still loading,
-    then ``vllm_omni.config`` re-entering ``pipeline_registry`` → PI0 →
-    DiffusionOutput. This child process must not preload those exports.
-    """
-    _assert_isolated_import_succeeds(
-        """
-import sys
-
-from vllm_omni.config.lora import LoRAConfig
-
-if "vllm_omni.config.pipeline_registry" in sys.modules:
-    raise SystemExit("LoRAConfig import loaded pipeline_registry")
-if "vllm_omni.diffusion.models.pi0.pipeline_pi0" in sys.modules:
-    raise SystemExit("LoRAConfig import loaded PI0_PIPELINE")
-
-from vllm_omni.diffusion.data import DiffusionOutput
-
-if DiffusionOutput is None or LoRAConfig is None:
-    raise SystemExit("expected LoRAConfig and DiffusionOutput")
-"""
-    )
-
-
-def test_pipeline_registry_does_not_import_pi0_runtime() -> None:
-    """Resolving the pi0 topology must not import its runtime pipeline.
-
-    ``pipeline_pi0`` imports ``diffusion.data``. Importing it while
-    ``diffusion.data`` is only partially initialized closes the Ascend plugin's
-    circular-import loop, so the registry must depend only on a lightweight
-    pipeline-config module.
-    """
-    _assert_isolated_import_succeeds(
-        """
-import sys
-
-from vllm_omni.config.pipeline_registry import resolve_pipeline_config
-
-if "vllm_omni.diffusion.models.pi0.pipeline_pi0" in sys.modules:
-    raise SystemExit("pipeline_registry imported the pi0 runtime pipeline")
-
-pipeline = resolve_pipeline_config("pi0")
-if pipeline is None or pipeline.model_type != "pi0":
-    raise SystemExit("expected the registered pi0 pipeline config")
-"""
-    )
@@ -5,20 +5,11 @@
 from __future__ import annotations

 from dataclasses import fields
-from inspect import Parameter, signature
 from pathlib import Path

-import msgspec
 import pytest
 from pydantic import ValidationError
-from pydantic.fields import FieldInfo
 from transformers import Qwen3OmniMoeConfig
-from vllm.config import CacheConfig as VllmCacheConfig
-from vllm.config import CompilationConfig as VllmCompilationConfig
-from vllm.config import LoadConfig as VllmLoadConfig
-from vllm.config import ParallelConfig as VllmParallelConfig
-from vllm.config import ProfilerConfig as VllmProfilerConfig
-from vllm.config import SchedulerConfig as VllmSchedulerConfig

 from tests.helpers.stage_config import get_deploy_config_path
 from vllm_omni.config import omni_config as omni_config_module
@@ -49,7 +40,6 @@ from vllm_omni.config.stage_config import (
    merge_pipeline_deploy,
 )
 from vllm_omni.diffusion.diffusion_kv.config import DiffusionKVCacheMode
-from vllm_omni.engine.stage_engine_startup import _serialize_stage_config
 from vllm_omni.engine.stage_init_utils import build_legacy_engine_args_dict

 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -253,29 +243,9 @@ def test_from_pipeline_config_rejects_unowned_deploy_engine_extras(engine_extras
    ],
    ids=["global", "ar-stage", "generation-stage"],
 )
-def test_from_pipeline_config_accepts_upstream_cache_dtype_for_llm_stages(cli_overrides, stage_id):
-    stage = _from_pipeline_key("qwen3_tts", cli_overrides=cli_overrides).stage_by_id(stage_id)
-
-    assert stage.cache_config.cache_dtype == "fp8"
-    assert "cache_dtype" in stage.cache_config._omni_explicit_fields
-
-
-@pytest.mark.parametrize("field_name", ["data_parallel_master_ip", "data_parallel_address"])
-def test_from_pipeline_config_normalizes_nested_parallel_config_aliases(field_name):
-    pipeline = _resolve_pipeline_or_skip("qwen3_tts")
-    deploy = DeployConfig(
-        stages=[
-            StageDeployConfig(
-                stage_id=0,
-                engine_extras={"parallel_config": {field_name: "10.0.0.1"}},
-            )
-        ]
-    )
-
-    stage = VllmOmniConfig.from_pipeline_config(pipeline, user_deploy_config=deploy).stage_by_id(0)
-
-    assert stage.parallel_config.data_parallel_master_ip == "10.0.0.1"
-    assert "data_parallel_master_ip" in stage.parallel_config._omni_explicit_fields
+def test_from_pipeline_config_rejects_diffusion_only_cli_fields_for_llm_stages(cli_overrides, stage_id):
+    with pytest.raises(ValueError, match=rf"Stage {stage_id} .*no structured config owner: kv_cache_dtype"):
+        _from_pipeline_key("qwen3_tts", cli_overrides=cli_overrides)


 def test_from_pipeline_config_accepts_diffusion_only_cli_fields_for_diffusion_stage():
@@ -467,8 +437,6 @@ def test_vllm_omni_stage_config_public_fields_use_typed_stage_realizations():
        "connector_config",
        "runtime_config",
        "parallel_config",
-        "compilation_config",
-        "profiler_config",
        "quantization_config",
    }
    assert "diffusion_config" not in public_fields
@@ -487,47 +455,10 @@ def test_runtime_config_fields_match_structured_runtime_scope():
        "num_gpus",
        "log_level",
        "log_stats",
+        "profiler_config",
    }


-def test_upstream_config_compatible_fields_materialize_mapping_inputs():
-    compilation_config = {"backend": "eager"}
-    profiler_config = {"profiler": "cuda"}
-
-    stage_config = VllmOmniARStageConfig(
-        stage_pipeline_config=_resolve_pipeline_or_skip("qwen3_tts").stages[0],
-        compilation_config=compilation_config,
-        profiler_config=profiler_config,
-    )
-
-    assert isinstance(stage_config.compilation_config, VllmCompilationConfig)
-    assert stage_config.compilation_config.backend == "eager"
-    assert isinstance(stage_config.profiler_config, VllmProfilerConfig)
-    assert stage_config.profiler_config.profiler == "cuda"
-
-
-def test_upstream_config_compatible_fields_validate_mapping_inputs():
-    with pytest.raises(ValidationError):
-        VllmOmniARStageConfig(
-            stage_pipeline_config=_resolve_pipeline_or_skip("qwen3_tts").stages[0],
-            profiler_config={"delay_iterations": -1},
-        )
-
-
-def test_upstream_config_compatible_fields_keep_prebuilt_config_types():
-    compilation_config = VllmCompilationConfig(backend="eager")
-    profiler_config = VllmProfilerConfig(profiler="cuda")
-
-    stage_config = VllmOmniARStageConfig(
-        stage_pipeline_config=_resolve_pipeline_or_skip("qwen3_tts").stages[0],
-        compilation_config=compilation_config,
-        profiler_config=profiler_config,
-    )
-
-    assert stage_config.compilation_config is compilation_config
-    assert stage_config.profiler_config is profiler_config
-
-
 def test_sub_config_fields_match_structured_scopes():
    assert {f.name for f in fields(OmniStageModelConfig)} == {
        "model",
@@ -558,6 +489,7 @@ def test_sub_config_fields_match_structured_scopes():
        "enforce_eager",
        "max_cudagraph_capture_size",
        "enable_flashinfer_autotune",
+        "compilation_config",
        "enable_multithread_weight_load",
        "num_weight_load_threads",
        "disable_autocast",
@@ -567,25 +499,29 @@ def test_sub_config_fields_match_structured_scopes():
        "model_subdir",
        "tokenizer_subdir",
    }
-    vllm_load_fields = {f.name for f in fields(VllmLoadConfig)}
-    assert issubclass(OmniStageLoadConfig, VllmLoadConfig)
-    assert {f.name for f in fields(OmniStageLoadConfig)} == vllm_load_fields | {
+    assert {f.name for f in fields(OmniStageLoadConfig)} == {
        "tokenizer",
+        "download_dir",
        "skip_tokenizer_init",
+        "load_format",
        "tokenizer_mode",
        "config_format",
        "skip_mm_profiling",
    }
-    assert OmniStageLoadConfig(load_format="PT").load_format == "pt"
-    assert issubclass(OmniStageCacheConfig, VllmCacheConfig)
-    assert {f.name for f in fields(OmniStageCacheConfig)} == {f.name for f in fields(VllmCacheConfig)} | {
+    assert {f.name for f in fields(OmniStageCacheConfig)} == {
+        "kv_cache_memory_bytes",
+        "gpu_memory_utilization",
+        "enable_prefix_caching",
        "disable_hybrid_kv_cache_manager",
        "mm_processor_cache_gb",
        "mamba_ssm_cache_dtype",
    }
-    assert issubclass(OmniStageSchedulerConfig, VllmSchedulerConfig)
-    assert {f.name for f in fields(OmniStageSchedulerConfig)} == {f.name for f in fields(VllmSchedulerConfig)} | {
+    assert {f.name for f in fields(OmniStageSchedulerConfig)} == {
+        "max_num_seqs",
+        "max_num_batched_tokens",
        "max_model_len",
+        "enable_chunked_prefill",
+        "async_scheduling",
    }
    assert {f.name for f in fields(OmniStageConnectorConfig)} == {
        "async_chunk",
@@ -594,10 +530,17 @@ def test_sub_config_fields_match_structured_scopes():
        "output_connectors",
        "input_connectors",
    }
-    vllm_parallel_fields = {f.name for f in fields(VllmParallelConfig)}
-    assert issubclass(OmniStageParallelConfig, VllmParallelConfig)
-    assert {f.name for f in fields(OmniStageParallelConfig)} == vllm_parallel_fields
-    assert {f.name for f in fields(OmniStageDiffusionParallelConfig)} == vllm_parallel_fields | {
+    assert {f.name for f in fields(OmniStageParallelConfig)} == {
+        "pipeline_parallel_size",
+        "data_parallel_size",
+        "tensor_parallel_size",
+        "enable_expert_parallel",
+        "world_size",
+    }
+    assert {f.name for f in fields(OmniStageDiffusionParallelConfig)} == {
+        "pipeline_parallel_size",
+        "data_parallel_size",
+        "tensor_parallel_size",
        "sequence_parallel_size",
        "ulysses_degree",
        "ring_degree",
@@ -611,86 +554,20 @@ def test_sub_config_fields_match_structured_scopes():
        "mask_sp_padding",
        "hsdp_shard_size",
        "hsdp_replicate_size",
+        "enable_expert_parallel",
+        "world_size",
    }


-def test_inherited_sub_configs_initialize_transport_safe_derived_fields():
-    scheduler_config = OmniStageSchedulerConfig(max_num_batched_tokens=4096)
-    assert scheduler_config.max_num_encoder_input_tokens == 4096
-    assert scheduler_config.encoder_cache_size == 4096
-
-    unresolved_scheduler_config = OmniStageSchedulerConfig()
-    assert unresolved_scheduler_config.max_num_encoder_input_tokens is None
-    assert unresolved_scheduler_config.encoder_cache_size is None
-
-    parallel_config = OmniStageParallelConfig(data_parallel_size=3, data_parallel_rank=2)
-    assert parallel_config.data_parallel_index == 2
-
-    diffusion_parallel_config = OmniStageDiffusionParallelConfig(data_parallel_size=3, data_parallel_rank=2)
-    assert diffusion_parallel_config.data_parallel_index == 2
-
-
-@pytest.mark.parametrize(
-    "config_cls",
-    [
-        OmniStageLoadConfig,
-        OmniStageCacheConfig,
-        OmniStageSchedulerConfig,
-        OmniStageParallelConfig,
-        OmniStageDiffusionParallelConfig,
-    ],
-)
-def test_inherited_sub_configs_are_keyword_only(config_cls):
-    assert all(parameter.kind is Parameter.KEYWORD_ONLY for parameter in signature(config_cls).parameters.values())
-
-    with pytest.raises(TypeError):
-        config_cls(1)
-
-
-@pytest.mark.parametrize(
-    "config_cls",
-    [
-        OmniStageLoadConfig,
-        OmniStageCacheConfig,
-        OmniStageSchedulerConfig,
-        OmniStageParallelConfig,
-        OmniStageDiffusionParallelConfig,
-    ],
-)
-def test_inherited_sub_configs_remain_msgpack_transport_safe(config_cls):
-    config = config_cls()
-
-    uninitialized_fields = [
-        config_field.name
-        for config_field in fields(config)
-        if isinstance(getattr(config, config_field.name), FieldInfo)
-    ]
-    assert uninitialized_fields == []
-
-    msgspec.msgpack.encode(_serialize_stage_config(config))
-
-
-def test_structured_llm_stage_registration_payloads_remain_msgpack_transport_safe():
-    omni_config = _from_pipeline_key("qwen3_tts")
-
-    for stage_config in omni_config.stage_configs:
-        payload = {
-            "stage_id": stage_config.stage_id,
-            "stage_config": _serialize_stage_config(stage_config),
-        }
-        msgspec.msgpack.encode(payload)
-
-
 def test_diffusion_parallel_config_fields_cover_legacy_surface():
    from vllm_omni.diffusion.data import DiffusionParallelConfig

    legacy_fields = {f.name for f in fields(DiffusionParallelConfig)}
    structured_fields = {f.name for f in fields(OmniStageDiffusionParallelConfig)}
    expected_upstream_fields = {"mask_sp_padding"}
-    vllm_parallel_fields = {f.name for f in fields(VllmParallelConfig)}

    assert legacy_fields | expected_upstream_fields <= structured_fields
-    assert structured_fields - legacy_fields - expected_upstream_fields - vllm_parallel_fields == set()
+    assert structured_fields - legacy_fields - expected_upstream_fields == {"world_size"}


 def test_diffusion_parallel_config_keeps_current_diffusion_parallel_surface():
@@ -908,52 +785,6 @@ def test_structured_diffusion_config_rejects_invalid_compile_granularity():
        omni_config_module._DiffusionConfigProjection(diffusion_compile_granularity="block")


-@pytest.mark.parametrize(
-    ("enable_dlo", "use_allgather"),
-    [(False, False), (True, True)],
-)
-@pytest.mark.parametrize(
-    ("field_name", "field_value"),
-    [
-        ("dlo_use_host_weight_cache", True),
-        ("dlo_host_weight_cache_pin_limit_gib", 1.0),
-    ],
-)
-def test_structured_diffusion_config_restricts_host_weight_cache_to_no_allgather_dlo(
-    enable_dlo,
-    use_allgather,
-    field_name,
-    field_value,
-):
-    with pytest.raises(ValueError, match="requires no-AllGather distributed layerwise offload"):
-        omni_config_module._DiffusionConfigProjection(
-            enable_distributed_layerwise_offload=enable_dlo,
-            dlo_use_allgather=use_allgather,
-            **{field_name: field_value},
-        )
-
-
-def test_structured_diffusion_config_requires_cache_for_registration_ceiling():
-    with pytest.raises(ValueError, match="requires dlo_use_host_weight_cache=True"):
-        omni_config_module._DiffusionConfigProjection(
-            enable_distributed_layerwise_offload=True,
-            dlo_use_allgather=False,
-            dlo_host_weight_cache_pin_limit_gib=1.0,
-        )
-
-
-@pytest.mark.parametrize(
-    "field_name",
-    [
-        "dlo_host_weight_cache_lock_timeout",
-        "dlo_host_weight_cache_pin_limit_gib",
-    ],
-)
-def test_structured_diffusion_config_rejects_non_finite_host_weight_cache_values(field_name):
-    with pytest.raises(ValueError, match=rf"{field_name} must be finite"):
-        omni_config_module._DiffusionConfigProjection(**{field_name: float("inf")})
-
-
 def test_from_pipeline_config_matches_stage_config_to_omegaconf_behavior_for_representative_stage():
    pipeline = _resolve_pipeline_or_skip("qwen3_tts")
    legacy_stage = merge_pipeline_deploy(pipeline, _load_default_deploy(pipeline))[0]
@@ -1247,21 +1078,3 @@ def test_diffusion_config_projection_keeps_mapping_quantization_config_serializa
    cfg = omni_config_module._DiffusionConfigProjection.from_kwargs(quantization_config=quantization_config)

    assert cfg.quantization_config == quantization_config
-
-
-def test_diffusion_quantization_mapping_reaches_terminal_config(monkeypatch):
-    from vllm_omni.diffusion.data import OmniDiffusionConfig
-
-    quantization_config = {"method": "int8", "activation_scheme": "dynamic"}
-    cfg = omni_config_module._DiffusionConfigProjection.from_kwargs(
-        quantization_config=quantization_config,
-    )
-
-    # Exercise the terminal construction performed by the future typed startup
-    # path without probing ports or loading remote model metadata.
-    monkeypatch.setattr(OmniDiffusionConfig, "_resolve_master_port", lambda _self: 29500)
-    monkeypatch.setattr(OmniDiffusionConfig, "enrich_config", lambda _self: None)
-    cfg.enrich_config()
-
-    assert cfg.quantization_config is not None
-    assert cfg.quantization_config.get_name() == "int8"
@@ -4,10 +4,7 @@ from types import SimpleNamespace
 import pytest
 from vllm.v1.engine import FinishReason

-from vllm_omni.core.sched.omni_scheduler_mixin import (
-    DEFAULT_INPUT_WAIT_TIMEOUT_S,
-    OmniSchedulerMixin,
-)
+from vllm_omni.core.sched.omni_scheduler_mixin import OmniSchedulerMixin
 from vllm_omni.core.sched.output import OmniChunkRecvHandle

 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -25,25 +22,13 @@ def test_schedule_lifecycle_helpers_process_and_restore_both_input_paths():
    scheduler.requests = {"request": object()}
    scheduler._consume_pending_connector_output = lambda mode: calls.append(("consume", mode))
    scheduler._process_pending_input_timeouts = lambda: calls.append(("timeouts",))
-
-    def _collect_timed_out(timeout_s):
-        calls.append(("chunk-timeouts", timeout_s))
-        return set()
-
-    def _collect_failed_sends():
-        calls.append(("failed-sends",))
-        return {}
-
    scheduler.chunk_transfer_adapter = SimpleNamespace(
-        receives_chunks=True,
        process_pending_chunks=lambda waiting, running, scheduler_requests: calls.append(
            ("process", waiting, running, scheduler_requests)
        ),
        restore_queues=lambda waiting, running, scheduler_requests: calls.append(
            ("restore-chunks", waiting, running, scheduler_requests)
        ),
-        collect_timed_out_request_ids=_collect_timed_out,
-        collect_failed_send_request_ids=_collect_failed_sends,
    )
    scheduler.input_coordinator = SimpleNamespace(
        restore_queues=lambda waiting: calls.append(("restore-full", waiting))
@@ -56,10 +41,6 @@ def test_schedule_lifecycle_helpers_process_and_restore_both_input_paths():
        ("consume", "ar"),
        ("timeouts",),
        ("process", scheduler.waiting, scheduler.running, scheduler.requests),
-        # The chunk deadline runs after chunks are applied, so a chunk that
-        # arrived this cycle resets the clock before it is measured (R1.1).
-        ("chunk-timeouts", DEFAULT_INPUT_WAIT_TIMEOUT_S),
-        ("failed-sends",),
        ("restore-chunks", scheduler.waiting, scheduler.running, scheduler.requests),
        ("restore-full", scheduler.waiting),
    ]
@@ -1,10 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Unit coverage for the two stage-input timeout hooks.
-
-``_process_pending_input_timeouts`` covers the full-payload transport;
-``_process_pending_chunk_timeouts`` covers async-chunk. Both share
-``VLLM_OMNI_INPUT_WAIT_TIMEOUT_S``.
+"""Unit coverage for _process_pending_input_timeouts.

 Verifies that the mixin correctly *delegates* timed-out requests to the
 base scheduler's ``finish_requests`` API with ``RequestStatus.FINISHED_ERROR``.
@@ -16,10 +12,6 @@ this file only asserts the wiring from the mixin to that API.

 from __future__ import annotations

-import importlib
-import logging
-import math
-import os
 from types import SimpleNamespace

 import pytest
@@ -110,248 +102,5 @@ def test_process_pending_input_timeouts_disabled_when_timeout_zero(monkeypatch):
    assert scheduler.finish_calls == []


-class _FakeChunkAdapter:
-    """Stands in for OmniChunkTransferAdapter's timeout-collection surface."""
-
-    def __init__(self, timed_out_ids, *, receives_chunks=True):
-        self._timed_out_ids = set(timed_out_ids)
-        self.receives_chunks = receives_chunks
-        self.calls = []
-
-    def collect_timed_out_request_ids(self, timeout_s):
-        self.calls.append(timeout_s)
-        return set(self._timed_out_ids)
-
-
-class _FakeChunkScheduler(OmniSchedulerMixin):
-    def __init__(self, requests, adapter):
-        self.requests = requests
-        self.chunk_transfer_adapter = adapter
-        self.finish_calls = []
-
-    def finish_requests(self, req_ids, status):
-        self.finish_calls.append((set(req_ids), status))
-
-
-def test_process_pending_chunk_timeouts_delegates_to_finish_requests():
-    """A request stalled in WAITING_FOR_CHUNK is failed, not left parked.
-
-    This is the async-chunk half of the safety net (RFC #4855 R1.1); before it
-    existed, the chunk path had no deadline at all and issue #3833 could park a
-    request forever.
-    """
-    req_id = "stalled-stream"
-    adapter = _FakeChunkAdapter(timed_out_ids={req_id})
-    scheduler = _FakeChunkScheduler({req_id: SimpleNamespace(request_id=req_id)}, adapter)
-
-    scheduler._process_pending_chunk_timeouts()
-
-    assert len(adapter.calls) == 1
-    assert adapter.calls[0] > 0
-    assert len(scheduler.finish_calls) == 1
-    finished_ids, status = scheduler.finish_calls[0]
-    assert finished_ids == {req_id}
-    assert getattr(status, "name", str(status)).endswith("FINISHED_ERROR")
-
-
-def test_process_pending_chunk_timeouts_skips_already_freed_request():
-    adapter = _FakeChunkAdapter(timed_out_ids={"already-freed"})
-    scheduler = _FakeChunkScheduler(requests={}, adapter=adapter)
-
-    scheduler._process_pending_chunk_timeouts()
-
-    assert scheduler.finish_calls == []
-
-
-def test_process_pending_chunk_timeouts_noop_on_a_producer_stage():
-    """Stage 0 does not receive chunks, so it has nothing to expire."""
-    adapter = _FakeChunkAdapter(timed_out_ids={"r1"}, receives_chunks=False)
-    scheduler = _FakeChunkScheduler({"r1": SimpleNamespace(request_id="r1")}, adapter)
-
-    scheduler._process_pending_chunk_timeouts()
-
-    assert adapter.calls == []
-    assert scheduler.finish_calls == []
-
-
-def test_process_pending_chunk_timeouts_noop_without_adapter():
-    class _NoAdapter(OmniSchedulerMixin):
-        def __init__(self):
-            self.requests = {}
-            self.chunk_transfer_adapter = None
-            self.finish_calls = []
-
-        def finish_requests(self, req_ids, status):
-            self.finish_calls.append((set(req_ids), status))
-
-    scheduler = _NoAdapter()
-    scheduler._process_pending_chunk_timeouts()
-    assert scheduler.finish_calls == []
-
-
-def test_process_pending_chunk_timeouts_disabled_when_timeout_zero(monkeypatch):
-    """The shared knob disables both transports' nets together."""
-    from vllm_omni.core.sched import omni_scheduler_mixin
-
-    monkeypatch.setattr(omni_scheduler_mixin, "DEFAULT_INPUT_WAIT_TIMEOUT_S", 0.0)
-
-    adapter = _FakeChunkAdapter(timed_out_ids={"r1"})
-    scheduler = _FakeChunkScheduler({"r1": SimpleNamespace(request_id="r1")}, adapter)
-    scheduler._process_pending_chunk_timeouts()
-
-    assert adapter.calls == []
-    assert scheduler.finish_calls == []
-
-
-class _FakeSendFailureAdapter:
-    """Stands in for the adapter's drained send-failure map (R1.2)."""
-
-    def __init__(self, failures):
-        self._failures = dict(failures)
-        self.collect_calls = 0
-
-    def collect_failed_send_request_ids(self):
-        self.collect_calls += 1
-        failures, self._failures = self._failures, {}
-        return failures
-
-
-class _FakeSendFailureScheduler(OmniSchedulerMixin):
-    def __init__(self, requests, adapter):
-        self.requests = requests
-        self.chunk_transfer_adapter = adapter
-
-
-def test_log_failed_chunk_sends_reports_a_live_request(caplog):
-    adapter = _FakeSendFailureAdapter({"req-live": "OSError: No space left on device"})
-    scheduler = _FakeSendFailureScheduler({"req-live": SimpleNamespace(request_id="req-live")}, adapter)
-
-    with caplog.at_level(logging.ERROR):
-        scheduler._log_failed_chunk_sends()
-
-    assert "req-live" in caplog.text
-    assert "No space left on device" in caplog.text
-
-
-def test_log_failed_chunk_sends_still_reports_a_departed_request(caplog):
-    """``collect_*`` drains the map, so a failure for a request that already left
-    this scheduler is the one place a drop could still go unrecorded."""
-    adapter = _FakeSendFailureAdapter({"req-gone": "OSError: No space left on device"})
-    scheduler = _FakeSendFailureScheduler({}, adapter)
-
-    with caplog.at_level(logging.WARNING):
-        scheduler._log_failed_chunk_sends()
-
-    assert adapter.collect_calls == 1
-    assert "req-gone" in caplog.text
-    assert "No space left on device" in caplog.text
-    assert "finished or aborted" in caplog.text
-
-
-def test_log_failed_chunk_sends_noop_without_adapter():
-    scheduler = _FakeSendFailureScheduler({}, None)
-    scheduler._log_failed_chunk_sends()
-
-
-# --- R1.4: VLLM_OMNI_INPUT_WAIT_TIMEOUT_S is parsed at import time, so each of
-# these reloads the module under a different environment. One knob now gates the
-# deadline on both transports, which is why a silent 0 matters more than it did
-# when only the full-payload path used it.
-
-
-@pytest.fixture
-def reload_with(monkeypatch):
-    """Reload the mixin module under a given env value, then put it back.
-
-    ``importlib.reload`` re-executes into the live module dict, so whatever the
-    last test parsed is what every later test in the session sees -- ``-k
-    zero_warns`` alone would leave ``DEFAULT_INPUT_WAIT_TIMEOUT_S = 0.0`` and
-    silently disable both deadlines. The rejection cases are worse: the
-    ValueError fires *after* the rebind, so ``nan`` lingers in the module.
-    ``monkeypatch`` restores the environment but not the module, so reload once
-    more on the way out.
-    """
-    import vllm_omni.core.sched.omni_scheduler_mixin as mod
-
-    def _reload(raw):
-        if raw is None:
-            monkeypatch.delenv("VLLM_OMNI_INPUT_WAIT_TIMEOUT_S", raising=False)
-        else:
-            monkeypatch.setenv("VLLM_OMNI_INPUT_WAIT_TIMEOUT_S", raw)
-        return importlib.reload(mod)
-
-    yield _reload
-
-    monkeypatch.undo()
-    importlib.reload(mod)
-
-
-@pytest.mark.parametrize(
-    "raw,expected",
-    [
-        (None, 600.0),
-        ("30", 30.0),
-        ("0.5", 0.5),
-        ("0", 0.0),  # explicit opt-out is still honoured
-    ],
-)
-def test_timeout_accepts_valid_values(reload_with, raw, expected):
-    mod = reload_with(raw)
-    assert mod.DEFAULT_INPUT_WAIT_TIMEOUT_S == expected
-
-
-@pytest.mark.parametrize("raw", ["-1", "-600", "-0.5"])
-def test_negative_timeout_is_rejected(reload_with, raw):
-    """`-1` is a common "no limit" idiom; substituting 600 would give an operator
-    who meant "never time out" mysterious failures ten minutes in. Fail loudly."""
-    with pytest.raises(ValueError, match="negative"):
-        reload_with(raw)
-
-
-@pytest.mark.parametrize("raw", ["nan", "inf", "-inf", "1e999"])
-def test_non_finite_timeout_is_rejected(reload_with, raw):
-    """nan compares false against every deadline check and inf is unreachable;
-    both disable the net while looking like a number."""
-    with pytest.raises(ValueError, match="not finite"):
-        reload_with(raw)
-
-
-def test_zero_warns_and_names_both_transports(reload_with, caplog):
-    with caplog.at_level(logging.WARNING):
-        mod = reload_with("0")
-    assert mod.DEFAULT_INPUT_WAIT_TIMEOUT_S == 0.0
-    text = caplog.text
-    assert "DISABLED" in text
-    assert "WAITING_FOR_INPUT" in text and "WAITING_FOR_CHUNK" in text
-
-
-def test_unparsable_timeout_falls_back(reload_with):
-    mod = reload_with("not-a-number")
-    assert mod.DEFAULT_INPUT_WAIT_TIMEOUT_S == 600.0
-
-
-def test_a_rejected_value_does_not_outlive_its_test(monkeypatch):
-    """The reason ``reload_with`` reloads on the way out, done here by hand.
-
-    ``nan`` is bound before the ValueError fires, so the reload that rejects it
-    still leaves it in the module -- and nan compares false against every
-    deadline check, which is exactly the silent-disable this item exists to
-    stop. Restoring the environment is not enough; the module has to be reloaded
-    too.
-    """
-    import vllm_omni.core.sched.omni_scheduler_mixin as mod
-
-    ambient = float(os.environ.get("VLLM_OMNI_INPUT_WAIT_TIMEOUT_S", "600"))
-
-    monkeypatch.setenv("VLLM_OMNI_INPUT_WAIT_TIMEOUT_S", "nan")
-    with pytest.raises(ValueError, match="not finite"):
-        importlib.reload(mod)
-    assert math.isnan(mod.DEFAULT_INPUT_WAIT_TIMEOUT_S)
-
-    monkeypatch.undo()
-    importlib.reload(mod)
-    assert mod.DEFAULT_INPUT_WAIT_TIMEOUT_S == ambient
-
-
 if __name__ == "__main__":
    pytest.main([__file__, "-v"])
@@ -10,7 +10,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Single-stage BAGEL (TP=1, CFG=1), t2i Task",
@@ -61,7 +61,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Single-stage BAGEL (TP=1, CFG=1), i2i Task",
@@ -113,7 +113,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Two-stage BAGEL (bagel.yaml), TP=1, CFG=1, t2i Task",
@@ -164,7 +164,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Two-stage BAGEL (bagel.yaml), TP=1, CFG=1, i2i Task",
@@ -10,7 +10,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "tts"
    ],
    "server_params": {
@@ -10,7 +10,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Single-device baseline with enforce-eager (no torch.compile)",
@@ -64,7 +64,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Single-device with torch.compile (default, no enforce-eager)",
@@ -117,7 +117,7 @@
          "num_cards": 2
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "CFG-parallel=2 with enforce-eager",
@@ -172,7 +172,7 @@
          "num_cards": 2
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "CFG-parallel=2 with torch.compile",
@@ -226,7 +226,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "CacheDiT with enforce-eager",
@@ -10,7 +10,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Single-device baseline (two input images)",
@@ -76,7 +76,7 @@
          "num_cards": 4
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Ulysses SP=2 + CFG=2 + VAE patch parallel=4",
@@ -125,7 +125,7 @@
          "num_cards": 4
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Ulysses SP=2 + CFG=2 + CacheDiT",
@@ -10,7 +10,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Single-device baseline",
@@ -73,7 +73,7 @@
          "num_cards": 4
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Ulysses SP=2 + CFG=2 + VAE patch parallel=4",
@@ -121,7 +121,7 @@
          "num_cards": 4
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Ulysses SP=2 + CFG=2 + CacheDiT",
@@ -10,7 +10,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "diffusion"
    ],
    "description": "Single-device baseline",
@@ -10,7 +10,7 @@
          "num_cards": 1
        }
      },
-      "local_model",
+      "full_model",
      "tts"
    ],
    "server_params": {
@@ -7,10 +7,12 @@ The detailed RFC document is maintained outside this repository (local/internal
 - Qwen3 tests: `test_reliability_qwen3_omni.py`
 - Wan2.2 tests: `test_reliability_wan22.py`
 - HunyuanImage tests: `test_reliability_hunyuan_image.py`
+- VoxCPM2 tests: `test_reliability_voxcpm2.py`

 ```bash
 pytest --collect-only tests/dfx/reliability
 pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m slow
 pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m slow
 pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m slow
+pytest -s -v tests/dfx/reliability/test_reliability_voxcpm2.py -m slow
 ```
@@ -976,68 +976,3 @@ def test_voices_delete_invalid_requests(
            "err_message": err_message,
        }
    )
-
-
-# ─── POST /v1/audio/speech · MiniMax Music 3 ───
-
-# Text-to-music on the speech endpoint. Most of the speech contract does not
-# apply: there is no speaker, no reference audio and no temperature, and the
-# lyrics alone are not a request. Each case below pins a rejection so a caller
-# who assumes the usual TTS shape finds out immediately.
-_MINIMAX_MUSIC3_SPEECH = [
-    pytest.param(
-        OmniServerParams(
-            model="MiniMaxAI/MiniMax-Music3",
-            stage_config_path=get_deploy_config_path("minimax_music3.yaml"),
-            server_args=_SPEECH_SERVER_ARGS,
-        ),
-        id="minimax_music3",
-        marks=hardware_marks(res={"cuda": "H100"}),
-    ),
-]
-
-_MINIMAX_MUSIC3_LYRICS = "[Verse]\nWalking down the empty street at midnight"
-_MINIMAX_MUSIC3_CAPTION = "A melancholic lo-fi hip-hop track at 85 BPM in F minor."
-# 25 frames per second, capped at six minutes of song.
-_MINIMAX_MUSIC3_MAX_FRAMES = 9000
-
-
-@pytest.mark.parametrize(
-    "overrides, err_message",
-    [
-        pytest.param({"instructions": None}, ("instructions",), id="caption_missing"),
-        pytest.param({"instructions": "   "}, ("instructions",), id="caption_blank"),
-        pytest.param({"input": "   "}, ("input",), id="lyrics_blank"),
-        pytest.param({"voice": "alloy"}, ("voice",), id="voice_rejected"),
-        pytest.param({"speed": 1.5}, ("speed",), id="speed_rejected"),
-        pytest.param(
-            {"max_new_tokens": _MINIMAX_MUSIC3_MAX_FRAMES + 1},
-            ("max_new_tokens", str(_MINIMAX_MUSIC3_MAX_FRAMES)),
-            id="max_new_tokens_over_cap",
-        ),
-        pytest.param({"max_new_tokens": 0}, ("max_new_tokens",), id="max_new_tokens_zero"),
-    ],
-)
-@pytest.mark.parametrize("omni_server", _MINIMAX_MUSIC3_SPEECH, indirect=True)
-def test_minimax_music3_speech_invalid_field_values(
-    omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
-    overrides: dict[str, object],
-    err_message: str | tuple[str, ...],
-) -> None:
-    body: dict[str, Any] = {
-        "model": omni_server.model,
-        "input": _MINIMAX_MUSIC3_LYRICS,
-        "instructions": _MINIMAX_MUSIC3_CAPTION,
-        "seed": 1,
-        "max_new_tokens": 250,
-        "response_format": "wav",
-    }
-    for key, value in overrides.items():
-        if value is None:
-            body.pop(key, None)
-        else:
-            body[key] = value
-    openai_client.send_audio_speech_http_request(
-        {"json": body, "timeout": 120, "err_code": 400, "err_message": err_message}
-    )
@@ -71,7 +71,7 @@ def _default_oom_device_spec() -> str:

