Merge remote-tracking branch 'origin/main' into HEAD

[BugFix][Higgs-Audio-V3] Fix transcript and concurrent sampling tests (#6422 )
Signed-off-by: akshatvishu <akshatnayak197@gmail.com>
2026-08-30 03:36:08 -07:00 · 2026-08-30 08:52:15 +00:00 · 2026-08-30 06:51:24 +00:00 · 2026-08-30 04:14:21 +00:00 · 2026-08-30 03:25:43 +00:00 · 2026-08-29 13:05:24 -07:00
227 changed files with 13420 additions and 1356 deletions
@@ -175,7 +175,7 @@ steps:
      grade: Blocking
      commands:
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+        - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py -m 'advanced_model and cuda' --run-level 'advanced_model'

    - label: "Qwen3-TTS CustomVoice E2E Test · Shard %N/%t"
      agent_pool: mi300_1
@@ -122,6 +122,7 @@ steps:
      - label: "Omni · Qwen3-Omni Test"
        source_file_dependencies:
          - tests/e2e/offline_inference/test_qwen3_omni.py
+          - tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py
          - tests/e2e/online_serving/test_qwen3_omni.py
          - tests/helpers/runtime.py
          - vllm_omni/entrypoints/openai/serving_chat.py
@@ -132,9 +133,10 @@ steps:
          - vllm_omni/model_executor/models/qwen3_tts/cuda_graph_decoder_wrapper.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
          - vllm_omni/deploy/qwen3_omni_moe.yaml
+          - vllm_omni/deploy/qwen3_omni_moe_thinking.yaml
        timeout_in_minutes: 50
        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+          - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py -m 'advanced_model and cuda' --run-level 'advanced_model'
        mirror_hardwares: h100_2

      - label: "Omni · MiniCPM-o 4.5 Test"
@@ -12,16 +12,22 @@ steps:
    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "omni-test"
    steps:

-      - label: ":full_moon: Omni · Function Test with H100"
-        timeout_in_minutes: 120
+      - label: ":full_moon: Omni · Function Test with H100 · Single-GPU"
+        timeout_in_minutes: 90
        commands:
-          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py --ignore=tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni and cards_1" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+        mirror_hardwares: h100_1
+
+      - label: ":full_moon: Omni · Function Test with H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni and cards_2" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py
        mirror_hardwares: h100_2

      - label: ":full_moon: Omni · MiniCPM-o 4.5 Duplex Test"
        timeout_in_minutes: 50
        commands:
-          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py -m "full_model and cuda and H100 and omni" --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py -m "full_model and cuda and H100 and omni and cards_1" --run-level "full_model"
        mirror_hardwares: h100_1

      - label: ":full_moon: Omni · Doc Test with L4"
@@ -31,10 +37,10 @@ steps:
          - pytest -s -v tests/examples/ -m "full_model and omni and L4" --run-level "full_model"
        mirror_hardwares: l4_4

-      - label: ":full_moon: Omni · Doc Test with H100"
+      - label: ":full_moon: Omni · Doc Test with H100 · 2-GPU"
        timeout_in_minutes: 90
        commands:
-          - pytest -s -v tests/examples/ -m "full_model and omni and H100" --run-level "full_model"
+          - pytest -s -v tests/examples/ -m "full_model and omni and H100 and cards_2" --run-level "full_model"
        mirror_hardwares: h100_2

      - label: ":full_moon: Omni · Accuracy Test"
@@ -136,15 +142,29 @@ steps:
          - pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy
        mirror_hardwares: l4_1

-      - label: ":full_moon: TTS · Perf Test"
-        key: nightly-tts-performance
+      - label: ":full_moon: TTS · Perf Test · Single-GPU"
+        key: nightly-tts-performance-single
        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_tts.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_tts.json -m "H100 and cards_1"
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        mirror_hardwares: h100_1
+
+      - label: ":full_moon: TTS · Perf Test · 2-GPU"
+        key: nightly-tts-performance-2gpu
+        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_tts.json -m "H100 and cards_2"
            EXIT=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
            exit $$EXIT
@@ -179,7 +199,6 @@ 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.

      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Single-GPU"
        timeout_in_minutes: 120
@@ -187,24 +206,24 @@ steps:
          - >-
            pytest -sv
            tests/e2e/online_serving/test_qwen_image_expansion.py
-            -m "full_model and diffusion and H100 and not distributed_cuda"
+            -m "full_model and diffusion and H100 and cards_1"
            --run-level "full_model"
        mirror_hardwares: h100_1

-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Qwen-Image"
+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · 2-GPU"
        timeout_in_minutes: 120
        commands:
          - >-
            pytest -sv
            tests/e2e/online_serving/test_qwen_image_expansion.py
-            -m "full_model and diffusion and H100 and distributed_cuda"
+            -m "full_model and diffusion and H100 and cards_2"
            --run-level "full_model"
        mirror_hardwares: h100_2

-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4"
+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4 · Multi-GPU"
        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 hunyuan" -m "full_model and diffusion and L4 and not cards_1" --run-level "full_model" --ignore=tests/e2e/accuracy
        mirror_hardwares: l4_4

      - label: ":full_moon: Diffusion X2I(&A&T) · Doc Test"
@@ -235,8 +254,8 @@ steps:
          - pytest -s -v tests/e2e/accuracy/test_hidream_o1_image.py -m full_model --run-level full_model
        mirror_hardwares: h100_2

-      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image"
-        key: nightly-diffusion-x2iat-performance-qwen-image
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image · Single-GPU"
+        key: nightly-diffusion-x2iat-performance-qwen-image-single
        timeout_in_minutes: 180
        commands:
          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
@@ -244,7 +263,23 @@ steps:
          - 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_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json -m "H100 and cards_1"
+            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 · Qwen-Image · 4-GPU"
+        key: nightly-diffusion-x2iat-performance-qwen-image-4gpu
+        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_vllm_omni.json -m "H100 and cards_4"
            EXIT=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
@@ -260,16 +295,28 @@ steps:
      build.pull_request.labels includes "diffusion-x2v-test"
    steps:

-      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test · Single-GPU"
        timeout_in_minutes: 90
        commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda" --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda and H100 and cards_1" --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda and H100 and cards_2" --run-level "full_model"
        mirror_hardwares: h100_2

-      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test · Single-GPU"
        timeout_in_minutes: 90
        commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda" --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda and H100 and cards_1" --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda and H100 and cards_2" --run-level "full_model"
        mirror_hardwares: h100_2

      - label: ":full_moon: Diffusion X2V · HunyuanVideo-1.5 Function Test"
@@ -310,17 +357,32 @@ steps:
          - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_i2v/test_hunyuanvideo15_i2v_video_similarity.py -m full_model --run-level full_model
        mirror_hardwares: h100_2

-      - label: ":full_moon: Diffusion X2V · Perf Test"
-        key: nightly-diffusion-x2v-performance
+      - label: ":full_moon: Diffusion X2V · Perf Test · Single-GPU"
+        key: nightly-diffusion-x2v-performance-single
        timeout_in_minutes: 180
        commands:
          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
          - |
            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json -m "H100 and cards_1"
+            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 X2V · Perf Test · 2-GPU"
+        key: nightly-diffusion-x2v-performance-2gpu
+        timeout_in_minutes: 180
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json -m "H100 and cards_2"
            EXIT1=$$?
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_cosmos3_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_cosmos3_vllm_omni.json -m "H100 and cards_2"
            EXIT2=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
@@ -371,10 +433,16 @@ steps:
          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and L4' --run-level "full_model"
        mirror_hardwares: l4_4

-      - label: ":full_moon: Diffusion · Distributed Test with H100"
+      - label: ":full_moon: Diffusion · Distributed Test with H100 · 2-GPU"
        timeout_in_minutes: 60
        commands:
-          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100' --run-level "full_model"
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100 and cards_2' --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: ":full_moon: Diffusion · Distributed Test with H100 · 4-GPU"
+        timeout_in_minutes: 60
+        commands:
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100 and cards_4' --run-level "full_model"
        mirror_hardwares: h100_4

      - label: ":full_moon: Diffusion · Offloader Test"
@@ -409,19 +477,25 @@ steps:
    depends_on:
      - nightly-omni-performance-no-async-chunk
      - nightly-omni-performance-async-chunk
-      - nightly-tts-performance
-      - nightly-diffusion-x2iat-performance-qwen-image
-      - nightly-diffusion-x2v-performance
+      - nightly-tts-performance-single
+      - nightly-tts-performance-2gpu
+      - nightly-diffusion-x2iat-performance-qwen-image-single
+      - nightly-diffusion-x2iat-performance-qwen-image-4gpu
+      - nightly-diffusion-x2v-performance-single
+      - nightly-diffusion-x2v-performance-2gpu
    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-single
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance-2gpu
      - 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-diffusion-x2iat-performance-qwen-image
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-single
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-4gpu
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance-single
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance-2gpu
      - 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"
@@ -44,10 +44,16 @@ steps:
          - pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m "slow"
        mirror_hardwares: h100_4

-      - label: "Reliability Test · Invalid parameters · H100"
-        timeout_in_minutes: 120
+      - label: "Reliability Test · Invalid parameters · H100 · Single-GPU"
+        timeout_in_minutes: 90
        commands:
-          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100"
+          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100 and cards_1"
+        mirror_hardwares: h100_1
+
+      - label: "Reliability Test · Invalid parameters · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100 and cards_2"
        mirror_hardwares: h100_2

      - label: "Reliability Test · Invalid parameters · L4"
@@ -105,27 +111,37 @@ steps:
    if: build.env("NON_CRITICAL") == "1"
    steps:

-      - label: "Omni · L4"
+      # Qwen2.5 expansion (CUDA 4-card / online 2-card) stays on this l4_4 job.
+      - label: "Omni · L4 · Multi-GPU"
        timeout_in_minutes: 90
        commands:
          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-          - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py --run-level "core_model"
+          - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py -m "slow and L4 and omni and not cards_1" --run-level "core_model"
          - >-
            pytest -s -v tests/e2e/
            --ignore=tests/e2e/online_serving/test_qwen2_5_omni_expansion.py
            --ignore=tests/e2e/offline_inference/test_qwen2_5_omni_expansion.py
-            -m "slow and L4 and omni"
+            -m "slow and L4 and omni and not cards_1"
            --run-level "full_model"
        mirror_hardwares: l4_4

-      - label: "Omni · H100"
+      - label: "Omni · H100 · Single-GPU"
        timeout_in_minutes: 90
        commands:
          - >-
            pytest -sv
-            tests/e2e/ -m "slow and H100 and omni"
+            tests/e2e/ -m "slow and H100 and omni and cards_1"
            --run-level "full_model"
-        mirror_hardwares: h100_2
+        mirror_hardwares: h100_1
+
+      - label: "Omni · H100 · 4-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and H100 and omni and cards_4"
+            --run-level "full_model"
+        mirror_hardwares: h100_4

      - label: "TTS · L4"
        timeout_in_minutes: 120
@@ -169,12 +185,21 @@ steps:
            --run-level "full_model"
        mirror_hardwares: l4_1

-      - label: "Diffusion · L4"
+      - label: "Diffusion · L4 · Single-GPU"
        timeout_in_minutes: 60
        commands:
          - >-
            pytest -sv
-            tests/e2e/ -m "slow and diffusion and L4"
+            tests/e2e/ -m "slow and diffusion and L4 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: l4_1
+
+      - label: "Diffusion · L4 · Multi-GPU"
+        timeout_in_minutes: 60
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and diffusion and L4 and not cards_1"
            --run-level "full_model"
        mirror_hardwares: l4_4