 OOM_INJECTION_CONFIG = {
    "device": _default_oom_device_spec(),
-    "target_mem_ratio": 0.95,
+    "target_mem_ratio": 0.92,
    "hold_seconds": 0,
    "startup_timeout_sec": 20,
    "strict": False,
@@ -0,0 +1,508 @@
+"""
+VoxCPM2 reliability integration tests.
+
+Scenarios follow ``docs/user_guide/fault_injection_reliability_matrix.md``.
+"""
+
+from __future__ import annotations
+
+import concurrent.futures
+import os
+import time
+from pathlib import Path
+from typing import Any
+
+import pytest
+
+from tests.dfx.conftest import (
+    assert_fault_exception,
+    create_reliability_omni_server_params,
+    resolve_oom_device_spec,
+)
+from tests.dfx.reliability.helpers import (
+    FaultInjector,
+    assert_no_server_tree_process_residual_and_gpu_release,
+    extract_openai_error_contract_from_bytes,
+    get_health_raw,
+    inject_gpu_oom,
+    make_process_kill_fault_injector,
+    make_server_root_kill_fault_injector,
+    make_server_tree_kill_fault_injector,
+    post_json_raw,
+    run_fault_injection_with_rate_load,
+    stop_gpu_oom_hogs,
+    worker_residual_timeout_after_kill_signal,
+)
+from tests.helpers.mark import hardware_test
+from tests.helpers.media import load_test_audio_data_url
+from tests.helpers.runtime import OpenAIClientHandler
+
+RELIABILITY_SCENARIOS: list[dict[str, Any]] = [
+    {
+        "test_name": "voxcpm2_reliability_default",
+        "server_params": {
+            "model": "openbmb/VoxCPM2",
+            "stage_config_name": "voxcpm2.yaml",
+            # VoxCPM2 uses a custom HF tokenizer; CLI default trust_remote_code=False
+            # would otherwise override deploy YAML settings during server startup.
+            "server_args": ["--trust-remote-code", "--disable-log-stats"],
+        },
+    },
+]
+
+DEPLOY_CONFIGS_DIR = Path(__file__).resolve().parent.parent.parent.parent / "vllm_omni" / "deploy"
+OOM_INJECTION_CONFIG = {
+    "target_mem_ratio": 0.96,
+    "hold_seconds": 0,
+    "startup_timeout_sec": 20,
+    "strict": False,
+}
+FAULT_ERROR_KEYWORDS = (
+    "the request failed",
+    "oom",
+    "out of memory",
+    "cuda",
+    "orchestrator",
+    "timeout",
+    "connection",
+    "500",
+    "503",
+)
+PROCESS_KILL_ERROR_KEYWORDS = (
+    "timeout",
+    "did not complete within",
+    "connection",
+    "engine",
+    "orchestrator",
+    "dead",
+    "internal",
+    "500",
+    "503",
+)
+RUNTIME_WORKER_PATTERN = "VLLM::"
+SERVE_SIGNAL_PARAMS = [
+    pytest.param("SIGTERM", id="sigterm"),
+    pytest.param("SIGINT", id="sigint"),
+    pytest.param("SIGKILL", id="sigkill"),
+]
+SERVE_WITH_LOAD_SIGNAL_PARAMS = [
+    pytest.param("SIGTERM", id="sigterm"),
+    pytest.param("SIGINT", id="sigint"),
+    pytest.param("SIGKILL", id="sigkill"),
+]
+TREE_SIGNAL_PARAMS = [
+    pytest.param("SIGTERM", id="sigterm"),
+    pytest.param("SIGKILL", id="sigkill"),
+]
+TREE_WITH_LOAD_SIGNAL_PARAMS = [
+    pytest.param("SIGTERM", id="sigterm"),
+    pytest.param("SIGKILL", id="sigkill"),
+]
+WORKER_SIGNAL_FAULT_PARAMS = [
+    pytest.param(
+        make_process_kill_fault_injector(
+            grep_patterns=RUNTIME_WORKER_PATTERN,
+            signal_name="SIGTERM",
+            limit=1,
+            post_kill_wait_seconds=2.0,
+        ),
+        id="runtime_process_chain_sigterm",
+    ),
+    pytest.param(
+        make_process_kill_fault_injector(
+            grep_patterns=RUNTIME_WORKER_PATTERN,
+            signal_name="SIGKILL",
+            limit=1,
+            post_kill_wait_seconds=2.0,
+        ),
+        id="runtime_process_chain_sigkill",
+    ),
+]
+
+VOXCPM2_PARAMS = create_reliability_omni_server_params(RELIABILITY_SCENARIOS, DEPLOY_CONFIGS_DIR)
+INFLIGHT_INJECTION_REQUEST_RATE = 0.3
+INFLIGHT_INJECTION_REQUEST_COUNT = 10
+_SPEECH_FF_HTTP_TIMEOUT_SEC = 30
+_SPEECH_FF_MAX_ELAPSED_SEC = 30
+
+
+def _speech_request_config(omni_server: Any) -> dict[str, Any]:
+    return {
+        "model": omni_server.model,
+        "input": "The weather is nice today, perfect for a walk in the park.",
+        "stream": False,
+        "timeout": 120.0,
+        "response_format": "wav",
+        "voice": "default",
+    }
+
+
+def _submit_speech_request(openai_client: OpenAIClientHandler, omni_server: Any) -> None:
+    openai_client.send_audio_speech_request(_speech_request_config(omni_server), request_num=1)
+
+
+def _oom_speech_request_config(omni_server: Any) -> dict[str, Any]:
+    base = "The weather is nice today, perfect for a walk in the park. "
+    return {
+        "model": omni_server.model,
+        "input": base * 80,
+        "max_new_tokens": 4096,
+        "ref_audio": load_test_audio_data_url("qwen3_tts/clone_2.wav"),
+        "stream": False,
+        "timeout": 300.0,
+        "response_format": "wav",
+        "voice": "default",
+    }
+
+
+def _submit_oom_speech_request(openai_client: OpenAIClientHandler, omni_server: Any) -> None:
+    openai_client.send_audio_speech_request(_oom_speech_request_config(omni_server), request_num=1)
+
+
+def _assert_post_fault_speech_fast_fail(host: str, port: int, *, model: str, scenario: str) -> None:
+    payload = {
+        "model": model,
+        "input": "post-fault fast-fail check",
+        "voice": "default",
+    }
+    start = time.monotonic()
+    try:
+        status, body = post_json_raw(host, port, "/v1/audio/speech", payload, timeout_sec=20)
+        elapsed = time.monotonic() - start
+        assert elapsed < 15, f"[{scenario} fast_fail] /v1/audio/speech did not fail fast: {elapsed:.2f}s"
+        assert status >= 500, (
+            f"[{scenario} fast_fail] expected server-side error after fault, got status={status}, body={body[:200]!r}"
+        )
+    except Exception:  # noqa: BLE001
+        elapsed = time.monotonic() - start
+        assert elapsed < 15, f"[{scenario} fast_fail] exception was too slow after fault: {elapsed:.2f}s"
+
+
+def _assert_post_fault_health_terminal(host: str, port: int, *, scenario: str) -> None:
+    deadline = time.monotonic() + 20.0
+    last_observation = ""
+    while time.monotonic() < deadline:
+        try:
+            status, body = get_health_raw(host, port, timeout_sec=5)
+            last_observation = f"http={status}, body={body[:200]!r}"
+            if status == 503:
+                return
+        except Exception as exc:  # noqa: BLE001
+            last_observation = f"exception={exc!r}"
+            return
+        time.sleep(0.5)
+    pytest.fail(f"[{scenario} health] no terminal post-fault health observed: {last_observation}")
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@pytest.mark.skip(reason="issue#4285")
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@pytest.mark.parametrize("omni_server_function", VOXCPM2_PARAMS, indirect=True)
+def test_reliability_fault_gpu_oom_speech_request_failure(
+    omni_server_function,
+    openai_client_function,
+) -> None:
+    stage_config_path = getattr(omni_server_function, "stage_config_path", None)
+    device_spec = resolve_oom_device_spec(OOM_INJECTION_CONFIG, stage_config_path)
+    handle = inject_gpu_oom(
+        device=device_spec,
+        target_mem_ratio=OOM_INJECTION_CONFIG["target_mem_ratio"],
+        hold_seconds=OOM_INJECTION_CONFIG["hold_seconds"],
+        startup_timeout_sec=OOM_INJECTION_CONFIG["startup_timeout_sec"],
+        strict=OOM_INJECTION_CONFIG["strict"],
+    )
+    try:
+        try:
+            _submit_oom_speech_request(openai_client_function, omni_server_function)
+        except Exception as exc:
+            assert_fault_exception(exc, FAULT_ERROR_KEYWORDS)
+        else:
+            pytest.fail("expected /v1/audio/speech request failure during GPU OOM injection")
+    finally:
+        stop_gpu_oom_hogs(handle)
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@pytest.mark.skip(reason="issue#4285")
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@pytest.mark.parametrize("omni_server_function", VOXCPM2_PARAMS, indirect=True)
+def test_reliability_fault_gpu_oom_speech_error_contract(
+    omni_server_function,
+) -> None:
+    stage_config_path = getattr(omni_server_function, "stage_config_path", None)
+    device_spec = resolve_oom_device_spec(OOM_INJECTION_CONFIG, stage_config_path)
+    handle = inject_gpu_oom(
+        device=device_spec,
+        target_mem_ratio=OOM_INJECTION_CONFIG["target_mem_ratio"],
+        hold_seconds=OOM_INJECTION_CONFIG["hold_seconds"],
+        startup_timeout_sec=OOM_INJECTION_CONFIG["startup_timeout_sec"],
+        strict=OOM_INJECTION_CONFIG["strict"],
+    )
+    host = omni_server_function.host
+    port = omni_server_function.port
+    oom_config = _oom_speech_request_config(omni_server_function)
+    payload = {
+        "model": oom_config["model"],
+        "input": oom_config["input"],
+        "voice": oom_config["voice"],
+        "max_new_tokens": oom_config["max_new_tokens"],
+        "ref_audio": oom_config["ref_audio"],
+    }
+    try:
+        status, body = post_json_raw(host, port, "/v1/audio/speech", payload, timeout_sec=300)
+    finally:
+        stop_gpu_oom_hogs(handle)
+
+    assert status >= 500, f"expected speech runtime error under OOM, got status={status}"
+    speech_error = extract_openai_error_contract_from_bytes(body)
+    assert speech_error is not None, f"speech error payload not OpenAI-compatible: {body[:300]!r}"
+    assert "code" in speech_error, f"speech error lacks code field: {speech_error!r}"
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@pytest.mark.skipif(os.name == "nt", reason="process-kill injection helper is POSIX-only")
+@pytest.mark.parametrize(
+    "fault_injector",
+    WORKER_SIGNAL_FAULT_PARAMS,
+    indirect=True,
+)
+@pytest.mark.parametrize("omni_server_function", VOXCPM2_PARAMS, indirect=True)
+def test_reliability_fault_process_kill_request_failure(
+    omni_server_after_fault_function,
+    openai_client_function,
+) -> None:
+    try:
+        _submit_speech_request(openai_client_function, omni_server_after_fault_function)
+    except Exception as exc:
+        assert_fault_exception(exc, PROCESS_KILL_ERROR_KEYWORDS)
+    else:
+        pytest.fail("expected /v1/audio/speech request failure after process-kill injection")
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@pytest.mark.skipif(os.name == "nt", reason="process-kill injection helper is POSIX-only")
+@pytest.mark.parametrize(
+    "fault_injector",
+    WORKER_SIGNAL_FAULT_PARAMS,
+    indirect=True,
+)
+@pytest.mark.parametrize("omni_server_function", VOXCPM2_PARAMS, indirect=True)
+def test_reliability_fault_process_kill_fast_fail_and_concurrent(
+    omni_server_after_fault_function,
+    openai_client_function,
+) -> None:
+    host = omni_server_after_fault_function.host
+    port = omni_server_after_fault_function.port
+    model = omni_server_after_fault_function.model
+    payload = {
+        "model": model,
+        "input": "Say hello in one short sentence.",
+        "voice": "default",
+    }
+
+    start = time.monotonic()
+    try:
+        status, body = post_json_raw(
+            host,
+            port,
+            "/v1/audio/speech",
+            payload,
+            timeout_sec=_SPEECH_FF_HTTP_TIMEOUT_SEC,
+        )
+        elapsed = time.monotonic() - start
+        assert elapsed <= _SPEECH_FF_MAX_ELAPSED_SEC + 1.0, (
+            f"[process_kill fast_fail] /v1/audio/speech did not fail fast after fault: {elapsed:.2f}s"
+        )
+        assert status >= 500, (
+            "[process_kill fast_fail] expected server-side failure after fault, "
+            f"got status={status}, body={body[:200]!r}"
+        )
+    except Exception:
+        elapsed = time.monotonic() - start
+        assert elapsed <= _SPEECH_FF_MAX_ELAPSED_SEC + 1.0, (
+            f"[process_kill fast_fail] /v1/audio/speech exception was too slow after fault: {elapsed:.2f}s"
+        )
+
+    start = time.monotonic()
+    with concurrent.futures.ThreadPoolExecutor(max_workers=3) as executor:
+        futures = [
+            executor.submit(_submit_speech_request, openai_client_function, omni_server_after_fault_function)
+            for _ in range(3)
+        ]
+        done, pending = concurrent.futures.wait(
+            futures,
+            timeout=30,
+            return_when=concurrent.futures.ALL_COMPLETED,
+        )
+
+    elapsed = time.monotonic() - start
+    assert not pending, f"[process_kill concurrent] some fault-time speech requests hung: pending={len(pending)}"
+    assert elapsed < 30, f"[process_kill concurrent] fault-time speech request convergence is too slow: {elapsed:.2f}s"
+
+    fault_observed = False
+    for future in done:
+        try:
+            future.result()
+        except Exception as exc:
+            fault_observed = True
+            assert_fault_exception(exc, PROCESS_KILL_ERROR_KEYWORDS)
+    assert fault_observed, (
+        "[process_kill concurrent] expected at least one /v1/audio/speech request to fail after fault"
+    )
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@pytest.mark.skipif(os.name == "nt", reason="process-kill injection helper is POSIX-only")
+@pytest.mark.parametrize(
+    "fault_injector",
+    WORKER_SIGNAL_FAULT_PARAMS,
+    indirect=True,
+)
+@pytest.mark.parametrize("omni_server_function", VOXCPM2_PARAMS, indirect=True)
+def test_reliability_fault_process_kill_worker_with_load_request_failure(
+    omni_server_function,
+    openai_client_function,
+    fault_injector: FaultInjector,
+) -> None:
+    scenario = "kill_worker_with_load"
+    load_result = run_fault_injection_with_rate_load(
+        submit_request=lambda: _submit_speech_request(openai_client_function, omni_server_function),
+        inject_fault=lambda: fault_injector(omni_server_function),
+        num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
+        request_rate=INFLIGHT_INJECTION_REQUEST_RATE,
+        completion_timeout_sec=120.0,
+    )
+    assert load_result["failure_observed"], (
+        f"[{scenario}] expected at least one load request failure after fault; load_result={load_result}"
+    )
+    host = omni_server_function.host
+    port = omni_server_function.port
+    _assert_post_fault_speech_fast_fail(host, port, model=omni_server_function.model, scenario=scenario)
+    _assert_post_fault_health_terminal(host, port, scenario=scenario)
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@pytest.mark.skipif(os.name == "nt", reason="process-kill injection helper is POSIX-only")
+@pytest.mark.parametrize("signal_name", SERVE_SIGNAL_PARAMS)
+@pytest.mark.parametrize("omni_server_function", VOXCPM2_PARAMS, indirect=True)
+def test_reliability_fault_process_kill_serve_root_no_load_fast_fail_and_cleanup(
+    omni_server_function,
+    signal_name: str,
+) -> None:
+    scenario = f"kill_serve_root_no_load_{signal_name.lower()}"
+    injector = make_server_root_kill_fault_injector(signal_name=signal_name, post_kill_wait_seconds=2.0)
+    injector(omni_server_function)
+    host = omni_server_function.host
+    port = omni_server_function.port
+    _assert_post_fault_speech_fast_fail(host, port, model=omni_server_function.model, scenario=scenario)
+    _assert_post_fault_health_terminal(host, port, scenario=scenario)
+    assert_no_server_tree_process_residual_and_gpu_release(
+        omni_server_function,
+        scenario=scenario,
+        timeout_sec=worker_residual_timeout_after_kill_signal(signal_name),
+    )
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@pytest.mark.skipif(os.name == "nt", reason="process-kill injection helper is POSIX-only")
+@pytest.mark.parametrize("signal_name", SERVE_WITH_LOAD_SIGNAL_PARAMS)
+@pytest.mark.parametrize("omni_server_function", VOXCPM2_PARAMS, indirect=True)
+def test_reliability_fault_process_kill_serve_root_with_load_fast_fail_and_cleanup(
+    omni_server_function,
+    openai_client_function,
+    signal_name: str,
+) -> None:
+    scenario = f"kill_serve_root_with_load_{signal_name.lower()}"
+    injector = make_server_root_kill_fault_injector(signal_name=signal_name, post_kill_wait_seconds=2.0)
+    load_result = run_fault_injection_with_rate_load(
+        submit_request=lambda: _submit_speech_request(openai_client_function, omni_server_function),
+        inject_fault=lambda: injector(omni_server_function),
+        num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
+        request_rate=INFLIGHT_INJECTION_REQUEST_RATE,
+        completion_timeout_sec=300.0,
+    )
+    assert load_result["failure_observed"], (
+        f"[{scenario}] expected at least one load request failure after fault; load_result={load_result}"
+    )
+    host = omni_server_function.host
+    port = omni_server_function.port
+    _assert_post_fault_speech_fast_fail(host, port, model=omni_server_function.model, scenario=scenario)
+    _assert_post_fault_health_terminal(host, port, scenario=scenario)
+    assert_no_server_tree_process_residual_and_gpu_release(
+        omni_server_function,
+        scenario=scenario,
+        timeout_sec=worker_residual_timeout_after_kill_signal(signal_name),
+    )
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@pytest.mark.skipif(os.name == "nt", reason="process-kill injection helper is POSIX-only")
+@pytest.mark.parametrize("signal_name", TREE_SIGNAL_PARAMS)
+@pytest.mark.parametrize("omni_server_function", VOXCPM2_PARAMS, indirect=True)
+def test_reliability_fault_process_kill_tree_no_load_fast_fail_and_cleanup(
+    omni_server_function,
+    signal_name: str,
+) -> None:
+    scenario = f"kill_serve_tree_no_load_{signal_name.lower()}"
+    injector = make_server_tree_kill_fault_injector(
+        signal_name=signal_name,
+        post_kill_wait_seconds=2.0,
+        inter_kill_wait_seconds=0.1,
+    )
+    injector(omni_server_function)
+    host = omni_server_function.host
+    port = omni_server_function.port
+    _assert_post_fault_speech_fast_fail(host, port, model=omni_server_function.model, scenario=scenario)
+    _assert_post_fault_health_terminal(host, port, scenario=scenario)
+    assert_no_server_tree_process_residual_and_gpu_release(
+        omni_server_function,
+        scenario=scenario,
+        timeout_sec=worker_residual_timeout_after_kill_signal(signal_name),
+    )
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@pytest.mark.skipif(os.name == "nt", reason="process-kill injection helper is POSIX-only")
+@pytest.mark.parametrize("signal_name", TREE_WITH_LOAD_SIGNAL_PARAMS)
+@pytest.mark.parametrize("omni_server_function", VOXCPM2_PARAMS, indirect=True)
+def test_reliability_fault_process_kill_tree_with_load_fast_fail_and_cleanup(
+    omni_server_function,
+    openai_client_function,
+    signal_name: str,
+) -> None:
+    scenario = f"kill_serve_tree_with_load_{signal_name.lower()}"
+    injector = make_server_tree_kill_fault_injector(
+        signal_name=signal_name,
+        post_kill_wait_seconds=2.0,
+        inter_kill_wait_seconds=0.1,
+    )
+    load_result = run_fault_injection_with_rate_load(
+        submit_request=lambda: _submit_speech_request(openai_client_function, omni_server_function),
+        inject_fault=lambda: injector(omni_server_function),
+        num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
+        request_rate=INFLIGHT_INJECTION_REQUEST_RATE,
+        completion_timeout_sec=120.0,
+    )
+    assert load_result["failure_observed"], (
+        f"[{scenario}] expected at least one load request failure after fault; load_result={load_result}"
+    )
+    host = omni_server_function.host
+    port = omni_server_function.port
+    _assert_post_fault_speech_fast_fail(host, port, model=omni_server_function.model, scenario=scenario)
+    _assert_post_fault_health_terminal(host, port, scenario=scenario)
+    assert_no_server_tree_process_residual_and_gpu_release(
+        omni_server_function,
+        scenario=scenario,
+        timeout_sec=worker_residual_timeout_after_kill_signal(signal_name),
+    )
@@ -0,0 +1,69 @@
+"""
+VoxCPM2 stability: OmniServer + ``vllm bench serve --omni`` for a fixed duration.
+
+Configuration: ``tests/dfx/stability/tests/test_voxcpm2.json``.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from tests.dfx.conftest import (
+    create_benchmark_indices,
+    create_test_parameter_mapping,
+    create_unique_server_params,
+    load_configs,
+)
+from tests.dfx.stability.helpers import _run_one_vllm_bench_batch, run_stability_benchmark_loop
+
+STABILITY_DIR = Path(__file__).resolve().parent.parent
+DEPLOY_CONFIGS_DIR = STABILITY_DIR / "deploy"
+CONFIG_FILE_PATH = str(STABILITY_DIR / "tests" / "test_voxcpm2.json")
+DEFAULT_NUM_PROMPTS_PER_BATCH = 20
+STABILITY_SERVER_TIMEOUT_ARGS = ["--stage-init-timeout", "600"]
+
+try:
+    BENCHMARK_CONFIGS = load_configs(CONFIG_FILE_PATH)
+except FileNotFoundError:
+    BENCHMARK_CONFIGS = []
+
+test_params = create_unique_server_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
+server_to_benchmark_mapping = create_test_parameter_mapping(BENCHMARK_CONFIGS) if BENCHMARK_CONFIGS else {}
+benchmark_indices = create_benchmark_indices(BENCHMARK_CONFIGS, server_to_benchmark_mapping)
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+@pytest.mark.parametrize("stability_benchmark_params", benchmark_indices, indirect=True)
+def test_stability_voxcpm2(omni_server, stability_benchmark_params):
+    test_name = stability_benchmark_params["test_name"]
+    params = stability_benchmark_params["params"]
+    duration_sec = params.get("duration_sec", 300)
+    num_prompts_per_batch = params.get("num_prompts_per_batch", DEFAULT_NUM_PROMPTS_PER_BATCH)
+    request_rate = params.get("request_rate")
+    max_concurrency = params.get("max_concurrency")
+
+    bench_params = {
+        k: v
+        for k, v in params.items()
+        if k not in ("duration_sec", "request_rate", "max_concurrency", "num_prompts_per_batch")
+    }
+
+    result = run_stability_benchmark_loop(
+        host=omni_server.host,
+        port=omni_server.port,
+        model=omni_server.model,
+        duration_sec=duration_sec,
+        params=bench_params,
+        request_rate=request_rate,
+        max_concurrency=max_concurrency,
+        result_dir=str(STABILITY_DIR),
+        num_prompts_per_batch=num_prompts_per_batch,
+        run_one_batch=_run_one_vllm_bench_batch,
+    )
+
+    assert result.get("failed", 0) == 0, f"[{test_name}] Failed requests detected: {result.get('errors', [])}"
+    assert result.get("completed", 0) > 0, f"[{test_name}] No requests completed"
@@ -18,7 +18,6 @@ from vllm_omni.experimental.ar_diffusion.capability import (
 )
 from vllm_omni.experimental.ar_diffusion.kv_cache import ARDiffusionKVConfig
 from vllm_omni.experimental.ar_diffusion.runner import ARDiffusionModelRunner
-from vllm_omni.experimental.ar_diffusion.tick_protocol import ARDiffusionTickRequest

 BLOCK = 16
 POS = "positive"
@@ -57,21 +56,6 @@ def dreamzero_like_spec(*, capacity: int = 2) -> ARDiffusionKVCacheSpec:
    )


-def tiny_spec(*, capacity: int = 2) -> ARDiffusionKVCacheSpec:
-    return ARDiffusionKVCacheSpec(
-        num_layers=1,
-        num_kv_heads=1,
-        head_size=1,
-        tokens_per_frame=1,
-        frames_per_block=3,
-        window_frames=3,
-        sink_frames=3,
-        kv_branches=(ARDiffusionKVBranchSpec("main", 0),),
-        session_capacity=capacity,
-        cross_attention=(ARDiffusionCrossAttentionKVSpec("text", 2),),
-    )
-
-
 class CapablePipeline:
    def __init__(self, spec: ARDiffusionKVCacheSpec) -> None:
        self.spec = spec
@@ -119,7 +103,6 @@ def make_runner(
    *,
    available_bytes: int = 1 << 28,
    step_execution: bool = False,
-    gpu_memory_fraction: float = 0.1,
 ) -> ARDiffusionModelRunner:
    runner = object.__new__(ARDiffusionModelRunner)
    runner.od_config = SimpleNamespace(
@@ -130,10 +113,7 @@ def make_runner(
    )
    runner.device = torch.device("cpu")
    runner.pipeline = pipeline
-    runner.ar_diffusion_kv_config = ARDiffusionKVConfig(
-        enable=True,
-        gpu_memory_fraction=gpu_memory_fraction,
-    )
+    runner.ar_diffusion_kv_config = ARDiffusionKVConfig(enable=True)
    runner.kv_cache = None
    runner._ar_diffusion_capability = None
    runner._ar_diffusion_kv_cache_spec = None
@@ -152,16 +132,6 @@ def commit_one_frame(runner: ARDiffusionModelRunner, session_id: str, kv_branch:
    return state


-def commit_one_block(runner: ARDiffusionModelRunner, session_id: str, kv_branch: str):
-    state = runner._get_or_create_session(session_id)
-    assert runner.kv_cache is not None
-    seq_len = BLOCK * runner.kv_cache.frames_per_block
-    ctx = state.get_kv_caches(kv_branch, seq_len=seq_len, commit_current=True)[0].forward_ctx
-    ctx.ensure_video_slots(torch.device("cpu"))
-    state.commit_paged_context(kv_branch)
-    return state
-
-
 def test_ar_runner_rejects_pipeline_without_capability():
    runner = object.__new__(ARDiffusionModelRunner)
    runner.od_config = SimpleNamespace(max_num_seqs=1)
@@ -170,43 +140,6 @@ def test_ar_runner_rejects_pipeline_without_capability():
        runner._preallocate_kv_cache(available_bytes=1 << 20)


-def test_runner_uses_typed_tick_as_authoritative_session_contract():
-    tick = ARDiffusionTickRequest(
-        session_id="world-7",
-        request_id="request-3",
-        chunk_index=3,
-        reset=True,
-    )
-    req = SimpleNamespace(
-        request_id="request-3",
-        sampling_params=SimpleNamespace(extra_args=tick.to_extra_args()),
-    )
-
-    session_id, extra_args, parsed = ARDiffusionModelRunner._request_session(req)
-
-    assert session_id == "world-7"
-    assert extra_args == tick.to_extra_args()
-    assert parsed == tick
-
-
-def test_runner_keeps_engine_request_id_separate_from_tick_correlation_id():
-    tick = ARDiffusionTickRequest(
-        session_id="world-7",
-        request_id="client-request-3",
-        chunk_index=3,
-    )
-    req = SimpleNamespace(
-        request_id="engine-request-uuid",
-        sampling_params=SimpleNamespace(extra_args=tick.to_extra_args()),
-    )
-
-    session_id, _, parsed = ARDiffusionModelRunner._request_session(req)
-
-    assert session_id == "world-7"
-    assert req.request_id == "engine-request-uuid"
-    assert parsed.request_id == "client-request-3"
-
-
 def test_lingbot_like_single_branch_session_reuse_reset_and_close():
    pipeline = CapablePipeline(lingbot_like_spec())
    runner = make_runner(pipeline)
@@ -237,31 +170,6 @@ def test_lingbot_like_single_branch_session_reuse_reset_and_close():
    assert pipeline.closes == ["s1"]


-def test_lingbot_like_interleaved_sessions_keep_independent_kv_partitions():
-    runner = make_runner(CapablePipeline(lingbot_like_spec(capacity=2)))
-    kv = runner.kv_cache
-    assert kv is not None
-
-    session_a = commit_one_block(runner, "world-a", "main")
-    a0_blocks = kv.window_block_ids(session_a.adapter("main"))
-    session_b = commit_one_block(runner, "world-b", "main")
-    b0_blocks = kv.window_block_ids(session_b.adapter("main"))
-    session_a_again = commit_one_block(runner, "world-a", "main")
-    a1_blocks = kv.window_block_ids(session_a_again.adapter("main"))
-
-    assert session_a_again is session_a
-    assert session_a.adapter("main").request_id == "ar::world-a::main"
-    assert session_b.adapter("main").request_id == "ar::world-b::main"
-    assert session_a.adapter("main").completed_chunks == 6
-    assert session_b.adapter("main").completed_chunks == 3
-    assert len(a0_blocks) == 3
-    assert len(b0_blocks) == 3
-    assert len(a1_blocks) == 6
-    assert set(a0_blocks) <= set(a1_blocks)
-    assert set(a1_blocks).isdisjoint(b0_blocks)
-    assert tuple(runner._sessions) == ("world-b", "world-a")
-
-
 def test_lingbot_like_sink_survives_sliding_window_eviction():
    runner = make_runner(CapablePipeline(lingbot_like_spec()))
    kv = runner.kv_cache
@@ -290,67 +198,24 @@ def test_dreamzero_like_two_branches_are_independent():


 def test_lru_eviction_releases_blocks_and_notifies_pipeline():
-    pipeline = CapablePipeline(lingbot_like_spec(capacity=2))
+    pipeline = CapablePipeline(lingbot_like_spec(capacity=8))
    runner = make_runner(pipeline)
    kv = runner.kv_cache
    assert kv is not None
-    assert runner._session_capacity == 2
-    assert kv.session_capacity == 2
+    assert runner._session_capacity == 1
+    assert kv.session_capacity == 1
    free_total = kv.manager.block_pool.get_num_free_blocks()
    old = commit_one_frame(runner, "old", "main")
    k = torch.randn(1, 8, 4, 64)
    old.populate_cross_attention("main", "text", [(k, k)] * old.num_layers)
    assert kv.manager.block_pool.get_num_free_blocks() < free_total