@@ -183,15 +208,24 @@ steps:
        commands:
          - >-
            pytest -sv
-            tests/e2e/ -m "slow and diffusion and H100 and not distributed_cuda"
+            tests/e2e/ -m "slow and diffusion and H100 and cards_1"
            --run-level "full_model"
        mirror_hardwares: h100_1

-      - label: "Diffusion · H100 · Multi-GPU"
+      - label: "Diffusion · H100 · 2-GPU"
        timeout_in_minutes: 90
        commands:
          - >-
            pytest -sv
-            tests/e2e/ -m "slow and diffusion and H100 and distributed_cuda"
+            tests/e2e/ -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · H100 · 4-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and diffusion and H100 and cards_4"
            --run-level "full_model"
        mirror_hardwares: h100_4
@@ -10,7 +10,7 @@ steps:
          DOCKER_BUILDKIT: "1"
          # Buildkite will automatically replace this with the actual commit hash
          VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}"
-          VLLM_VERSION: "v0.27.0"
+          VLLM_VERSION: "v0.28.0"
        priority: 100
        timeout_in_minutes: 60
        soft_fail: false
@@ -5,7 +5,12 @@ set -ex

 omni_source_dir=$(git rev-parse --show-toplevel)

-base_image_name="xpu/vllm-omni-ci-base:${VLLM_VERSION:?VLLM_VERSION must be set}"
+: "${VLLM_VERSION:?VLLM_VERSION must be set}"
+# Official upstream XPU image. It is published from vllm's own
+# docker/Dockerfile.xpu, which is exactly what build_vllm_base() clones and
+# builds below, so pulling it is equivalent to building the fallback base.
+upstream_base_image="vllm/vllm-openai-xpu:${VLLM_VERSION}"
+local_base_image="xpu/vllm-omni-ci-base:${VLLM_VERSION}"
 image_name="xpu/vllm-omni-ci:${BUILDKITE_COMMIT:?BUILDKITE_COMMIT must be set}"
 container_name="xpu_${BUILDKITE_COMMIT}_$(
    tr -dc A-Za-z0-9 </dev/urandom | head -c 10
@@ -49,9 +54,36 @@ build_vllm_base() (
        "${vllm_source_dir}"
 )

-if [ -z "$(docker images -q "${base_image_name}")" ]; then
-    build_vllm_base
-fi
+# Resolve the vLLM base image, in priority order:
+#   1. the published upstream image for VLLM_VERSION, pulled fresh (so moving
+#      tags like nightly are refreshed rather than served stale from disk);
+#   2. a local copy of that same image, when the registry is unreachable;
+#   3. VLLM_BASE, if that image is already on disk;
+#   4. otherwise build VLLM_BASE from the matching vLLM source tag (~30min).
+# VLLM_BASE names the fallback base and defaults to ${local_base_image}; set it
+# to reuse or produce a base under a different tag.
+resolve_vllm_base() {
+    if docker pull "${upstream_base_image}"; then
+        base_image_name="${upstream_base_image}"
+        return
+    fi
+
+    # Registry unreachable, but a previous run may have left the image behind.
+    if [ -n "$(docker images -q "${upstream_base_image}")" ]; then
+        echo "WARNING: could not pull ${upstream_base_image}; using local copy" >&2
+        base_image_name="${upstream_base_image}"
+        return
+    fi
+
+    echo "WARNING: ${upstream_base_image} unavailable remotely and locally" >&2
+    base_image_name="${VLLM_BASE:-${local_base_image}}"
+    if [ -z "$(docker images -q "${base_image_name}")" ]; then
+        echo "WARNING: building ${base_image_name} from vLLM ${VLLM_VERSION} source" >&2
+        build_vllm_base
+    fi
+}
+
+resolve_vllm_base

 # Try building the docker image
 docker_build "${image_name}" --build-arg "VLLM_BASE=${base_image_name}" --build-arg "VLLM_VERSION=${VLLM_VERSION}"
@@ -10,6 +10,18 @@ steps:
        -m 'core_model and npu and A3 and omni'
        --run-level 'core_model'

+  - label: "Diffusion NPU layer tests"
+    key: ready-diffusion-npu-layer-tests
+    depends_on: upload-ready-pipeline
+    mirror_hardwares: a3_npu_2
+    commands:
+      - >-
+        pytest -s -v
+        tests/diffusion/layers/test_activation.py
+        tests/diffusion/layers/test_fused_qk_norm_rope_npu.py
+        -m 'core_model and npu and A3'
+        --run-level 'core_model'
+
  - group: ":card_index_dividers: Omni Model Test"
    key: ready-omni-group
    depends_on: upload-ready-pipeline
@@ -338,7 +338,7 @@ Classify the model's **CI priority** first:

 **L2 & L3 online — same file, dual marks on the baseline smoke:** The **first / simplest** case in `test_{slug}.py` (default deploy, minimal steps, single prompt) should carry **both** `@pytest.mark.core_model` and `@pytest.mark.advanced_model` on the **same** function so L2 (`test-ready.yml`) and L3 (`test-merge.yml`) share one smoke test. Heavier deploy variants or API paths in the same file use **`advanced_model` only**. When L3 moves to nightly, migrate those heavier cases into `test_{slug}_expansion.py` with `full_model` and remove the dedicated `test-merge.yml` job (see `test_longcat_image_expansion.py`, `test_qwen_image_expansion.py`).

-**L4 design (high priority):** Combine multiple supported features (Cache-DiT, TP, USP, CFG, HSDP, CPU offload, quantization) into **few parametrized** `OmniServerParams` rows so each feature appears in at least one case without exploding GPU jobs. Shard single-GPU vs multi-GPU cases across the nightly X2I/X2V function steps (1-GPU vs `distributed_cuda`).
+**L4 design (high priority):** Combine multiple supported features (Cache-DiT, TP, USP, CFG, HSDP, CPU offload, quantization) into **few parametrized** `OmniServerParams` rows so each feature appears in at least one case without exploding GPU jobs. Shard single-GPU vs multi-GPU cases across the nightly X2I/X2V function steps (`cards_1` vs `not cards_1`).

 **L4 design (medium / low):** One or two parametrized rows covering the best quality/perf trade-off; skip perf JSON unless the model is high priority.

@@ -34,7 +34,7 @@ but the changes are correct — stage the modified files and re-commit.
 | `check-forbidden-imports` | Stdlib `re`/`base64`, pickle, Hugging Face Hub API, or direct Triton/TileLang | `import regex as re` and `pybase64`; use `vllm.transformers_utils.repo_utils`. Do **not** add the file to `allowed_files` without review |
 | `check-torch-cuda-call` | New `torch.cuda.*` call site | Use `current_omni_platform` / `OmniPlatform`; do not grow `ALLOWED_FILES` without review |
 | `check-tts-adapter-migration` | New `self._tts_model_type` branch in `serving_speech.py` | Put per-model logic in `tts_adapters/`. Lower `MAX_MODEL_TYPE_BRANCHES` when removing branches; do not raise it without review |
-| `check-test-ci-coverage` | New `tests/**/test_*.py` missing level or hardware mark | Add `core_model`/`advanced_model`/… plus `cpu`/`cuda`/`hardware_test(` |
+| `check-mark` | New `tests/**/test_*.py` missing level/hardware mark, or direct `pytest.mark.H100` | Add `core_model`/`advanced_model`/… plus `cpu`/`cuda`/`hardware_test(`; SKU marks only via helpers |
 | `mypy-3.10` | Type error on changed `vllm_omni/` files | Fix the types; model trees are excluded. Extra versions: `pre-commit run --hook-stage manual mypy-3.12` |
 | `markdownlint-cli2` | Docs/README markdown lint | Accept auto-fix under `docs/` / `recipes/` / root README |
 | `check-buildkite` | Invalid `.buildkite/*.yml` | Fix the YAML; do not grow `SKIP_FILES` without review |
@@ -52,8 +52,8 @@ Always attach markers deliberately:
  - `tts` — speech synthesis / TTS-only models (`/v1/audio/speech`, voice clone, etc.)
  - `diffusion` — generative diffusion models (image / audio / text / video from noise)
 - **Cross-cutting area** (when relevant): `parallel`, `cache`, `example`, `benchmark`
- **Hardware**: `cpu`, `gpu`, `cuda`, `rocm`, `npu`, `L4`, `H100`, `distributed_cuda`, …
- Optional: `slow`, distributed markers when multi-card is required
+- **Hardware**: `cpu`, `gpu`, `cuda`, `rocm`, `npu`, …; SKU (`H100`, `L4`, …) and `cards_{n}` via `@hardware_test(...)` / `hardware_marks(...)` only — never `pytest.mark.H100` (`check-mark` rejects it)
+- Optional: `slow`. Multi-card: `num_cards=N` on the helper (filter with `cards_N` / `not cards_1`)

 **Baseline smoke (L2 + L3):** The simplest e2e case per model — default deploy, minimal request — should usually carry **both** `@pytest.mark.core_model` and `@pytest.mark.advanced_model` on the **same** test function so `test-ready.yml` and `test-merge.yml` share one test. `send_*_request` picks validation depth from `--run-level`. References: `test_voxcpm2_tts.py::test_text_to_audio_001`, `test_qwen3_tts_customvoice.py::test_text_to_audio_001`. Heavier scenarios use **`advanced_model` only**; L4 expansion uses **`full_model`**.

@@ -274,7 +274,7 @@ def test_text_to_video_001(omni_server, online_client) -> None:

 **3. Cross-cutting rules**

- New `.py` files need the Omni SPDX header (`Copyright contributors to the vLLM-Omni project`, not `vLLM project`). `tests/` may keep stdlib `re`/`base64`; pickle is still banned unless the file is already on the pickle allowlist. Every collected `tests/**/test_*.py` needs a CI **level** mark and a **hardware** mark/helper (`check-test-ci-coverage` runs locally; GHA skips it). Do not grow `allowed_files` to land a test. Policy: [docs/contributing/README.md](../../../docs/contributing/README.md#linting).
+- New `.py` files need the Omni SPDX header (`Copyright contributors to the vLLM-Omni project`, not `vLLM project`). `tests/` may keep stdlib `re`/`base64`; pickle is still banned unless the file is already on the pickle allowlist. Every collected `tests/**/test_*.py` needs a CI **level** mark and a **hardware** platform mark/helper (`check-mark` runs locally; GHA skips it). Do not write `pytest.mark.H100` / `pytest.mark.L4`; use the helpers so `cards_{n}` is attached. Do not grow `allowed_files` to land a test. Policy: [docs/contributing/README.md](../../../docs/contributing/README.md#linting).
 - Reuse existing fixtures for the chosen scenario; do not mix “online client” assumptions into offline `OmniRunner` tests without a clear reason.
 - Avoid external network dependency in assertions unless the scenario is explicitly “online serving” or doc examples that require a model hub (then align with CI secrets/cache).
 - Keep **one test function = one intent** (one modality combo, one endpoint contract, or one acceleration combo).
@@ -79,7 +79,7 @@ E2E (L2+) should not rely on mocks unless documenting a rare exception; prefer r
   - `full_model` — L4 nightly (`test-nightly.yml`); some expansion tests still carry both `advanced_model` and `full_model` during migration
 2. **Model type** (pick one): `omni`, `tts`, or `diffusion`
 3. **Cross-cutting** when relevant: `parallel`, `cache`, `example`, `benchmark`
-4. **Hardware**: `cpu`, `cuda`, `rocm`, `npu`, `L4`, `H100`, `distributed_cuda`, …
+4. **Hardware**: `cpu`, `cuda`, `rocm`, `npu`, `L4`, `H100`, `cards_1`, …
 5. Multi-card: `@hardware_test(...)` in `tests/helpers/mark.py`

 ## Command Templates
@@ -244,7 +244,7 @@ pytest -s -v dfx/reliability/invalid_param_test/test_invalid_image_generation.py

 ```bash
 cd tests
-pytest -s -v -m "core_model and distributed_cuda and L4" --run-level=core_model
+pytest -s -v -m "core_model and L4 and not cards_1" --run-level=core_model
 ```

 ### Concrete e2e paths (common in-tree)
@@ -27,4 +27,4 @@ jobs:
        # Don't run historical-debt hooks against the whole tree in CI yet.
        # They still run locally on changed files.
        env:
-          SKIP: check-test-ci-coverage,markdownlint-cli2,shellcheck,check-spdx-header,mypy-3.10
+          SKIP: check-mark,markdownlint-cli2,shellcheck,check-spdx-header,mypy-3.10
@@ -105,8 +105,8 @@ repos:
      <<: *mypy_common
      stages: [manual]

-    - id: check-test-ci-coverage
-      name: Ensure test files have pytestmarks
+    - id: check-mark
+      name: Ensure test files have CI marks (no direct SKU)
      entry: python tools/pre_commit/check_test_marks.py
      language: python
      types: [python]
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.27.0
+ARG VLLM_ASCEND_TAG=v0.28.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}

 # WORKDIR /vllm-workspace/vllm
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.27.0-a3
+ARG VLLM_ASCEND_TAG=v0.28.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}

 # WORKDIR /vllm-workspace/vllm
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.27.0
+ARG VLLM_ASCEND_TAG=v0.28.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}

 ARG APP_DIR=/vllm-workspace/vllm-omni
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.27.0-a3
+ARG VLLM_ASCEND_TAG=v0.28.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}

 ARG APP_DIR=/vllm-workspace/vllm-omni
@@ -1,140 +1,33 @@
-# Argument to configure vllm base image if pre-built
-ARG VLLM_BASE=vllm-base
-
-FROM intel/deep-learning-essentials:2025.3.2-0-devel-ubuntu24.04 AS vllm-base
-
-WORKDIR /workspace/
-
-ARG PYTHON_VERSION=3.12
-ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/xpu"
-
-RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
-    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list
-
-RUN apt clean && apt-get update -y && \
-    apt-get install -y --no-install-recommends --fix-missing \
-    curl \
-    ffmpeg \
-    git \
-    libsndfile1 \
-    libsm6 \
-    libxext6 \
-    libgl1 \
-    lsb-release \
-    libaio-dev \
-    numactl \
-    wget \
-    vim \
-    python3.12 \
-    python3.12-dev \
-    python3-pip
-
-# Add oneAPI repo, pin oneAPI to 2025.3, then install pinned packages in one layer.
-RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
-    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
-    printf '%s\n' \
-    'Package: intel-oneapi-* intel-deep-learning-essentials* intel-pti*' \
-    'Pin: version 2025.3*' \
-    'Pin-Priority: 1001' \
-    > /etc/apt/preferences.d/oneapi-2025.3.pref && \
-    apt-get update -y && \
-    apt-get install -y --no-install-recommends \
-    intel-oneapi-compiler-dpcpp-cpp-2025.3 \
-    intel-oneapi-mkl-devel-2025.3 \
-    intel-oneapi-dnnl-devel-2025.3 && \
-    rm -rf /var/lib/apt/lists/*
-
-# Install UMD with fixed version
-RUN mkdir neo && \
-    cd neo && \
-    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-core-2_2.24.8+20344_amd64.deb && \
-    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-opencl-2_2.24.8+20344_amd64.deb && \
-    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/intel-ocloc_25.48.36300.8-0_amd64.deb && \
-    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/intel-opencl-icd_25.48.36300.8-0_amd64.deb && \
-    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/libigdgmm12_22.8.2_amd64.deb && \
-    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/libze-intel-gpu1_25.48.36300.8-0_amd64.deb && \
-    wget https://github.com/oneapi-src/level-zero/releases/download/v1.26.0/level-zero_1.26.0+u24.04_amd64.deb && \
-    dpkg -i *.deb && \
-    cd .. && \
-    rm -rf neo
-
-ENV PATH="/root/.local/bin:$PATH"
-ENV VIRTUAL_ENV="/opt/venv"
-ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh
-RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
-ENV PATH="$VIRTUAL_ENV/bin:$PATH"
-
-# This oneccl contains the BMG support which is not the case for default version of oneapi 2025.3.
-ARG ONECCL_INSTALLER="intel-oneccl-2021.15.9.14_offline.sh"
-RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.9/${ONECCL_INSTALLER}" && \
-    bash "${ONECCL_INSTALLER}" -a --silent --eula accept && \
-    rm "${ONECCL_INSTALLER}" && \
-    echo "source /opt/intel/oneapi/setvars.sh --force" >> /root/.bashrc && \
-    echo "source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force" >> /root/.bashrc
-RUN rm -f /opt/intel/oneapi/ccl/latest && \
-    ln -s /opt/intel/oneapi/ccl/2021.15 /opt/intel/oneapi/ccl/latest
-
-SHELL ["bash", "-c"]
-CMD ["bash", "-c", "source /root/.bashrc && exec bash"]
-
-WORKDIR /workspace/
-
-ENV UV_HTTP_TIMEOUT=500
-
-# Configure package index for XPU
-ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
-ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-ENV UV_LINK_MODE="copy"
-
-ARG VLLM_VERSION=v0.27.0
-RUN git clone -b ${VLLM_VERSION} https://github.com/vllm-project/vllm
-WORKDIR /workspace/vllm
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --upgrade pip && \
-    uv pip install -r requirements/xpu.txt && \
-    uv pip install setuptools_scm &&\
-    uv pip install grpcio-tools protobuf nanobind && \
-    source /opt/intel/oneapi/setvars.sh --force && \
-    source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force && \
-    export CMAKE_PREFIX_PATH="$(python3 -c 'import site; print(site.getsitepackages()[0])'):${CMAKE_PREFIX_PATH}"
-
-ENV LD_LIBRARY_PATH="/opt/intel/oneapi/ccl/2021.15/lib:$LD_LIBRARY_PATH:/usr/local/lib/"
-
-ENV VLLM_TARGET_DEVICE=xpu
-ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip uninstall triton triton-xpu && \
-    uv pip install triton-xpu==3.7.2 && \
-    uv pip install --no-build-isolation --no-deps .
-
-CMD ["/bin/bash"]
-
-FROM vllm-base AS vllm-openai
-
-# install nixl from source code
-RUN uv pip install nixl
-
-# install additional dependencies for openai api server
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install accelerate hf_transfer pytest pytest_asyncio lm_eval[api] modelscope
-
-# install development dependencies (for testing)
-RUN uv pip install -e tests/vllm_test_utils
-
-# ensure vllm is properly installed
-RUN python -c "import vllm, inspect; print(vllm.__file__)"
-RUN uv pip show vllm
-
-CMD ["/bin/bash"]
-
-ENTRYPOINT []
+# Target vLLM version: v0.28.0. Keep VLLM_VERSION aligned with the target vLLM
+# release on every rebase.
+#
+# This file layers vLLM-Omni on top of the published upstream XPU image, which
+# already ships a self-consistent XPU runtime (torch-xpu, oneCCL/UCX+NIXL, the
+# pinned UMD/compute-runtime stack, and vLLM itself) built from
+# https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.xpu. Docker
+# pulls that base automatically, so there is nothing to build here beyond the
+# Omni layers. Mirrors how docker/Dockerfile.ci consumes vllm/vllm-openai.
+#
+# If the tag does not exist yet (targeting an unreleased vLLM commit), build the
+# base from upstream's own Dockerfile and point VLLM_BASE at the result:
+#
+#   git clone --depth 1 --branch $VLLM_VERSION https://github.com/vllm-project/vllm /tmp/vllm
+#   docker build -t vllm-openai-xpu:local --target vllm-openai \
+#       -f /tmp/vllm/docker/Dockerfile.xpu /tmp/vllm
+#   docker build -f docker/Dockerfile.xpu --build-arg VLLM_BASE=vllm-openai-xpu:local .
+#
+# .buildkite/intel/scripts/run-xpu-test.sh automates that pull-then-build fallback.
+ARG VLLM_BASE_IMAGE=vllm/vllm-openai-xpu
+ARG VLLM_VERSION=v0.28.0
+ARG VLLM_BASE=${VLLM_BASE_IMAGE}:${VLLM_VERSION}

 FROM ${VLLM_BASE} AS vllm-omni

+# The base image ships ENTRYPOINT ["vllm", "serve"]; clear it so this stage stays
+# usable as a plain shell/test container regardless of which VLLM_BASE was used.
+ENTRYPOINT []
+CMD ["/bin/bash"]
+
 RUN apt-get update && \
    apt-get install -y espeak-ng ffmpeg git jq && \
    apt-get clean && \
@@ -49,7 +49,7 @@ vLLM-Omni's `pre-commit` hooks will now run automatically every time you commit.
    `--all-files` run stays green while historical debt is cleaned up. The
    current `SKIP` list in
    [`.github/workflows/pre-commit.yml`](https://github.com/vllm-project/vllm-omni/blob/main/.github/workflows/pre-commit.yml)
-    is `check-test-ci-coverage`, `markdownlint-cli2`, `shellcheck`,
+    is `check-mark`, `markdownlint-cli2`, `shellcheck`,
    `check-spdx-header`, and `mypy-3.10`. Those hooks still run on **your
    commit** for changed files. A passing GitHub pre-commit check does not mean
    they passed locally. Hooks **not** on that list (forbidden imports,
@@ -66,7 +66,7 @@ already there). `check-pickle-imports` is gone: pickle is now one rule inside
 | `markdownlint-cli2` | Markdown in `docs/`, `recipes/`, `README.md`, `CONTRIBUTING.md` (not `.claude/` / `.cursor/` / `CLAUDE.md`) | skipped |
 | `mypy-3.10` | Type-check changed `vllm_omni/` files (model trees excluded). `tests/` uses `--follow-imports skip` | skipped |
 | `mypy-3.11` / `3.12` / `3.13` | Same checker, extra Python versions | not installed; `pre-commit run --hook-stage manual mypy-3.12` |
-| `check-test-ci-coverage` | Every `tests/**/test_*.py` has a CI level mark and a hardware mark/helper | skipped |
+| `check-mark` | Every `tests/**/test_*.py` has a CI level mark and a hardware platform mark or helper; no direct `pytest.mark.H100` / SKU | skipped |
 | `check-tts-adapter-migration` | `self._tts_model_type` branches in `serving_speech.py` must not increase | runs |
 | `shellcheck` | `*.sh` quoting / undefined vars | skipped |
 | `check-spdx-header` | Omni SPDX header on `.py` / `.pyi` / `.sh` / `.rs` / `.proto` | skipped |
@@ -144,11 +144,13 @@ pre-commit run --hook-stage manual mypy-3.12

 #### Test CI marks

-`check-test-ci-coverage` requires each collected test module under `tests/` to
+`check-mark` requires each collected test module under `tests/` to
 have at least one CI **level** mark (`core_model`, `advanced_model`,
-`full_model`, `local_model`, or `slow`) and a **hardware** mark (`cpu`, `cuda`,
-`H100`, …) or helper (`hardware_test(` / `hardware_marks(`). GitHub Actions
-skips this hook; local commit does not. See the
+`full_model`, `local_model`, or `slow`) and a **hardware** platform mark
+(`cpu`, `cuda`, …) or helper (`hardware_test(` / `hardware_marks(`). SKU marks
+(`H100`, `L4`, …) must come from those helpers so `cards_{n}` is attached;
+direct `pytest.mark.H100` is rejected. GitHub Actions skips this hook; local
+commit does not. See the
 [test writing guide](./ci/test_writing_guide.md).