-    runner._get_or_create_session("new")
-    assert tuple(runner._sessions) == ("old", "new")
-    runner._get_or_create_session("newest")
-
-    assert tuple(runner._sessions) == ("new", "newest")
-    assert pipeline.closes == ["old"]
-    assert "old" not in kv._cross_sessions
-    assert kv.manager.block_pool.get_num_free_blocks() == free_total
-
-
-def test_budget_reduced_capacity_drives_runner_lru():
-    pipeline = CapablePipeline(tiny_spec(capacity=2))
-    # One tiny LingBot-like session requires 128 bytes; two require 192.
-    runner = make_runner(
-        pipeline,
-        available_bytes=128,
-        gpu_memory_fraction=1.0,
-    )
-    kv = runner.kv_cache
-    assert kv is not None
-    assert kv.requested_session_capacity == 2
-    assert kv.session_capacity == 1
-    assert runner._session_capacity == 1
-
-    runner._get_or_create_session("old")
    runner._get_or_create_session("new")

    assert tuple(runner._sessions) == ("new",)
    assert pipeline.closes == ["old"]
-
-
-def test_dreamzero_like_requested_capacity_is_capped_by_budget():
-    spec = dreamzero_like_spec(capacity=64)
-    # Per all-layer self-KV page: 65,536 bytes. Two resident sessions need:
-    # managed=(2 * (2 * 6 + 4) + 2)=34 pages, scratch=8 pages,
-    # cross-attention=1 page/session, for 44 pages total.
-    page_bytes = 65_536
-    pipeline = CapablePipeline(spec)
-    runner = make_runner(
-        pipeline,
-        available_bytes=44 * page_bytes,
-        gpu_memory_fraction=1.0,
-    )
-    kv = runner.kv_cache
-    assert kv is not None
-    assert kv.requested_session_capacity == 64
-    assert kv.session_capacity == 2
-    assert runner._session_capacity == 2
-    assert kv.cross_attention_reserved_bytes == 2 * page_bytes
+    assert "old" not in kv._cross_sessions
+    assert kv.manager.block_pool.get_num_free_blocks() == free_total


 def test_forward_exception_releases_pending_allocation_and_model_state(monkeypatch):
@@ -367,10 +232,7 @@ def test_forward_exception_releases_pending_allocation_and_model_state(monkeypat
        raise RuntimeError("layer exploded")

    monkeypatch.setattr(DiffusionModelRunner, "execute_model", boom)
-    request = SimpleNamespace(
-        request_id="broken-request",
-        sampling_params=SimpleNamespace(extra_args={"session_id": "broken"}),
-    )
+    request = SimpleNamespace(sampling_params=SimpleNamespace(extra_args={"session_id": "broken"}))

    with pytest.raises(RuntimeError, match="layer exploded"):
        runner.execute_model(request)
@@ -401,10 +263,7 @@ def test_synchronize_exception_uses_forward_cleanup_path(monkeypatch):
    monkeypatch.setattr(DiffusionModelRunner, "execute_model", return_after_allocation)
    monkeypatch.setattr(torch.accelerator, "synchronize", synchronize_boom)
    runner.device = torch.device("cuda")
-    request = SimpleNamespace(
-        request_id="broken-request",
-        sampling_params=SimpleNamespace(extra_args={"session_id": "broken"}),
-    )
+    request = SimpleNamespace(sampling_params=SimpleNamespace(extra_args={"session_id": "broken"}))

    with pytest.raises(RuntimeError, match="asynchronous kernel failed"):
        runner.execute_model(request)
@@ -1,231 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-from __future__ import annotations
-
-from collections.abc import AsyncIterator, Sequence
-from typing import Any
-
-import pytest
-
-from vllm_omni.experimental.ar_diffusion.consumer import ARDiffusionOmniTickConsumer
-from vllm_omni.experimental.ar_diffusion.session import (
-    ARDiffusionSession,
-    ARDiffusionSessionEvent,
-)
-from vllm_omni.experimental.ar_diffusion.tick_protocol import (
-    ARDiffusionChunkMetadata,
-    ARDiffusionControlInput,
-    ARDiffusionTickRequest,
-)
-from vllm_omni.inputs.data import (
-    OmniDiffusionSamplingParams,
-    OmniPromptType,
-    OmniSamplingParams,
-)
-from vllm_omni.outputs import OmniRequestOutput
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-class FakeStagePool:
-    def __init__(self, num_replicas: int = 1) -> None:
-        self.num_replicas = num_replicas
-
-
-class FakeEngine:
-    def __init__(self, num_replicas: int = 1) -> None:
-        self.stage_pools = [FakeStagePool(num_replicas)]
-
-
-class FakeGenerateClient:
-    def __init__(
-        self,
-        *,
-        include_metadata: bool = True,
-        applied_event_ids: list[int] | None = None,
-        output_error: str | None = None,
-        num_replicas: int = 1,
-    ) -> None:
-        self.engine = FakeEngine(num_replicas)
-        self.include_metadata = include_metadata
-        self.applied_event_ids = applied_event_ids
-        self.output_error = output_error
-        self.calls: list[dict[str, Any]] = []
-
-    async def generate(
-        self,
-        prompt: OmniPromptType,
-        *,
-        request_id: str,
-        sampling_params_list: Sequence[OmniSamplingParams],
-        output_modalities: list[str] | None = None,
-    ) -> AsyncIterator[OmniRequestOutput]:
-        self.calls.append(
-            {
-                "prompt": prompt,
-                "request_id": request_id,
-                "sampling_params_list": sampling_params_list,
-                "output_modalities": output_modalities,
-            }
-        )
-        params = sampling_params_list[0]
-        assert isinstance(params, OmniDiffusionSamplingParams)
-        tick = ARDiffusionTickRequest.from_extra_args(params.extra_args)
-        assert tick is not None
-        chunk_metadata = ARDiffusionChunkMetadata.from_tick(tick).to_dict()
-        if self.applied_event_ids is not None:
-            chunk_metadata["applied_event_ids"] = self.applied_event_ids
-        metadata = {"metadata": {"ar_diffusion": chunk_metadata}} if self.include_metadata else {}
-        yield OmniRequestOutput(
-            request_id=request_id,
-            stage_id=0,
-            final_output_type="video",
-            images=["chunk"],  # type: ignore[list-item]
-            _multimodal_output=metadata,
-            finished=True,
-            error=self.output_error,
-        )
-
-
-class FakeLifecycle:
-    async def reset_session(self, session_id: str) -> None:
-        pass
-
-    async def close_session(self, session_id: str) -> None:
-        pass
-
-
-def make_tick() -> ARDiffusionTickRequest:
-    return ARDiffusionTickRequest(
-        session_id="world-1",
-        request_id="world-1-chunk-3",
-        chunk_index=3,
-        applied_event_ids=(7, 8),
-        prompt="turn left",
-        controls=(
-            ARDiffusionControlInput(
-                track="camera",
-                schema="example.camera.v1",
-                data={"samples": [{"x": 1.0}]},
-            ),
-        ),
-    )
-
-
-def test_consumer_rejects_replicated_ar_diffusion_stage() -> None:
-    client = FakeGenerateClient(num_replicas=2)
-
-    with pytest.raises(
-        ValueError,
-        match=r"require exactly one replica.*num_replicas=2",
-    ):
-        ARDiffusionOmniTickConsumer(
-            client,
-            prompt_provider=lambda tick: tick.prompt or "",
-            sampling_params_list=[OmniDiffusionSamplingParams()],
-            diffusion_stage_id=0,
-        )
-
-    assert client.calls == []
-
-
-@pytest.mark.asyncio
-async def test_consumer_submits_typed_tick_and_returns_standard_output() -> None:
-    client = FakeGenerateClient()
-    template = OmniDiffusionSamplingParams(
-        output_type="latent",
-        extra_args={"deployment_option": "kept"},
-    )
-    consumer = ARDiffusionOmniTickConsumer(
-        client,
-        prompt_provider=lambda tick: {
-            "prompt": tick.prompt,
-            "multi_modal_data": {"image": "initial-image"},
-        },
-        sampling_params_list=[template],
-        diffusion_stage_id=0,
-        output_modalities=["video"],
-    )
-    tick = make_tick()
-
-    output = await consumer.execute_tick(tick)
-
-    assert output.images == ["chunk"]
-    assert consumer.chunk_metadata(output) == ARDiffusionChunkMetadata.from_tick(tick)
-    assert len(client.calls) == 1
-    call = client.calls[0]
-    assert call["request_id"] == tick.request_id
-    assert call["output_modalities"] == ["video"]
-    assert call["prompt"] == {
-        "prompt": "turn left",
-        "multi_modal_data": {"image": "initial-image"},
-    }
-    submitted_params = call["sampling_params_list"][0]
-    assert submitted_params.extra_args["deployment_option"] == "kept"
-    assert ARDiffusionTickRequest.from_extra_args(submitted_params.extra_args) == tick
-    assert "ar_diffusion_tick" not in template.extra_args
-
-
-@pytest.mark.asyncio
-async def test_consumer_rejects_missing_chunk_metadata() -> None:
-    consumer = ARDiffusionOmniTickConsumer(
-        FakeGenerateClient(include_metadata=False),
-        prompt_provider=lambda tick: tick.prompt or "",
-        sampling_params_list=[OmniDiffusionSamplingParams()],
-        diffusion_stage_id=0,
-    )
-
-    with pytest.raises(ValueError, match="multimodal_output.metadata"):
-        await consumer.execute_tick(make_tick())
-
-
-@pytest.mark.asyncio
-async def test_consumer_rejects_non_monotonic_applied_event_metadata() -> None:
-    consumer = ARDiffusionOmniTickConsumer(
-        FakeGenerateClient(applied_event_ids=[8, 7]),
-        prompt_provider=lambda tick: tick.prompt or "",
-        sampling_params_list=[OmniDiffusionSamplingParams()],
-        diffusion_stage_id=0,
-    )
-
-    with pytest.raises(ValueError, match="unique and strictly increasing"):
-        await consumer.execute_tick(make_tick())
-
-
-@pytest.mark.asyncio
-async def test_consumer_preserves_engine_failure_before_metadata_validation() -> None:
-    consumer = ARDiffusionOmniTickConsumer(
-        FakeGenerateClient(include_metadata=False, output_error="model failed"),
-        prompt_provider=lambda tick: tick.prompt or "",
-        sampling_params_list=[OmniDiffusionSamplingParams()],
-        diffusion_stage_id=0,
-    )
-
-    with pytest.raises(RuntimeError, match="model failed"):
-        await consumer.execute_tick(make_tick())
-
-
-@pytest.mark.asyncio
-async def test_session_commits_only_after_standard_engine_output_metadata() -> None:
-    client = FakeGenerateClient()
-    consumer = ARDiffusionOmniTickConsumer(
-        client,
-        prompt_provider=lambda tick: tick.prompt or "",
-        sampling_params_list=[OmniDiffusionSamplingParams()],
-        diffusion_stage_id=0,
-    )
-    session = ARDiffusionSession(
-        "world-1",
-        tick_consumer=consumer,
-        lifecycle=FakeLifecycle(),
-        max_pending_events=4,
-        request_id_factory=lambda session_id, chunk_index: f"{session_id}-chunk-{chunk_index}",
-    )
-    await session.accept_event(ARDiffusionSessionEvent(event_id=7, prompt="turn left"))
-
-    output = await session.next_chunk()
-
-    assert output.request_id == "world-1-chunk-0"
-    assert session.chunk_index == 1
-    assert session.pending_event_count == 0
-    assert client.calls[0]["request_id"] == "world-1-chunk-0"
@@ -30,8 +30,6 @@ pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]

 BLOCK = 16

-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-

 def make_spec(*, chunk_size=BLOCK, window_chunks=2, sink_chunks=0, reset_at_boundary=False):
    return ChunkWindowSpec(
@@ -97,6 +95,7 @@ def test_sliding_replace_snaps_to_chunk_boundary():
 def test_sink_chunks_protected():
    # sink = 1 chunk (16 tokens) is never skipped.
    skip = chunk_window_skipped_tokens(80, chunk_size=16, sliding_window=32, sink_chunks=1, reset_at_boundary=False)
+    # raw overflow = 80 - 32 - 16 = 32 -> snapped 32
    assert skip == 32


@@ -202,149 +201,12 @@ def test_cross_attn_pool_deducted_from_self_attn_budget():
        kv_branches=(ARDiffusionKVBranchSpec("positive", 0), ARDiffusionKVBranchSpec("negative", 0)),
        session_capacity=1,
        cross_attention_lengths={"text": L},
-    )
-    cross_bytes = 2 * 2 * L * 4 * 64 * torch.float16.itemsize * 2  # K+V, pos+neg, layers
-    page_bytes = kv.spec.page_size_bytes * 2
-    scratch_bytes = kv.scratch_num_blocks * page_bytes
-    expected = (int(avail * 0.5) - cross_bytes - scratch_bytes) // page_bytes
-    assert kv.num_blocks == expected
-    assert kv.cross_attention_reserved_bytes == cross_bytes
-    assert (kv.num_blocks_total * page_bytes) + cross_bytes <= int(avail * 0.5)
-    assert expected > 1 * (2 + 1) + 2  # above the local-slot minimum, so the cross deduction is what's tested
-
-
-def _make_tiny_capacity_kv(
-    *,
-    requested_capacity: int,
-    available_bytes: int,
-    gpu_memory_fraction: float = 1.0,
-    model_owned_state_bytes_per_session: int = 0,
-) -> ARDiffusionKVCache:
-    return ARDiffusionKVCache(
-        ARDiffusionKVConfig(
-            enable=True,
-            chunk_size=1,
-            window_chunks=3,
-            sink_chunks=3,
-            gpu_memory_fraction=gpu_memory_fraction,
-        ),
-        num_layers=1,
-        num_kv_heads=1,
-        head_size=1,
-        dtype=torch.float32,
-        block_size=1,
-        max_model_len=128,
-        available_bytes=available_bytes,
-        kv_branches=(ARDiffusionKVBranchSpec("main", 0),),
-        session_capacity=requested_capacity,
-        cross_attention_lengths={"text": 2},
-        frames_per_block=3,
-        model_owned_state_bytes_per_session=model_owned_state_bytes_per_session,
        device=torch.device("cpu"),
    )
-
-
-def test_capacity_two_retains_both_windows_and_allocates_next_block():
-    # LingBot-like single-branch geometry: page=8 bytes; capacity=2 requires
-    # 17 managed + 3 scratch blocks, plus two 16-byte cross-attention
-    # reservations: 192 bytes exactly.
-    kv = _make_tiny_capacity_kv(requested_capacity=2, available_bytes=192)
-    assert kv.requested_session_capacity == 2
-    assert kv.session_capacity == 2
-    assert kv.managed_num_blocks == 17
-
-    adapters = [kv.begin_request(session_id) for session_id in ("first", "second")]
-    for adapter in adapters:
-        for _ in range(6):  # three sink + three window blocks
-            kv.allocate_token_slots(adapter, 1)
-            kv.commit_chunk(adapter)
-        assert len(kv.window_block_ids(adapter)) == 6
-
-    # Only one request is in flight, so its three-frame block is counted once.
-    kv.allocate_token_slots(adapters[0], 3)
-
-
-def test_requested_capacity_is_capped_and_cross_reservation_uses_effective_capacity():
-    kv = _make_tiny_capacity_kv(requested_capacity=64, available_bytes=192)
-
-    assert kv.requested_session_capacity == 64
-    assert kv.session_capacity == 2
-    assert kv.cross_attention_bytes_per_session == 16
-    assert kv.cross_attention_reserved_bytes == 32
-    assert kv.num_blocks_total * 8 + kv.cross_attention_reserved_bytes == kv.memory_budget_bytes
-
-
-def test_capacity_rejects_budget_that_cannot_fit_one_session():
-    # One session needs 11 managed + 3 scratch blocks and 16 cross bytes:
-    # 14 * 8 + 16 = 128 bytes.
-    with pytest.raises(ValueError, match="cannot fit one session.*available=127 bytes.*required=128 bytes"):
-        _make_tiny_capacity_kv(requested_capacity=1, available_bytes=127)
-
-
-def test_fraction_is_soft_floor_when_one_session_fits_actual_memory():
-    kv = _make_tiny_capacity_kv(
-        requested_capacity=2,
-        available_bytes=1280,
-        gpu_memory_fraction=0.05,
-    )
-
-    assert kv.configured_memory_budget_bytes == 64
-    assert kv.memory_budget_bytes == 128
-    assert kv.session_capacity == 1
-
-
-def test_model_owned_state_reduces_effective_session_capacity():
-    kv = _make_tiny_capacity_kv(
-        requested_capacity=2,
-        available_bytes=200,
-        model_owned_state_bytes_per_session=16,
-    )
-
-    assert kv.session_capacity == 1
-    assert kv.model_owned_state_reserved_bytes == 16
-    assert (
-        kv.num_blocks_total * 8 + kv.cross_attention_reserved_bytes + kv.model_owned_state_reserved_bytes
-        <= kv.memory_budget_bytes
-    )
-
-
-def _make_shipped_lingbot_geometry(*, gpu_memory_fraction: float) -> ARDiffusionKVCache:
-    return ARDiffusionKVCache(
-        ARDiffusionKVConfig(
-            enable=True,
-            chunk_size=1560,
-            window_chunks=9,
-            sink_chunks=9,
-            gpu_memory_fraction=gpu_memory_fraction,
-        ),
-        num_layers=40,
-        num_kv_heads=40,
-        head_size=128,
-        dtype=torch.bfloat16,
-        block_size=1560,
-        max_model_len=1 << 20,
-        available_bytes=100 * (1 << 30),
-        kv_branches=(ARDiffusionKVBranchSpec("main", 0),),
-        session_capacity=2,
-        cross_attention_lengths={"text": 512},
-        frames_per_block=3,
-        model_owned_state_bytes_per_session=7_488_000,
-    )
-
-
-def test_shipped_lingbot_geometry_default_fraction_admits_one_session():
-    kv = _make_shipped_lingbot_geometry(gpu_memory_fraction=0.1)
-
-    assert kv.configured_memory_budget_bytes == 10 * (1 << 30)
-    assert kv.memory_budget_bytes == 33_653_670_400
-    assert kv.session_capacity == 1
-
-
-def test_shipped_lingbot_geometry_tuned_fraction_admits_two_sessions():
-    kv = _make_shipped_lingbot_geometry(gpu_memory_fraction=0.6)
-
-    assert kv.memory_budget_bytes == 60 * (1 << 30)
-    assert kv.session_capacity == 2
+    cross_bytes = 2 * 2 * L * 4 * 64 * torch.float16.itemsize * 2  # K+V, pos+neg, layers
+    expected = compute_num_blocks(avail - cross_bytes, 0.5, kv.spec.page_size_bytes * 2)
+    assert kv.num_blocks == expected
+    assert expected > 1 * (2 + 1) + 2  # above the local-slot floor, so the cross deduction is what's tested


 def _make_kv(
@@ -359,24 +221,15 @@ def _make_kv(
        if local_branches == 1
        else (ARDiffusionKVBranchSpec("positive", 0), ARDiffusionKVBranchSpec("negative", 1))
    )
-    page_bytes = 2 * BLOCK * 4 * 64 * torch.float32.itemsize
-    declared_scratch_blocks = (max_scratch_tokens_per_branch + BLOCK - 1) // BLOCK
-    scratch_blocks = local_branches * (num_frame_per_block + declared_scratch_blocks)
-    managed_blocks = local_branches * (window_chunks + num_frame_per_block) + 2
    return ARDiffusionKVCache(
-        ARDiffusionKVConfig(
-            enable=True,
-            chunk_size=BLOCK,
-            window_chunks=window_chunks,
-            gpu_memory_fraction=1.0,
-        ),
+        ARDiffusionKVConfig(enable=True, chunk_size=BLOCK, window_chunks=window_chunks),
        num_layers=1,
        num_kv_heads=4,
        head_size=64,
        dtype=torch.float32,
        block_size=BLOCK,
        max_model_len=4096,
-        available_bytes=(managed_blocks + scratch_blocks) * page_bytes,
+        available_bytes=1 << 16,  # tiny -> the floor binds
        device=torch.device("cpu"),
        kv_branches=kv_branches,
        session_capacity=1,
@@ -54,15 +54,13 @@ def test_full_request_lifecycle_and_eviction():
        used = {int(s) // kv.block_size for s in slots}
        assert kv.null_block_id not in used
        assert used <= set(block_table)
-        # Allocation may transiently add one in-flight chunk.
+        # Resident window stays bounded by window + the in-flight chunk.
        assert len(kv.window_block_ids(adapter)) <= kv.spec.window_chunks + 1
-        kv.commit_chunk(adapter)
-        # A successful commit immediately prunes physical ownership.
-        assert len(kv.window_block_ids(adapter)) <= kv.spec.window_chunks
        free_after.append(kv.manager.block_pool.get_num_free_blocks())
+        kv.commit_chunk(adapter)

    # Pool memory plateaus once the window is full (eviction recycles blocks).
-    assert free_after[-1] == free_after[kv.spec.window_chunks - 1]
+    assert free_after[-1] == free_after[kv.spec.window_chunks]

    kv.end_request(adapter)
    assert kv.manager.block_pool.get_num_free_blocks() == free_total
@@ -332,7 +332,7 @@ def test_block_table_padded_to_fixed_width():

 def test_prepare_is_idempotent_and_layers_share_metadata():
    device = torch.device("cpu")
-    kv, st = make_state(num_layers=2)
+    _, st = make_state(num_layers=2)
    contexts = st.get_kv_caches(POS, seq_len=BLOCK, commit_current=False)
    fctx = contexts[0].forward_ctx
    fctx.prepare(device=device, action_len=0, query_len=BLOCK)
@@ -349,10 +349,6 @@ def test_prepare_is_idempotent_and_layers_share_metadata():
    assert i0.block_table is i1.block_table
    assert i0.seq_lens is i1.seq_lens
    assert i0.video_slots is i1.video_slots
-    assert i0.key_pool is kv._k_pools[0]
-    assert i0.value_pool is kv._v_pools[0]
-    assert i1.key_pool is kv._k_pools[1]
-    assert i1.value_pool is kv._v_pools[1]


 def test_layer_inputs_before_prepare_raises():
@@ -174,19 +174,6 @@ def test_window_reset_can_keep_named_cross_attention_cache():
    assert not st.is_cross_attention_populated(POS, "text")