 #### TTS adapter ratchet
@@ -402,7 +402,7 @@ loads the model once per mode.
 Before splitting, confirm each half actually collects at least one test under the job's forwarded filters — pytest exits with code 5 ("no tests collected") if a half is empty, which the script reports as a failure, red-ing a job that used to pass as one combined invocation.

 Also check the tests' `num_cards` against the job's `mirror_hardwares` preset.
-`cuda_marks` turns `num_cards > 1` into `skipif(device_count() < num_cards)`, so a
+`_cuda_marks` turns `num_cards > 1` into `skipif(device_count() < num_cards)`, so a
 test asking for more GPUs than the preset provides is silently skipped and its
 coverage artifact comes back empty. Splitting such a job produces a file that
 looks fine and measures nothing — "Diffusion · Z Image Test" is the current
@@ -81,8 +81,8 @@ Supported form:

 | Form | Example | Effect |
 | ---- | ------- | ------ |
-| Array (single platform) | `"mark": [{"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 2}}, "full_model", "diffusion"]` | Hardware marks from `hardware_marks(...)`; `num_cards` > 1 adds `distributed_*` (+ CUDA `skipif_cuda` when applicable). String entries become extra pytest marks. |
-| Array (multi-platform) | `"mark": [{"hardware_marks": {"res": {"cuda": "H100", "rocm": "MI325", "npu": "A2"}, "num_cards": {"cuda": 2, "rocm": 2, "npu": 1}}}, "full_model", "omni"]` | Same as above, but declares **multiple platforms** in one case. Filter examples: `-m "full_model and H100 and cuda"`, `-m "full_model and MI325 and rocm"`. |
+| Array (single platform) | `"mark": [{"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 2}}, "full_model", "diffusion"]` | Hardware marks from `hardware_marks(...)`; `num_cards` adds `cards_{n}` (including `cards_1` when omitted). String entries become extra pytest marks. |
+| Array (multi-platform) | `"mark": [{"hardware_marks": {"res": {"cuda": "H100", "rocm": "MI325"}, "num_cards": 2}}, "full_model", "omni"]` | Same as above, but declares **multiple platforms** on one item. All platforms must share the same `num_cards` (int or identical dict values). Different counts need `@hardware_test` or one `hardware_marks()` per `pytest.param`. Filter examples: `-m "full_model and H100 and cards_2 and cuda"`, `-m "full_model and MI325 and cards_2 and rocm"`. |

 Recommended for L4 perf cases:

@@ -157,7 +157,7 @@ pytest -s -v tests/dfx/perf/scripts/run_benchmark.py \
 ```

 <!-- markdownlint-disable-next-line MD051 -->
-See also [Markers for Tests](#markers-for-tests) for registered hardware markers (`H100`, `L4`, `cuda`, `distributed_cuda`, …).
+See also [Markers for Tests](#markers-for-tests) for registered hardware markers (`H100`, `L4`, `cuda`, `cards_1`, …).

 #### `server_params` Configuration

@@ -88,7 +88,7 @@ Failed jobs: 1/2
    ```bash
    cd tests
    pytest -s -v test_xxxx.py --run-level=core_model
-    pytest -s -v -m "core_model and distributed_cuda and L4" --run-level=core_model
+    pytest -s -v -m "core_model and L4 and not cards_1" --run-level=core_model
    ```

 === "L3 level"
@@ -119,7 +119,7 @@ Failed jobs: 1/2

    ```bash
    pytest -s -v test_xxxx.py --run-level=advanced_model
-    pytest -s -v -m "advanced_model and distributed_cuda and L4" --run-level=advanced_model
+    pytest -s -v -m "advanced_model and L4 and not cards_1" --run-level=advanced_model
    ```

 === "L4 level"
@@ -153,12 +153,12 @@ Failed jobs: 1/2
    ```bash
    cd tests
    pytest -s -v test_xxxx.py --run-level=full_model
-    pytest -s -v -m "full_model and distributed_cuda and L4" --run-level=full_model
+    pytest -s -v -m "full_model and L4 and not cards_1" --run-level=full_model
    pytest -s -v -m "full_model and (omni or tts) and H100" --run-level=full_model
    ```
    If you only want to run specific test cases on a particular platform, you can use:
    ```bash
-    pytest -s -v -m "full_model and distributed_cuda and L4"  --run-level=full_model
+    pytest -s -v -m "full_model and L4 and not cards_1"  --run-level=full_model
    ```
    Note: ``run_benchmark.py`` and ``run_diffusion_benchmark.py`` accept an optional ``--test-config-file``. If omitted, each loads every ``*.json`` under ``tests/dfx/perf/tests/`` (omni/tts vs diffusion split by ``is_diffusion_perf_config``) and pytest ``-m`` filters by each case's JSON ``mark``:
    ```bash
@@ -48,17 +48,12 @@ Defined in `pyproject.toml`:
 | `MI325`            | Tests that require MI325 GPU (AMD/ROCm) *                                                               |
 | `A2`               | Tests that require A2 NPU *                                                                             |
 | `A3`               | Tests that require A3 NPU *                                                                             |
-| `distributed_cuda` | Tests that require multi cards on CUDA platform *                                                       |
-| `distributed_rocm` | Tests that require multi cards on ROCm platform  *                                                      |
-| `distributed_npu`  | Tests that require multi cards on NPU platform  *                                                       |
-| `skipif_cuda`      | Skip if the num of CUDA cards is less than the required *                                               |
-| `skipif_rocm`      | Skip if the num of ROCm cards is less than the required *                                               |
-| `skipif_npu`       | Skip if the num of NPU cards is less than the required *                                                |
+| `cards_{n}`        | Tests that require *n* accelerator cards (`cards_1` … `cards_8`); auto-added from `num_cards` *         |
 | `slow`             | Slow tests (may skip in quick CI)                                                                       |
 | `benchmark`        | Benchmark tests (decorator on runner test functions; perf JSON uses `full_model` + type marker instead) |
 | `local_model`      | Tests requiring local / non-HF-hub model weights                                                        |

-\* Means those markers are auto-added by `@hardware_test` (parametrization decorator) or `hardware_marks` (only returning the list of marks for flexibility).
+\* Means those markers are auto-added by `@hardware_test` (parametrization decorator) or `hardware_marks` (only returning the list of marks for flexibility). Do **not** write `pytest.mark.H100` / `pytest.mark.L4` (or other SKU names) by hand: `check-mark` rejects that so `cards_{n}` cannot be skipped.

 #### Example usage for markers

@@ -81,25 +76,23 @@ def test_video_to_audio()
 This decorator is intended to make hardware-aware, cross-platform test authoring easier and more robust for CI/CD environments. The `hardware_test` decorator in `vllm-omni/tests/helpers/mark.py` performs the following actions:

 1. **Applies platform and resource markers**  
-   Adds the appropriate pytest markers for each specified hardware platform (e.g., `cuda`, `rocm`, `xpu`, `npu`) and resource type (e.g., `L4`, `H100`, `MI325`, `B60`, `A2`, `A3`).
+   Adds the appropriate pytest markers for each specified hardware platform (e.g., `cuda`, `rocm`, `xpu`, `npu`) and resource type (e.g., `L4`, `H100`, `H200`, `B200`, `MI325`, `B60`, `A2`, `A3`).
   ```python
   @pytest.mark.cuda
   @pytest.mark.L4
   ```
-2. **Handles multi-card (distributed) scenarios**  
-   For tests requiring multiple cards, it automatically adds distributed markers such as `distributed_cuda`, `distributed_rocm`, or `distributed_npu`.
-   ```python
-   @pytest.mark.distributed_cuda(num_cards=num_cards)
-   ```
+2. **Handles card-count filtering**  
+   Adds a filterable marker `cards_{n}` (including `cards_1` when `num_cards` is omitted / 1). Use this with the SKU and platform, e.g. `-m "H100 and cards_4 and cuda"` vs `-m "H100 and cards_2 and cuda"`. Any multi-card CUDA test is `-m "cuda and not cards_1"`.
+
+   When `res` lists more than one platform **and** `num_cards` differs by platform, `@hardware_test` collects **one variant per platform** so those `cards_{n}` marks do not stack. `-m "H100 and cards_4 and cuda"` then selects only the CUDA/H100 4-card variant, not a ROCm 2-card sibling. The same count on every platform stays a single item.
 3. **Supports flexible card requirements**  
   Accepts `num_cards` as either a single integer for all platforms or as a dictionary with per-platform values. If not specified, defaults to 1 card per platform.

 4. **Integrates resource validation**  
-   On CUDA, adds a skip marker (`skipif_cuda`) if the system does not have the required number of devices.
-   Support for `skipif_rocm` and `skipif_npu` will be implemented later.
+   On CUDA and XPU, when `num_cards > 1`, attaches pytest's built-in `skipif` if the machine has fewer devices than requested. This is not a named `skipif_*` marker and is not used for `-m` filtering.

 5. **Works with pytest filtering**  
-   Allows tests to be filtered and selected at runtime using standard pytest marker expressions (e.g., `-m "distributed_cuda and L4"`).
+   Allows tests to be filtered and selected at runtime using standard pytest marker expressions (e.g., `-m "H100 and cards_2"`, `-m "L4 and not cards_1"`).

 ##### Example usage for decorator

@@ -117,13 +110,14 @@ This decorator is intended to make hardware-aware, cross-platform test authoring
        num_cards=2,
    )
    ```
- `res` must be a dict; supported resources: CUDA (L4/H100), ROCm (MI325), XPU (B60), MUSA (S5000), NPU (A2/A3)
+- `res` must be a dict. SKU values come from `pyproject.toml` markers tagged `[hardware-resource]` plus the platform key (`[cuda]`, `[rocm]`, `[xpu]`, `[npu]`, `[musa]`).
 - `num_cards` can be int (all platforms) or dict (per platform); defaults to 1 when missing
- Distributed markers (`distributed_cuda`, `distributed_rocm`, `distributed_npu`) are auto-added for multi-card cases
 - Filtering examples:
-    - CUDA only: `pytest -m "distributed_cuda and L4"`
-    - ROCm only: `pytest -m "distributed_rocm and MI325"`
-    - NPU only: `pytest -m "distributed_npu"`
+    - CUDA H100 4-card: `pytest -m "H100 and cards_4 and cuda"`
+    - CUDA H100 2-card: `pytest -m "H100 and cards_2 and cuda"`
+    - CUDA multi-card (any count): `pytest -m "L4 and not cards_1"`
+    - ROCm multi-card: `pytest -m "MI325 and not cards_1"`
+    - NPU multi-card: `pytest -m "npu and not cards_1"`

 #### Function: `hardware_marks`

@@ -179,37 +173,30 @@ If you want to add support for a new platform (e.g., "tpu" for a new accelerator
   [tool.pytest.ini_options]
   markers = [
       # ... existing markers ...
-       "tpu: Tests that require TPU device",
-       "TPU_V3: Tests that require TPU v3 hardware",
-       "distributed_tpu: Tests that require multiple TPU devices",
+       "tpu: [hardware-platform] Tests that require TPU device",
+       # Add [gpu] next to [hardware-platform] if this platform should also get pytest.mark.gpu.
+       "TPU_V3: [hardware-resource] [tpu] Tests that require TPU v3 hardware",
+       "cards_1: [hardware-cards] Tests that require 1 accelerator card",
+       "cards_2: [hardware-cards] Tests that require 2 accelerator cards",
   ]
   ```
+   `check-mark` treats `[hardware-platform]` names (including `cpu`/`gpu`) as allowed `pytest.mark.<name>` values. `hardware_test(res=...)` still rejects `cpu`/`gpu` as `res` keys.
 2. **Implement a marker construction function for your platform** in `vllm-omni/tests/helpers/mark.py`:
   ```python
   # In vllm-omni/tests/helpers/mark.py

-   def tpu_marks(*, res: str, num_cards: int):
-       test_platform = pytest.mark.tpu
-       if res == "TPU_V3":
-           test_resource = pytest.mark.TPU_V3
-       else:
-           raise ValueError(
-               f"Invalid TPU resource type: {res}. Supported: TPU_V3")
-
-       if num_cards == 1:
-           return [test_platform, test_resource]
-       else:
-           test_distributed = pytest.mark.distributed_tpu(num_cards=num_cards)
-           # Optionally: add skipif_tpu when implemented
-           return [test_platform, test_resource, test_distributed]
+   def _tpu_marks(*, res: str, num_cards: int):
+       _require_sku("tpu", res)
+       return [pytest.mark.tpu, getattr(pytest.mark, res), _cards_mark(num_cards)]
   ```
-3. **Update `hardware_marks` to recognize your new platform**:
-    In the relevant place (see the `hardware_marks` implementation), add:
+3. **Register the builder** in `_marks_for_platform`:
    ```python
-    if platform == "tpu":
-        marks = tpu_marks(res=resource, num_cards=cards)
+    builders = {
+        # ... existing platforms ...
+        "tpu": _tpu_marks,
+    }
    ```
-    (`hardware_test` calls `hardware_marks` internally, so both will pick up the change.)
+    (`hardware_test` / `hardware_marks` both go through `_marks_for_platform`.)
 4. **(Recommended) Add a test using your new markers**:
   ```python
   @hardware_test(
@@ -223,11 +210,11 @@ If you want to add support for a new platform (e.g., "tpu" for a new accelerator
 **Summary**:  

 - Add pytest markers for your new platform/resources  
- Implement a marker function (`xxx_marks`)  
- Plug into `hardware_marks`  
- You're done: tests using `@hardware_test` or `hardware_marks` with your platform now automatically get the correct markers, distribution, and isolation!
+- Implement a private marker function (`_xxx_marks`)
+- Register it in `_marks_for_platform`
+- You're done: tests using `@hardware_test` or `hardware_marks` with your platform now automatically get the correct markers and `cards_{n}` filtering.

-See code in `vllm-omni/tests/helpers/mark.py` for existing examples (`cuda_marks`, `rocm_marks`, `npu_marks`).
+See code in `vllm-omni/tests/helpers/mark.py` for existing examples (`_cuda_marks`, `_rocm_marks`, `_npu_marks`).

 ## Test case style

@@ -438,7 +438,7 @@ selects the scheduler and worker family:
 Key topology fields belong to `StagePipelineConfig`:

 | Field | Description |
-|-------|-------------|
+| ------- | ------------- |
 | `model_stage` | Logical stage name (`ar_stage`, `decoder`, etc.). |
 | `execution_type` | Runtime family for this stage. |
 | `input_sources` | Upstream stage IDs that provide input. |
@@ -960,7 +960,7 @@ Common failures and fixes:
 | `check-forbidden-imports` | Stdlib `re`/`base64`, pickle, Hugging Face Hub API, or direct Triton/TileLang | `import regex as re` and `pybase64`; do not grow the allowlist without review |
 | `check-torch-cuda-call` | New `torch.cuda.*` call site | Use `current_omni_platform`; do not grow `ALLOWED_FILES` without review |
 | `check-tts-adapter-migration` | New `_tts_model_type` branch in `serving_speech.py` | Put logic in `tts_adapters/`; lower `MAX_MODEL_TYPE_BRANCHES` when removing branches |
-| `check-test-ci-coverage` | Test file missing level or hardware mark | Add `core_model`/… plus `cpu`/`cuda`/`hardware_test(` |
+| `check-mark` | Test file missing level/hardware mark, or direct `pytest.mark.H100` | Add `core_model`/… plus `cpu`/`cuda`/`hardware_test(`; SKU marks only via helpers |
 | `shellcheck` | Native `shellcheck` missing, or script warning | Install via apt/dnf/brew (or `shellcheck.exe` on Windows). See [Linting](../README.md#linting) |
 | `mypy-3.10` / `markdownlint-cli2` / `check-buildkite` | Types, docs markdown, or Buildkite YAML | See [Linting](../README.md#linting) |

@@ -26,6 +26,8 @@ Legend: ✅ supported, ⚠️ compatibility path or limited validation, ❌ unsu
 | **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. |
+| **Online INT8 linears** | ✅ Ordinary loader finalizes the int8 weight (contiguous or transposed view) and fp32 scale before DLO sharding. | ✅ Ordinary loader retains complete rank-local tensors. |
+| **Online MXFP8 linears** | ✅ Ordinary loader finalizes the fp8 weight (contiguous) and e8m0 block scale (contiguous or transposed view) 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`. |
@@ -438,11 +440,15 @@ 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. Per-tensor online FP8 linears can use DLO AllGather after the
-ordinary loader finalizes their runtime weights and scales; DLO then shards and
-reconstructs those tensors with their recorded layouts. Other online methods
-must use `--dlo-no-use-allgather` or disable online quantization until their
-runtime layouts are validated.
+ordinary loader. Per-tensor online FP8, online INT8, and online MXFP8 linears
+can use DLO AllGather after the ordinary loader finalizes their runtime
+weights and scales; DLO then shards and reconstructs those tensors with their
+recorded layouts (they all keep plain 1-byte dtypes over ordinary strided
+views: online INT8 has an int8 weight plus fp32 scale, online MXFP8 has an
+fp8 weight plus e8m0 block scale, each with a contiguous or transposed-view
+finalized tensor, all covered by the same physical-order packing as online
+FP8). Other online methods must use `--dlo-no-use-allgather` or disable
+online quantization until their runtime layouts are validated.

 The Host Weight Runtime representation, publication, and no-AllGather consumer
 contracts are merged; see
@@ -459,8 +465,8 @@ Current source-level validation includes:
 - 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;
- ordinary-loader fallback for per-tensor online FP8 linears followed by DLO
-  sharding of finalized weights and scales;
+- ordinary-loader fallback for per-tensor online FP8, online INT8, and online
+  MXFP8 linears followed by DLO sharding of finalized weights and scales;
 - exact loader-to-backend plan transfer and ordinary-loader fallback;
 - ordered publication hashing, overlapped durability, unordered parallel
  checksum fallback, and node-local source-digest reuse/invalidation;
@@ -22,59 +22,60 @@ We use **Qwen-Image** as the reference implementation for decode parallel, and *

 **VAE Patch Parallelism** is an acceleration technique for both **encoding** and **decoding**. Instead of processing the entire tensor at once, the tensor is:

-+ Split into multiple spatial tiles
+- Split into multiple spatial tiles

-+ Distributed across multiple ranks
+- Distributed across multiple ranks

-+ Encoded/Decoded in parallel
+- Encoded/Decoded in parallel

-+ Merged to reconstruct the final output
+- Merged to reconstruct the final output

 This approach:

-+ Distributes computation across multiple devices
+- Distributes computation across multiple devices

-+ Reduces peak memory usage per device
+- Reduces peak memory usage per device

-+ Accelerates encoding/decoding latency
+- Accelerates encoding/decoding latency

 ### When to Use Encode vs Decode Parallel

-| Operation | Use Case | Example |
-|-----------|----------|---------|
-| **Decode Parallel** | Text-to-Image, Text-to-Video | Latent → Image/Video |
-| **Encode Parallel** | Image-to-Video (I2V) | Image → Latent (for conditioning) |
+| Operation           | Use Case                     | Example                           |
+| ------------------- | ---------------------------- | --------------------------------- |
+| **Decode Parallel** | Text-to-Image, Text-to-Video | Latent → Image/Video              |
+| **Encode Parallel** | Image-to-Video (I2V)         | Image → Latent (for conditioning) |

 ### Architecture
+
 We introduce **DistributedVaeExecutor** as the core component responsible for distributed VAE encoding/decoding.

 The executor is model-agnostic and accepts three function parameters:

-+ split – Partition the latent into tiles
+- split – Partition the latent into tiles

-+ exec – Decode a single tile
+- exec – Decode a single tile

-+ merge – Combine decoded tiles into the final output
+- merge – Combine decoded tiles into the final output

 #### Execution Flow

-+ Call split(z) to generate a list of TileTask and a GridSpec
+- Call split(z) to generate a list of TileTask and a GridSpec

-+ Dispatch tasks across ranks using workload-based balancing
+- Dispatch tasks across ranks using workload-based balancing

-+ Each rank executes exec(task) on its assigned tiles
+- Each rank executes exec(task) on its assigned tiles

-+ Gather decoded tile results to rank 0
+- Gather decoded tile results to rank 0

-+ Rank 0 performs merge(...)
+- Rank 0 performs merge(...)

-+ (Optional) Broadcast final result to all ranks
+- (Optional) Broadcast final result to all ranks

 This design separates:

-+ Distributed execution logic
+- Distributed execution logic

-+ Model-specific tiling and merging logic
+- Model-specific tiling and merging logic

 #### Why split / exec / merge is necessary?

@@ -82,23 +83,23 @@ The latent tensor cannot be arbitrarily partitioned.

 During decoding:

-+ Each output pixel may depend on neighboring pixels
+- Each output pixel may depend on neighboring pixels

-+ The receptive field is model-dependent
+- The receptive field is model-dependent

 Therefore:

-+ Tiles must include overlap
+- Tiles must include overlap

-+ Merge must perform blending to avoid seams
+- Merge must perform blending to avoid seams

 ## Step-by-Step Implementation (Decode)

 ### Step 1: Implement DistributedAutoencoderKLQwenImage
+
 `QwenImagePipeline` use `AutoencoderKLQwenImage` for vae, so implement a distributed version:

-
-```
+```python
 class DistributedAutoencoderKLQwenImage(AutoencoderKLQwenImage, DistributedVaeMixin):
    @classmethod
    def from_pretrained(cls, *args: Any, **kwargs: Any):
@@ -106,28 +107,35 @@ class DistributedAutoencoderKLQwenImage(AutoencoderKLQwenImage, DistributedVaeMi
        model.init_distributed()
        return model
 ```
+
 **Key points**:
-+ Inherit both AutoencoderKLQwenImage and DistributedVaeMixin
-+ Call init_distributed() after loading weights
+
+- Inherit both AutoencoderKLQwenImage and DistributedVaeMixin
+- Call init_distributed() after loading weights

 ### Step 2: Implement split/exec/merge
+
 Reuse `AutoencoderKLQwenImage.tiled_decode` logic and divide it into three stages. And we need return tiles with `GridSpec` and `TileTask`:
-```
+
+```python
 class GridSpec:
    split_dims: tuple[int, ...]  # Tensor dimensions being split (e.g., (2, 3) for (B, C, H, W))
    grid_shape: tuple[int, ...]  # Tile grid layout (num_rows, num_cols)
    tile_spec: dict = field(default_factory=dict) # Metadata required for merging
    output_dtype: torch.dtype | None = None # Final output dtype
 ```
-```
+
+```python
 class TileTask:
    tile_id: int # task id
    grid_coord: tuple[int, ...]  # Tile position in grid
    tensor: torch.Tensor | list[torch.Tensor]  # The tile tensor
    workload: int | float = 1 # Used for load balancing (e.g., tile area)
 ```
+
 And tiled base split/exec/merge as follow:
-```
+
+```python
 def tile_split(self, z: torch.Tensor) -> tuple[list[TileTask], GridSpec]:
    # mostly copy from AutoencoderKL
    _, _, num_frames, height, width = z.shape
@@ -210,11 +218,14 @@ def tile_merge(self, coord_tensor_map: dict[tuple[int, ...], torch.Tensor], grid
 ```

 ### Step 3: Override tiled_decode
+
 We need to override tiled_decode, the main logic is:
-+ check distributed is enabled
-+ select split/exec/merge
-+ Invoke self.distributed_executor.execute to decode
-```
+
+- check distributed is enabled
+- select split/exec/merge
+- Invoke self.distributed_executor.execute to decode
+
+```python
 def tiled_decode(self, z: torch.Tensor, return_dict: bool = True):
    if not self.is_distributed_enabled():
        return super().tiled_decode(z, return_dict=return_dict)
@@ -230,11 +241,14 @@ def tiled_decode(self, z: torch.Tensor, return_dict: bool = True):

    return DecoderOutput(sample=result)
 ```
+
 `broadcast_result` is set to True or False depending on the model; when enabled, the result will be used even on ranks other than 0.

 ### Step 4: Modify Pipeline
+
 Change vae model from AutoencoderKLQwenImage to DistributedAutoencoderKLQwenImage
-```
+
+```diff
 class YourModelPipeline(nn.Module):
    def __init__(
        self,
@@ -259,8 +273,8 @@ For models that require VAE encoding (e.g., Image-to-Video), you can also parall

 Similar to decode, split the input tensor into tiles. Key considerations:

-+ **Patchify handling**: If the model uses `patch_size`, scale tile parameters accordingly
-+ **Temporal chunking**: Video VAEs may have temporal compression (e.g., 4x)
+- **Patchify handling**: If the model uses `patch_size`, scale tile parameters accordingly
+- **Temporal chunking**: Video VAEs may have temporal compression (e.g., 4x)

 ```python
 def encode_tile_split(self, x: torch.Tensor) -> tuple[list[TileTask], GridSpec]:
@@ -406,23 +420,28 @@ def tiled_encode(self, x: torch.Tensor) -> torch.Tensor:
 **Key differences from decode parallel:**

 | Aspect | Decode Parallel | Encode Parallel |
-|--------|-----------------|-----------------|
+| -------- | ----------------- | ----------------- |
 | `broadcast_result` | Often `False` (only rank 0 needs output) | `True` (all ranks need latents for diffusion) |
 | Patchify | Applied in merge (unpatchify) | Handled by parent `_encode()` before `tiled_encode()` |
 | Temporal chunking | Frame-by-frame | Chunk-based (e.g., 1 + 4n frames) |

 ## Testing
-Verify numerical consistency between:
-+ vae_patch_parallel_size = 1

-+ vae_patch_parallel_size = N
+Verify numerical consistency between:
+
+- vae_patch_parallel_size = 1
+- vae_patch_parallel_size = N

 Example:
+
+```python
 torch.allclose(output_1, output_n, atol=1e-5)
+```

 Testing requirements:
-+ Fix random seed
-+ Use identical tiling strategy
+
+- Fix random seed
+- Use identical tiling strategy

 ```python
 m = Omni(
@@ -434,6 +453,7 @@ m = Omni(
        ),
    )
 ```
+
 When vae_patch_parallel_size is larger than the DiT world size, it will automatically fall back to using the DiT world size instead.

 ## Reference Implementations
@@ -441,7 +461,7 @@ When vae_patch_parallel_size is larger than the DiT world size, it will automati
 Complete examples in the codebase:

 | Model | Path | Decode Parallel | Encode Parallel |
-|-------|------|-----------------|-----------------|
+| ------- | ------ | ----------------- | ----------------- |
 | **Z-Image** | `vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl.py` | ✅ | ❌ |
 | **Wan2.2** | `vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_wan.py` | ✅ | ✅ |
 | **Qwen-Image** | `vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_qwenimage.py` | ✅ | ❌ |
@@ -456,7 +476,7 @@ The tile-parallel executor above assigns independent spatial **tiles** to ranks.
 ### How it differs from tile parallel

 | Aspect | Tile parallel (`"tile"`) | Spatially-sharded (`"spatial_shard_height"`/`"spatial_shard_width"`) |
-|--------|--------------------------|-----------------------------------------------|
+| -------- | -------------------------- | ----------------------------------------------- |
 | Unit of work | Independent overlapping tiles | A single global feature map sharded along H or W |
 | Cross-rank communication | Gather tiles to rank 0, stitch + blend | Per-conv **halo exchange** of boundary rows/cols (P2P) |
 | Output assembly | Blend overlapping tiles | All-gather shards on rank 0, trim padding (matches `broadcast_result=False`) |
@@ -481,7 +501,7 @@ serve.py (--vae-parallel-mode) / OmniEngineArgs
 ### Notes

 - The decoder is patched **in place** the first time spatial-shard decode runs and is bound to a single split dimension for the lifetime of the VAE instance.
- Numerical correctness vs. single-GPU decode is covered by `tests/diffusion/distributed/test_wan_spatial_shard.py::test_spatial_shard_decode_matches_reference` (multi-GPU, nightly `full_model` + `distributed_cuda`).
+- Numerical correctness vs. single-GPU decode is covered by `tests/diffusion/distributed/test_wan_spatial_shard.py::test_spatial_shard_decode_matches_reference` (multi-GPU, nightly `full_model` + `cards_2`).

 ---

@@ -37,6 +37,7 @@ validation_paths:
  - tests/engine/test_async_omni_engine_input.py
  - tests/engine/test_async_omni_engine_outputs.py
  - tests/engine/test_async_omni_engine_abort.py
+  - tests/e2e/offline_inference/test_qwen3_omni.py
  - tests/engine/test_async_omni_engine_stage_init.py
  - tests/engine/test_orchestrator.py
  - tests/engine/test_orchestrator_error_handling.py
@@ -113,6 +114,15 @@ Numbers become append-only after normative promotion.

 Test routing, output ordering, queue correlation, cancellation, failure
 propagation, shutdown ordering, and representative multi-stage execution.
+AR abort of a final-stage LLM request must deliver a terminal output with
+`finish_reason="abort"` and the generated prefix so collocated training can
+resume; diffusion abort remains whole-sample retry. Aborting a parallel-sampling
+child must drop `parent_requests` once no children remain. EngineCore control
+RPCs (`sleep` / `wake_up` / `pause_scheduler` / `resume_scheduler`) must
+propagate worker exceptions rather than returning `{"supported": False}`.
+When a request has multiple final output stages, only the last abort message
+is request-terminal. Output-processor abort state is committed only after the
+physical EngineCore abort succeeds.
 Changes that cross into stage runtime or public error behavior require review
 from that contract's owners.

@@ -43,6 +43,7 @@ depends_on:
 validation_paths:
  - tests/entrypoints/test_omni_entrypoints.py
  - tests/entrypoints/test_async_omni.py
+  - tests/entrypoints/test_async_omni_pause_sleep_routing.py
  - tests/entrypoints/test_async_omni_duplex.py
  - tests/entrypoints/test_serve.py
  - tests/entrypoints/test_stream_finish_reason.py
@@ -119,7 +120,13 @@ promotion.

 Test request validation, protocol conversion, model-adapter routing,
 streaming, session identity, disconnect/cancellation, and error mapping for
-each affected offline or serving entrypoint.
+each affected offline or serving entrypoint. Sleep must wait for in-flight
+`generate()` admission before EngineCore offload; `wake_up` does not resume
+admission — callers must `resume_generation()`. Sleeping tags are tracked per
+stage so `wake_up(stage_ids=[0])` does not skip a later `wake_up(stage_ids=[1])`.
+Streaming input pumps take an admission slot immediately before each EngineCore
+ADD or update, not while waiting for the next client chunk. Frontend abort
+keeps `request_states` until `generate()` consumes the terminal output.

 ## Promotion gate

@@ -23,6 +23,19 @@ vLLM-Omni currently recommends using the Docker image setup steps below.
 DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile.xpu -t vllm-omni-xpu --shm-size=4g .
 ```

+This layers vLLM-Omni on top of the published `vllm/vllm-openai-xpu:<VLLM_VERSION>`
+base image, which Docker pulls automatically. To target a different vLLM release,
+pass `--build-arg VLLM_VERSION=<tag>`. If that tag has not been published yet, build
+the base from upstream's own Dockerfile first and point the build at it:
+
+```bash
+git clone --depth 1 --branch "$VLLM_VERSION" https://github.com/vllm-project/vllm /tmp/vllm
+DOCKER_BUILDKIT=1 docker build -t vllm-openai-xpu:local --target vllm-openai \
+  -f /tmp/vllm/docker/Dockerfile.xpu /tmp/vllm
+DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile.xpu -t vllm-omni-xpu --shm-size=4g \
+  --build-arg VLLM_BASE=vllm-openai-xpu:local .
+```
+
 #### Launch the docker image