-def test_clear_cross_attention_preserves_self_attention_history():
-    kv, st = make_state(num_layers=1, cross_attn_length=8)
-    _prepare_and_commit(st, POS, 2)
-    k = torch.randn(1, 8, N_HEADS, HEAD_DIM)
-    st.populate_cross_attention(POS, "text", [(k, k)])
-    resident_before = kv.window_block_ids(st.adapter(POS))
-
-    st.clear_cross_attention()
-
-    assert kv.window_block_ids(st.adapter(POS)) == resident_before
-    assert not st.is_cross_attention_populated(POS, "text")
-
-
 def test_cross_attention_uses_logical_branch_when_local_slot_is_shared():
    _, st = make_state(num_layers=2, cross_attn_length=8, shared_local_index=True)
    assert st.kv_cache.num_local_kv_branches == 1
@@ -1,505 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-from __future__ import annotations
-
-import asyncio
-from typing import Any
-
-import pytest
-
-from vllm_omni.experimental.ar_diffusion.session import (
-    ARDiffusionChunkMetadataError,
-    ARDiffusionEventAcceptanceStatus,
-    ARDiffusionEventQueueFullError,
-    ARDiffusionPreparedControls,
-    ARDiffusionSession,
-    ARDiffusionSessionError,
-    ARDiffusionSessionEvent,
-    ARDiffusionSessionManager,
-    ARDiffusionSessionStateError,
-    ARDiffusionSessionStatus,
-    ARDiffusionStaleEventError,
-    ARDiffusionWorkerLifecycle,
-)
-from vllm_omni.experimental.ar_diffusion.tick_protocol import (
-    ARDiffusionChunkMetadata,
-    ARDiffusionControlInput,
-    ARDiffusionTickRequest,
-)
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-class FakeTickConsumer:
-    def __init__(self) -> None:
-        self.ticks: list[ARDiffusionTickRequest] = []
-        self.failures_remaining = 0
-        self.started: asyncio.Event | None = None
-        self.release: asyncio.Event | None = None
-        self.metadata_override: ARDiffusionChunkMetadata | None = None
-
-    async def execute_tick(self, tick: ARDiffusionTickRequest) -> ARDiffusionChunkMetadata:
-        self.ticks.append(tick)
-        if self.started is not None:
-            self.started.set()
-        if self.release is not None:
-            await self.release.wait()
-        if self.failures_remaining:
-            self.failures_remaining -= 1
-            raise RuntimeError("fake tick failed")
-        return self.metadata_override or ARDiffusionChunkMetadata.from_tick(tick)
-
-    def chunk_metadata(self, output: ARDiffusionChunkMetadata) -> ARDiffusionChunkMetadata:
-        return output
-
-
-class FakeLifecycle:
-    def __init__(self) -> None:
-        self.resets: list[str] = []
-        self.closes: list[str] = []
-        self.close_failures_remaining = 0
-
-    async def reset_session(self, session_id: str) -> None:
-        self.resets.append(session_id)
-
-    async def close_session(self, session_id: str) -> None:
-        self.closes.append(session_id)
-        if self.close_failures_remaining:
-            self.close_failures_remaining -= 1
-            raise RuntimeError("fake close failed")
-
-
-class FailingCommitReducer:
-    def __init__(self) -> None:
-        self.commit_calls = 0
-        self.reset_calls = 0
-
-    def prepare(
-        self,
-        *,
-        current_controls: dict[str, ARDiffusionControlInput],
-        events: tuple[ARDiffusionSessionEvent, ...],
-        chunk_index: int,
-    ) -> ARDiffusionPreparedControls:
-        del current_controls, events, chunk_index
-        return ARDiffusionPreparedControls(controls=())
-
-    def commit(self, prepared: ARDiffusionPreparedControls) -> None:
-        del prepared
-        self.commit_calls += 1
-        raise RuntimeError("fake reducer commit failed")
-
-    def reset(self) -> None:
-        self.reset_calls += 1
-
-
-def make_session(
-    *,
-    max_pending_events: int = 8,
-    consumer: FakeTickConsumer | None = None,
-    lifecycle: FakeLifecycle | None = None,
-) -> tuple[ARDiffusionSession, FakeTickConsumer, FakeLifecycle]:
-    resolved_consumer = consumer or FakeTickConsumer()
-    resolved_lifecycle = lifecycle or FakeLifecycle()
-    session = ARDiffusionSession(
-        "world-1",
-        tick_consumer=resolved_consumer,
-        lifecycle=resolved_lifecycle,
-        max_pending_events=max_pending_events,
-        request_id_factory=lambda session_id, chunk_index: f"{session_id}-request-{chunk_index}",
-    )
-    return session, resolved_consumer, resolved_lifecycle
-
-
-@pytest.mark.asyncio
-async def test_events_are_accepted_then_sorted_and_applied_at_chunk_boundary() -> None:
-    session, consumer, _ = make_session()
-    event_two = ARDiffusionSessionEvent(event_id=2, prompt="turn")
-    event_one = ARDiffusionSessionEvent(
-        event_id=1,
-        controls=(
-            ARDiffusionControlInput(
-                track="camera",
-                schema="example.camera.v1",
-                data={"samples": [{"x": 1.0}]},
-            ),
-        ),
-    )
-
-    accepted_two = await session.accept_event(event_two)
-    accepted_one = await session.accept_event(event_one)
-    duplicate = await session.accept_event(event_two)
-
-    assert accepted_two.status is ARDiffusionEventAcceptanceStatus.ACCEPTED
-    assert accepted_one.status is ARDiffusionEventAcceptanceStatus.ACCEPTED
-    assert duplicate.status is ARDiffusionEventAcceptanceStatus.DUPLICATE
-    assert consumer.ticks == []
-
-    metadata = await session.next_chunk()
-
-    assert metadata.applied_event_ids == (1, 2)
-    assert consumer.ticks[0] == ARDiffusionTickRequest(
-        session_id="world-1",
-        request_id="world-1-request-0",
-        chunk_index=0,
-        applied_event_ids=(1, 2),
-        prompt="turn",
-        controls=(
-            ARDiffusionControlInput(
-                track="camera",
-                schema="example.camera.v1",
-                data={"samples": [{"x": 1.0}]},
-            ),
-        ),
-    )
-    assert session.pending_event_count == 0
-
-    applied_duplicate = await session.accept_event(event_two)
-    assert applied_duplicate.status is ARDiffusionEventAcceptanceStatus.DUPLICATE
-
-    next_metadata = await session.next_chunk()
-    assert next_metadata.chunk_index == 1
-    assert next_metadata.applied_event_ids == ()
-    assert consumer.ticks[1].prompt == "turn"
-    assert consumer.ticks[1].controls == consumer.ticks[0].controls
-
-
-@pytest.mark.asyncio
-async def test_composite_event_is_isolated_and_queue_admission_is_atomic() -> None:
-    session, consumer, _ = make_session(max_pending_events=1)
-    payload = {"samples": [{"x": 1.0}]}
-    composite = ARDiffusionSessionEvent(
-        event_id=5,
-        prompt="new prompt",
-        controls=(
-            ARDiffusionControlInput(
-                track="camera",
-                schema="example.camera.v1",
-                data=payload,
-            ),
-            ARDiffusionControlInput(
-                track="action",
-                schema="example.action.v1",
-                data={"keys": ["left"]},
-            ),
-        ),
-    )
-    await session.accept_event(composite)
-    payload["samples"][0]["x"] = 99.0
-
-    duplicate = await session.accept_event(composite)
-    assert duplicate.status is ARDiffusionEventAcceptanceStatus.DUPLICATE
-    with pytest.raises(ARDiffusionEventQueueFullError, match="queue is full"):
-        await session.accept_event(ARDiffusionSessionEvent(event_id=6, prompt="rejected"))
-
-    await session.next_chunk()
-    assert consumer.ticks[0].applied_event_ids == (5,)
-    assert consumer.ticks[0].prompt == "new prompt"
-    assert consumer.ticks[0].controls[0].data["keys"] == ("left",)
-    assert consumer.ticks[0].controls[1].data["samples"] == ({"x": 1.0},)
-    assert consumer.ticks[0].controls[0].to_dict()["data"] == {"keys": ["left"]}
-    assert consumer.ticks[0].controls[1].to_dict()["data"] == {"samples": [{"x": 1.0}]}
-
-
-def test_event_rejects_duplicate_tracks_and_non_transport_payloads() -> None:
-    with pytest.raises(ValueError, match="at most once"):
-        ARDiffusionSessionEvent(
-            event_id=1,
-            controls=(
-                ARDiffusionControlInput(track="camera", schema="one", data={}),
-                ARDiffusionControlInput(track="camera", schema="two", data={}),
-            ),
-        )
-
-    with pytest.raises(ValueError, match="transport-safe"):
-        ARDiffusionControlInput(
-            track="camera",
-            schema="example.camera.v1",
-            data={"object": object()},
-        )
-
-
-@pytest.mark.asyncio
-async def test_failed_tick_closes_worker_and_requires_reset() -> None:
-    consumer = FakeTickConsumer()
-    consumer.failures_remaining = 1
-    lifecycle = FakeLifecycle()
-    session, _, _ = make_session(consumer=consumer, lifecycle=lifecycle)
-    await session.accept_event(ARDiffusionSessionEvent(event_id=3, prompt="retry me"))
-
-    with pytest.raises(RuntimeError, match="fake tick failed"):
-        await session.next_chunk()
-
-    assert session.status is ARDiffusionSessionStatus.FAILED
-    assert session.chunk_index == 0
-    assert session.pending_event_count == 1
-    assert lifecycle.closes == ["world-1"]
-    with pytest.raises(ARDiffusionSessionStateError, match="failed"):
-        await session.next_chunk()
-
-    await session.reset()
-    assert session.status is ARDiffusionSessionStatus.ACTIVE
-    assert session.pending_event_count == 0
-    assert lifecycle.resets == ["world-1"]
-    await session.accept_event(ARDiffusionSessionEvent(event_id=4, prompt="after reset"))
-    metadata = await session.next_chunk()
-    assert metadata.chunk_index == 0
-    assert metadata.applied_event_ids == (4,)
-    assert [tick.applied_event_ids for tick in consumer.ticks] == [(3,), (4,)]
-
-
-@pytest.mark.asyncio
-async def test_inflight_snapshot_is_immutable_and_new_event_waits_for_next_chunk() -> None:
-    consumer = FakeTickConsumer()
-    consumer.started = asyncio.Event()
-    consumer.release = asyncio.Event()
-    session, _, _ = make_session(consumer=consumer)
-    await session.accept_event(ARDiffusionSessionEvent(event_id=10, prompt="first"))
-
-    first_chunk = asyncio.create_task(session.next_chunk())
-    await consumer.started.wait()
-    accepted = await session.accept_event(ARDiffusionSessionEvent(event_id=11, prompt="second"))
-    assert accepted.status is ARDiffusionEventAcceptanceStatus.ACCEPTED
-    assert consumer.ticks[0].prompt == "first"
-    assert consumer.ticks[0].applied_event_ids == (10,)
-
-    with pytest.raises(ARDiffusionStaleEventError, match="event floor is 10"):
-        await session.accept_event(ARDiffusionSessionEvent(event_id=9, prompt="stale"))
-
-    consumer.release.set()
-    first_metadata = await first_chunk
-    second_metadata = await session.next_chunk()
-    assert first_metadata.applied_event_ids == (10,)
-    assert second_metadata.applied_event_ids == (11,)
-    assert consumer.ticks[1].prompt == "second"
-
-
-@pytest.mark.asyncio
-async def test_metadata_mismatch_does_not_commit_snapshot() -> None:
-    consumer = FakeTickConsumer()
-    consumer.metadata_override = ARDiffusionChunkMetadata(
-        session_id="another-world",
-        request_id="wrong",
-        chunk_index=0,
-    )
-    lifecycle = FakeLifecycle()
-    session, _, _ = make_session(consumer=consumer, lifecycle=lifecycle)
-    await session.accept_event(ARDiffusionSessionEvent(event_id=1, prompt="pending"))
-
-    with pytest.raises(ARDiffusionChunkMetadataError, match="does not match"):
-        await session.next_chunk()
-
-    assert session.chunk_index == 0
-    assert session.pending_event_count == 1
-    assert session.status is ARDiffusionSessionStatus.FAILED
-    assert lifecycle.closes == ["world-1"]
-
-
-@pytest.mark.asyncio
-async def test_reducer_commit_failure_is_fail_closed_without_partial_session_commit() -> None:
-    consumer = FakeTickConsumer()
-    lifecycle = FakeLifecycle()
-    reducer = FailingCommitReducer()
-    session = ARDiffusionSession(
-        "world-1",
-        tick_consumer=consumer,
-        lifecycle=lifecycle,
-        max_pending_events=4,
-        request_id_factory=lambda session_id, chunk_index: f"{session_id}-request-{chunk_index}",
-        control_reducer=reducer,
-    )
-    await session.accept_event(ARDiffusionSessionEvent(event_id=1, prompt="pending"))
-
-    with pytest.raises(RuntimeError, match="fake reducer commit failed"):
-        await session.next_chunk()
-
-    assert reducer.commit_calls == 1
-    assert session.status is ARDiffusionSessionStatus.FAILED
-    assert session.chunk_index == 0
-    assert session.pending_event_count == 1
-    assert lifecycle.closes == ["world-1"]
-    with pytest.raises(ARDiffusionSessionStateError, match="failed"):
-        await session.next_chunk()
-
-    await session.reset()
-    assert reducer.reset_calls == 1
-    assert session.status is ARDiffusionSessionStatus.ACTIVE
-    assert session.chunk_index == 0
-    assert session.pending_event_count == 0
-
-
-@pytest.mark.asyncio
-async def test_reset_close_and_disconnect_reach_lifecycle_boundary() -> None:
-    consumer = FakeTickConsumer()
-    lifecycle = FakeLifecycle()
-    manager = ARDiffusionSessionManager(
-        tick_consumer=consumer,
-        lifecycle=lifecycle,
-        max_pending_events=4,
-        request_id_factory=lambda session_id, chunk_index: f"{session_id}-{chunk_index}",
-    )
-    session = await manager.create_session("world-reset")
-    await session.accept_event(ARDiffusionSessionEvent(event_id=4, prompt="before reset"))
-    await session.next_chunk()
-    await manager.reset_session("world-reset")
-
-    assert lifecycle.resets == ["world-reset"]
-    assert session.status is ARDiffusionSessionStatus.ACTIVE
-    assert session.chunk_index == 0
-    duplicate = await session.accept_event(ARDiffusionSessionEvent(event_id=4, prompt="duplicate"))
-    assert duplicate.status is ARDiffusionEventAcceptanceStatus.DUPLICATE
-
-    reset_tick = await session.next_chunk()
-    assert reset_tick.chunk_index == 0
-    assert consumer.ticks[-1].prompt is None
-    assert consumer.ticks[-1].controls == ()
-
-    await manager.close_session("world-reset")
-    assert lifecycle.closes == ["world-reset"]
-    assert session.status is ARDiffusionSessionStatus.CLOSED
-    with pytest.raises(ARDiffusionSessionStateError, match="does not exist"):
-        await manager.get_session("world-reset")
-
-    disconnected = await manager.create_session("world-disconnect")
-    await manager.disconnect("world-disconnect")
-    assert disconnected.status is ARDiffusionSessionStatus.CLOSED
-    assert lifecycle.closes[-1] == "world-disconnect"
-
-
-@pytest.mark.asyncio
-async def test_manager_keeps_interleaved_sessions_independent() -> None:
-    consumer = FakeTickConsumer()
-    manager = ARDiffusionSessionManager(
-        tick_consumer=consumer,
-        lifecycle=FakeLifecycle(),
-        max_pending_events=4,
-        request_id_factory=lambda session_id, chunk_index: f"{session_id}-request-{chunk_index}",
-    )
-    session_a = await manager.create_session("world-a")
-    session_b = await manager.create_session("world-b")
-
-    await session_a.accept_event(ARDiffusionSessionEvent(event_id=10, prompt="a0"))
-    await session_a.next_chunk()
-    await session_b.accept_event(ARDiffusionSessionEvent(event_id=1, prompt="b0"))
-    await session_b.next_chunk()
-    await session_a.accept_event(ARDiffusionSessionEvent(event_id=11, prompt="a1"))
-    await session_a.next_chunk()
-
-    assert [
-        (
-            tick.session_id,
-            tick.request_id,
-            tick.chunk_index,
-            tick.applied_event_ids,
-        )
-        for tick in consumer.ticks
-    ] == [
-        ("world-a", "world-a-request-0", 0, (10,)),
-        ("world-b", "world-b-request-0", 0, (1,)),
-        ("world-a", "world-a-request-1", 1, (11,)),
-    ]
-    assert session_a.chunk_index == 2
-    assert session_b.chunk_index == 1
-    with pytest.raises(ARDiffusionStaleEventError, match="event floor is 11"):
-        await session_a.accept_event(ARDiffusionSessionEvent(event_id=2, prompt="stale only for a"))
-    assert (
-        await session_b.accept_event(ARDiffusionSessionEvent(event_id=2, prompt="b1"))
-    ).status is ARDiffusionEventAcceptanceStatus.ACCEPTED
-
-
-@pytest.mark.asyncio
-async def test_failed_close_retains_tombstone_until_cleanup_retry_succeeds() -> None:
-    consumer = FakeTickConsumer()
-    lifecycle = FakeLifecycle()
-    lifecycle.close_failures_remaining = 1
-    manager = ARDiffusionSessionManager(
-        tick_consumer=consumer,
-        lifecycle=lifecycle,
-        max_pending_events=4,
-    )
-    session = await manager.create_session("world-close-retry")
-
-    with pytest.raises(RuntimeError, match="fake close failed"):
-        await manager.close_session("world-close-retry")
-
-    assert session.status is ARDiffusionSessionStatus.CLEANUP_FAILED
-    assert await manager.get_session("world-close-retry") is session
-    with pytest.raises(ARDiffusionSessionStateError, match="already exists"):
-        await manager.create_session("world-close-retry")
-    with pytest.raises(ARDiffusionSessionStateError, match="cleanup_failed"):
-        await manager.reset_session("world-close-retry")
-
-    await manager.close_session("world-close-retry")
-
-    assert lifecycle.closes == ["world-close-retry", "world-close-retry"]
-    assert session.status is ARDiffusionSessionStatus.CLOSED
-    with pytest.raises(ARDiffusionSessionStateError, match="does not exist"):
-        await manager.get_session("world-close-retry")
-    recreated = await manager.create_session("world-close-retry")
-    assert recreated is not session
-
-
-@pytest.mark.asyncio
-async def test_failed_disconnect_retains_session_for_explicit_cleanup_retry() -> None:
-    lifecycle = FakeLifecycle()
-    lifecycle.close_failures_remaining = 1
-    manager = ARDiffusionSessionManager(
-        tick_consumer=FakeTickConsumer(),
-        lifecycle=lifecycle,
-        max_pending_events=4,
-    )
-    session = await manager.create_session("world-disconnect-retry")
-
-    with pytest.raises(RuntimeError, match="fake close failed"):
-        await manager.disconnect("world-disconnect-retry")
-
-    assert session.status is ARDiffusionSessionStatus.CLEANUP_FAILED
-    assert await manager.get_session("world-disconnect-retry") is session
-
-    await manager.close_session("world-disconnect-retry")
-    assert session.status is ARDiffusionSessionStatus.CLOSED
-
-
-class FakeRPCClient:
-    def __init__(self, result: list[Any]) -> None:
-        self.result = result
-        self.calls: list[dict[str, Any]] = []
-
-    async def collective_rpc(self, **kwargs: Any) -> list[Any]:
-        self.calls.append(kwargs)
-        return self.result
-
-
-@pytest.mark.asyncio
-async def test_worker_lifecycle_routes_to_selected_stage_rpc() -> None:
-    client = FakeRPCClient([[True]])
-    lifecycle = ARDiffusionWorkerLifecycle(client, stage_ids=[2], timeout=3.0)
-
-    await lifecycle.reset_session("world-1")
-    await lifecycle.close_session("world-1")
-
-    assert client.calls == [
-        {
-            "method": "reset_ar_diffusion_session",
-            "timeout": 3.0,
-            "args": ("world-1",),
-            "stage_ids": [2],
-        },
-        {
-            "method": "close_ar_diffusion_session",
-            "timeout": 3.0,
-            "args": ("world-1",),
-            "stage_ids": [2],
-        },
-    ]
-
-
-@pytest.mark.asyncio
-async def test_worker_lifecycle_rejects_unsupported_stage_result() -> None:
-    lifecycle = ARDiffusionWorkerLifecycle(
-        FakeRPCClient([{"supported": False, "error": "not an AR stage"}]),
-        stage_ids=[1],
-    )
-
-    with pytest.raises(ARDiffusionSessionError, match="not an AR stage"):
-        await lifecycle.close_session("world-1")
@@ -1,108 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-import pytest
-
-from vllm_omni.experimental.ar_diffusion.tick_protocol import (
-    ARDiffusionChunkMetadata,
-    ARDiffusionControlInput,
-    ARDiffusionTickRequest,
-)
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-def test_tick_round_trip_preserves_identity_and_control_snapshot() -> None:
-    tick = ARDiffusionTickRequest(
-        session_id="world-1",
-        request_id="chunk-request-7",
-        chunk_index=7,
-        applied_event_ids=(10, 11),
-        prompt="turn left",
-        controls=(
-            ARDiffusionControlInput(
-                track="camera",
-                schema="lingbot.camera_trajectory.v1",
-                data={"poses": [[1.0, 0.0]], "intrinsics": [[2.0]]},
-            ),
-        ),
-    )
-
-    restored = ARDiffusionTickRequest.from_extra_args(tick.to_extra_args())
-
-    assert restored == tick
-    assert ARDiffusionChunkMetadata.from_tick(tick).to_dict() == {
-        "session_id": "world-1",
-        "request_id": "chunk-request-7",
-        "chunk_index": 7,
-        "applied_event_ids": [10, 11],
-    }
-
-
-@pytest.mark.parametrize(
-    ("field", "value", "message"),
-    [
-        ("session_id", "", "session_id"),
-        ("request_id", "", "request_id"),
-        ("chunk_index", -1, "chunk_index"),
-        ("applied_event_ids", (2, 1), "strictly increasing"),
-        ("applied_event_ids", (1, 1), "strictly increasing"),
-    ],
-)
-def test_tick_rejects_invalid_identity(
-    field: str,
-    value: object,
-    message: str,
-) -> None:
-    values = {
-        "session_id": "world-1",
-        "request_id": "request-1",
-        "chunk_index": 0,
-        "applied_event_ids": (),
-    }
-    values[field] = value
-
-    with pytest.raises(ValueError, match=message):
-        ARDiffusionTickRequest(**values)
-
-
-def test_tick_snapshot_is_detached_and_deeply_immutable() -> None:
-    payload = {"samples": [{"keys": ["w"]}]}
-    event_ids = [1, 2]
-    controls = [
-        ARDiffusionControlInput(
-            track="camera",
-            schema="example.camera.v1",
-            data=payload,
-        )
-    ]
-    tick = ARDiffusionTickRequest(
-        session_id="world-1",
-        request_id="request-a",
-        chunk_index=0,
-        applied_event_ids=event_ids,  # type: ignore[arg-type]
-        controls=controls,  # type: ignore[arg-type]
-    )
-    serialized = tick.to_extra_args()
-
-    payload["samples"][0]["keys"].append("j")
-    event_ids.append(3)
-    controls.clear()
-    serialized["ar_diffusion_tick"]["controls"][0]["data"]["samples"][0]["keys"].append("k")
-
-    assert tick.applied_event_ids == (1, 2)
-    assert isinstance(tick.controls, tuple)
-    assert tick.to_extra_args()["ar_diffusion_tick"]["controls"] == [
-        {
-            "track": "camera",
-            "schema": "example.camera.v1",
-            "data": {"samples": [{"keys": ["w"]}]},
-        }
-    ]
-    with pytest.raises(TypeError):
-        tick.controls[0].data["samples"] = ()  # type: ignore[index]
-    with pytest.raises(AttributeError):
-        tick.controls[0].data["samples"][0]["keys"].append("j")
-
-
-def test_legacy_request_has_no_typed_tick() -> None:
-    assert ARDiffusionTickRequest.from_extra_args({"session_id": "legacy"}) is None
@@ -1,53 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-import pytest
-
-from vllm_omni.diffusion.worker.diffusion_worker import DiffusionWorker
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-class FakeARDiffusionRunner:
-    def __init__(self) -> None:
-        self.resets: list[str] = []
-        self.closes: list[str] = []
-
-    def reset_session(self, session_id: str) -> None:
-        self.resets.append(session_id)
-
-    def close_session(self, session_id: str) -> None:
-        self.closes.append(session_id)
-
-
-class FakeRegularRunner:
-    pass
-
-
-def make_worker(runner: object) -> DiffusionWorker:
-    worker = object.__new__(DiffusionWorker)
-    worker.model_runner = runner
-    return worker
-
-
-def test_worker_delegates_ar_session_lifecycle_to_runner() -> None:
-    runner = FakeARDiffusionRunner()
-    worker = make_worker(runner)
-
-    assert worker.reset_ar_diffusion_session("world-1") is True
-    assert worker.close_ar_diffusion_session("world-1") is True
-    assert runner.resets == ["world-1"]
-    assert runner.closes == ["world-1"]
-
-
-def test_worker_reports_unsupported_runner_without_model_assumptions() -> None:
-    worker = make_worker(FakeRegularRunner())
-
-    assert worker.reset_ar_diffusion_session("world-1") is False
-    assert worker.close_ar_diffusion_session("world-1") is False
-
-
-def test_worker_rejects_empty_session_id() -> None:
-    worker = make_worker(FakeARDiffusionRunner())
-
-    with pytest.raises(ValueError, match="session_id"):
-        worker.close_ar_diffusion_session("")
@@ -1,297 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-from types import SimpleNamespace
-from unittest.mock import Mock
-
-import pytest
-import torch
-from vllm.v1.kv_cache_interface import FullAttentionSpec
-
-import vllm_omni.diffusion.diffusion_kv.initialization as initialization
-import vllm_omni.diffusion.vllm_config as diffusion_vllm_config
-from vllm_omni.diffusion.diffusion_kv.config import DiffusionKVCacheMode
-from vllm_omni.diffusion.sched.request_scheduler import RequestScheduler
-from vllm_omni.quantization import build_quant_config
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
-
-
-class _Executor:
-    def __init__(self) -> None:
-        self.specs = [{"layer0": object()}, {"layer0": object()}]
-        self.memory = [1024, 1024]
-        self.configs = None
-        self.profile_requests = None
-
-    def get_kv_cache_specs(self):
-        return self.specs
-
-    def determine_available_kv_memory(self, profile_requests):
-        self.profile_requests = profile_requests
-        return self.memory
-
-    def set_kv_cache_configs(self, configs) -> None:
-        self.configs = configs
-
-
-def _od_config(**overrides):
-    values = dict(
-        model="test-diffusion",
-        diffusion_kv_mode=DiffusionKVCacheMode.PAGED_SCHEDULER,
-        dtype=torch.bfloat16,
-        max_model_len=64,
-        quantization_config=None,
-        tf_model_config=None,
-        enforce_eager=True,
-        is_moe=False,
-        additional_config={},
-        profiler_config=None,
-        parallel_config=SimpleNamespace(
-            tensor_parallel_size=1,
-            pipeline_parallel_size=1,
-            data_parallel_size=1,
-            enable_expert_parallel=False,
-        ),
-        gpu_memory_utilization=0.9,
-        kv_cache_memory_bytes=None,
-        max_num_batched_tokens=64,
-        num_gpus=1,
-    )
-    values.update(overrides)
-    return SimpleNamespace(**values)
-
-
-def test_control_plane_builds_worker_and_scheduler_configs(monkeypatch) -> None:
-    executor = _Executor()
-    od_config = _od_config(num_gpus=2)
-    vllm_config = SimpleNamespace(
-        model_config=SimpleNamespace(max_model_len=4096),
-        max_in_flight_tokens=128,
-    )
-    worker_configs = [object(), object()]
-    scheduler_kv_cache_config = object()
-    profile_requests = [object()]
-    monkeypatch.setattr(initialization, "create_diffusion_vllm_config", lambda *args, **kwargs: vllm_config)
-    build = Mock(return_value=(worker_configs, scheduler_kv_cache_config))
-    monkeypatch.setattr(
-        initialization,
-        "build_native_kv_cache_configs",
-        build,
-    )
-    monkeypatch.setattr(initialization, "resolve_kv_cache_block_sizes", lambda *args: (16, 16))
-
-    result = initialization.initialize_diffusion_kv_control_plane(
-        executor,
-        od_config,
-        profile_requests=profile_requests,
-        device=torch.device("cpu"),
-    )
-
-    build.assert_called_once_with(vllm_config, executor.specs, executor.memory)
-    assert result == (scheduler_kv_cache_config, 16, 16, vllm_config)
-    assert executor.configs == worker_configs
-    assert executor.profile_requests is profile_requests
-
-
-def test_dense_mode_skips_worker_kv_initialization(monkeypatch) -> None:
-    executor = _Executor()
-    od_config = _od_config(diffusion_kv_mode=DiffusionKVCacheMode.DENSE_LEGACY)
-    get_device = Mock(side_effect=AssertionError("dense mode must not resolve a KV device"))
-    monkeypatch.setattr(initialization.current_omni_platform, "get_torch_device", get_device)
-
-    result = initialization.initialize_diffusion_kv_control_plane(executor, od_config)
-
-    assert result is None
-    assert executor.configs is None
-    get_device.assert_not_called()
-
-
-def test_control_plane_rejects_rank_count_mismatch(monkeypatch) -> None:
-    executor = _Executor()
-    executor.memory = [1024]
-    od_config = _od_config(num_gpus=2)
-    monkeypatch.setattr(initialization, "create_diffusion_vllm_config", lambda *args, **kwargs: object())
-
-    with pytest.raises(ValueError, match="rank count mismatch"):
-        initialization.initialize_diffusion_kv_control_plane(
-            executor,
-            od_config,
-            profile_requests=[object()],
-            device=torch.device("cpu"),
-        )
-
-
-def test_paged_control_plane_requires_prepared_profile_request() -> None:
-    with pytest.raises(ValueError, match="prepared memory profile requests"):
-        initialization.initialize_diffusion_kv_control_plane(
-            _Executor(),
-            _od_config(),
-            device=torch.device("cpu"),
-        )
-
-
-def test_native_config_builder_produces_matching_worker_and_scheduler_pools() -> None:
-    od_config = _od_config()
-    vllm_config = diffusion_vllm_config.create_diffusion_vllm_config(torch.device("cpu"), od_config)
-    spec = FullAttentionSpec(
-        block_size=vllm_config.cache_config.block_size,
-        num_kv_heads=2,
-        head_size=8,
-        dtype=torch.bfloat16,
-        non_causal=True,
-    )
-    num_blocks = 32
-
-    worker_configs, scheduler_kv_cache_config = initialization.build_native_kv_cache_configs(
-        vllm_config,
-        [{"layer0": spec}],
-        [spec.page_size_bytes * num_blocks],
-    )
-
-    assert len(worker_configs) == 1
-    assert worker_configs[0].num_blocks == num_blocks
-    assert scheduler_kv_cache_config.num_blocks == num_blocks
-    assert worker_configs[0].kv_cache_groups == scheduler_kv_cache_config.kv_cache_groups
-
-
-def test_native_control_plane_config_initializes_scheduler() -> None:
-    od_config = _od_config()
-    vllm_config = diffusion_vllm_config.create_diffusion_vllm_config(torch.device("cpu"), od_config)
-    spec = FullAttentionSpec(
-        block_size=vllm_config.cache_config.block_size,
-        num_kv_heads=2,
-        head_size=8,
-        dtype=torch.bfloat16,
-        non_causal=True,
-    )
-    _, scheduler_kv_cache_config = initialization.build_native_kv_cache_configs(
-        vllm_config,
-        [{"layer0": spec}],
-        [spec.page_size_bytes * 32],
-    )
-
-    scheduler = RequestScheduler()
-    scheduler.initialize(
-        od_config,
-        kv_cache_config=scheduler_kv_cache_config,
-        scheduler_block_size=spec.block_size,
-        hash_block_size=spec.block_size,
-        kv_vllm_config=vllm_config,
-    )
-
-    assert scheduler._diffusion_kv_manager is not None
-    scheduler.close()
-
-
-def test_dense_config_keeps_new_paged_sizing_inputs_inactive() -> None:
-    od_config = _od_config(
-        diffusion_kv_mode=DiffusionKVCacheMode.DENSE_LEGACY,
-        parallel_config=SimpleNamespace(
-            tensor_parallel_size=1,
-            pipeline_parallel_size=4,
-            data_parallel_size=1,
-            enable_expert_parallel=False,
-        ),
-        gpu_memory_utilization=0.5,
-        kv_cache_memory_bytes=4096,
-        max_num_batched_tokens=64,
-    )
-    vllm_config = diffusion_vllm_config.create_base_diffusion_vllm_config(torch.device("cpu"), od_config)
-    original_pp_size = vllm_config.parallel_config.pipeline_parallel_size
-    original_memory_utilization = vllm_config.cache_config.gpu_memory_utilization
-    original_memory_bytes = vllm_config.cache_config.kv_cache_memory_bytes
-    original_token_budget = vllm_config.scheduler_config.max_num_batched_tokens
-
-    diffusion_vllm_config.configure_diffusion_vllm_config(vllm_config, od_config)
-
-    assert vllm_config.parallel_config.pipeline_parallel_size == original_pp_size
-    assert vllm_config.cache_config.gpu_memory_utilization == original_memory_utilization
-    assert vllm_config.cache_config.kv_cache_memory_bytes == original_memory_bytes
-    assert vllm_config.scheduler_config.max_num_batched_tokens == original_token_budget
-
-
-def test_paged_config_forwards_gpu_memory_utilization_to_native_cache_config() -> None:
-    vllm_config = diffusion_vllm_config.create_diffusion_vllm_config(
-        torch.device("cpu"),
-        _od_config(gpu_memory_utilization=0.42),
-    )
-
-    assert vllm_config.cache_config.gpu_memory_utilization == 0.42
-
-
-def test_diffusion_vllm_model_config_supplies_dtype_for_quant_methods() -> None:
-    quantization_config = build_quant_config(
-        {
-            "quant_method": "modelopt",
-            "quant_algo": "FP8",
-            "ignore": [],
-        }
-    )
-    model_config = diffusion_vllm_config._make_diffusion_vllm_model_config(
-        _od_config(
-            diffusion_kv_mode=DiffusionKVCacheMode.DENSE_LEGACY,
-            quantization_config=quantization_config,
-        )
-    )
-
-    assert model_config.dtype is torch.bfloat16
-    assert model_config.quantization == "modelopt"
-    assert model_config.quantization_config is quantization_config
-    assert model_config.is_quantized
-    assert model_config.is_moe is False
-    assert model_config.is_nvfp4_quantized() is False
-    assert model_config.original_max_model_len == model_config.max_model_len
-
-    unquantized_model_config = diffusion_vllm_config._make_diffusion_vllm_model_config(
-        _od_config(
-            diffusion_kv_mode=DiffusionKVCacheMode.DENSE_LEGACY,
-            quantization_config=None,
-        )
-    )
-    assert not unquantized_model_config.is_quantized
-
-
-def test_minus_one_max_model_len_uses_model_limit() -> None:
-    od_config = SimpleNamespace(
-        max_model_len=-1,
-        tf_model_config=SimpleNamespace(max_position_embeddings=4096),
-        diffusion_kv_mode=DiffusionKVCacheMode.PAGED_SCHEDULER,
-    )
-
-    assert diffusion_vllm_config.resolve_diffusion_max_model_len(od_config) == 4096
-
-    model_config = diffusion_vllm_config._make_diffusion_vllm_model_config(
-        _od_config(
-            max_model_len=-1,
-            tf_model_config=SimpleNamespace(max_position_embeddings=4096),
-        )
-    )
-    assert model_config.max_model_len == 4096
-    assert model_config.original_max_model_len == -1
-
-
-def test_minus_one_max_model_len_is_auto_fitted_by_native_cache_sizing() -> None:
-    vllm_config = diffusion_vllm_config.create_diffusion_vllm_config(
-        torch.device("cpu"),
-        _od_config(
-            max_model_len=-1,
-            tf_model_config=SimpleNamespace(max_position_embeddings=4096),
-        ),
-    )
-    spec = FullAttentionSpec(
-        block_size=vllm_config.cache_config.block_size,
-        num_kv_heads=2,
-        head_size=8,
-        dtype=torch.bfloat16,
-        non_causal=True,
-    )
-
-    initialization.build_native_kv_cache_configs(
-        vllm_config,
-        [{"layer0": spec}],
-        [spec.page_size_bytes * 16],
-    )
-
-    assert vllm_config.model_config.original_max_model_len == -1
-    assert vllm_config.model_config.max_model_len == 256
@@ -1,211 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import pytest
-import torch
-import vllm.v1.core.single_type_kv_cache_manager as native_kv_managers
-from vllm.v1.kv_cache_interface import FullAttentionSpec, KVCacheConfig, KVCacheGroupSpec, KVCacheTensor
-
-from vllm_omni.diffusion.diffusion_kv.manager import DiffusionKVAdmissionError, DiffusionKVCacheManager
-from vllm_omni.diffusion.diffusion_kv.request import DiffusionKVContext, DiffusionKVRequest
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
-
-BLOCK_SIZE = 4
-
-
-def _config(num_blocks: int) -> KVCacheConfig:
-    native_kv_managers.register_all_kvcache_specs(None)
-    spec = FullAttentionSpec(
-        block_size=BLOCK_SIZE,
-        num_kv_heads=2,
-        head_size=8,
-        dtype=torch.bfloat16,
-    )
-    return KVCacheConfig(
-        num_blocks=num_blocks,
-        kv_cache_tensors=[KVCacheTensor(size=spec.page_size_bytes * num_blocks, shared_by=["layer0"])],
-        kv_cache_groups=[KVCacheGroupSpec(layer_names=["layer0"], kv_cache_spec=spec)],
-    )
-
-
-def _request(public_id: str, sequence_id: int, *, seq_len: int = 8, kv_contexts=()) -> DiffusionKVRequest:
-    return DiffusionKVRequest(
-        f"{public_id}/diffusion-kv/{sequence_id}",
-        sequence_id=sequence_id,
-        prefix_len=4,
-        target_len=4,
-        seq_len=seq_len,
-        kv_contexts=kv_contexts,
-    )
-
-
-def _manager(num_blocks: int, *, max_model_len: int = 64) -> DiffusionKVCacheManager:
-    return DiffusionKVCacheManager(
-        _config(num_blocks),
-        max_model_len=max_model_len,
-        scheduler_block_size=BLOCK_SIZE,
-        hash_block_size=BLOCK_SIZE,
-    )
-
-
-def test_reserve_and_free_multi_cfg_request() -> None:
-    manager = _manager(8)
-    free_before = manager.native_manager.block_pool.get_num_free_blocks()
-    requests = (_request("public", 0), _request("public", 1))
-
-    metadata = manager.reserve_request("public", requests)
-
-    assert metadata is not None
-    assert metadata.request_id == "public"
-    assert metadata.allocation_generation == 1
-    assert [sequence.sequence_id for sequence in metadata.sequences] == [0, 1]
-    assert [sequence.prefix_len for sequence in metadata.sequences] == [4, 4]
-    assert [sequence.target_len for sequence in metadata.sequences] == [4, 4]
-    assert [sequence.seq_len for sequence in metadata.sequences] == [8, 8]
-    assert [len(sequence.block_ids[0]) for sequence in metadata.sequences] == [2, 2]
-    assert all(sequence.context_ids == () for sequence in metadata.sequences)
-    assert metadata.contexts == ()
-    assert metadata == manager.get_metadata("public")
-    assert manager.native_manager.block_pool.get_num_free_blocks() == free_before - 4
-
-    manager.free_request("public")
-    assert manager.native_manager.block_pool.get_num_free_blocks() == free_before
-
-
-def test_impossible_cfg_allocation_rolls_back_and_fails_fast() -> None:
-    manager = _manager(3)
-    free_before = manager.native_manager.block_pool.get_num_free_blocks()
-
-    with pytest.raises(DiffusionKVAdmissionError, match="cannot fit even when the block pool is empty"):
-        manager.reserve_request(
-            "public",
-            (_request("public", 0), _request("public", 1)),
-        )
-
-    assert manager.has_request("public") is False
-    assert manager.native_manager.block_pool.get_num_free_blocks() == free_before
-
-
-def test_impossible_cfg_allocation_fails_fast_while_pool_is_busy() -> None:
-    manager = _manager(5)
-    assert manager.reserve_request("running", (_request("running", 0),)) is not None
-    free_before = manager.native_manager.block_pool.get_num_free_blocks()
-
-    with pytest.raises(DiffusionKVAdmissionError, match="required_blocks=6, available_blocks=4"):
-        manager.reserve_request(
-            "impossible",
-            tuple(_request("impossible", sequence_id) for sequence_id in range(3)),
-        )
-
-    assert manager.has_request("impossible") is False
-    assert manager.native_manager.block_pool.get_num_free_blocks() == free_before
-
-
-def test_temporary_capacity_pressure_returns_none() -> None:
-    manager = _manager(3)
-    assert manager.reserve_request("running", (_request("running", 0),)) is not None
-
-    allocation = manager.reserve_request("waiting", (_request("waiting", 0),))
-
-    assert allocation is None
-    assert manager.has_request("waiting") is False
-
-
-def test_partial_cfg_rollback_can_retry_after_capacity_is_released() -> None:
-    manager = _manager(5)
-    assert manager.reserve_request("running", (_request("running", 0),)) is not None
-    waiting = (_request("waiting", 0), _request("waiting", 1))
-
-    assert manager.reserve_request("waiting", waiting) is None
-    assert manager.has_request("waiting") is False
-
-    manager.free_request("running")
-    metadata = manager.reserve_request("waiting", waiting)
-
-    assert metadata is not None
-    assert [len(sequence.block_ids[0]) for sequence in metadata.sequences] == [2, 2]
-
-
-def test_reallocated_request_gets_a_new_allocation_generation() -> None:
-    manager = _manager(3)
-    request = (_request("public", 0),)
-
-    first = manager.reserve_request("public", request)
-    assert first is not None
-    manager.free_request("public")
-    second = manager.reserve_request("public", request)
-
-    assert second is not None
-    assert second.allocation_generation == first.allocation_generation + 1
-
-
-def test_cfg_allocation_exception_rolls_back(monkeypatch) -> None:
-    manager = _manager(8)
-    free_before = manager.native_manager.block_pool.get_num_free_blocks()
-    native_allocate = manager.native_manager.allocate_slots
-    num_calls = 0
-
-    def fail_second_allocation(*args, **kwargs):
-        nonlocal num_calls
-        num_calls += 1
-        if num_calls == 2:
-            raise RuntimeError("injected allocation failure")
-        return native_allocate(*args, **kwargs)
-
-    monkeypatch.setattr(manager.native_manager, "allocate_slots", fail_second_allocation)
-
-    with pytest.raises(RuntimeError, match="injected allocation failure"):
-        manager.reserve_request(
-            "public",
-            (_request("public", 0), _request("public", 1)),
-        )
-
-    assert manager.has_request("public") is False
-    assert manager.native_manager.block_pool.get_num_free_blocks() == free_before
-
-
-def test_rejects_independent_context_until_role_routing_is_implemented() -> None:
-    manager = _manager(8)
-    context = DiffusionKVContext(context_id="text", cache_role="cross.text", num_tokens=8)
-
-    with pytest.raises(DiffusionKVAdmissionError, match="DiffusionKVContext"):
-        manager.reserve_request(
-            "public",
-            (_request("public", 0, kv_contexts=(context,)),),
-        )
-
-
-def test_rejects_sequence_longer_than_admission_bound() -> None:
-    manager = _manager(8, max_model_len=8)
-
-    with pytest.raises(DiffusionKVAdmissionError, match="exceeds max_model_len"):
-        manager.reserve_request("public", (_request("public", 0, seq_len=9),))
-
-
-@pytest.mark.parametrize(
-    "requests",
-    [
-        (_request("public", 1),),
-        (_request("public", 1), _request("public", 0)),
-        (_request("public", 0), _request("public", 2)),
-    ],
-)
-def test_rejects_ambiguous_sequence_order(requests) -> None:
-    manager = _manager(8)
-
-    with pytest.raises(ValueError, match="ordered by contiguous sequence_id"):
-        manager.reserve_request("public", requests)
-
-
-def test_close_releases_every_public_request() -> None:
-    manager = _manager(8)
-    free_before = manager.native_manager.block_pool.get_num_free_blocks()
-    assert manager.reserve_request("first", (_request("first", 0),)) is not None
-    assert manager.reserve_request("second", (_request("second", 0),)) is not None
-
-    manager.close()
-
-    assert manager.native_manager.block_pool.get_num_free_blocks() == free_before
-    assert manager.has_request("first") is False
-    assert manager.has_request("second") is False
@@ -107,8 +107,11 @@ def test_request_rejects_non_context_values() -> None:
 def _manager(*, enable_caching: bool) -> KVCacheManager:
    # vLLM 0.26 lazily registers its built-in KV specs. Another test may have
    # already registered an out-of-tree spec, making the registry non-empty
-    # before FullAttentionSpec is registered. Initialize the native mappings.
-    native_kv_managers.register_all_kvcache_specs(None)
+    # before FullAttentionSpec is registered. Initialize the native mappings
+    # explicitly; vLLM 0.25 has no registration hook and needs no setup.
+    register_all = getattr(native_kv_managers, "register_all_kvcache_specs", None)
+    if register_all is not None:
+        register_all(None)

    spec = FullAttentionSpec(
        block_size=BLOCK_SIZE,
@@ -1,259 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-from contextlib import nullcontext
-from types import SimpleNamespace
-from unittest.mock import Mock
-
-import pytest
-import torch
-from torch import nn
-from vllm.v1.kv_cache_interface import FullAttentionSpec, KVCacheConfig, KVCacheGroupSpec, KVCacheTensor
-
-import vllm_omni.diffusion.worker.diffusion_worker as diffusion_worker_module
-from vllm_omni.diffusion.attention.layer import Attention
-from vllm_omni.diffusion.diffusion_kv.config import DiffusionKVCacheMode
-from vllm_omni.diffusion.worker.diffusion_model_runner import DiffusionModelRunner
-from vllm_omni.diffusion.worker.diffusion_worker import DiffusionWorker
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
-
-
-class _Backend:
-    @classmethod
-    def indexes_kv_by_block_stride(cls) -> bool:
-        return True
-
-
-def _attention(*, enabled: bool) -> Attention:
-    attention = Attention.__new__(Attention)
-    nn.Module.__init__(attention)
-    attention.paged_kv_cache_role = "primary" if enabled else None
-    attention.paged_kv_cache_dtype = torch.bfloat16
-    attention.num_kv_heads = 2
-    attention.head_size = 8
-    attention.causal = False
-    attention.attn_backend = _Backend
-    return attention
-
-
-def _runner(attention: Attention) -> DiffusionModelRunner:
-    runner = object.__new__(DiffusionModelRunner)
-    runner.od_config = SimpleNamespace(diffusion_kv_mode=DiffusionKVCacheMode.PAGED_SCHEDULER)
-    runner.vllm_config = SimpleNamespace(
-        cache_config=SimpleNamespace(block_size=4),
-        model_config=SimpleNamespace(dtype=torch.float16),
-    )
-    pipeline = nn.Module()
-    pipeline.image_attention = attention
-    runner.pipeline = pipeline
-    return runner
-
-
-def test_runner_discovers_native_spec_from_loaded_attention() -> None:
-    specs = _runner(_attention(enabled=True)).get_kv_cache_spec()
-
-    assert set(specs) == {"image_attention"}
-    spec = specs["image_attention"]
-    assert isinstance(spec, FullAttentionSpec)
-    assert spec.block_size == 4
-    assert spec.num_kv_heads == 2
-    assert spec.head_size == 8
-    assert spec.dtype is torch.bfloat16
-    assert spec.indexes_kv_by_block_stride is True
-    assert spec.non_causal is True
-
-
-def test_runner_rejects_paged_mode_without_cache_enabled_attention() -> None:
-    runner = _runner(_attention(enabled=False))
-
-    with pytest.raises(RuntimeError, match="no cache-enabled Attention"):
-        runner.get_kv_cache_spec()
-
-
-def test_runner_retains_matching_rank_local_config() -> None:
-    runner = _runner(_attention(enabled=True))
-    spec = runner.get_kv_cache_spec()["image_attention"]
-    config = KVCacheConfig(
-        num_blocks=8,
-        kv_cache_tensors=[KVCacheTensor(size=spec.page_size_bytes * 8, shared_by=["image_attention"])],
-        kv_cache_groups=[KVCacheGroupSpec(layer_names=["image_attention"], kv_cache_spec=spec)],
-    )
-
-    runner.set_kv_cache_config(config)
-
-    assert runner.kv_cache_config is config
-
-
-def test_runner_rejects_rank_local_config_for_different_layers() -> None:
-    runner = _runner(_attention(enabled=True))
-    spec = runner.get_kv_cache_spec()["image_attention"]
-    config = KVCacheConfig(
-        num_blocks=8,
-        kv_cache_tensors=[KVCacheTensor(size=spec.page_size_bytes * 8, shared_by=["other_attention"])],
-        kv_cache_groups=[KVCacheGroupSpec(layer_names=["other_attention"], kv_cache_spec=spec)],
-    )
-
-    with pytest.raises(ValueError, match="layer mismatch"):
-        runner.set_kv_cache_config(config)
-
-
-def test_worker_selects_its_rank_local_config() -> None:
-    worker = object.__new__(DiffusionWorker)
-    worker.rank = 1
-    worker.od_config = SimpleNamespace(num_gpus=2)
-    worker.model_runner = SimpleNamespace(set_kv_cache_config=lambda config: setattr(worker, "installed", config))
-    configs = [object(), object()]
-
-    worker.set_kv_cache_configs(configs)
-
-    assert worker.installed is configs[1]
-    with pytest.raises(ValueError, match="rank count mismatch"):
-        worker.set_kv_cache_configs(configs[:1])
-
-
-def test_worker_honors_explicit_kv_memory_budget(monkeypatch) -> None:
-    worker = object.__new__(DiffusionWorker)
-    worker.rank = 0
-    worker.vllm_config = SimpleNamespace(cache_config=SimpleNamespace(kv_cache_memory_bytes=4096))
-    worker.init_snapshot = SimpleNamespace(free_memory=16384)
-    worker.requested_memory = 8192
-    worker.model_runner = SimpleNamespace(profile_run=Mock(), model_memory_usage=1024)
-    monkeypatch.setattr(diffusion_worker_module, "_all_gather_rank_values", lambda value: [value])
-    profile_request = object()
-
-    assert worker.determine_available_kv_memory(profile_request) == [4096]
-    worker.model_runner.profile_run.assert_called_once_with(profile_request)
-
-
-def test_worker_treats_zero_kv_memory_budget_as_profiled_auto_sizing(monkeypatch) -> None:
-    worker = object.__new__(DiffusionWorker)
-    worker.rank = 0
-    worker.init_snapshot = SimpleNamespace(free_memory=900)
-    worker.requested_memory = 750
-    worker.vllm_config = SimpleNamespace(
-        cache_config=SimpleNamespace(
-            kv_cache_memory_bytes=0,
-            gpu_memory_utilization=0.75,
-        )
-    )
-    worker.model_runner = SimpleNamespace(profile_run=Mock(), model_memory_usage=100)
-    monkeypatch.setattr(diffusion_worker_module, "_all_gather_rank_values", lambda value: [value])
-    profile_result = SimpleNamespace(
-        non_kv_cache_memory=400,
-        after_profile=SimpleNamespace(free_memory=500),
-    )
-    monkeypatch.setattr(
-        diffusion_worker_module,
-        "memory_profiling",
-        lambda snapshot, weights_memory: nullcontext(profile_result),
-    )
-    profile_request = object()
-
-    assert worker.determine_available_kv_memory(profile_request) == [350]
-    worker.model_runner.profile_run.assert_called_once_with(profile_request)
-
-
-@pytest.mark.parametrize("non_kv_cache_memory", [750, 800])
-def test_worker_rejects_non_positive_profiled_kv_memory(monkeypatch, non_kv_cache_memory: int) -> None:
-    worker = object.__new__(DiffusionWorker)
-    worker.rank = 0
-    worker.init_snapshot = SimpleNamespace(free_memory=900)
-    worker.requested_memory = 750
-    worker.vllm_config = SimpleNamespace(
-        cache_config=SimpleNamespace(
-            kv_cache_memory_bytes=None,
-            gpu_memory_utilization=0.75,
-        )
-    )
-    worker.model_runner = SimpleNamespace(profile_run=Mock(), model_memory_usage=100)
-    monkeypatch.setattr(diffusion_worker_module, "_all_gather_rank_values", lambda value: [value])
-    profile_result = SimpleNamespace(
-        non_kv_cache_memory=non_kv_cache_memory,
-        after_profile=SimpleNamespace(free_memory=100),
-    )
-    monkeypatch.setattr(
-        diffusion_worker_module,
-        "memory_profiling",
-        lambda snapshot, weights_memory: nullcontext(profile_result),
-    )
-    profile_request = object()
-
-    with pytest.raises(
-        RuntimeError,
-        match=(
-            "rank 0: RuntimeError: No memory remains for Diffusion KV cache after profiling: "
-            f"requested_memory=750 bytes, non_kv_cache_memory={non_kv_cache_memory} bytes"
-        ),
-    ):
-        worker.determine_available_kv_memory(profile_request)
-
-    worker.model_runner.profile_run.assert_called_once_with(profile_request)
-
-
-def test_worker_profiles_activation_headroom_instead_of_current_residency(monkeypatch) -> None:
-    worker = object.__new__(DiffusionWorker)
-    worker.rank = 0
-    worker.init_snapshot = SimpleNamespace(free_memory=1200)
-    worker.requested_memory = 1000
-    worker.vllm_config = SimpleNamespace(
-        cache_config=SimpleNamespace(
-            kv_cache_memory_bytes=None,
-            gpu_memory_utilization=0.8,
-        )
-    )
-    worker.model_runner = SimpleNamespace(profile_run=Mock(), model_memory_usage=250)
-    monkeypatch.setattr(diffusion_worker_module, "_all_gather_rank_values", lambda value: [value])
-    profile_result = SimpleNamespace(
-        non_kv_cache_memory=650,
-        after_profile=SimpleNamespace(free_memory=550),
-    )
-    profiling_calls = []
-
-    def profile(snapshot, *, weights_memory):
-        profiling_calls.append((snapshot, weights_memory))
-        return nullcontext(profile_result)
-
-    monkeypatch.setattr(
-        diffusion_worker_module,
-        "memory_profiling",
-        profile,
-    )
-    profile_request = object()
-
-    assert worker.determine_available_kv_memory(profile_request) == [350]
-    assert profiling_calls == [(worker.init_snapshot, 250)]
-    worker.model_runner.profile_run.assert_called_once_with(profile_request)
-
-
-def test_worker_rejects_profile_without_preload_snapshot(monkeypatch) -> None:
-    worker = object.__new__(DiffusionWorker)
-    worker.init_snapshot = None
-    worker.requested_memory = None
-    worker.vllm_config = SimpleNamespace(cache_config=SimpleNamespace(kv_cache_memory_bytes=None))
-    worker.model_runner = SimpleNamespace()
-    monkeypatch.setattr(diffusion_worker_module, "_all_gather_rank_values", lambda value: [value])
-
-    with pytest.raises(RuntimeError, match="snapshot was not captured before model loading"):
-        worker.determine_available_kv_memory(object())
-
-
-def test_rank_probe_gathers_local_failure_before_raising(monkeypatch) -> None:
-    gathered = []
-
-    def gather(local_result):
-        gathered.append(local_result)
-        return [local_result, (True, "peer-result")]
-
-    def fail_local_probe():
-        raise ValueError("local failure")
-
-    monkeypatch.setattr(diffusion_worker_module, "_all_gather_rank_values", gather)
-
-    with pytest.raises(RuntimeError, match="rank 0: ValueError: local failure"):
-        diffusion_worker_module._run_and_gather_rank_values(
-            "test probe",
-            fail_local_probe,
-        )
-
-    assert gathered == [(False, "ValueError: local failure")]
@@ -85,15 +85,19 @@ def test_new_request_data_carries_scheduler_allocation_atomically() -> None:
    assert new_req.diffusion_kv_metadata is metadata


-def test_paged_request_without_metadata_fails_before_request_forward() -> None:
+def test_paged_request_without_metadata_reaches_request_forward() -> None:
    runner = make_runner(DiffusionKVCacheMode.PAGED_SCHEDULER)
-    runner._execute_request_list = Mock()
+    expected_output = DiffusionOutput(output=None)
+    runner._execute_request_list = Mock(
+        return_value=SimpleNamespace(
+            runner_outputs=[SimpleNamespace(result=expected_output)],
+        )
+    )
    req = SimpleNamespace(request_id="req-0")

-    with pytest.raises(ValueError, match="requires Diffusion KV metadata"):
-        runner.execute_model(req)
+    assert runner.execute_model(req) is expected_output

-    runner._execute_request_list.assert_not_called()
+    runner._execute_request_list.assert_called_once()


 def test_dense_request_rejects_metadata_before_request_forward() -> None:
@@ -137,18 +141,18 @@ def test_request_rpc_rejects_envelope_request_identity_mismatch() -> None:
    assert calls == []


-def test_batch_path_rejects_missing_metadata_before_forward() -> None:
+def test_batch_path_allows_missing_metadata_before_forward() -> None:
    runner = make_runner(DiffusionKVCacheMode.PAGED_SCHEDULER)
-    runner._execute_request_list = Mock()
+    expected_output = object()
+    runner._execute_request_list = Mock(return_value=expected_output)
    new_req = NewRequestData(
        request_id="req-0",
        req=SimpleNamespace(request_id="req-0"),
    )

-    with pytest.raises(ValueError, match="requires Diffusion KV metadata"):
-        runner.execute_model_batch(make_scheduler_output(new_req), runner.od_config)
+    assert runner.execute_model_batch(make_scheduler_output(new_req), runner.od_config) is expected_output

-    runner._execute_request_list.assert_not_called()
+    runner._execute_request_list.assert_called_once()


 def test_batch_path_rejects_envelope_request_identity_mismatch() -> None:
@@ -166,7 +170,7 @@ def test_batch_path_rejects_envelope_request_identity_mismatch() -> None:
    runner._execute_request_list.assert_not_called()


-def test_step_path_rejects_missing_metadata_before_step_support_check() -> None:
+def test_step_path_allows_missing_metadata_before_step_support_check() -> None:
    runner = make_runner(DiffusionKVCacheMode.PAGED_SCHEDULER)
    runner.pipeline = object()
    runner._supports_step_mode = Mock(return_value=False)
@@ -175,10 +179,10 @@ def test_step_path_rejects_missing_metadata_before_step_support_check() -> None:
        req=SimpleNamespace(request_id="req-0"),
    )

-    with pytest.raises(ValueError, match="requires Diffusion KV metadata"):
+    with pytest.raises(ValueError, match="does not support step execution"):
        runner.execute_stepwise(make_scheduler_output(new_req))

-    runner._supports_step_mode.assert_not_called()
+    runner._supports_step_mode.assert_called_once()


 def test_step_path_rejects_envelope_request_identity_mismatch() -> None:
@@ -284,42 +288,3 @@ def test_dense_worker_call_does_not_extend_model_runner_signature() -> None:
    worker.execute_model(req, SimpleNamespace())

    assert calls == [(req, None)]
-
-
-def test_dlo_worker_selects_request_and_metadata_from_same_envelope(monkeypatch) -> None:
-    calls: list[tuple] = []
-
-    class ModelRunner:
-        def execute_model(self, req, **kwargs):
-            calls.append((req, kwargs))
-            return DiffusionOutput(output=None)
-
-    worker = object.__new__(DiffusionWorker)
-    worker.model_runner = ModelRunner()
-    worker.lora_manager = None
-    worker._get_profiler = lambda: None
-    monkeypatch.setattr(
-        "vllm_omni.diffusion.distributed.parallel_state.get_data_parallel_rank",
-        lambda: 1,
-    )
-    req_0 = SimpleNamespace(request_id="req-0")
-    req_1 = SimpleNamespace(request_id="req-1")
-    metadata_0 = make_metadata("req-0")
-    metadata_1 = make_metadata("req-1")
-    envelopes = [
-        NewRequestData(request_id="req-0", req=req_0, diffusion_kv_metadata=metadata_0),
-        NewRequestData(request_id="req-1", req=req_1, diffusion_kv_metadata=metadata_1),
-    ]
-
-    result = worker.execute_model(envelopes, SimpleNamespace())
-
-    assert calls == [
-        (
-            req_1,
-            {
-                "kv_prefetch_job": None,
-                "diffusion_kv_metadata": metadata_1,
-            },
-        )
-    ]
-    assert result["dp_rank"] == 1
@@ -12,10 +12,9 @@ import torch
 import torch.nn as nn
 from huggingface_hub import snapshot_download
 from vllm.config.load import LoadConfig
-from vllm.model_executor.layers.linear import UnquantizedLinearMethod

 from vllm_omni.diffusion.config import get_current_diffusion_config, get_current_diffusion_config_or_none
-from vllm_omni.diffusion.data import DiffusionParallelConfig, OmniDiffusionConfig
+from vllm_omni.diffusion.data import OmniDiffusionConfig
 from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
 from vllm_omni.diffusion.model_loader.host_weight_plan import (
    HostWeightPlan,
@@ -92,17 +91,6 @@ def _make_loader_with_weights(weight_names: list[str]) -> DiffusersPipelineLoade
    return loader


-def _make_dlo_online_quant_config() -> OmniDiffusionConfig:
-    return OmniDiffusionConfig(
-        model="",
-        dtype=torch.float32,
-        quantization_config="fp8",
-        parallel_config=DiffusionParallelConfig(data_parallel_size=2),
-        enable_distributed_layerwise_offload=True,
-        dlo_use_allgather=True,
-    )
-
-
 def test_strict_check_only_validates_source_prefix_parameters():
    model = _DummyPipelineModel(source_prefix="transformer.")
    loader = _make_loader_with_weights(["transformer.weight"])
@@ -310,7 +298,6 @@ def test_dlo_transfers_loader_plan_and_skips_ordinary_weight_loading(monkeypatch

    assert loader.load_model(load_device="cpu") is model
    assert not loaded_ordinary_weights
-    assert plan.requires_backend
    assert loader.take_host_weight_plan() is plan
    assert loader.take_host_weight_plan() is None

@@ -348,7 +335,7 @@ def test_dlo_plan_loads_component_sources_outside_planned_dit(monkeypatch):
        backing_kind="checkpoint_mmap",
        bindings={
            "transformer.weight": TensorBinding(
-                storage_key="weight",
+                checkpoint_key="weight",
                file_path="unused",
            )
        },
@@ -376,9 +363,8 @@ def test_dlo_plan_loads_component_sources_outside_planned_dit(monkeypatch):
    assert loader.take_host_weight_plan() is plan


-def test_dlo_plan_fallback_retains_ordinary_loader_weights_by_default(monkeypatch):
+def test_dlo_plan_fallback_runs_ordinary_loader(monkeypatch):
    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
-    import vllm_omni.diffusion.model_loader.host_weight_cache as cache_mod

    od_config = SimpleNamespace(
        dtype=torch.float32,
@@ -403,274 +389,11 @@ def test_dlo_plan_fallback_retains_ordinary_loader_weights_by_default(monkeypatc
        lambda *_args, **_kwargs: HostWeightPlanResult(None, "not direct-compatible"),
    )

-    def build_cache(*_args, **_kwargs):
-        pytest.fail("host weight cache must require explicit selection")
-
-    monkeypatch.setattr(cache_mod, "build_host_weight_cache_plan", build_cache)
-
    assert loader.load_model(load_device="cpu") is model
    assert calls == ["load", "process"]
    assert loader.take_host_weight_plan() is None


-def test_dlo_host_weight_cache_is_built_after_final_post_load_mutation(monkeypatch, tmp_path):
-    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
-    import vllm_omni.diffusion.model_loader.host_weight_cache as cache_mod
-
-    od_config = SimpleNamespace(
-        dtype=torch.float32,
-        parallel_config=SimpleNamespace(
-            use_hsdp=False,
-            tensor_parallel_size=1,
-            sequence_parallel_size=1,
-            data_parallel_size=2,
-            cfg_parallel_size=1,
-            pipeline_parallel_size=1,
-            ulysses_degree=1,
-            ring_degree=1,
-            allgather_degree=1,
-            ulysses_mode="strict",
-        ),
-        quantization_config=None,
-        enable_distributed_layerwise_offload=True,
-        dlo_use_allgather=False,
-        dlo_use_host_weight_cache=True,
-        dlo_host_weight_cache_dir=str(tmp_path),
-        dlo_host_weight_cache_lock_timeout=1.0,
-        model="unused",
-        revision=None,
-        model_config={},
-    )
-    loader = DiffusersPipelineLoader(LoadConfig(), od_config)
-    model = nn.Module()
-    model.transformer = nn.Linear(2, 2, bias=False)
-    model.transformer.quant_method = UnquantizedLinearMethod()
-    events: list[str] = []
-    runtime_plan = HostWeightPlan(
-        backing_kind="host_weight_cache",
-        bindings={},
-        runtime_layout_key="layout-key",
-    )
-
-    def load_weights(_model):
-        events.append("load")
-        with torch.no_grad():
-            model.transformer.weight.fill_(1)
-
-    def process_weights(_model, _device):
-        events.append("process")
-        with torch.no_grad():
-            model.transformer.weight.add_(1)
-
-    def calibrate(_model):
-        events.append("calibrate")
-        with torch.no_grad():
-            model.transformer.weight.add_(1)
-
-    original_eval = model.eval
-
-    def eval_model():
-        events.append("eval")
-        return original_eval()
-
-    def build_cache(_pipeline, **kwargs):
-        events.append("cache")
-        assert torch.equal(model.transformer.weight, torch.full((2, 2), 3.0))
-        assert kwargs["tensor_parallel_rank"] == 0
-        assert "data_parallel_size" not in kwargs["sequence_parallel_guard"]
-        assert kwargs["quantization_config"] is None
-        return HostWeightPlanResult(runtime_plan)
-
-    loader._init_from_load_format = lambda *_args, **_kwargs: model  # type: ignore[method-assign]
-    loader.load_weights = load_weights  # type: ignore[method-assign]
-    loader._process_weights_after_loading = process_weights  # type: ignore[method-assign]
-    loader._apply_skip_softmax_calibration = calibrate  # type: ignore[method-assign]
-    model.eval = eval_model  # type: ignore[method-assign]
-    monkeypatch.setattr(
-        loader_mod,
-        "build_checkpoint_mmap_plan",
-        lambda *_args, **_kwargs: HostWeightPlanResult(None, "requires normalization"),
-    )
-    monkeypatch.setattr(cache_mod, "build_host_weight_cache_plan", build_cache)
-
-    assert loader.load_model(load_device="cpu") is model
-    assert events == ["load", "process", "calibrate", "eval", "cache"]
-    assert loader.take_host_weight_plan() is runtime_plan
-
-
-@pytest.mark.parametrize("pin_cpu_memory", [True, False])
-def test_dlo_host_weight_cache_prefers_final_layout_over_checkpoint_plan(monkeypatch, tmp_path, pin_cpu_memory):
-    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
-    import vllm_omni.diffusion.model_loader.host_weight_cache as cache_mod
-
-    od_config = SimpleNamespace(
-        dtype=torch.float32,
-        parallel_config=SimpleNamespace(
-            use_hsdp=False,
-            tensor_parallel_size=1,
-            sequence_parallel_size=1,
-            data_parallel_size=1,
-            cfg_parallel_size=1,
-            pipeline_parallel_size=1,
-        ),
-        quantization_config=None,
-        enable_distributed_layerwise_offload=True,
-        dlo_use_allgather=False,
-        dlo_use_host_weight_cache=True,
-        pin_cpu_memory=pin_cpu_memory,
-        dlo_host_weight_cache_dir=str(tmp_path),
-        dlo_host_weight_cache_pin_limit_gib=0.0,
-        model="unused",
-    )
-    loader = DiffusersPipelineLoader(LoadConfig(), od_config)
-    model = nn.Module()
-    model.transformer = nn.Linear(2, 2, bias=False)
-    events: list[str] = []
-    runtime_plan = HostWeightPlan(
-        backing_kind="host_weight_cache",
-        bindings={},
-        runtime_layout_key="layout-key",
-    )
-
-    loader._init_from_load_format = lambda *_args, **_kwargs: model  # type: ignore[method-assign]
-    loader.load_weights = lambda _model: events.append("load")  # type: ignore[method-assign]
-    loader._process_weights_after_loading = lambda *_args: events.append("process")  # type: ignore[method-assign]
-    loader._apply_skip_softmax_calibration = lambda _model: None  # type: ignore[method-assign]
-
-    def unexpected_checkpoint_plan(*_args, **_kwargs):
-        pytest.fail("selected host weight cache mode must not select the raw checkpoint mmap plan")
-
-    def build_cache(_pipeline, **_kwargs):
-        events.append("cache")
-        return HostWeightPlanResult(runtime_plan)
-
-    monkeypatch.setattr(loader_mod, "build_checkpoint_mmap_plan", unexpected_checkpoint_plan)
-    monkeypatch.setattr(cache_mod, "build_host_weight_cache_plan", build_cache)
-
-    assert loader.load_model(load_device="cpu") is model
-    assert events == ["load", "process", "cache"]
-    assert loader.take_host_weight_plan() is runtime_plan
-
-
-def test_dlo_host_weight_cache_failure_retains_ordinary_weights(monkeypatch, tmp_path):
-    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
-    import vllm_omni.diffusion.model_loader.host_weight_cache as cache_mod
-
-    od_config = SimpleNamespace(
-        dtype=torch.float32,
-        parallel_config=SimpleNamespace(
-            use_hsdp=False,
-            tensor_parallel_size=1,
-            sequence_parallel_size=1,
-            data_parallel_size=2,
-            cfg_parallel_size=1,
-            pipeline_parallel_size=1,
-        ),
-        quantization_config=None,
-        enable_distributed_layerwise_offload=True,
-        dlo_use_allgather=False,
-        dlo_use_host_weight_cache=True,
-        dlo_host_weight_cache_dir=str(tmp_path),
-        model="unused",
-    )
-    loader = DiffusersPipelineLoader(LoadConfig(), od_config)
-    model = nn.Module()
-    model.transformer = nn.Linear(2, 2, bias=False)
-    model.transformer.quant_method = object()
-    loader._init_from_load_format = lambda *_args, **_kwargs: model  # type: ignore[method-assign]
-    loader.load_weights = lambda _model: None  # type: ignore[method-assign]
-    loader._process_weights_after_loading = lambda *_args: None  # type: ignore[method-assign]
-    loader._apply_skip_softmax_calibration = lambda _model: None  # type: ignore[method-assign]
-    monkeypatch.setattr(
-        loader_mod,
-        "build_checkpoint_mmap_plan",
-        lambda *_args, **_kwargs: HostWeightPlanResult(None, "requires normalization"),
-    )
-
-    def build_cache(_pipeline, **kwargs):
-        assert kwargs["quantization_config"] == "module_quant_method"
-        return HostWeightPlanResult(None, "writer timed out", "lock_timeout")
-
-    monkeypatch.setattr(cache_mod, "build_host_weight_cache_plan", build_cache)
-
-    original_weight = model.transformer.weight.detach().clone()
-    assert loader.load_model(load_device="cpu") is model
-    assert torch.equal(model.transformer.weight, original_weight)
-    assert loader.take_host_weight_plan() is None
-
-
-def test_dlo_allgather_online_fp8_uses_ordinary_loader(monkeypatch):
-    from vllm.model_executor.layers.quantization.online.fp8 import (
-        Fp8PerTensorOnlineLinearMethod,
-    )
-
-    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
-
-    od_config = _make_dlo_online_quant_config()
-    loader = DiffusersPipelineLoader(LoadConfig(), od_config)
-    model = nn.Module()
-    model.transformer = nn.Linear(2, 2, bias=False)
-    model.transformer.quant_method = object.__new__(Fp8PerTensorOnlineLinearMethod)
-    model.transformer.quant_method.uses_meta_device = True
-    calls: list[object] = []
-    allowlist_models: list[nn.Module] = []
-
-    original_allowlist_check = loader._unsupported_dlo_allgather_online_quant_methods
-
-    def check_allowlist(candidate: nn.Module) -> tuple[str, ...]:
-        allowlist_models.append(candidate)
-        return original_allowlist_check(candidate)
-
-    loader._init_from_load_format = lambda *_args, **_kwargs: model  # type: ignore[method-assign]
-    monkeypatch.setattr(loader, "_unsupported_dlo_allgather_online_quant_methods", check_allowlist)
-    loader._request_offload_after_quant = lambda _model: 1  # type: ignore[method-assign]
-    loader.load_weights = (  # type: ignore[method-assign]
-        lambda _model, *, stream_online_quant_to_cpu=False: calls.append(("load", stream_online_quant_to_cpu))
-    )
-    loader._process_weights_after_loading = lambda *_args: calls.append("process")  # type: ignore[method-assign]
-    loader._apply_skip_softmax_calibration = lambda _model: None  # type: ignore[method-assign]
-    monkeypatch.setattr(
-        loader_mod,
-        "build_checkpoint_mmap_plan",
-        lambda *_args, **_kwargs: HostWeightPlanResult(
-            None,
-            "online quantization requires the ordinary loader",
-        ),
-    )
-
-    assert loader.load_model(load_device="cpu", device=torch.device("cpu")) is model
-    assert allowlist_models == [model]
-    assert calls == [("load", True), "process"]
-    assert loader.take_host_weight_plan() is None
-
-
-def test_dlo_allgather_rejects_unvalidated_online_quant_method(monkeypatch):
-    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
-
-    class UnsupportedOnlineMethod:
-        uses_meta_device = True
-
-    od_config = _make_dlo_online_quant_config()
-    loader = DiffusersPipelineLoader(LoadConfig(), od_config)
-    model = nn.Module()
-    model.transformer = nn.Linear(2, 2, bias=False)
-    model.transformer.quant_method = UnsupportedOnlineMethod()
-
-    loader._init_from_load_format = lambda *_args, **_kwargs: model  # type: ignore[method-assign]
-    loader._apply_skip_softmax_calibration = lambda _model: None  # type: ignore[method-assign]
-    monkeypatch.setattr(
-        loader_mod,
-        "build_checkpoint_mmap_plan",
-        lambda *_args, **_kwargs: HostWeightPlanResult(
-            None,
-            "online quantization requires the ordinary loader",
-        ),
-    )
-
-    with pytest.raises(ValueError, match="per-tensor FP8 linears"):
-        loader.load_model(load_device="cpu")
-
-
 def test_hsdp_processes_quantized_weights_before_sharding(mocker):
    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
    from vllm_omni.diffusion.offloader.module_collector import PipelineModules
@@ -1,260 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import fcntl
-import multiprocessing as mp
-import shutil
-from pathlib import Path
-
-import pytest
-import torch
-from torch import nn
-
-from vllm_omni.diffusion.model_loader.host_weight_cache import (
-    build_host_weight_cache_plan,
-    default_host_weight_cache_root,
-)
-
-pytestmark = [pytest.mark.diffusion, pytest.mark.cpu, pytest.mark.core_model]
-
-_SPAWN_TIMEOUT_SECONDS = 120
-
-
-class _LoaderMarker:
-    pass
-
-
-class _Transformer(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.proj = nn.Linear(3, 2)
-        self.register_buffer("persistent", torch.arange(2, dtype=torch.float32))
-        self.register_buffer("temporary", torch.ones(1), persistent=False)
-        with torch.no_grad():
-            self.proj.weight.copy_(torch.arange(6, dtype=torch.float32).reshape(2, 3))
-            self.proj.bias.copy_(torch.tensor([7.0, 8.0]))
-
-
-class _Pipeline(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self.transformer = _Transformer()
-
-
-def test_default_root_is_not_split_by_replica_vllm_cache(monkeypatch):
-    monkeypatch.setenv("VLLM_CACHE_ROOT", "/tmp/replica-specific-vllm-cache")
-
-    assert default_host_weight_cache_root().endswith("/.cache/vllm-omni/dlo-host-weights")
-
-
-def _build(pipeline: nn.Module, cache_root: Path, **overrides):
-    kwargs = {
-        "dit_modules": (("transformer", pipeline.transformer),),
-        "loader_type": _LoaderMarker,
-        "cache_root": cache_root,
-        "lock_timeout_seconds": 2.0,
-        "max_shard_bytes": 32,
-        "model_identity": "test/model",
-        "revision": "revision",
-        "runtime_dtype": torch.float32,
-        "load_format": "default",
-        "loader_inputs": {"model_config": {"hidden_size": 3}},
-        "tensor_parallel_size": 1,
-        "tensor_parallel_rank": 0,
-        "sequence_parallel_guard": {"sequence_parallel_size": 1, "backend": "none"},
-        "use_hsdp": False,
-        "enable_expert_parallel": False,
-        "quantization_config": None,
-        "cfg_parallel_size": 1,
-        "pipeline_parallel_size": 1,
-    }
-    kwargs.update(overrides)
-    return build_host_weight_cache_plan(pipeline, **kwargs)
-
-
-def _multiprocess_builder(cache_root: str, queue) -> None:
-    result = _build(_Pipeline(), Path(cache_root))
-    queue.put(
-        (
-            result.fallback_code,
-            result.plan.runtime_layout_key if result.plan is not None else None,
-            sorted(binding.file_path for binding in result.plan.bindings.values()) if result.plan is not None else [],
-        )
-    )
-
-
-def test_equivalent_process_layouts_reuse_one_entry(tmp_path):
-    first = _build(_Pipeline(), tmp_path)
-    second = _build(_Pipeline(), tmp_path)
-
-    assert first.plan is not None
-    assert second.plan is not None
-    assert first.plan.runtime_layout_key == second.plan.runtime_layout_key
-    assert not first.plan.requires_backend
-    assert {binding.file_path for binding in first.plan.bindings.values()} == {
-        binding.file_path for binding in second.plan.bindings.values()
-    }
-    assert set(first.plan.bindings) == {
-        "transformer.persistent",
-        "transformer.proj.bias",
-        "transformer.proj.weight",
-    }
-
-
-def test_equivalent_local_model_paths_reuse_one_entry(tmp_path, monkeypatch):
-    model_dir = tmp_path / "model"
-    model_dir.mkdir()
-    model_link = tmp_path / "model-link"
-    model_link.symlink_to(model_dir, target_is_directory=True)
-    cache_root = tmp_path / "cache"
-    monkeypatch.setenv("HOME", str(tmp_path))
-    monkeypatch.chdir(tmp_path)
-
-    identities = (str(model_dir), "~/model", "model", str(model_link))
-    results = [_build(_Pipeline(), cache_root, model_identity=identity) for identity in identities]
-
-    assert all(result.plan is not None for result in results)
-    assert len({result.plan.runtime_layout_key for result in results if result.plan is not None}) == 1
-    assert len(list((cache_root / "v1").iterdir())) == 1
-
-
-def test_unstable_loader_identity_fails_closed(tmp_path):
-    result = _build(_Pipeline(), tmp_path, loader_inputs={"opaque": object()})
-
-    assert result.plan is None
-    assert result.fallback_code == "unstable_identity"
-    assert result.fallback_reason == "host weight cache identity does not support values of type builtins.object"
-    assert list(tmp_path.rglob("*.safetensors")) == []
-
-
-def test_tp_coordinate_and_sp_implementation_guard_split_entries(tmp_path):
-    pipeline = _Pipeline()
-    tp0 = _build(pipeline, tmp_path, tensor_parallel_size=2, tensor_parallel_rank=0)
-    tp1 = _build(pipeline, tmp_path, tensor_parallel_size=2, tensor_parallel_rank=1)
-    sp2 = _build(
-        pipeline,
-        tmp_path,
-        sequence_parallel_guard={"sequence_parallel_size": 2, "backend": "ulysses"},
-    )
-
-    assert tp0.plan is not None and tp1.plan is not None and sp2.plan is not None
-    assert len({tp0.plan.runtime_layout_key, tp1.plan.runtime_layout_key, sp2.plan.runtime_layout_key}) == 3
-
-
-def test_final_runtime_content_is_authoritative(tmp_path):
-    first_pipeline = _Pipeline()
-    second_pipeline = _Pipeline()
-    with torch.no_grad():
-        second_pipeline.transformer.proj.weight[0, 0] += 1
-
-    first = _build(first_pipeline, tmp_path)
-    second = _build(second_pipeline, tmp_path)
-
-    assert first.plan is not None and second.plan is not None
-    assert first.plan.runtime_layout_key != second.plan.runtime_layout_key
-
-
-def test_expert_parallel_ownership_fails_closed(tmp_path):
-    result = _build(_Pipeline(), tmp_path, enable_expert_parallel=True)
-
-    assert result.plan is None
-    assert result.fallback_code == "unsupported_expert_parallel"
-    assert list(tmp_path.rglob("*.safetensors")) == []
-
-
-def test_pipeline_parallel_ownership_fails_closed(tmp_path):
-    result = _build(_Pipeline(), tmp_path, pipeline_parallel_size=2)
-
-    assert result.plan is None
-    assert result.fallback_code == "unsupported_pp"
-    assert list(tmp_path.rglob("*.safetensors")) == []
-
-
-def test_corrupt_published_entry_is_rebuilt_under_the_key_lock(tmp_path):
-    first = _build(_Pipeline(), tmp_path)
-    assert first.plan is not None
-    shard = Path(next(iter(first.plan.bindings.values())).file_path)
-    old_inode = shard.stat().st_ino
-    with shard.open("r+b") as handle:
-        handle.seek(-1, 2)
-        byte = handle.read(1)
-        handle.seek(-1, 2)
-        handle.write(bytes([byte[0] ^ 0xFF]))
-
-    second = _build(_Pipeline(), tmp_path)
-
-    assert second.plan is not None
-    assert second.plan.runtime_layout_key == first.plan.runtime_layout_key
-    assert Path(next(iter(second.plan.bindings.values())).file_path).stat().st_ino != old_inode
-
-
-def test_lock_timeout_keeps_the_ordinary_tensors(tmp_path):
-    initial = _build(_Pipeline(), tmp_path)
-    assert initial.plan is not None
-    cache_key = initial.plan.runtime_layout_key
-    entry_dir = Path(next(iter(initial.plan.bindings.values())).file_path).parent
-    shutil.rmtree(entry_dir)
-    lock_path = tmp_path / ".locks" / f"{cache_key}.lock"
-
-    with lock_path.open("a+b") as lock_handle:
-        fcntl.flock(lock_handle.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
-        result = _build(_Pipeline(), tmp_path, lock_timeout_seconds=0.01)
-
-    assert result.plan is None
-    assert result.fallback_code == "lock_timeout"
-
-
-def test_next_writer_removes_stale_temp_after_process_death(tmp_path):
-    initial = _build(_Pipeline(), tmp_path)
-    assert initial.plan is not None
-    cache_key = initial.plan.runtime_layout_key
-    entry_dir = Path(next(iter(initial.plan.bindings.values())).file_path).parent
-    shutil.rmtree(entry_dir)
-    stale_temp = tmp_path / ".tmp" / f"{cache_key}.stale-writer"
-    stale_temp.mkdir(parents=True)
-    (stale_temp / "partial").write_bytes(b"partial")
-
-    result = _build(_Pipeline(), tmp_path)
-
-    assert result.plan is not None
-    assert not stale_temp.exists()
-
-
-@pytest.mark.parametrize("case", ["alias", "external_alias", "noncontiguous"])
-def test_unsupported_tensor_layouts_fail_closed(tmp_path, case):
-    pipeline = _Pipeline()
-    if case == "alias":
-        pipeline.transformer.tied = pipeline.transformer.proj.weight
-        expected_code = "unsupported_alias"
-    elif case == "external_alias":
-        pipeline.encoder = nn.Module()
-        pipeline.encoder.shared = pipeline.transformer.proj.weight
-        expected_code = "unsupported_alias"
-    else:
-        pipeline.transformer.proj.weight = nn.Parameter(torch.arange(6, dtype=torch.float32).reshape(3, 2).t())
-        expected_code = "unsupported_tensor"
-
-    result = _build(pipeline, tmp_path)
-
-    assert result.plan is None
-    assert result.fallback_code == expected_code
-    assert list(tmp_path.rglob("*.safetensors")) == []
-
-
-def test_concurrent_processes_publish_one_immutable_entry(tmp_path):
-    context = mp.get_context("spawn")
-    queue = context.Queue()
-    processes = [context.Process(target=_multiprocess_builder, args=(str(tmp_path), queue)) for _ in range(2)]
-    for process in processes:
-        process.start()
-    # Spawn workers import the full vLLM stack before entering the target;
-    # loaded CI hosts can spend more than 30 seconds in that startup path.
-    results = [queue.get(timeout=_SPAWN_TIMEOUT_SECONDS) for _ in processes]
-    for process in processes:
-        process.join(timeout=_SPAWN_TIMEOUT_SECONDS)
-
-    assert all(process.exitcode == 0 for process in processes)
-    assert all(code is None for code, _, _ in results)
-    assert len({key for _, key, _ in results}) == 1
-    assert len({tuple(paths) for _, _, paths in results}) == 1
-    assert len(list((tmp_path / "v1").iterdir())) == 1
@@ -6,10 +6,7 @@ from types import SimpleNamespace