 ##### Launch with OpenAI API Server
@@ -26,15 +26,15 @@ th {
 | `InternVLAA1Pipeline` | InternVLA-A1 | `InternRobotics/InternVLA-A1-3B` | ✅︎ | ✅︎ | | | — |
 | `Gr00tN1d7Pipeline` | GR00T N1.7 | `nvidia/GR00T-N1.7-3B` | ✅︎ | | | | — |
 | `HunyuanImage3ForCausalMM` | HunyuanImage3.0 (DiT-only) | `tencent/HunyuanImage-3.0`, `tencent/HunyuanImage-3.0-Instruct` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` | ✅︎ | | | | [Published](https://recipes.vllm.ai/Qwen/Qwen-Image) |
-| `QwenImagePipeline` | Qwen-Image-2512 | `Qwen/Qwen-Image-2512` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Qwen/Qwen-Image-2512.md) |
+| `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` | ✅︎ | | | ✅︎ | [Published](https://recipes.vllm.ai/Qwen/Qwen-Image) |
+| `QwenImagePipeline` | Qwen-Image-2512 | `Qwen/Qwen-Image-2512` | ✅︎ | | | ✅︎ | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Qwen/Qwen-Image-2512.md) |
 | `QwenImageEditPipeline` | Qwen-Image-Edit | `Qwen/Qwen-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2509 | `Qwen/Qwen-Image-Edit-2509` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `QwenImageLayeredPipeline` | Qwen-Image-Layered | `Qwen/Qwen-Image-Layered` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2511 | `Qwen/Qwen-Image-Edit-2511` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `GlmImagePipeline` | GLM-Image | `zai-org/GLM-Image` | ✅︎ | ✅︎ | | | — |
 | `ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `Krea2Pipeline` | Krea 2 (Raw + Turbo) | `krea/Krea-2-Raw`, `krea/Krea-2-Turbo` | ✅︎ | | | | — |
+| `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 | | | | — |
@@ -42,7 +42,7 @@ th {
 | `SanaWmPipeline` | SANA-WM | `BBBBruce/SANA-WM_bidirectional-stage1-diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/NVIDIA/SANA-WM.md) |
 | `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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `Wan22VACEPipeline` | Wan2.2-VACE | `Pyros13/Wan2.2-VACE-Fun-A14B-Diffusers` | ✅︎ |   |   |   | — |
+| `Wan22VACEPipeline` | Wan2.2-VACE | `Pyros13/Wan2.2-VACE-Fun-A14B-Diffusers` | ✅︎ |   |   | ✅︎ | — |
 | `LTX2Pipeline` | LTX-2 / LTX-2.3 one-stage T2V and I2V | `Lightricks/LTX-2`, `diffusers/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | | — |
 | `LTX2TwoStagePipeline` | LTX-2 / LTX-2.3 ordinary two-stage T2V and I2V | `Lightricks/LTX-2`, `diffusers/LTX-2.3-Diffusers` + matching Lightricks LoRA and upsampler | ✅︎ | ✅︎ | | | — |
 | `LTX2DistilledOneStagePipeline` | LTX-2 / LTX-2.3 merged-distilled one-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled`, `diffusers/LTX-2.3-Distilled-Diffusers` | ✅︎ | ✅︎ | | | — |
@@ -62,12 +62,12 @@ th {
 | `LongcatImagePipeline` | LongCat-Image | `meituan-longcat/LongCat-Image` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `LongCatVideoAvatarPipeline` | LongCat-Video-Avatar-1.5 A2V/AI2V (native single-speaker and multi-speaker AI2V/AVC) | `meituan-longcat/LongCat-Video-Avatar-1.5` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/meituan-longcat/LongCat-Video-Avatar-1.5.md) |
-| `BooguImagePipeline` | Boogu-Image | `Boogu/Boogu-Image-0.1-Base`, `Boogu/Boogu-Image-0.1-Edit` | ✅︎ | | | | — |
+| `BooguImagePipeline` | Boogu-Image | `Boogu/Boogu-Image-0.1-Base`, `Boogu/Boogu-Image-0.1-Edit` | ✅︎ | | | ✅︎ | — |
 | `StableDiffusionXLPipeline` | Stable-Diffusion-XL | `stabilityai/stable-diffusion-xl-base-1.0` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `StableDiffusion3Pipeline` | Stable-Diffusion-3 | `stabilityai/stable-diffusion-3.5-medium` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `CosyVoice3Model` | CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | ✅︎ | ✅︎ | | ✅︎ | — |
-| `OmniVoiceModel` | OmniVoice | `k2-fsa/OmniVoice` | ✅︎ | | | | — |
-| `VoxCPM2TalkerForConditionalGeneration` | VoxCPM2 | `openbmb/VoxCPM2` | ✅︎ | | | | — |
+| `OmniVoiceModel` | OmniVoice | `k2-fsa/OmniVoice` | ✅︎ | | | ✅︎ | — |
+| `VoxCPM2TalkerForConditionalGeneration` | VoxCPM2 | `openbmb/VoxCPM2` | ✅︎ | | | ✅︎ | — |
 | `DotsTTSForConditionalGeneration` | dots.tts | `rednote-hilab/dots.tts-soar` | ✅︎ | | | | — |
 | `MammothModa2ForConditionalGeneration` | MammothModa2-Preview | `bytedance-research/MammothModa2-Preview` | ✅︎ | ✅︎ | | | — |
 | `MammothModa2ForConditionalGeneration` | MammothModa2-Dev (AR-only image understanding) | `bytedance-research/MammothModa2-Dev` | ✅︎ | | | | — |
@@ -78,10 +78,10 @@ th {
 | `OmniGen2Pipeline` | OmniGen2 | `OmniGen2/OmniGen2` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `StableAudioPipeline` | Stable-Audio-Open | `stabilityai/stable-audio-open-1.0` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `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-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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `MingTTSForConditionalGeneration` | Ming-omni-tts dense 0.5B | `inclusionAI/Ming-omni-tts-0.5B` | ✅︎ | | | | — |
+| `MingTTSForConditionalGeneration` | Ming-omni-tts dense 0.5B | `inclusionAI/Ming-omni-tts-0.5B` | ✅︎ | | | ✅︎ | — |
 | `GLMTTSForConditionalGeneration` | GLM-TTS | `zai-org/GLM-TTS` | ✅︎ | | | | — |
 | `MossTTSNanoForCausalLM` | MOSS-TTS-Nano | `OpenMOSS-Team/MOSS-TTS-Nano` | ✅︎ | | | | — |
 | `MossTTSDelayModel` | MOSS-TTS, MOSS-TTSD, MOSS-SoundEffect, MOSS-VoiceGenerator | `OpenMOSS-Team/MOSS-VoiceGenerator` | ✅︎ | | | | — |
@@ -101,10 +101,10 @@ th {
 | `HunyuanVideo15ImageToVideoPipeline` | HunyuanVideo-1.5-I2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v` | ✅︎ | ✅︎ | | | — |
 | `VoxtralTTSForConditionalGeneration` | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✅︎ | ✅︎ | | | — |
 | `CovoAudioForConditionalGeneration` | Covo-Audio-Chat | `tencent/Covo-Audio-Chat` | ✅︎ | | | | — |
-|`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | | — |
+|`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | ✅︎ | — |
 | `MiniCPMO45OmniForConditionalGeneration` | MiniCPM-o 4.5 | `openbmb/MiniCPM-o-4_5` | ✅︎ | | ✅︎ | | — |
 | `ErnieImagePipeline` | ERNIE-Image | `baidu/ERNIE-Image`, `baidu/ERNIE-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `GepardTalkerForConditionalGeneration` | Gepard-1.0 | `nineninesix/gepard-1.0` | ✅︎ | | | | — |
+| `GepardTalkerForConditionalGeneration` | Gepard-1.0 | `nineninesix/gepard-1.0` | ✅︎ | | | ✅︎ | — |
 |`HiDreamImagePipeline` | HiDream-I1-Full | `HiDream-ai/HiDream-I1-Full` | ✅︎ | ✅︎ | | | — |
 | `HiDreamO1ImagePipeline` | HiDream-O1-Image | `HiDream-ai/HiDream-O1-Image` | ✅︎ | | | | — |

@@ -17,6 +17,8 @@ globally.
 | --- | --- | --- |
 | `sparsity` | finite, `[0, 1]` | Nominal dropped-key-block fraction; default `0.8`; `0` disables sparsity |
 | `start_step` | integer, `>= 0` | Number of early denoise steps kept dense |
+| `end_step` | integer, `>= 0` | Number of final denoise steps kept dense (tail fallback) |
+| `precision` | `"bf16"`, `"fp8"`, `"mix"` | Kernel precision mode; default `"bf16"`. Requires MindIE-SD with `sparse_attention(precision=...)` |
 | `skip_layers` | selector such as `"0-3,38"` | DiT blocks kept dense |

 ```bash
@@ -52,6 +54,12 @@ platform raises. It is incompatible with ring sequence parallelism because
 the kernel needs the complete key sequence for ranking; use Ulysses sequence
 parallelism with `ring_degree=1`.

+The `precision` knob requires a MindIE-SD release whose `sparse_attention`
+accepts `precision=`; older releases accept it through `**kwargs` but
+silently ignore it and run the BF16 path. `RAINFUSION_ATTN` raises at runtime
+when a non-`bf16` precision is requested without that support, so pin a
+compatible MindIE-SD release when using `precision="mix"` or `"fp8"`.
+
 ## Geometry handling

 RainFusion handles arbitrary video grids by rearranging video spatially and,
@@ -12,15 +12,17 @@ LoRA adapters are lightweight, model-specific fine-tuning weights that can be ap
 ## LoRA Adapter Format

 ### PEFT (Parameter-Efficient Fine-Tuning) format (default)
+
 A typical PEFT-format LoRA adapter directory structure:

-```
+```text
 lora_adapter/
 ├── adapter_config.json
 └── adapter_model.safetensors
 ```

 The `adapter_config.json` file contains metadata about the LoRA adapter, including:
+
 - `r`: LoRA rank
 - `lora_alpha`: LoRA alpha scaling factor
 - `target_modules`: List of module names to apply LoRA to
@@ -67,7 +69,7 @@ For distilled few-step LoRAs, pass `lora_backend="distill"` together with one or
 For Qwen-Image and Wan pipelines, the distill backend calls `pipeline.load_lora_weights(...)` during worker initialization.

 | Pipeline | Supported distilled LoRA repo | Notes |
-|----------|--------------------------------|-------|
+| ---------- | -------------------------------- | ------- |
 | `QwenImagePipeline` | `lightx2v/Qwen-Image-2512-Lightning` | Used with Qwen-Image-2512 Lightning-style few-step inference. |
 | `Wan22Pipeline` | `lightx2v/Wan2.1-Distill-Loras`, `lightx2v/Wan2.2-Distill-Loras` | Wan2.1 uses one LoRA file. For dual-transformer Wan2.2 MoE, pass high-noise then low-noise LoRA files. |
 | `Wan22I2VPipeline` | `lightx2v/Wan2.2-Distill-Loras` | For dual-transformer Wan2.2 MoE, pass high-noise then low-noise LoRA files. |
@@ -240,6 +242,37 @@ Notes:
 - Output quality and speed depend on the replacement checkpoints and sampling params you choose.
 - If you only need to fuse distilled LoRAs into a Wan2.2 checkpoint at load time (without the full LightX2V convert + assemble pipeline), you can instead pass them directly via `--lora-backend distill --lora-path <high>.safetensors <low>.safetensors`. See the [Distill backend](#distill-backend-fuse-distilled-lora-at-init) section above.