 import pytest

-from vllm_omni.diffusion.models.dreamzero.pipeline_dreamzero import (
-    DREAMZERO_MODEL_OWNED_STATE_BYTES_PER_SESSION,
-    DreamZeroPipeline,
-)
+from vllm_omni.diffusion.models.dreamzero.pipeline_dreamzero import DreamZeroPipeline
 from vllm_omni.diffusion.models.dreamzero.state_dreamzero import DreamZeroState

 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -35,37 +32,6 @@ def test_dreamzero_pipeline_state_is_session_keyed() -> None:
    assert session_b.call_count == 3


-def test_dreamzero_kv_spec_accounts_for_model_owned_cuda_state(monkeypatch) -> None:
-    pipeline = DreamZeroPipeline.__new__(DreamZeroPipeline)
-    pipeline.transformer = SimpleNamespace(
-        frame_seqlen=16,
-        blocks=[
-            SimpleNamespace(
-                self_attn=SimpleNamespace(
-                    max_attention_size=96,
-                    tp_num_heads=4,
-                )
-            )
-        ],
-        text_len=8,
-        model_type="t2v",
-        num_layers=2,
-        num_heads=4,
-        dim=256,
-        num_frame_per_block=4,
-        num_action_per_block=2,
-        num_state_per_block=1,
-    )
-    monkeypatch.setattr(
-        "vllm_omni.diffusion.models.dreamzero.pipeline_dreamzero.get_classifier_free_guidance_world_size",
-        lambda: 1,
-    )
-
-    spec = pipeline.ar_diffusion_kv_cache_spec()
-
-    assert spec.model_owned_state_bytes_per_session == DREAMZERO_MODEL_OWNED_STATE_BYTES_PER_SESSION
-
-
 def test_dreamzero_pipeline_state_follows_runner_lifecycle_notifications() -> None:
    pipeline = _empty_pipeline()

@@ -118,14 +118,6 @@ def _call_mgr(
    )


-def test_cache_manager_registers_attention_without_adding_dense_state() -> None:
-    mgr = _make_cache_mgr()
-
-    assert isinstance(mgr, nn.Module)
-    assert dict(mgr.named_modules())["attn"] is mgr.attn
-    assert mgr.state_dict() == {}
-
-
 # ============================================================
 # Test 1: No AR KV — basic cache → reuse
 # ============================================================
@@ -1 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
@@ -1,13 +0,0 @@
-{
-  "_class_name": "LingBotWorldCausalDMDPipeline",
-  "_diffusers_version": "0.35.0.dev0",
-  "boundary_ratio": 0.9,
-  "image_encoder": [null, null],
-  "image_processor": [null, null],
-  "scheduler": ["diffusers", "UniPCMultistepScheduler"],
-  "text_encoder": ["transformers", "UMT5EncoderModel"],
-  "tokenizer": ["transformers", "T5TokenizerFast"],
-  "transformer": ["diffusers", "CausalLingBotWorldTransformer3DModel"],
-  "transformer_2": [null, null],
-  "vae": ["diffusers", "AutoencoderKLWan"]
-}
@@ -1,42 +0,0 @@
-{
-  "sources": {
-    "config": "https://huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast-diffusers/resolve/main/transformer/config.json",
-    "shard_1_header": "https://huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast-diffusers/resolve/main/transformer/diffusion_pytorch_model-00001-of-00008.safetensors",
-    "shard_8_header": "https://huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast-diffusers/resolve/main/transformer/diffusion_pytorch_model-00008-of-00008.safetensors"
-  },
-  "dtype": "F32",
-  "shard_representatives": {
-    "diffusion_pytorch_model-00001-of-00008.safetensors": "blocks.0.cam_injector_layer1.bias",
-    "diffusion_pytorch_model-00002-of-00008.safetensors": "blocks.4.cam_injector_layer1.bias",
-    "diffusion_pytorch_model-00003-of-00008.safetensors": "blocks.10.cam_injector_layer1.bias",
-    "diffusion_pytorch_model-00004-of-00008.safetensors": "blocks.14.cam_scale_layer.bias",
-    "diffusion_pytorch_model-00005-of-00008.safetensors": "blocks.20.cam_injector_layer1.bias",
-    "diffusion_pytorch_model-00006-of-00008.safetensors": "blocks.25.cam_injector_layer1.bias",
-    "diffusion_pytorch_model-00007-of-00008.safetensors": "blocks.30.cam_injector_layer1.bias",
-    "diffusion_pytorch_model-00008-of-00008.safetensors": "blocks.35.cam_injector_layer1.bias"
-  },
-  "parameter_shapes": {
-    "blocks.0.cam_injector_layer1.bias": [5120],
-    "blocks.4.cam_injector_layer1.bias": [5120],
-    "blocks.10.cam_injector_layer1.bias": [5120],
-    "blocks.14.cam_scale_layer.bias": [5120],
-    "blocks.20.cam_injector_layer1.bias": [5120],
-    "blocks.25.cam_injector_layer1.bias": [5120],
-    "blocks.30.cam_injector_layer1.bias": [5120],
-    "blocks.35.cam_injector_layer1.bias": [5120],
-    "patch_embedding.weight": [5120, 36, 1, 2, 2],
-    "patch_embedding_wancamctrl.weight": [5120, 1536],
-    "c2ws_hidden_states_layer1.weight": [5120, 5120],
-    "c2ws_hidden_states_layer2.weight": [5120, 5120],
-    "blocks.0.modulation": [1, 6, 5120],
-    "blocks.0.self_attn.q.weight": [5120, 5120],
-    "blocks.0.ffn.0.weight": [13824, 5120],
-    "blocks.0.ffn.2.weight": [5120, 13824],
-    "blocks.0.cam_injector_layer1.weight": [5120, 5120],
-    "blocks.0.cam_injector_layer2.weight": [5120, 5120],
-    "blocks.0.cam_scale_layer.weight": [5120, 5120],
-    "blocks.0.cam_shift_layer.weight": [5120, 5120],
-    "head.modulation": [1, 2, 5120],
-    "head.head.weight": [64, 5120]
-  }
-}
@@ -1,30 +0,0 @@
-{
-  "_class_name": "UniPCMultistepScheduler",
-  "_diffusers_version": "0.35.0.dev0",
-  "beta_end": 0.02,
-  "beta_schedule": "linear",
-  "beta_start": 0.0001,
-  "disable_corrector": [],
-  "dynamic_thresholding_ratio": 0.995,
-  "final_sigmas_type": "zero",
-  "flow_shift": 5.0,
-  "lower_order_final": true,
-  "num_train_timesteps": 1000,
-  "predict_x0": true,
-  "prediction_type": "flow_prediction",
-  "rescale_betas_zero_snr": false,
-  "sample_max_value": 1.0,
-  "solver_order": 2,
-  "solver_p": null,
-  "solver_type": "bh2",
-  "steps_offset": 0,
-  "thresholding": false,
-  "time_shift_type": "exponential",
-  "timestep_spacing": "linspace",
-  "trained_betas": null,
-  "use_beta_sigmas": false,
-  "use_dynamic_shifting": false,
-  "use_exponential_sigmas": false,
-  "use_flow_sigmas": true,
-  "use_karras_sigmas": false
-}
@@ -1,112 +0,0 @@
-{
-  "source": "https://huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast-diffusers/resolve/main/transformer/diffusion_pytorch_model.safetensors.index.json",
-  "index_parameter_count": 1421,
-  "canonical_name_set_sha256": "5d31b7f64fd5e5e271b789b2dbe50fb7bbf146bd8b750646bdc9abb4e5b8139d",
-  "canonical_name_to_shard_sha256": "e1addc27d2b1fad6f10226a23a265ee3e3c61e74dce4d5859727bdf180a67992",
-  "canonical_serialization": "sorted lines formatted as parameter_name=shard_filename with LF separators and no trailing LF",
-  "shards": {
-    "diffusion_pytorch_model-00001-of-00008.safetensors": {
-      "parameter_count": 180,
-      "parameter_names_sha256": "81a3d4778de604e75609aae7dac3d8dcbc081811e7748267fe953bbeb375e79d"
-    },
-    "diffusion_pytorch_model-00002-of-00008.safetensors": {
-      "parameter_count": 180,
-      "parameter_names_sha256": "4e36e2596a6468f29179fa5795396f2e9a7f7d8debc912b34a6f6c2c7c7c610d"
-    },
-    "diffusion_pytorch_model-00003-of-00008.safetensors": {
-      "parameter_count": 180,
-      "parameter_names_sha256": "6f74a7b2cf7444c2daa2b3fc32bd3787589b35543c230c5c13b89f0f0f2a0998"
-    },
-    "diffusion_pytorch_model-00004-of-00008.safetensors": {
-      "parameter_count": 180,
-      "parameter_names_sha256": "ec5c967ed5271ba7834337c4ee93a73f8ed0eead4492d4920624251c65c058ac"
-    },
-    "diffusion_pytorch_model-00005-of-00008.safetensors": {
-      "parameter_count": 180,
-      "parameter_names_sha256": "ad1fdab5830dd9ce43ea007f112b4be223327b7c8197c766d9cf9219e74d619b"
-    },
-    "diffusion_pytorch_model-00006-of-00008.safetensors": {
-      "parameter_count": 180,
-      "parameter_names_sha256": "fd58f0d86be42d8afcf55d3e07d8a22f7f6ef5064da931a600c890b71b6dfbae"
-    },
-    "diffusion_pytorch_model-00007-of-00008.safetensors": {
-      "parameter_count": 180,
-      "parameter_names_sha256": "305a1714b5f4fb54d5877c593b451273ab1490d90fd0896a299d96e140295047"
-    },
-    "diffusion_pytorch_model-00008-of-00008.safetensors": {
-      "parameter_count": 161,
-      "parameter_names_sha256": "4c18d94d96168682eb212f93e3cdf7313928776d8d9e6e4de0255fe899decb25"
-    }
-  },
-  "top_level_modules": [
-    "blocks",
-    "c2ws_hidden_states_layer1",
-    "c2ws_hidden_states_layer2",
-    "head",
-    "patch_embedding",
-    "patch_embedding_wancamctrl",
-    "text_embedding",
-    "time_embedding",
-    "time_projection"
-  ],
-  "non_block_parameters": [
-    "c2ws_hidden_states_layer1.bias",
-    "c2ws_hidden_states_layer1.weight",
-    "c2ws_hidden_states_layer2.bias",
-    "c2ws_hidden_states_layer2.weight",
-    "head.head.bias",
-    "head.head.weight",
-    "head.modulation",
-    "patch_embedding.bias",
-    "patch_embedding.weight",
-    "patch_embedding_wancamctrl.bias",
-    "patch_embedding_wancamctrl.weight",
-    "text_embedding.0.bias",
-    "text_embedding.0.weight",
-    "text_embedding.2.bias",
-    "text_embedding.2.weight",
-    "time_embedding.0.bias",
-    "time_embedding.0.weight",
-    "time_embedding.2.bias",
-    "time_embedding.2.weight",
-    "time_projection.1.bias",
-    "time_projection.1.weight"
-  ],
-  "block_parameter_suffixes": [
-    "cam_injector_layer1.bias",
-    "cam_injector_layer1.weight",
-    "cam_injector_layer2.bias",
-    "cam_injector_layer2.weight",
-    "cam_scale_layer.bias",
-    "cam_scale_layer.weight",
-    "cam_shift_layer.bias",
-    "cam_shift_layer.weight",
-    "cross_attn.k.bias",
-    "cross_attn.k.weight",
-    "cross_attn.norm_k.weight",
-    "cross_attn.norm_q.weight",
-    "cross_attn.o.bias",
-    "cross_attn.o.weight",
-    "cross_attn.q.bias",
-    "cross_attn.q.weight",
-    "cross_attn.v.bias",
-    "cross_attn.v.weight",
-    "ffn.0.bias",
-    "ffn.0.weight",
-    "ffn.2.bias",
-    "ffn.2.weight",
-    "modulation",
-    "norm3.bias",
-    "norm3.weight",
-    "self_attn.k.bias",
-    "self_attn.k.weight",
-    "self_attn.norm_k.weight",
-    "self_attn.norm_q.weight",
-    "self_attn.o.bias",
-    "self_attn.o.weight",
-    "self_attn.q.bias",
-    "self_attn.q.weight",
-    "self_attn.v.bias",
-    "self_attn.v.weight"
-  ]
-}
@@ -1,294 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-from __future__ import annotations
-
-import asyncio
-
-import pytest
-import torch
-
-from vllm_omni.diffusion.models.lingbot_world.actions import (
-    LINGBOT_CAMERA_ACTION_SCHEMA,
-    LINGBOT_CAMERA_TRAJECTORY_SCHEMA,
-    LingBotCameraControlReducer,
-    integrate_lingbot_camera_actions,
-    parse_lingbot_camera_action_frames,
-)
-from vllm_omni.experimental.ar_diffusion.session import (
-    ARDiffusionSession,
-    ARDiffusionSessionEvent,
-)
-from vllm_omni.experimental.ar_diffusion.tick_protocol import (
-    ARDiffusionChunkMetadata,
-    ARDiffusionControlInput,
-    ARDiffusionTickRequest,
-)
-
-pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
-
-
-def _action_control(mode: str, **data) -> ARDiffusionControlInput:
-    return ARDiffusionControlInput(
-        track="camera",
-        schema=LINGBOT_CAMERA_ACTION_SCHEMA,
-        data={"mode": mode, **data},
-    )
-
-
-def _prepared_frames(prepared) -> tuple[tuple[str, ...], ...]:
-    assert len(prepared.controls) == 1
-    return prepared.controls[0].data["frames"]
-
-
-def test_state_controls_hold_release_and_preserve_short_pulse() -> None:
-    reducer = LingBotCameraControlReducer()
-    pressed = reducer.prepare(
-        current_controls={},
-        events=(
-            ARDiffusionSessionEvent(
-                event_id=1,
-                controls=(
-                    _action_control(
-                        "state",
-                        transitions=[{"actions": ["W"], "client_ts_ms": 10}],
-                    ),
-                ),
-            ),
-        ),
-        chunk_index=0,
-    )
-    assert _prepared_frames(pressed) == (("w",), ("w",), ("w",))
-    assert pressed.controls[0].to_dict()["data"]["frames"] == [["w"], ["w"], ["w"]]
-    reducer.commit(pressed)
-
-    held = reducer.prepare(
-        current_controls={"camera": pressed.controls[0]},
-        events=(),
-        chunk_index=1,
-    )
-    assert _prepared_frames(held) == (("w",), ("w",), ("w",))
-    reducer.commit(held)
-
-    pulse = reducer.prepare(
-        current_controls={"camera": held.controls[0]},
-        events=(
-            ARDiffusionSessionEvent(
-                event_id=2,
-                controls=(
-                    _action_control(
-                        "state",
-                        transitions=[
-                            {"actions": ["j"], "client_ts_ms": 20},
-                            {"actions": [], "client_ts_ms": 21},
-                        ],
-                    ),
-                ),
-            ),
-        ),
-        chunk_index=2,
-    )
-    assert _prepared_frames(pulse) == (("j",), (), ())
-    reducer.commit(pulse)
-
-    released = reducer.prepare(
-        current_controls={"camera": pulse.controls[0]},
-        events=(),
-        chunk_index=3,
-    )
-    assert _prepared_frames(released) == ((), (), ())
-
-
-def test_script_is_transactional_consumed_once_and_neutral_padded() -> None:
-    reducer = LingBotCameraControlReducer()
-    event = ARDiffusionSessionEvent(
-        event_id=1,
-        controls=(
-            _action_control(
-                "script",
-                frames=[["w"], ["d"], ["l"], ["i"]],
-            ),
-        ),
-    )
-    first_attempt = reducer.prepare(
-        current_controls={},
-        events=(event,),
-        chunk_index=0,
-    )
-    assert _prepared_frames(first_attempt) == (("w",), ("d",), ("l",))
-
-    # An uncommitted failed attempt must not consume the reducer's script.
-    second_attempt = reducer.prepare(
-        current_controls={},
-        events=(event,),
-        chunk_index=0,
-    )
-    assert _prepared_frames(second_attempt) == (("w",), ("d",), ("l",))
-    reducer.commit(second_attempt)
-
-    tail = reducer.prepare(
-        current_controls={"camera": second_attempt.controls[0]},
-        events=(),
-        chunk_index=1,
-    )
-    assert _prepared_frames(tail) == (("i",), (), ())
-    reducer.commit(tail)
-    exhausted = reducer.prepare(
-        current_controls={"camera": tail.controls[0]},
-        events=(),
-        chunk_index=2,
-    )
-    assert _prepared_frames(exhausted) == ((), (), ())
-
-
-def test_trajectory_control_remains_an_expert_override() -> None:
-    reducer = LingBotCameraControlReducer()
-    trajectory = ARDiffusionControlInput(
-        track="camera",
-        schema=LINGBOT_CAMERA_TRAJECTORY_SCHEMA,
-        data={"poses": [[[1.0]]], "intrinsics": [[1.0]]},
-    )
-    prepared = reducer.prepare(
-        current_controls={},
-        events=(
-            ARDiffusionSessionEvent(
-                event_id=1,
-                controls=(trajectory,),
-            ),
-        ),
-        chunk_index=0,
-    )
-    assert prepared.controls == (trajectory,)
-
-
-def test_action_integrator_matches_lingbot_motion_and_calibration() -> None:
-    trajectory, pitch = integrate_lingbot_camera_actions(
-        [["w"], ["d"], ["l"]],
-        width=832,
-        height=480,
-    )
-
-    assert pitch == 0.0
-    torch.testing.assert_close(
-        trajectory.poses[0, :3, 3],
-        torch.tensor([0.0, 0.0, 0.05], dtype=torch.float64),
-        atol=1e-6,
-        rtol=0,
-    )
-    torch.testing.assert_close(
-        trajectory.poses[1, :3, 3],
-        torch.tensor([0.05, 0.0, 0.05], dtype=torch.float64),
-        atol=1e-6,
-        rtol=0,
-    )
-    expected_yaw = torch.tensor(
-        [
-            [0.9945219, 0.0, 0.10452846],
-            [0.0, 1.0, 0.0],
-            [-0.10452846, 0.0, 0.9945219],
-        ],
-        dtype=torch.float64,
-    )
-    torch.testing.assert_close(
-        trajectory.poses[2, :3, :3],
-        expected_yaw,
-        atol=1e-6,
-        rtol=0,
-    )
-    torch.testing.assert_close(
-        trajectory.intrinsics[0],
-        torch.tensor([500.0, 500.0, 416.0, 240.0]),
-    )
-
-    resized, _ = integrate_lingbot_camera_actions(
-        [[]],
-        width=416,
-        height=240,
-    )
-    scaled = resized.intrinsics[0] * torch.tensor([416 / 832, 240 / 480, 416 / 832, 240 / 480])
-    torch.testing.assert_close(
-        scaled,
-        torch.tensor([500.0, 500.0, 208.0, 120.0]),
-    )
-
-
-@pytest.mark.parametrize(
-    ("data", "message"),
-    [
-        ({"mode": "frames", "frames": [["x"], [], []]}, "W/A/S/D/I/J/K/L"),
-        ({"mode": "frames", "frames": [["w"]]}, "exactly one action list"),
-        ({"mode": "state", "frames": [[], [], []]}, "mode='frames'"),
-    ],
-)
-def test_model_tick_action_payload_validation(data, message: str) -> None:
-    with pytest.raises(ValueError, match=message):
-        parse_lingbot_camera_action_frames(data, expected_frames=3)
-
-
-class _TickConsumer:
-    def __init__(self) -> None:
-        self.ticks: list[ARDiffusionTickRequest] = []
-
-    async def execute_tick(self, tick: ARDiffusionTickRequest) -> ARDiffusionChunkMetadata:
-        self.ticks.append(tick)
-        return ARDiffusionChunkMetadata.from_tick(tick)
-
-    def chunk_metadata(self, output: ARDiffusionChunkMetadata) -> ARDiffusionChunkMetadata:
-        return output
-
-
-class _Lifecycle:
-    async def reset_session(self, session_id: str) -> None:
-        del session_id
-
-    async def close_session(self, session_id: str) -> None:
-        del session_id
-
-
-def test_session_reducer_keeps_prompt_and_actions_atomic_and_resets() -> None:
-    async def exercise() -> None:
-        consumer = _TickConsumer()
-        reducer = LingBotCameraControlReducer()
-        session = ARDiffusionSession(
-            "world-actions",
-            tick_consumer=consumer,
-            lifecycle=_Lifecycle(),
-            max_pending_events=8,
-            request_id_factory=lambda session_id, chunk_index: f"{session_id}-{chunk_index}",
-            control_reducer=reducer,
-        )
-        await session.accept_event(
-            ARDiffusionSessionEvent(
-                event_id=1,
-                prompt="turn toward the window",
-                controls=(
-                    _action_control(
-                        "state",
-                        transitions=[{"actions": ["d"]}],
-                    ),
-                ),
-            )
-        )
-        await session.next_chunk()
-
-        tick = consumer.ticks[0]
-        assert tick.prompt == "turn toward the window"
-        assert tick.applied_event_ids == (1,)
-        assert tick.controls[0].data["frames"] == (("d",), ("d",), ("d",))
-
-        await session.reset()
-        await session.accept_event(
-            ARDiffusionSessionEvent(
-                event_id=2,
-                controls=(
-                    _action_control(
-                        "state",
-                        transitions=[{"actions": []}],
-                    ),
-                ),
-            )
-        )
-        await session.next_chunk()
-        assert consumer.ticks[1].chunk_index == 0
-        assert consumer.ticks[1].controls[0].data["frames"] == ((), (), ())
-
-    asyncio.run(exercise())
@@ -1,543 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-from __future__ import annotations
-
-import importlib.util
-import math
-import sys
-from pathlib import Path
-from types import ModuleType
-
-import pytest
-import torch
-import torch.nn.functional as F
-from torch import nn
-
-pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
-
-_MODULE_PATH = Path(__file__).parents[4] / "vllm_omni/diffusion/models/lingbot_world/transformer.py"
-_MODULE_NAME = "_lingbot_world_attention_under_test"
-_MISSING_MODULE = object()
-_STUBBED_MODULE_NAMES = (
-    "vllm",
-    "vllm.distributed",
-    "vllm.model_executor",
-    "vllm.model_executor.layers",
-    "vllm.model_executor.layers.conv",
-    "vllm.model_executor.layers.linear",
-    "vllm.model_executor.model_loader",
-    "vllm.model_executor.model_loader.weight_utils",
-    "vllm.model_executor.utils",
-    "vllm_omni",
-    "vllm_omni.diffusion",
-    "vllm_omni.diffusion.attention",
-    "vllm_omni.diffusion.attention.layer",
-    "vllm_omni.diffusion.layers",
-    "vllm_omni.diffusion.layers.norm",
-    "vllm_omni.diffusion.layers.rope",
-    "vllm_omni.experimental",
-    "vllm_omni.experimental.ar_diffusion",
-    "vllm_omni.experimental.ar_diffusion.kv_cache",
-    "vllm_omni.experimental.ar_diffusion.kv_cache.paged_attention",
-)
-
-
-@pytest.fixture(autouse=True)
-def _inference_context():
-    with torch.inference_mode():
-        yield
-
-
-def _install_vllm_stubs() -> None:
-    for name in _STUBBED_MODULE_NAMES:
-        sys.modules[name] = ModuleType(name)
-
-    distributed = sys.modules["vllm.distributed"]
-    distributed.get_tensor_model_parallel_rank = lambda: 0
-    distributed.get_tensor_model_parallel_world_size = lambda: 1
-    distributed.tensor_model_parallel_all_reduce = lambda value: value
-
-    def set_weight_attrs(weight: torch.Tensor, attrs: dict) -> None:
-        for name, value in attrs.items():
-            setattr(weight, name, value)
-
-    sys.modules["vllm.model_executor.utils"].set_weight_attrs = set_weight_attrs
-    sys.modules["vllm.model_executor.model_loader.weight_utils"].default_weight_loader = lambda param, loaded_weight: (
-        param.data.copy_(loaded_weight)
-    )
-
-    class _Linear(nn.Module):
-        def __init__(
-            self,
-            input_size: int,
-            output_size: int,
-            *,
-            bias: bool = True,
-            return_bias: bool = False,
-            **kwargs,
-        ) -> None:
-            super().__init__()
-            del kwargs
-            self.return_bias = return_bias
-            self.weight = nn.Parameter(torch.empty(output_size, input_size))
-            self.bias = nn.Parameter(torch.empty(output_size)) if bias else None
-            nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
-            if self.bias is not None:
-                nn.init.zeros_(self.bias)
-            self.calls = 0
-
-        def forward(self, value: torch.Tensor):
-            self.calls += 1
-            output = F.linear(value, self.weight, self.bias)
-            return (output, self.bias) if self.return_bias else output
-
-    linear = sys.modules["vllm.model_executor.layers.linear"]
-    linear.ColumnParallelLinear = _Linear
-    linear.RowParallelLinear = _Linear
-
-    class _QKVLinear(nn.Module):
-        def __init__(
-            self,
-            hidden_size: int,
-            head_size: int,
-            total_num_heads: int,
-            total_num_kv_heads: int | None = None,
-            *,
-            bias: bool = True,
-            **kwargs,
-        ) -> None:
-            super().__init__()
-            del kwargs
-            self.num_heads = total_num_heads
-            self.num_kv_heads = total_num_kv_heads or total_num_heads
-            output_size = (self.num_heads + 2 * self.num_kv_heads) * head_size
-            self.weight = nn.Parameter(torch.empty(output_size, hidden_size))
-            self.bias = nn.Parameter(torch.empty(output_size)) if bias else None
-            nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
-            if self.bias is not None:
-                nn.init.zeros_(self.bias)
-            self.calls = 0
-
-            def load_shard(param: torch.Tensor, loaded_weight: torch.Tensor, shard_id: str) -> None:
-                shard_index = {"q": 0, "k": 1, "v": 2}[shard_id]
-                shard_size = param.shape[0] // 3
-                param.data.narrow(0, shard_index * shard_size, shard_size).copy_(loaded_weight)
-
-            self.weight.weight_loader = load_shard
-            if self.bias is not None:
-                self.bias.weight_loader = load_shard
-
-        def forward(self, value: torch.Tensor):
-            self.calls += 1
-            return F.linear(value, self.weight, self.bias), None
-
-    linear.QKVParallelLinear = _QKVLinear
-
-    class _Conv3dLayer(nn.Conv3d):
-        def __init__(self, in_channels: int, out_channels: int, kernel_size, stride, **kwargs) -> None:
-            super().__init__(in_channels, out_channels, kernel_size=kernel_size, stride=stride, **kwargs)
-
-    sys.modules["vllm.model_executor.layers.conv"].Conv3dLayer = _Conv3dLayer
-
-    class _LayerNorm(nn.LayerNorm):
-        def __init__(self, dim: int, eps: float = 1e-6, elementwise_affine: bool = True) -> None:
-            super().__init__(dim, eps=eps, elementwise_affine=elementwise_affine)
-
-    sys.modules["vllm_omni.diffusion.layers.norm"].LayerNorm = _LayerNorm
-
-    class _Attention(nn.Module):
-        def __init__(self, *args, softmax_scale: float, **kwargs) -> None:
-            super().__init__()
-            del args
-            self.softmax_scale = softmax_scale
-            self.causal = kwargs["causal"]
-            self.calls: list[tuple[torch.Tensor, torch.Tensor, torch.Tensor]] = []
-
-        def forward(
-            self,
-            query: torch.Tensor,
-            key: torch.Tensor,
-            value: torch.Tensor,
-            attn_metadata=None,
-        ) -> torch.Tensor:
-            del attn_metadata
-            self.calls.append((query.detach().clone(), key.detach().clone(), value.detach().clone()))
-            scores = torch.einsum("bqhd,bkhd->bhqk", query, key) * self.softmax_scale
-            weights = scores.softmax(dim=-1)
-            return torch.einsum("bhqk,bkhd->bqhd", weights, value)
-
-    sys.modules["vllm_omni.diffusion.attention.layer"].Attention = _Attention
-
-    class _RotaryEmbeddingWan(nn.Module):
-        def __init__(self, *args, **kwargs) -> None:
-            super().__init__()
-            del args, kwargs
-            self.calls: list[tuple[torch.Tensor, torch.Tensor]] = []
-
-        def forward(self, value: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
-            self.calls.append((cos.detach().clone(), sin.detach().clone()))
-            return value
-
-    sys.modules["vllm_omni.diffusion.layers.rope"].RotaryEmbeddingWan = _RotaryEmbeddingWan
-
-    paged_attention = sys.modules["vllm_omni.experimental.ar_diffusion.kv_cache.paged_attention"]
-
-    class _PagedLayerContext:
-        pass
-
-    class _PagedLayerInputs:
-        pass
-
-    paged_attention.ARDiffusionPagedLayerContext = _PagedLayerContext
-    paged_attention.ARDiffusionPagedLayerInputs = _PagedLayerInputs
-    paged_attention.ar_diffusion_paged_attention = lambda *args, **kwargs: (_ for _ in ()).throw(
-        AssertionError("CUDA paged attention was not expected in this CPU test")
-    )
-    paged_attention.paged_write_attn = lambda *args, **kwargs: (_ for _ in ()).throw(
-        AssertionError("paged path was not expected in this direct-cache test")
-    )
-
-
-def _load_module():
-    assert _MODULE_PATH.exists(), "LingBot attention module has not been implemented"
-    module_names = (*_STUBBED_MODULE_NAMES, _MODULE_NAME)
-    previous = {name: sys.modules.get(name, _MISSING_MODULE) for name in module_names}
-    try:
-        _install_vllm_stubs()
-        spec = importlib.util.spec_from_file_location(_MODULE_NAME, _MODULE_PATH)
-        assert spec is not None and spec.loader is not None
-        module = importlib.util.module_from_spec(spec)
-        sys.modules[spec.name] = module
-        spec.loader.exec_module(module)
-        return module
-    finally:
-        for name, original in previous.items():
-            if original is _MISSING_MODULE:
-                sys.modules.pop(name, None)
-            else:
-                sys.modules[name] = original
-
-
-def test_load_module_restores_process_module_state() -> None:
-    module_names = (*_STUBBED_MODULE_NAMES, _MODULE_NAME)
-    previous = {name: sys.modules.get(name, _MISSING_MODULE) for name in module_names}
-
-    _load_module()
-
-    for name, original in previous.items():
-        assert sys.modules.get(name, _MISSING_MODULE) is original
-
-
-def test_allocate_lingbot_cache_creates_request_local_layer_storage() -> None:
-    module = _load_module()
-
-    cache = module.allocate_lingbot_cache(
-        batch_size=2,
-        num_layers=3,
-        max_tokens=7,
-        num_local_heads=2,
-        head_dim=4,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-    assert len(cache.self_attention) == 3
-    assert len(cache.cross_attention) == 3
-    assert cache.cross_attention == [None, None, None]
-    assert cache.self_attention[0].key.shape == (2, 7, 2, 4)
-    assert cache.self_attention[0].value.shape == (2, 7, 2, 4)
-    assert cache.self_attention[0].key.dtype == torch.float32
-    assert cache.self_attention[0].end == 0
-    assert cache.self_attention[0].key.data_ptr() != cache.self_attention[1].key.data_ptr()
-
-
-def _allocate_single_layer(module, *, max_tokens: int):
-    return module.allocate_lingbot_cache(
-        batch_size=1,
-        num_layers=1,
-        max_tokens=max_tokens,
-        num_local_heads=1,
-        head_dim=2,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-
-def _set_identity_attention(attention: nn.Module) -> None:
-    qkv = getattr(attention, "qkv", None)
-    if qkv is not None:
-        projections = (qkv, attention.o)
-    else:
-        projections = (attention.q, attention.k, attention.v, attention.o)
-    for projection in projections:
-        with torch.no_grad():
-            identity = torch.eye(2)
-            projection.weight.copy_(torch.cat((identity, identity, identity)) if projection is qkv else identity)
-            if projection.bias is not None:
-                projection.bias.zero_()
-    attention.norm_q = nn.Identity()
-    attention.norm_k = nn.Identity()
-
-
-def _tokens(*values: float) -> torch.Tensor:
-    return torch.tensor([[[value, 0.0] for value in values]])
-
-
-def _record_inputs(module: nn.Module) -> list[tuple]:
-    records: list[tuple] = []
-    module.register_forward_pre_hook(lambda _module, args: records.append(args))
-    return records
-
-
-def _record_outputs(module: nn.Module) -> list[torch.Tensor]:
-    records: list[torch.Tensor] = []
-
-    def record(_module, _args, output) -> None:
-        records.append(output[0] if isinstance(output, tuple) else output)
-
-    module.register_forward_hook(record)
-    return records
-
-
-def _cache_snapshot(cache) -> tuple[torch.Tensor, torch.Tensor, tuple[int, int, int | None, int]]:
-    return (
-        cache.key.detach().view(torch.uint8).clone(),
-        cache.value.detach().view(torch.uint8).clone(),
-        (cache.end, cache.absolute_end, cache.last_start, cache.sink_end),
-    )
-
-
-def _assert_cache_unchanged(cache, snapshot) -> None:
-    key, value, metadata = snapshot
-    assert torch.equal(cache.key.detach().view(torch.uint8), key)
-    assert torch.equal(cache.value.detach().view(torch.uint8), value)
-    assert (cache.end, cache.absolute_end, cache.last_start, cache.sink_end) == metadata
-
-
-def test_self_attention_repeated_offset_overwrites_then_later_offset_appends() -> None:
-    module = _load_module()
-    attention = module.LingBotSelfAttention(dim=2, num_heads=1)
-    _set_identity_attention(attention)
-    cache = _allocate_single_layer(module, max_tokens=6).self_attention[0]
-
-    attention(_tokens(1, 2), cache=cache, current_start=0, sink_tokens=0)
-    attention(_tokens(10, 20), cache=cache, current_start=0, sink_tokens=0)
-
-    assert cache.end == 2
-    assert cache.absolute_end == 2
-    torch.testing.assert_close(cache.key[0, : cache.end, 0, 0], torch.tensor([10.0, 20.0]))
-
-    attention(_tokens(3, 4), cache=cache, current_start=2, sink_tokens=0)
-
-    assert cache.end == 4
-    assert cache.absolute_end == 4
-    torch.testing.assert_close(cache.key[0, : cache.end, 0, 0], torch.tensor([10.0, 20.0, 3.0, 4.0]))
-
-
-def test_self_attention_uses_one_fused_qkv_projection() -> None:
-    module = _load_module()
-    attention = module.LingBotSelfAttention(dim=2, num_heads=1)
-    cache = _allocate_single_layer(module, max_tokens=2).self_attention[0]
-
-    attention(_tokens(1, 2), cache=cache, current_start=0, sink_tokens=0)
-
-    assert attention.qkv.calls == 1
-
-
-def test_self_attention_routes_paged_inputs_through_engine_owned_cache() -> None:
-    module = _load_module()
-    attention = module.LingBotSelfAttention(dim=2, num_heads=1)
-    _set_identity_attention(attention)
-    paged_inputs = module.ARDiffusionPagedLayerInputs()
-    calls = []
-
-    def paged_write(inputs, query, key, value, k_act, v_act, scale):
-        calls.append((inputs, query.clone(), key.clone(), value.clone(), k_act, v_act, scale))
-        return query
-
-    module.paged_write_attn = paged_write
-    output = attention(
-        _tokens(1, 2),
-        cache=paged_inputs,
-        current_start=0,
-        sink_tokens=0,
-        update_cache=False,
-    )
-
-    assert output.shape == (1, 2, 2)
-    assert len(calls) == 1
-    assert calls[0][0] is paged_inputs
-    assert calls[0][4:6] == (None, None)
-    assert attention.attn.calls == []
-    assert not any(hasattr(attention, name) for name in ("q", "k", "v"))
-
-
-def test_self_attention_is_chunk_causal_without_masking_inside_current_chunk() -> None:
-    module = _load_module()
-    attention = module.LingBotSelfAttention(dim=2, num_heads=1)
-    _set_identity_attention(attention)
-    cache = _allocate_single_layer(module, max_tokens=8).self_attention[0]
-    attention_inputs = _record_inputs(attention.attn)
-
-    attention(_tokens(1, 2), cache=cache, current_start=0, sink_tokens=0)
-    first_keys = attention_inputs[-1][1].detach().clone()
-    attention(_tokens(3, 4), cache=cache, current_start=2, sink_tokens=0)
-    second_keys = attention_inputs[-1][1].detach().clone()
-
-    assert attention.attn.causal is False
-    torch.testing.assert_close(first_keys[0, :, 0, 0], torch.tensor([1.0, 2.0]))
-    torch.testing.assert_close(second_keys[0, :, 0, 0], torch.tensor([1.0, 2.0, 3.0, 4.0]))
-
-
-def test_self_attention_retains_sink_and_latest_local_history_after_eviction() -> None:
-    module = _load_module()
-    attention = module.LingBotSelfAttention(dim=2, num_heads=1)
-    _set_identity_attention(attention)
-    cache = _allocate_single_layer(module, max_tokens=4).self_attention[0]
-    attention_inputs = _record_inputs(attention.attn)
-
-    attention(_tokens(1, 2), cache=cache, current_start=0, sink_tokens=1)
-    attention(_tokens(3, 4), cache=cache, current_start=2, sink_tokens=1)
-    attention(_tokens(5, 6), cache=cache, current_start=4, sink_tokens=1)
-
-    visible_keys = attention_inputs[-1][1].detach().clone()
-    assert cache.end == 4
-    assert cache.absolute_end == 6
-    torch.testing.assert_close(visible_keys[0, :, 0, 0], torch.tensor([1.0, 4.0, 5.0, 6.0]))
-    torch.testing.assert_close(cache.key[0, : cache.end, 0, 0], torch.tensor([1.0, 4.0, 5.0, 6.0]))
-
-
-def test_cross_attention_projects_encoder_kv_once_per_request() -> None:
-    module = _load_module()
-    attention = module.LingBotCrossAttention(dim=2, num_heads=1)
-    _set_identity_attention(attention)
-    key_outputs = _record_outputs(attention.k)
-    value_outputs = _record_outputs(attention.v)
-    attention_inputs = _record_inputs(attention.attn)
-
-    output, cache = attention(_tokens(1), _tokens(2, 3), cache=None)
-    second_output, reused_cache = attention(_tokens(4), _tokens(20, 30), cache=cache)
-
-    assert output.shape == second_output.shape == (1, 1, 2)
-    assert reused_cache is cache
-    assert len(key_outputs) == 1
-    assert len(value_outputs) == 1
-    assert cache.end == 2
-    torch.testing.assert_close(cache.key[0, :, 0, 0], torch.tensor([2.0, 3.0]))
-    torch.testing.assert_close(attention_inputs[-1][1], cache.key)
-
-
-def test_self_attention_cache_is_isolated_between_requests() -> None:
-    module = _load_module()
-    attention = module.LingBotSelfAttention(dim=2, num_heads=1)
-    _set_identity_attention(attention)
-    first = _allocate_single_layer(module, max_tokens=4).self_attention[0]
-    second = _allocate_single_layer(module, max_tokens=4).self_attention[0]
-
-    attention(_tokens(1, 2), cache=first, current_start=0, sink_tokens=0)
-
-    assert first.end == 2
-    assert second.end == 0
-    assert torch.count_nonzero(second.key) == 0
-    attention(_tokens(8), cache=second, current_start=0, sink_tokens=0)
-    torch.testing.assert_close(first.key[0, : first.end, 0, 0], torch.tensor([1.0, 2.0]))
-    torch.testing.assert_close(second.key[0, : second.end, 0, 0], torch.tensor([8.0]))
-
-
-def test_cross_attention_keeps_checkpoint_qkvo_and_norm_parameter_names() -> None:
-    module = _load_module()
-    attention = module.LingBotCrossAttention(dim=4, num_heads=2)
-
-    names = set(attention.state_dict())
-
-    assert {
-        "q.weight",
-        "k.weight",
-        "v.weight",
-        "o.weight",
-        "norm_q.weight",
-        "norm_k.weight",
-    } <= names
-    assert callable(attention.norm_q.weight.weight_loader)
-
-
-def test_tp_world_size_one_attention_output_shapes() -> None:
-    module = _load_module()
-    self_attention = module.LingBotSelfAttention(dim=4, num_heads=2)
-    cross_attention = module.LingBotCrossAttention(dim=4, num_heads=2)
-    cache = module.allocate_lingbot_cache(
-        batch_size=2,
-        num_layers=1,
-        max_tokens=5,
-        num_local_heads=2,
-        head_dim=2,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-    hidden_states = torch.randn(2, 3, 4)
-    encoder_hidden_states = torch.randn(2, 5, 4)
-
-    self_output = self_attention(hidden_states, cache=cache.self_attention[0], current_start=0, sink_tokens=0)
-    cross_output, cache.cross_attention[0] = cross_attention(
-        hidden_states,
-        encoder_hidden_states,
-        cache=cache.cross_attention[0],
-    )
-
-    assert self_output.shape == hidden_states.shape
-    assert cross_output.shape == hidden_states.shape
-    assert cache.cross_attention[0] is not None
-
-
-def test_self_attention_applies_rotary_embedding_to_current_query_and_key() -> None:
-    module = _load_module()
-    attention = module.LingBotSelfAttention(dim=2, num_heads=1)
-    cache = _allocate_single_layer(module, max_tokens=2).self_attention[0]
-    cos = torch.ones(2, 1)
-    sin = torch.zeros(2, 1)
-    rotary_inputs = _record_inputs(attention.rotary_embedding)
-
-    attention(_tokens(1, 2), cache=cache, current_start=0, rotary_emb=(cos, sin), sink_tokens=0)
-
-    assert len(rotary_inputs) == 2
-    torch.testing.assert_close(rotary_inputs[0][1], cos)
-    torch.testing.assert_close(rotary_inputs[1][2], sin)
-
-
-def test_attention_block_owns_checkpoint_parent_namespaces() -> None:
-    module = _load_module()
-    block = module.LingBotAttentionBlock(dim=4, num_heads=2, prefix="blocks.0")
-
-    names = set(block.state_dict())
-    expected = {
-        *(f"self_attn.{child}.weight" for child in ("qkv", "o", "norm_q", "norm_k")),
-        *(f"cross_attn.{child}.weight" for child in ("q", "k", "v", "o", "norm_q", "norm_k")),
-    }
-
-    assert expected <= names
-
-
-def test_cross_attention_caches_are_isolated_between_requests() -> None:
-    module = _load_module()
-    attention = module.LingBotCrossAttention(dim=2, num_heads=1)
-
-    _, first = attention(_tokens(1), _tokens(2, 3), cache=None)
-    _, second = attention(_tokens(1), _tokens(8, 9), cache=None)
-
-    assert first.key.untyped_storage().data_ptr() != second.key.untyped_storage().data_ptr()
-    assert first.value.untyped_storage().data_ptr() != second.value.untyped_storage().data_ptr()
-    second_key = second.key.clone()
-    first.key.zero_()
-    torch.testing.assert_close(second.key, second_key)
-
-
-def test_tp_rmsnorm_weight_loader_selects_rank_shard(monkeypatch: pytest.MonkeyPatch) -> None:
-    module = _load_module()
-    monkeypatch.setattr(module, "get_tensor_model_parallel_world_size", lambda: 2)
-    monkeypatch.setattr(module, "get_tensor_model_parallel_rank", lambda: 1)
-    norm = module._LingBotRMSNorm(2, eps=1e-6)
-
-    norm.weight.weight_loader(norm.weight, torch.tensor([10.0, 20.0, 30.0, 40.0]))
-
-    torch.testing.assert_close(norm.weight, torch.tensor([30.0, 40.0]))
@@ -1,525 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-from __future__ import annotations
-
-import importlib.util
-import math
-import subprocess
-import sys
-from pathlib import Path
-
-import numpy as np
-import pytest
-import torch
-
-_MODULE_PATH = Path(__file__).parents[4] / "vllm_omni/diffusion/models/lingbot_world/camera.py"
-_SPEC = importlib.util.spec_from_file_location("_lingbot_world_camera_under_test", _MODULE_PATH)
-assert _SPEC is not None and _SPEC.loader is not None
-_CAMERA_MODULE = importlib.util.module_from_spec(_SPEC)
-sys.modules[_SPEC.name] = _CAMERA_MODULE
-_SPEC.loader.exec_module(_CAMERA_MODULE)
-
-CameraTrajectory = _CAMERA_MODULE.CameraTrajectory
-build_plucker_embedding = _CAMERA_MODULE.build_plucker_embedding
-interpolate_camera_trajectory = _CAMERA_MODULE.interpolate_camera_trajectory
-load_camera_trajectory = _CAMERA_MODULE.load_camera_trajectory
-resolve_trusted_action_directory = _CAMERA_MODULE.resolve_trusted_action_directory
-
-pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
-
-
-def _identity_poses(num_frames: int) -> np.ndarray:
-    return np.repeat(np.eye(4, dtype=np.float64)[None], num_frames, axis=0)
-
-
-def _write_trajectory(
-    directory: Path,
-    poses: np.ndarray,
-    intrinsics: np.ndarray,
-) -> None:
-    np.save(directory / "poses.npy", poses)
-    np.save(directory / "intrinsics.npy", intrinsics)
-
-
-def _trusted(directory: Path):
-    return resolve_trusted_action_directory(".", directory)
-
-
-def _trajectory(
-    poses: torch.Tensor,
-    intrinsics: torch.Tensor | None = None,
-) -> CameraTrajectory:
-    if intrinsics is None:
-        intrinsics = torch.tensor([[832.0 / 3.0, 160.0, 416.0, 240.0]]).repeat(poses.shape[0], 1)
-    return CameraTrajectory(poses=poses, intrinsics=intrinsics)
-
-
-@pytest.mark.parametrize("present_file", [None, "poses.npy"])
-def test_load_camera_trajectory_rejects_missing_files(tmp_path: Path, present_file: str | None) -> None:
-    if present_file == "poses.npy":
-        np.save(tmp_path / present_file, _identity_poses(1))
-
-    missing_file = "poses.npy" if present_file is None else "intrinsics.npy"
-    with pytest.raises(FileNotFoundError, match=missing_file):
-        load_camera_trajectory(_trusted(tmp_path))
-
-
-@pytest.mark.parametrize(
-    ("poses", "intrinsics", "message"),
-    [
-        (np.zeros((2, 3, 4)), np.zeros((2, 4)), "poses.npy"),
-        (_identity_poses(2), np.zeros((2, 3)), "intrinsics.npy"),
-        (_identity_poses(0), np.zeros((0, 4)), "1 and 4096 source frames"),
-        (_identity_poses(2), np.zeros((1, 4)), "same number of frames"),
-    ],
-)
-def test_load_camera_trajectory_rejects_invalid_shapes(
-    tmp_path: Path,
-    poses: np.ndarray,
-    intrinsics: np.ndarray,
-    message: str,
-) -> None:
-    _write_trajectory(tmp_path, poses, intrinsics)
-
-    with pytest.raises(ValueError, match=message):
-        load_camera_trajectory(_trusted(tmp_path))
-
-
-@pytest.mark.parametrize("array_name", ["poses", "intrinsics"])
-def test_load_camera_trajectory_rejects_non_finite_values(tmp_path: Path, array_name: str) -> None:
-    poses = _identity_poses(1)
-    intrinsics = np.ones((1, 4), dtype=np.float64)
-    if array_name == "poses":
-        poses[0, 0, 0] = np.nan
-    else:
-        intrinsics[0, 0] = np.inf
-    _write_trajectory(tmp_path, poses, intrinsics)
-
-    with pytest.raises(ValueError, match="finite"):
-        load_camera_trajectory(_trusted(tmp_path))
-
-
-def test_load_camera_trajectory_preserves_raw_c2w_poses(tmp_path: Path) -> None:
-    first_c2w = np.array(
-        [
-            [0.0, -1.0, 0.0, 2.0],
-            [1.0, 0.0, 0.0, 3.0],
-            [0.0, 0.0, 1.0, 4.0],
-            [0.0, 0.0, 0.0, 1.0],
-        ]
-    )
-    second_c2w = np.array(
-        [
-            [0.0, 0.0, 1.0, 5.0],
-            [0.0, 1.0, 0.0, 4.0],
-            [-1.0, 0.0, 0.0, 6.0],
-            [0.0, 0.0, 0.0, 1.0],
-        ]
-    )
-    poses = np.stack((first_c2w, second_c2w))
-    intrinsics = np.array([[100.0, 110.0, 120.0, 130.0]] * 2)
-    _write_trajectory(tmp_path, poses, intrinsics)
-
-    trajectory = load_camera_trajectory(_trusted(tmp_path))
-
-    assert trajectory.poses.dtype == torch.float32
-    assert trajectory.intrinsics.dtype == torch.float32
-    torch.testing.assert_close(trajectory.poses, torch.from_numpy(poses).float())
-    torch.testing.assert_close(trajectory.intrinsics, torch.from_numpy(intrinsics).float())
-
-
-def test_load_camera_trajectory_rejects_object_dtype_before_materializing(tmp_path: Path) -> None:
-    poses = _identity_poses(1).astype(object)
-    intrinsics = np.ones((1, 4), dtype=np.float32)
-    _write_trajectory(tmp_path, poses, intrinsics)
-
-    with pytest.raises(ValueError, match="poses.npy.*real numeric"):
-        load_camera_trajectory(_trusted(tmp_path))
-
-
-def test_load_camera_trajectory_accepts_official_length_and_materializes_request_prefix(tmp_path: Path) -> None:
-    _write_trajectory(
-        tmp_path,
-        _identity_poses(269),
-        np.arange(269 * 4, dtype=np.float32).reshape(269, 4),
-    )
-
-    trajectory = load_camera_trajectory(_trusted(tmp_path))
-
-    assert trajectory.poses.shape == (117, 4, 4)
-    assert trajectory.intrinsics.shape == (117, 4)
-    np.testing.assert_array_equal(
-        trajectory.intrinsics.numpy(),
-        np.arange(269 * 4, dtype=np.float32).reshape(269, 4)[:117],
-    )
-
-
-def test_load_camera_trajectory_rejects_unbounded_source_frame_count(tmp_path: Path) -> None:
-    _write_trajectory(
-        tmp_path,
-        _identity_poses(4097),
-        np.ones((4097, 4), dtype=np.float32),
-    )
-
-    with pytest.raises(ValueError, match="poses.npy.*bounded camera-array limit|4096 source frames"):
-        load_camera_trajectory(_trusted(tmp_path))
-
-
-def test_load_camera_trajectory_rejects_truncated_payload_before_loading(tmp_path: Path) -> None:
-    _write_trajectory(tmp_path, _identity_poses(1), np.ones((1, 4), dtype=np.float32))
-    poses_path = tmp_path / "poses.npy"
-    poses_path.write_bytes(poses_path.read_bytes()[:-1])
-
-    with pytest.raises(ValueError, match="poses.npy byte size"):
-        load_camera_trajectory(_trusted(tmp_path))
-
-
-def test_load_camera_trajectory_rejects_oversized_npy_header(tmp_path: Path) -> None:
-    header_size = 20_000
-    (tmp_path / "poses.npy").write_bytes(
-        np.lib.format.magic(2, 0) + header_size.to_bytes(4, "little") + b" " * (header_size - 1) + b"\n"
-    )
-    np.save(tmp_path / "intrinsics.npy", np.ones((1, 4), dtype=np.float32))
-
-    with pytest.raises(ValueError, match="poses.npy.*oversized NPY header"):
-        load_camera_trajectory(_trusted(tmp_path))
-
-
-def test_load_camera_trajectory_rejects_symlink_swap_after_resolution(tmp_path: Path) -> None:
-    trusted_root = tmp_path / "trusted"
-    action_directory = trusted_root / "action"
-    outside = tmp_path / "outside"
-    action_directory.mkdir(parents=True)
-    outside.mkdir()
-    _write_trajectory(action_directory, _identity_poses(1), np.ones((1, 4), dtype=np.float32))
-    _write_trajectory(outside, _identity_poses(1), np.ones((1, 4), dtype=np.float32))
-    resolved = resolve_trusted_action_directory("action", trusted_root)
-    action_directory.rename(trusted_root / "original-action")
-    action_directory.symlink_to(outside, target_is_directory=True)
-
-    with pytest.raises(ValueError, match="securely open.*action directory"):
-        load_camera_trajectory(resolved)
-
-
-def test_load_camera_trajectory_rejects_symlinked_camera_file(tmp_path: Path) -> None:
-    trusted_root = tmp_path / "trusted"
-    action_directory = trusted_root / "action"
-    outside = tmp_path / "outside"
-    action_directory.mkdir(parents=True)
-    outside.mkdir()
-    _write_trajectory(action_directory, _identity_poses(1), np.ones((1, 4), dtype=np.float32))
-    np.save(outside / "poses.npy", _identity_poses(1))
-    resolved = resolve_trusted_action_directory("action", trusted_root)
-    (action_directory / "poses.npy").unlink()
-    (action_directory / "poses.npy").symlink_to(outside / "poses.npy")
-
-    with pytest.raises(ValueError, match="poses.npy.*opened securely"):
-        load_camera_trajectory(resolved)
-
-
-@pytest.mark.skipif(not hasattr(__import__("os"), "mkfifo"), reason="requires POSIX FIFO support")
-def test_load_camera_trajectory_rejects_fifo_without_blocking(tmp_path: Path) -> None:
-    import os
-
-    os.mkfifo(tmp_path / "poses.npy")
-    np.save(tmp_path / "intrinsics.npy", np.ones((1, 4), dtype=np.float32))
-    script = """
-import importlib.util
-import sys
-
-spec = importlib.util.spec_from_file_location("_lingbot_camera_fifo_test", sys.argv[1])
-module = importlib.util.module_from_spec(spec)
-sys.modules[spec.name] = module
-spec.loader.exec_module(module)
-action = module.resolve_trusted_action_directory(".", sys.argv[2])
-try:
-    module.load_camera_trajectory(action)
-except ValueError as error:
-    print(error)
-else:
-    raise AssertionError("FIFO camera input was accepted")
-"""
-
-    completed = subprocess.run(
-        [sys.executable, "-c", script, str(_MODULE_PATH), str(tmp_path)],
-        capture_output=True,
-        text=True,
-        timeout=3,
-        check=False,
-    )
-
-    assert completed.returncode == 0, completed.stderr
-    assert "poses.npy must be a regular NPY file" in completed.stdout
-
-
-def test_load_camera_trajectory_decodes_the_validated_file_snapshot(tmp_path: Path, monkeypatch) -> None:
-    original_poses = _identity_poses(1)
-    replacement_poses = original_poses.copy()
-    replacement_poses[0, 0, 3] = 123.0
-    _write_trajectory(tmp_path, original_poses, np.ones((1, 4), dtype=np.float32))
-    poses_path = tmp_path / "poses.npy"
-    original_load = _CAMERA_MODULE.np.load
-    mutation_count = 0
-
-    def mutate_source_then_decode(source, *args, **kwargs):
-        nonlocal mutation_count
-        if mutation_count == 0:
-            mutation_count += 1
-            np.save(poses_path, replacement_poses)
-        return original_load(source, *args, **kwargs)
-
-    monkeypatch.setattr(_CAMERA_MODULE.np, "load", mutate_source_then_decode)
-
-    trajectory = load_camera_trajectory(_trusted(tmp_path))
-
-    assert mutation_count == 1
-    torch.testing.assert_close(trajectory.poses, torch.from_numpy(original_poses).float())
-
-
-def test_load_camera_trajectory_ignores_non_camera_event_files(tmp_path: Path) -> None:
-    poses = _identity_poses(1)
-    intrinsics = np.ones((1, 4), dtype=np.float32)
-    _write_trajectory(tmp_path, poses, intrinsics)
-    np.save(tmp_path / "wasd_events.npy", np.array([{"key": "w"}], dtype=object))
-
-    trajectory = load_camera_trajectory(_trusted(tmp_path))
-
-    torch.testing.assert_close(trajectory.poses, torch.from_numpy(poses).float())
-
-
-def test_interpolate_camera_trajectory_uses_linear_and_spherical_interpolation() -> None:
-    poses = torch.eye(4).repeat(2, 1, 1)
-    poses[1, :3, :3] = torch.tensor(
-        [
-            [0.0, -1.0, 0.0],
-            [1.0, 0.0, 0.0],
-            [0.0, 0.0, 1.0],
-        ]
-    )
-    poses[1, :3, 3] = torch.tensor([2.0, 4.0, 6.0])
-    intrinsics = torch.tensor([[10.0, 20.0, 30.0, 40.0], [20.0, 40.0, 60.0, 80.0]])
-    trajectory = CameraTrajectory(poses=poses, intrinsics=intrinsics)
-
-    result = interpolate_camera_trajectory(trajectory, 3)
-
-    torch.testing.assert_close(result.poses[0], poses[0])
-    torch.testing.assert_close(result.poses[-1], poses[-1])
-    torch.testing.assert_close(result.poses[1, :3, 3], torch.tensor([1.0, 2.0, 3.0]))
-    expected_mid_rotation = torch.tensor(
-        [
-            [math.sqrt(0.5), -math.sqrt(0.5), 0.0],
-            [math.sqrt(0.5), math.sqrt(0.5), 0.0],
-            [0.0, 0.0, 1.0],
-        ]
-    )
-    torch.testing.assert_close(result.poses[1, :3, :3], expected_mid_rotation)
-    torch.testing.assert_close(result.intrinsics[1], torch.tensor([15.0, 30.0, 45.0, 60.0]))
-
-
-def test_build_plucker_embedding_has_forward_center_ray() -> None:
-    trajectory = _trajectory(torch.eye(4).unsqueeze(0))
-
-    result = build_plucker_embedding(
-        trajectory,
-        height=3,
-        width=3,
-        target_height=3,
-        target_width=3,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-    torch.testing.assert_close(result[0, :3, 1, 1], torch.zeros(3))
-    torch.testing.assert_close(result[0, 3:, 1, 1], torch.tensor([0.0, 0.0, 1.0]))
-
-
-def test_build_plucker_embedding_scales_reference_intrinsics() -> None:
-    trajectory = _trajectory(
-        torch.eye(4).unsqueeze(0),
-        torch.tensor([[416.0, 240.0, 416.0, 240.0]]),
-    )
-
-    result = build_plucker_embedding(
-        trajectory,
-        height=4,
-        width=8,
-        target_height=4,
-        target_width=8,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-    expected_corner_direction = torch.tensor([-0.875, -0.75, 1.0])
-    expected_corner_direction /= torch.linalg.vector_norm(expected_corner_direction)
-    torch.testing.assert_close(result[0, 3:, 0, 0], expected_corner_direction)
-
-
-def test_build_plucker_embedding_samples_target_pixel_centers() -> None:
-    trajectory = _trajectory(
-        torch.eye(4).unsqueeze(0),
-        torch.tensor([[416.0, 240.0, 416.0, 240.0]]),
-    )
-
-    result = build_plucker_embedding(
-        trajectory,
-        height=6,
-        width=10,
-        target_height=2,
-        target_width=5,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-    # Target centers map through the source scale: (0.5 * 2, 0.5 * 3).
-    expected_direction = torch.tensor([-0.8, -0.5, 1.0])
-    expected_direction /= torch.linalg.vector_norm(expected_direction)
-    torch.testing.assert_close(result[0, 3:, 0, 0], expected_direction)
-
-
-def test_build_plucker_embedding_uses_framewise_deltas_for_rotations_and_translations() -> None:
-    poses = torch.eye(4).repeat(3, 1, 1)
-    poses[0, :3, 3] = torch.tensor([5.0, 7.0, 11.0])
-    poses[1, :3, :3] = torch.tensor(
-        [
-            [0.0, -1.0, 0.0],
-            [1.0, 0.0, 0.0],
-            [0.0, 0.0, 1.0],
-        ]
-    )
-    poses[1, :3, 3] = torch.tensor([5.0, 9.0, 11.0])
-    poses[2, :3, :3] = torch.tensor(
-        [
-            [-1.0, 0.0, 0.0],
-            [0.0, -1.0, 0.0],
-            [0.0, 0.0, 1.0],
-        ]
-    )
-    poses[2, :3, 3] = torch.tensor([3.0, 9.0, 11.0])
-    trajectory = _trajectory(
-        poses,
-        torch.tensor([[416.0, 480.0, 208.0, 240.0]]).repeat(3, 1),
-    )
-
-    result = build_plucker_embedding(
-        trajectory,
-        height=1,
-        width=2,
-        target_height=1,
-        target_width=2,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-    expected_direction = torch.tensor([0.0, 1.0, 1.0]) / math.sqrt(2.0)
-    torch.testing.assert_close(result[0, :3, 0, 1], torch.zeros(3))
-    torch.testing.assert_close(result[1, :3, 0, 1], torch.tensor([0.0, 1.0, 0.0]))
-    torch.testing.assert_close(result[2, :3, 0, 1], torch.tensor([0.0, 1.0, 0.0]))
-    torch.testing.assert_close(result[1, 3:, 0, 1], expected_direction)
-    torch.testing.assert_close(result[2, 3:, 0, 1], expected_direction)
-
-
-def test_build_plucker_embedding_normalizes_framewise_translation_by_max_norm() -> None:
-    poses = torch.eye(4).repeat(3, 1, 1)
-    poses[:, :3, 3] = torch.tensor(
-        [
-            [10.0, -2.0, 5.0],
-            [13.0, 2.0, 5.0],
-            [19.0, 10.0, 5.0],
-        ]
-    )
-
-    result = build_plucker_embedding(
-        _trajectory(poses),
-        height=1,
-        width=1,
-        target_height=1,
-        target_width=1,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-    expected_origins = torch.tensor([[0.0, 0.0, 0.0], [0.3, 0.4, 0.0], [0.6, 0.8, 0.0]])
-    torch.testing.assert_close(result[:, :3, 0, 0], expected_origins)
-
-
-def test_build_plucker_embedding_interpolates_raw_poses_before_framewise_conversion() -> None:
-    poses = torch.eye(4).repeat(2, 1, 1)
-    poses[0, :3, 3] = torch.tensor([20.0, 4.0, -2.0])
-    poses[1, :3, 3] = torch.tensor([26.0, 12.0, -2.0])
-    raw = _trajectory(poses)
-
-    interpolated = interpolate_camera_trajectory(raw, 3)
-    result = build_plucker_embedding(
-        interpolated,
-        height=1,
-        width=1,
-        target_height=1,
-        target_width=1,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-    expected_delta = torch.tensor([0.6, 0.8, 0.0])
-    torch.testing.assert_close(result[0, :3, 0, 0], torch.zeros(3))
-    torch.testing.assert_close(result[1, :3, 0, 0], expected_delta)
-    torch.testing.assert_close(result[2, :3, 0, 0], expected_delta)
-
-
-def test_build_plucker_embedding_uses_half_pixel_centers_and_origin_direction_channel_order() -> None:
-    poses = torch.eye(4).repeat(2, 1, 1)
-    poses[1, :3, 3] = torch.tensor([2.0, 0.0, 0.0])
-    trajectory = _trajectory(
-        poses,
-        torch.tensor([[832.0, 480.0, 0.0, 0.0]]).repeat(2, 1),
-    )
-
-    result = build_plucker_embedding(
-        trajectory,
-        height=1,
-        width=2,
-        target_height=1,
-        target_width=2,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-    expected_direction = torch.tensor([0.25, 0.5, 1.0])
-    expected_direction /= torch.linalg.vector_norm(expected_direction)
-    torch.testing.assert_close(result[1, :3, 0, 0], torch.tensor([1.0, 0.0, 0.0]))
-    torch.testing.assert_close(result[1, 3:, 0, 0], expected_direction)
-
-
-def test_build_plucker_embedding_honors_shape_dtype_and_device() -> None:
-    trajectory = _trajectory(torch.eye(4).repeat(2, 1, 1))
-
-    result = build_plucker_embedding(
-        trajectory,
-        height=12,
-        width=20,
-        target_height=3,
-        target_width=5,
-        device=torch.device("cpu"),
-        dtype=torch.float64,
-    )
-
-    assert result.shape == (2, 6, 3, 5)
-    assert result.dtype == torch.float64
-    assert result.device == torch.device("cpu")
-
-
-def test_build_plucker_embedding_is_deterministic() -> None:
-    trajectory = _trajectory(torch.eye(4).repeat(2, 1, 1))
-    kwargs = {
-        "height": 12,
-        "width": 20,
-        "target_height": 3,
-        "target_width": 5,
-        "device": torch.device("cpu"),
-        "dtype": torch.float32,
-    }
-
-    first = build_plucker_embedding(trajectory, **kwargs)
-    second = build_plucker_embedding(trajectory, **kwargs)
-
-    torch.testing.assert_close(first, second, rtol=0.0, atol=0.0)
@@ -1,685 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-from __future__ import annotations
-
-import hashlib
-import inspect
-import json
-import math
-from pathlib import Path
-
-import pytest
-import torch
-
-from tests.diffusion.models.lingbot_world import test_lingbot_world_attention as attention_tests
-
-pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
-
-_FIXTURE_PATH = Path(__file__).with_name("fixtures") / "lingbot_world_weight_index_names.json.fixture"
-_SHAPE_FIXTURE_PATH = Path(__file__).with_name("fixtures") / "lingbot_world_official_shapes.json.fixture"
-
-
-@pytest.fixture(autouse=True)
-def _inference_context():
-    with torch.inference_mode():
-        yield
-
-
-def _tiny_model(
-    module,
-    *,
-    num_layers: int = 2,
-    num_frames_per_block: int = 1,
-    sliding_window_num_frames: int = 3,
-):
-    return module.CausalLingBotWorldTransformer3DModel(
-        patch_size=(1, 2, 2),
-        num_attention_heads=2,
-        attention_head_dim=2,
-        in_channels=36,
-        out_channels=2,
-        text_dim=6,
-        freq_dim=4,
-        ffn_dim=8,
-        num_layers=num_layers,
-        cross_attn_norm=True,
-        eps=1e-6,
-        rope_max_seq_len=16,
-        sink_size=1,
-        num_frames_per_block=num_frames_per_block,
-        sliding_window_num_frames=sliding_window_num_frames,
-        local_attn_size=-1,
-    )
-
-
-def _cache(module, model, *, max_tokens: int | None = None):
-    if max_tokens is None:
-        window_frames = (
-            model.config.local_attn_size
-            if model.config.local_attn_size != -1
-            else model.config.sliding_window_num_frames
-        )
-        max_tokens = window_frames * 2 * 2
-    return module.allocate_lingbot_cache(
-        batch_size=1,
-        num_layers=len(model.blocks),
-        max_tokens=max_tokens,
-        num_local_heads=2,
-        head_dim=2,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-
-def _checkpoint_parameter_specs(model) -> dict[str, tuple[torch.Size, torch.dtype]]:
-    specs: dict[str, tuple[torch.Size, torch.dtype]] = {}
-    for name, parameter in model.named_parameters():
-        marker = ".self_attn.qkv."
-        if marker not in name:
-            specs[name] = (parameter.shape, parameter.dtype)
-            continue
-        shard_shape = torch.Size((parameter.shape[0] // 3, *parameter.shape[1:]))
-        for projection in ("q", "k", "v"):
-            specs[name.replace(marker, f".self_attn.{projection}.")] = (shard_shape, parameter.dtype)
-    return specs
-
-
-def _checkpoint_weights_for_model(model) -> tuple[dict[str, torch.Tensor], dict[str, torch.Tensor]]:
-    checkpoint_weights: dict[str, torch.Tensor] = {}
-    expected_parameters: dict[str, torch.Tensor] = {}
-    next_value = 1
-    for name, parameter in model.named_parameters():
-        marker = ".self_attn.qkv."
-        if marker not in name:
-            value = torch.full_like(parameter, next_value)
-            checkpoint_weights[name] = value
-            expected_parameters[name] = value
-            next_value += 1
-            continue
-        shards = []
-        shard_shape = (parameter.shape[0] // 3, *parameter.shape[1:])
-        for projection in ("q", "k", "v"):
-            value = torch.full(shard_shape, next_value, dtype=parameter.dtype, device=parameter.device)
-            checkpoint_weights[name.replace(marker, f".self_attn.{projection}.")] = value
-            shards.append(value)
-            next_value += 1
-        expected_parameters[name] = torch.cat(shards)
-    return checkpoint_weights, expected_parameters
-
-
-def _self_cache_snapshot(cache) -> list[tuple[torch.Tensor, torch.Tensor, tuple[int, int, int | None, int]]]:
-    return [attention_tests._cache_snapshot(layer_cache) for layer_cache in cache.self_attention]
-
-
-def _assert_self_cache_unchanged(cache, snapshot) -> None:
-    for layer_cache, layer_snapshot in zip(cache.self_attention, snapshot, strict=True):
-        attention_tests._assert_cache_unchanged(layer_cache, layer_snapshot)
-
-
-def test_tiny_transformer_runs_four_chunks_with_explicit_cache_commit_and_camera_path() -> None:
-    torch.manual_seed(7)
-    module = attention_tests._load_module()
-    model = _tiny_model(module).eval()
-    cache = _cache(module, model)
-    encoder_hidden_states = torch.randn(1, 3, 6)
-    camera_hidden_states = torch.randn(1, 6 * 8 * 8, 1, 4, 4)
-    cross_key_outputs = [attention_tests._record_outputs(block.cross_attn.k) for block in model.blocks]
-    text_projection_outputs = attention_tests._record_outputs(model.text_embedding)
-    outputs = []
-
-    assert model.patch_embedding.in_channels == 36
-    assert model.patch_embedding_wancamctrl.in_features == 6 * 8 * 8 * 1 * 2 * 2
-
-    for start_frame in range(4):
-        hidden_states = torch.randn(1, 36, 1, 4, 4)
-        timestep = torch.tensor([float(start_frame + 1)])
-        snapshot = _self_cache_snapshot(cache)
-
-        transient_output = model(
-            hidden_states,
-            timestep,
-            encoder_hidden_states,
-            camera_hidden_states,
-            cache=cache,
-            start_frame=start_frame,
-            update_cache=False,
-        )
-
-        assert transient_output.shape == (1, 2, 1, 4, 4)
-        _assert_self_cache_unchanged(cache, snapshot)
-        assert all(layer_cache is not None for layer_cache in cache.cross_attention)
-
-        committed_output = model(
-            hidden_states,
-            timestep,
-            encoder_hidden_states,
-            camera_hidden_states,
-            cache=cache,
-            start_frame=start_frame,
-            update_cache=True,
-        )
-        torch.testing.assert_close(committed_output, transient_output)
-        outputs.append(committed_output)
-
-        expected_token_end = (start_frame + 1) * 2 * 2
-        assert all(layer_cache.absolute_end == expected_token_end for layer_cache in cache.self_attention)
-
-    assert len(outputs) == 4
-    assert all(len(outputs) == 1 for outputs in cross_key_outputs)
-    assert len(text_projection_outputs) == 1
-
-    alternate_cache = _cache(module, model)
-    alternate_output = model(
-        torch.zeros(1, 36, 1, 4, 4),
-        torch.tensor([1.0]),
-        encoder_hidden_states,
-        torch.zeros_like(camera_hidden_states),
-        cache=alternate_cache,
-        start_frame=0,
-        update_cache=False,
-    )
-    camera_output = model(
-        torch.zeros(1, 36, 1, 4, 4),
-        torch.tensor([1.0]),
-        encoder_hidden_states,
-        torch.ones_like(camera_hidden_states),
-        cache=_cache(module, model),
-        start_frame=0,
-        update_cache=False,
-    )
-    assert not torch.equal(camera_output, alternate_output)
-
-
-def test_partial_cross_attention_cache_projects_text_only_for_missing_layers() -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(module, num_layers=2).eval()
-    cache = _cache(module, model)
-    hidden_states = torch.randn(1, 36, 1, 4, 4)
-    timestep = torch.tensor([1.0])
-    encoder_hidden_states = torch.randn(1, 3, 6)
-    camera_hidden_states = torch.randn(1, 6 * 8 * 8, 1, 4, 4)
-
-    model(
-        hidden_states,
-        timestep,
-        encoder_hidden_states,
-        camera_hidden_states,
-        cache=cache,
-        start_frame=0,
-        update_cache=False,
-    )
-    cache.cross_attention[1] = None
-    text_projection_outputs = attention_tests._record_outputs(model.text_embedding)
-    cross_key_outputs = [attention_tests._record_outputs(block.cross_attn.k) for block in model.blocks]
-
-    model(
-        hidden_states,
-        timestep,
-        encoder_hidden_states,
-        camera_hidden_states,
-        cache=cache,
-        start_frame=0,
-        update_cache=False,
-    )
-
-    assert len(text_projection_outputs) == 1
-    assert [len(outputs) for outputs in cross_key_outputs] == [0, 1]
-    assert all(layer_cache is not None for layer_cache in cache.cross_attention)
-
-
-@pytest.mark.parametrize("frames", [1, 6], ids=("partial", "multiple_blocks"))
-def test_forward_rejects_chunks_that_do_not_equal_configured_block_size(frames: int) -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(
-        module,
-        num_layers=1,
-        num_frames_per_block=3,
-        sliding_window_num_frames=6,
-    )
-    cache = _cache(module, model)
-    snapshot = _self_cache_snapshot(cache)
-    patch_inputs = attention_tests._record_inputs(model.patch_embedding)
-
-    with pytest.raises(ValueError, match="exactly 3 post-patch frames"):
-        model(
-            torch.randn(1, 36, frames, 4, 4),
-            torch.tensor([1.0]),
-            torch.randn(1, 3, 6),
-            torch.randn(1, 6 * 8 * 8, frames, 4, 4),
-            cache=cache,
-            start_frame=0,
-            update_cache=True,
-        )
-
-    assert patch_inputs == []
-    _assert_self_cache_unchanged(cache, snapshot)
-    assert cache.cross_attention == [None]
-
-
-def test_forward_accepts_exactly_one_configured_frame_block() -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(
-        module,
-        num_layers=1,
-        num_frames_per_block=3,
-        sliding_window_num_frames=6,
-    )
-
-    output = model(
-        torch.randn(1, 36, 3, 4, 4),
-        torch.tensor([1.0]),
-        torch.randn(1, 3, 6),
-        torch.randn(1, 6 * 8 * 8, 3, 4, 4),
-        cache=_cache(module, model),
-        start_frame=0,
-        update_cache=False,
-    )
-
-    assert output.shape == (1, 2, 3, 4, 4)
-
-
-def test_transformer_allocates_request_cache_from_its_configured_geometry() -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(
-        module,
-        num_layers=2,
-        num_frames_per_block=3,
-        sliding_window_num_frames=6,
-    )
-
-    cache = model.allocate_cache(
-        batch_size=2,
-        latent_height=8,
-        latent_width=12,
-        device=torch.device("cpu"),
-        dtype=torch.float32,
-    )
-
-    assert len(cache.self_attention) == 2
-    assert len(cache.cross_attention) == 2
-    assert cache.cross_attention == [None, None]
-    assert all(layer.key.shape == (2, 6 * 4 * 6, 2, 2) for layer in cache.self_attention)
-    assert all(layer.value.shape == (2, 6 * 4 * 6, 2, 2) for layer in cache.self_attention)
-
-
-def test_video_patch_embedding_uses_temporal_height_width_token_order() -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(module, num_layers=1, num_frames_per_block=2)
-    hidden_states = torch.zeros(1, 36, 2, 2, 4)
-    hidden_states[0, 0] = torch.tensor(
-        [
-            [[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]],
-            [[9.0, 10.0, 11.0, 12.0], [13.0, 14.0, 15.0, 16.0]],
-        ]
-    )
-    with torch.no_grad():
-        model.patch_embedding.weight.zero_()
-        model.patch_embedding.bias.zero_()
-        model.patch_embedding.weight[0, 0, 0, 0, 0] = 1
-
-    tokens = model.patch_embedding(hidden_states).flatten(2).transpose(1, 2)
-
-    torch.testing.assert_close(tokens[0, :, 0], torch.tensor([1.0, 3.0, 9.0, 11.0]))
-
-
-def test_unpatchify_restores_two_frame_channel_and_spatial_order() -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(module, num_layers=1, num_frames_per_block=2)
-    hidden_states = (
-        torch.stack(
-            (
-                torch.arange(0, 8),
-                torch.arange(10, 18),
-                torch.arange(20, 28),
-                torch.arange(30, 38),
-            )
-        )
-        .unsqueeze(0)
-        .float()
-    )
-    expected = torch.tensor(
-        [
-            [
-                [[0, 2, 10, 12], [4, 6, 14, 16]],
-                [[20, 22, 30, 32], [24, 26, 34, 36]],
-            ],
-            [
-                [[1, 3, 11, 13], [5, 7, 15, 17]],
-                [[21, 23, 31, 33], [25, 27, 35, 37]],
-            ],
-        ],
-        dtype=torch.float32,
-    ).unsqueeze(0)
-
-    output = model._unpatchify(
-        hidden_states,
-        batch_size=1,
-        frames=2,
-        height=1,
-        width=2,
-    )
-
-    torch.testing.assert_close(output, expected)
-
-
-def test_head_modulation_broadcasts_distinct_condition_per_frame() -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(module, num_layers=1, num_frames_per_block=2)
-    model.head.norm = torch.nn.Identity()
-    with torch.no_grad():
-        model.head.modulation.zero_()
-        model.head.head.weight.zero_()
-        model.head.head.bias.zero_()
-        model.head.head.weight[:, 0] = 1
-    hidden_states = torch.ones(1, 4, model.dim)
-    timestep_embedding = torch.tensor([[[10.0, 0.0, 0.0, 0.0], [20.0, 0.0, 0.0, 0.0]]])
-    expected = torch.tensor([21.0, 21.0, 41.0, 41.0]).view(1, 4, 1).expand(1, 4, 8)
-
-    output = model.head(hidden_states, timestep_embedding)
-
-    torch.testing.assert_close(output, expected)
-
-
-def test_constructor_defaults_match_official_checkpoint_config() -> None:
-    module = attention_tests._load_module()
-    parameters = inspect.signature(module.CausalLingBotWorldTransformer3DModel.__init__).parameters
-    expected = {
-        "in_channels": 36,
-        "out_channels": 16,
-        "num_layers": 40,
-        "num_attention_heads": 40,
-        "attention_head_dim": 128,
-        "patch_size": (1, 2, 2),
-        "text_dim": 4096,
-        "ffn_dim": 13824,
-        "freq_dim": 256,
-        "rope_max_seq_len": 1024,
-        "eps": 1e-6,
-        "cross_attn_norm": True,
-        "sink_size": 9,
-        "num_frames_per_block": 3,
-        "sliding_window_num_frames": 18,
-        "local_attn_size": -1,
-    }
-
-    assert {name: parameters[name].default for name in expected} == expected
-
-
-def test_transformer_exposes_parameter_dtype_for_pipeline_runtime() -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(module)
-
-    assert model.dtype == next(model.parameters()).dtype
-
-
-def test_transformer_declares_regional_compile_block() -> None:
-    module = attention_tests._load_module()
-
-    assert module.CausalLingBotWorldTransformer3DModel._repeated_blocks == ["LingBotAttentionBlock"]
-
-
-def test_lingbot_rms_norm_uses_global_tp_square_mean(monkeypatch) -> None:
-    module = attention_tests._load_module()
-    reduced_values: list[torch.Tensor] = []
-
-    monkeypatch.setattr(module, "get_tensor_model_parallel_world_size", lambda: 2)
-
-    def all_reduce(value: torch.Tensor) -> torch.Tensor:
-        reduced_values.append(value.detach().clone())
-        return value * 2
-
-    monkeypatch.setattr(module, "tensor_model_parallel_all_reduce", all_reduce)
-    norm = module._LingBotRMSNorm(hidden_size=2, eps=0.0)
-    value = torch.tensor([[3.0, 4.0]])
-
-    output = norm(value)
-
-    torch.testing.assert_close(output, value / math.sqrt(12.5))
-    assert len(reduced_values) == 1
-    torch.testing.assert_close(reduced_values[0], torch.tensor([[25.0]]))
-
-
-def test_attention_rejects_heads_not_divisible_by_tp_size(monkeypatch) -> None:
-    module = attention_tests._load_module()
-    monkeypatch.setattr(module, "get_tensor_model_parallel_world_size", lambda: 3)
-
-    with pytest.raises(ValueError, match="num_heads=2.*tp_size=3"):
-        module.LingBotSelfAttention(dim=4, num_heads=2)
-
-
-def test_constructor_rejects_unsupported_qk_norm_with_config_error() -> None:
-    module = attention_tests._load_module()
-
-    with pytest.raises(ValueError, match="qk_norm.*rms_norm_across_heads"):
-        module.CausalLingBotWorldTransformer3DModel(qk_norm="layer_norm")
-
-
-@pytest.mark.parametrize("field", ["image_dim", "added_kv_proj_dim", "pos_embed_seq_len"])
-def test_constructor_rejects_non_null_image_embedding_fields(field: str) -> None:
-    module = attention_tests._load_module()
-
-    with pytest.raises(ValueError, match=field):
-        module.CausalLingBotWorldTransformer3DModel(**{field: 4})
-
-
-def test_constructor_rejects_unsupported_quantization_with_runtime_error() -> None:
-    module = attention_tests._load_module()
-
-    with pytest.raises(RuntimeError, match="quant_config.*not supported"):
-        module.CausalLingBotWorldTransformer3DModel(quant_config=object())
-
-
-def test_from_config_accepts_diffusers_metadata_and_normalizes_patch_size() -> None:
-    module = attention_tests._load_module()
-
-    class ConfigProbe(module.CausalLingBotWorldTransformer3DModel):
-        def __init__(self, **kwargs) -> None:
-            self.received_kwargs = kwargs
-
-    config = {
-        "_class_name": "CausalLingBotWorldTransformer3DModel",
-        "_diffusers_version": "0.35.0.dev0",
-        "patch_size": [1, 2, 2],
-        "num_attention_heads": 40,
-        "attention_head_dim": 128,
-        "in_channels": 36,
-        "out_channels": 16,
-        "text_dim": 4096,
-        "freq_dim": 256,
-        "ffn_dim": 13824,
-        "num_layers": 40,
-        "cross_attn_norm": True,
-        "eps": 1e-6,
-        "image_dim": None,
-        "added_kv_proj_dim": None,
-        "rope_max_seq_len": 1024,
-        "pos_embed_seq_len": None,
-        "qk_norm": "rms_norm_across_heads",
-        "sink_size": 9,
-        "num_frames_per_block": 3,
-        "sliding_window_num_frames": 18,
-        "local_attn_size": -1,
-    }
-
-    model = ConfigProbe.from_config(config, prefix="transformer")
-
-    assert model.received_kwargs["patch_size"] == (1, 2, 2)
-    assert model.received_kwargs["num_layers"] == 40
-    assert model.received_kwargs["prefix"] == "transformer"
-
-
-@pytest.mark.parametrize(
-    ("field", "value"),
-    [
-        ("in_channels", 35),
-        ("num_layers", 39),
-        ("num_attention_heads", 32),
-        ("attention_head_dim", 64),
-        ("ffn_dim", 10240),
-        ("num_frames_per_block", 4),
-        ("sliding_window_num_frames", 24),
-        ("sink_size", 0),
-    ],
-)
-def test_from_config_rejects_checkpoint_topology_drift(field: str, value: object) -> None:
-    module = attention_tests._load_module()
-    config = {
-        "_class_name": "CausalLingBotWorldTransformer3DModel",
-        "_diffusers_version": "0.35.0.dev0",
-        "patch_size": [1, 2, 2],
-        "num_attention_heads": 40,
-        "attention_head_dim": 128,
-        "in_channels": 36,
-        "out_channels": 16,
-        "text_dim": 4096,
-        "freq_dim": 256,
-        "ffn_dim": 13824,
-        "num_layers": 40,
-        "cross_attn_norm": True,
-        "eps": 1e-6,
-        "image_dim": None,
-        "added_kv_proj_dim": None,
-        "rope_max_seq_len": 1024,
-        "pos_embed_seq_len": None,
-        "qk_norm": "rms_norm_across_heads",
-        "sink_size": 9,
-        "num_frames_per_block": 3,
-        "sliding_window_num_frames": 18,
-        "local_attn_size": -1,
-    }
-    config[field] = value
-
-    with pytest.raises(ValueError, match=field):
-        module.CausalLingBotWorldTransformer3DModel.from_config(config)
-
-
-def test_from_config_ignores_non_semantic_checkpoint_metadata() -> None:
-    module = attention_tests._load_module()
-    config = {
-        "_class_name": "CausalLingBotWorldTransformer3DModel",
-        "architectures": ["CausalLingBotWorldTransformer3DModel"],
-        "future_diffusers_metadata": {"version": 1},
-    }
-
-    with torch.device("meta"):
-        model = module.CausalLingBotWorldTransformer3DModel.from_config(config)
-
-    assert model.config.in_channels == 36
-    assert model.config.num_layers == 40
-
-
-def test_load_weights_uses_parameter_loaders_and_rejects_unknown_model_keys() -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(module, num_layers=1)
-    checkpoint_weights, expected_parameters = _checkpoint_weights_for_model(model)
-    loader_calls: list[str] = []
-    qkv_weight = model.blocks[0].self_attn.qkv.weight
-
-    def record_loader(param: torch.Tensor, loaded_weight: torch.Tensor, shard_id: str) -> None:
-        loader_calls.append(shard_id)
-        shard_index = {"q": 0, "k": 1, "v": 2}[shard_id]
-        shard_size = param.shape[0] // 3
-        param.data.narrow(0, shard_index * shard_size, shard_size).copy_(loaded_weight)
-
-    qkv_weight.weight_loader = record_loader
-    loaded = model.load_weights(iter(checkpoint_weights.items()))
-
-    assert set(checkpoint_weights) <= loaded
-    assert set(dict(model.named_parameters())) <= loaded
-    assert loader_calls == ["q", "k", "v"]
-    for name, param in model.named_parameters():
-        torch.testing.assert_close(param, expected_parameters[name])
-
-    first_name = next(iter(checkpoint_weights))
-    partial = _tiny_model(module, num_layers=1)
-    partial_loaded = partial.load_weights([(first_name, checkpoint_weights[first_name])])
-    expected_partial = {first_name}
-    for projection in ("q", "k", "v"):
-        marker = f".self_attn.{projection}."
-        if marker in first_name:
-            expected_partial.add(first_name.replace(marker, ".self_attn.qkv."))
-            break
-    assert partial_loaded == expected_partial
-
-    with pytest.raises(KeyError, match="unexpected_model.weight"):
-        model.load_weights([("unexpected_model.weight", torch.ones(1))])
-
-
-def test_load_weights_consumes_checkpoint_iterator_incrementally() -> None:
-    module = attention_tests._load_module()
-    model = _tiny_model(module, num_layers=1)
-    first_name, first_param = next(iter(model.named_parameters()))
-    loaded_weight = torch.full_like(first_param, 17)
-
-    def weights():
-        yield first_name, loaded_weight
-        torch.testing.assert_close(first_param, loaded_weight)
-
-    assert model.load_weights(weights()) == {first_name}
-
-
-def test_real_auto_weights_loader_covers_fused_model_parameter_namespace() -> None:
-    utils = pytest.importorskip("vllm.model_executor.models.utils")
-    module = attention_tests._load_module()
-    transformer = _tiny_model(module, num_layers=1)
-    checkpoint_weights, _ = _checkpoint_weights_for_model(transformer)
-
-    class Pipeline(torch.nn.Module):
-        def __init__(self) -> None:
-            super().__init__()
-            self.transformer = transformer
-
-    pipeline = Pipeline()
-    loaded = utils.AutoWeightsLoader(pipeline).load_weights(
-        (f"transformer.{name}", value) for name, value in checkpoint_weights.items()
-    )
-
-    assert set(dict(pipeline.named_parameters())) <= loaded
-
-
-def test_checkpoint_weight_index_fixture_matches_model_namespaces() -> None:
-    module = attention_tests._load_module()
-    fixture = json.loads(_FIXTURE_PATH.read_text())
-    model = _tiny_model(module, num_layers=1)
-    parameter_names = set(_checkpoint_parameter_specs(model))
-
-    assert {name.split(".", 1)[0] for name in parameter_names} == set(fixture["top_level_modules"])
-    assert {name for name in parameter_names if not name.startswith("blocks.")} == set(fixture["non_block_parameters"])
-    assert {name.removeprefix("blocks.0.") for name in parameter_names if name.startswith("blocks.0.")} == set(
-        fixture["block_parameter_suffixes"]
-    )
-    expanded_names = set(fixture["non_block_parameters"])
-    expanded_names.update(
-        f"blocks.{layer}.{suffix}" for layer in range(40) for suffix in fixture["block_parameter_suffixes"]
-    )
-    canonical_names = "\n".join(sorted(expanded_names))
-
-    assert len(expanded_names) == fixture["index_parameter_count"] == 1421
-    assert hashlib.sha256(canonical_names.encode()).hexdigest() == fixture["canonical_name_set_sha256"]
-    assert fixture["canonical_name_to_shard_sha256"] == (
-        "e1addc27d2b1fad6f10226a23a265ee3e3c61e74dce4d5859727bdf180a67992"
-    )
-    assert list(fixture["shards"]) == [
-        f"diffusion_pytorch_model-{index:05}-of-00008.safetensors" for index in range(1, 9)
-    ]
-    assert sum(shard["parameter_count"] for shard in fixture["shards"].values()) == 1421
-    assert all(len(shard["parameter_names_sha256"]) == 64 for shard in fixture["shards"].values())
-
-
-def test_official_default_shapes_match_public_safetensors_header_fixture() -> None:
-    module = attention_tests._load_module()
-    fixture = json.loads(_SHAPE_FIXTURE_PATH.read_text())
-
-    with torch.device("meta"):
-        model = module.CausalLingBotWorldTransformer3DModel()
-    parameter_specs = _checkpoint_parameter_specs(model)
-    expected_dtype = {"F32": torch.float32}[fixture["dtype"]]
-
-    assert len(fixture["shard_representatives"]) == 8
-    assert set(fixture["shard_representatives"].values()) <= set(fixture["parameter_shapes"])
-
-    for name, expected_shape in fixture["parameter_shapes"].items():
-        shape, dtype = parameter_specs[name]
-        assert tuple(shape) == tuple(expected_shape), name
-        assert dtype == expected_dtype
@@ -1,39 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-from __future__ import annotations
-
-import ast
-from pathlib import Path
-
-import pytest
-
-pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
-
-_ROOT = Path(__file__).parents[4]
-_PRODUCTION_FILES = (
-    _ROOT / "vllm_omni/diffusion/models/lingbot_world/camera.py",
-    _ROOT / "vllm_omni/diffusion/models/lingbot_world/transformer.py",
-    _ROOT / "vllm_omni/diffusion/models/lingbot_world/pipeline.py",
-)
-
-
-def test_all_lingbot_production_functions_have_complete_annotations() -> None:
-    missing: list[str] = []
-    for path in _PRODUCTION_FILES:
-        tree = ast.parse(path.read_text())
-        for node in ast.walk(tree):
-            if not isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
-                continue
-            arguments = (*node.args.posonlyargs, *node.args.args, *node.args.kwonlyargs)
-            for argument in arguments:
-                if argument.arg not in {"self", "cls"} and argument.annotation is None:
-                    missing.append(f"{path.name}:{node.lineno}:{node.name}:{argument.arg}")
-            if node.args.vararg is not None and node.args.vararg.annotation is None:
-                missing.append(f"{path.name}:{node.lineno}:{node.name}:*{node.args.vararg.arg}")
-            if node.args.kwarg is not None and node.args.kwarg.annotation is None:
-                missing.append(f"{path.name}:{node.lineno}:{node.name}:**{node.args.kwarg.arg}")
-            if node.returns is None:
-                missing.append(f"{path.name}:{node.lineno}:{node.name}:return")
-
-    assert missing == []
Author	SHA1	Message	Date
Hongsheng Liu	65c86dec44	[Docs] Clarify TP1 DLO mmap support scope Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-16 10:54:24 +08:00
Hongsheng Liu	5ce549bc57	Merge branch 'main' into codex-dlo-host-weight-plan	2026-08-16 00:13:33 +08:00
Hongsheng Liu	fa2f5b4571	[Diffusion] Preserve non-DiT loading with DLO plans Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-15 23:57:03 +08:00
Hongsheng Liu	11687f73bd	[Diffusion] Rebase DLO plan and generalize lifecycle example Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-15 23:40:24 +08:00
Hongsheng Liu	88614175c0	[Diffusion] Address DLO host-weight review feedback Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-15 23:32:27 +08:00
Hongsheng Liu	e2796b6168	[Diffusion] Add loader-owned host-weight plans for DLO Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-15 23:32:27 +08:00