+## MiniMax-H3 adapter-declared schedules
+
+MiniMax-H3 supports two few-step mechanisms that must not be conflated:
+
+- **Checkpoint-pinned schedule**: a merged release writes `base_schedule` into
+  `model_index.json`. Requests must pass `num_inference_steps` as the interval
+  count (for example `4` for `[1.0, 0.7, 0.4, 0.15, 0.0]`).
+- **Runtime Turbo LoRA**: LightX2V Turbo artifacts request five sigma points and
+  enforce `flow_shift=6`, `audio_flow_shift=3`.
+- **Runtime native LoRA**: FlashGen-style artifacts declare
+  `key_format=minimax-h3-native` and embed `base_schedule` in safetensors
+  metadata. When active, the adapter schedule overrides the base checkpoint and
+  requests must use the interval-count contract (`num_inference_steps=4`).
+  Request-mode generation may omit the field and take the count from the adapter
+  schedule; step execution requires it explicitly, because the step scheduler
+  derives the total step count from the request before the adapter schedule is
+  known.
+
+Native artifacts also declare `qkv_layout=grouped`. The H3 loader reorders fused
+`qkv_proj` LoRA rows with the same `_reorder_grouped_qkv_to_qkv` helper used for
+base-weight loading, then binds the packed Q/K/V slices through the legacy PEFT
+manager without modifying `DiffusionLoRAManager`.
+
+Both runtime adapters run with distributed layerwise offload in request-mode
+generation, where the manager keeps the LoRA A/B buffers resident on the compute
+device while DLO streams the base blocks. MiniMax-H3 step execution still
+rejects DLO, so `--step-execution` cannot be combined with
+`--enable-distributed-layerwise-offload`. Model-level CPU offload and standard
+layerwise offload are rejected in every mode, because the dynamic LoRA tensors
+are neither parameters nor registered buffers and therefore do not participate
+in those weight lifecycles.

 ## See Also

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """
 Image-to-Video generation example using Wan2.2 I2V/TI2V models, LTX2/LTX-2.3,
@@ -111,6 +111,13 @@ def build_image_to_video_prompt(
    return result


+def _validate_video_output_type(output_type: str) -> None:
+    if output_type not in {"image", "video"}:
+        raise ValueError(
+            f"Unexpected output type '{output_type}', expected 'video' or legacy 'image' for video generation."
+        )
+
+
 def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description=(
@@ -665,10 +672,7 @@ def main():
        frames = frames[0] if frames else None

    if isinstance(frames, OmniRequestOutput):
-        if frames.final_output_type != "image":
-            raise ValueError(
-                f"Unexpected output type '{frames.final_output_type}', expected 'image' for video generation."
-            )
+        _validate_video_output_type(frames.final_output_type)
        if frames.multimodal_output and "audio" in frames.multimodal_output:
            audio = frames.multimodal_output["audio"]
            audio_sample_rate = frames.multimodal_output.get("audio_sample_rate", audio_sample_rate)
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 import argparse
 import json
@@ -174,6 +174,13 @@ def build_text_to_video_prompt(prompt: str, negative_prompt: str | None) -> dict
    return result


+def _validate_video_output_type(output_type: str) -> None:
+    if output_type not in {"image", "video"}:
+        raise ValueError(
+            f"Unexpected output type '{output_type}', expected 'video' or legacy 'image' for video generation."
+        )
+
+
 def _normalize_float_tensor(tensor: torch.Tensor, source_range: str) -> torch.Tensor:
    """Normalize decoded frames according to the pipeline's output contract."""
    if not tensor.is_floating_point():
@@ -620,10 +627,7 @@ def main():
        frames = frames[0] if frames else None

    if isinstance(frames, OmniRequestOutput):
-        if frames.final_output_type != "image":
-            raise ValueError(
-                f"Unexpected output type '{frames.final_output_type}', expected 'image' for video generation."
-            )
+        _validate_video_output_type(frames.final_output_type)
        if frames.multimodal_output and "audio" in frames.multimodal_output:
            audio = frames.multimodal_output["audio"]
            audio_sample_rate = frames.multimodal_output.get("audio_sample_rate", audio_sample_rate)
@@ -248,11 +248,11 @@ addopts = [
    "--dist=loadgroup"
 ]
 markers = [
-    # ci/cd required
-    "core_model: L1&L2 tests (run in each PR)",
-    "advanced_model: L3 level tests (run on each merge)",
-    "full_model: L4 level tests (run nightly)",
-    "local_model: tests (run locally)",
+    # ci/cd required — tag with [ci-level] for tests.helpers.mark.get_level_markers().
+    "core_model: [ci-level] L1&L2 tests (run in each PR)",
+    "advanced_model: [ci-level] L3 level tests (run on each merge)",
+    "full_model: [ci-level] L4 level tests (run nightly)",
+    "local_model: [ci-level] tests (run locally)",
    # function module markers
    "diffusion: Diffusion model tests",
    "omni: Omni model tests",
@@ -262,38 +262,41 @@ markers = [
    "sp: Sequence parallelism tests",
    "example: Doc example code tests",
    # platform markers
-    "cpu: Tests that run on CPU",
-    "gpu: Tests that run on GPU (auto-added)",
-    "cuda: Tests that run on CUDA (auto-added)",
-    "rocm: Tests that run on AMD/ROCm (auto-added)",
-    "xpu: Tests that run on XPU (auto-added)",
-    "npu: Tests that run on NPU/Ascend (auto-added)",
-    "musa: Tests that run on MUSA/Moore Threads (auto-added)",
+    # Tag with [hardware-platform] for tests.helpers.mark.get_supported_platforms().
+    # Add [gpu] on accelerator platforms that should also get pytest.mark.gpu.
+    # cpu/gpu are pytest marks; hardware_test(res=...) rejects them as keys.
+    "cpu: [hardware-platform] Tests that run on CPU",
+    "gpu: [hardware-platform] Tests that run on GPU (auto-added)",
+    "cuda: [hardware-platform] [gpu] Tests that run on CUDA (auto-added)",
+    "rocm: [hardware-platform] [gpu] Tests that run on AMD/ROCm (auto-added)",
+    "xpu: [hardware-platform] [gpu] Tests that run on XPU (auto-added)",
+    "npu: [hardware-platform] Tests that run on NPU/Ascend (auto-added)",
+    "musa: [hardware-platform] [gpu] Tests that run on MUSA/Moore Threads (auto-added)",
    # specified computation resources marks (auto-added)
-    # Tag with [hardware-resource] for tests.helpers.mark.get_hardware_mark_list().
-    "H100: [hardware-resource] Tests that require H100 GPU",
-    "H800: [hardware-resource] Tests that require H800 GPU",
-    "H200: [hardware-resource] Tests that require H200 GPU",
-    "L4: [hardware-resource] Tests that require L4 GPU",
-    "B200: [hardware-resource] Tests that require B200 GPU",
-    "MI325: [hardware-resource] Tests that require MI325 GPU (AMD/ROCm)",
-    "B60: [hardware-resource] Tests that require Intel Arc Pro B60 XPU",
-    "S5000: [hardware-resource] Tests that require S5000 GPU (Moore Threads/MUSA)",
-    "A2: [hardware-resource] Tests that require A2 NPU",
-    "A3: [hardware-resource] Tests that require A3 NPU",
-    "distributed_cuda: Tests that require multi cards on CUDA platform",
-    "distributed_rocm: Tests that require multi cards on ROCm platform",
-    "distributed_xpu: Tests that require multi cards on XPU platform",
-    "distributed_npu: Tests that require multi cards on NPU platform",
-    "distributed_musa: Tests that require multi cards on MUSA platform",
-    "skipif_cuda: Skip if the num of CUDA cards is less than the required",
-    "skipif_rocm: Skip if the num of ROCm cards is less than the required",
-    "skipif_xpu: Skip if the num of XPU cards is less than the required",
-    "skipif_npu: Skip if the num of NPU cards is less than the required",
-    "skipif_musa: Skip if the num of MUSA cards is less than the required",
+    # [hardware-resource] + [cuda]/[rocm]/[xpu]/[npu]/[musa] for hardware_test(res=...).
+    # Do not put [gpu] on SKU lines: that would make gpu a res key.
+    "H100: [hardware-resource] [cuda] Tests that require H100 GPU",
+    "H800: [hardware-resource] [cuda] Tests that require H800 GPU",
+    "H200: [hardware-resource] [cuda] Tests that require H200 GPU",
+    "L4: [hardware-resource] [cuda] Tests that require L4 GPU",
+    "B200: [hardware-resource] [cuda] Tests that require B200 GPU",
+    "MI325: [hardware-resource] [rocm] Tests that require MI325 GPU (AMD/ROCm)",
+    "B60: [hardware-resource] [xpu] Tests that require Intel Arc Pro B60 XPU",
+    "S5000: [hardware-resource] [musa] Tests that require S5000 GPU (Moore Threads/MUSA)",
+    "A2: [hardware-resource] [npu] Tests that require A2 NPU",
+    "A3: [hardware-resource] [npu] Tests that require A3 NPU",
+    # Tag with [hardware-cards] for tests.helpers.mark.get_supported_card_counts().
+    "cards_1: [hardware-cards] Tests that require 1 accelerator card",
+    "cards_2: [hardware-cards] Tests that require 2 accelerator cards",
+    "cards_3: [hardware-cards] Tests that require 3 accelerator cards",
+    "cards_4: [hardware-cards] Tests that require 4 accelerator cards",
+    "cards_5: [hardware-cards] Tests that require 5 accelerator cards",
+    "cards_6: [hardware-cards] Tests that require 6 accelerator cards",
+    "cards_7: [hardware-cards] Tests that require 7 accelerator cards",
+    "cards_8: [hardware-cards] Tests that require 8 accelerator cards",
    # more detailed markers
    "sp: Sequence parallelism tests (multi-GPU)",
-    "slow: Slow tests (may skip in quick CI)",
+    "slow: [ci-level] Slow tests (may skip in quick CI)",
    "benchmark: Benchmark tests",
    "xdist: xdist-parallelizable tests",
 ]
@@ -860,6 +860,63 @@ fixed HBM usage. Model-level and standard layerwise offload remain unsupported.
 The five requested sigma points produce the four denoiser evaluations expected
 by the Turbo artifact.

+### FlashGen native LoRA
+
+The FlashGen 4-step T2VA artifact uses the native MiniMax-H3 module layout and
+declares its distilled sigma schedule in safetensors metadata. It is published on
+[ModelScope](https://modelscope.cn/models/FlashGen/Minimax-H3-4step-lora-flashgen):
+
+```text
+FlashGen/Minimax-H3-4step-lora-flashgen/minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors
+```
+
+Download only that file:
+
+```bash
+python -m pip install modelscope
+export FLASHGEN_DIR=/path/to/minimax-h3-flashgen-lora
+export FLASHGEN_FILE=minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors
+modelscope download FlashGen/Minimax-H3-4step-lora-flashgen \
+  --local_dir "${FLASHGEN_DIR}" \
+  --include "${FLASHGEN_FILE}"
+export FLASHGEN_LORA="${FLASHGEN_DIR}/${FLASHGEN_FILE}"
+```
+
+Start from a non-offloaded or DLO FL2VA server command and add
+`--task-type fl2va --lora-backend peft --lora-path "${FLASHGEN_LORA}"`.
+Each request must use T2VA and the distilled interval-count contract:
+
+```bash
+-F 'num_inference_steps=4' \
+-F 'extra_params={"task":"t2va","duration":5.2}' \
+-F "lora={\"name\":\"h3-flashgen-v1.0\",\"path\":\"${FLASHGEN_LORA}\",\"scale\":1.0}"
+```
+
+This path rejects Ref2VA and checkpoints that already pin `base_schedule` in
+`model_index.json`. The adapter metadata carries
+`base_schedule=1.0,0.7,0.4,0.15,0.0`, so `num_inference_steps=4` means four
+denoiser evaluations, not five sigma points. Request-mode generation may omit
+the field and take the count from the adapter schedule; `--step-execution`
+requires it explicitly, because the step scheduler reads the total step count
+off the request at admission, before the adapter schedule is known.
+
+DLO is supported in request-mode generation on the same terms as the Turbo
+adapter: the request-switchable LoRA A/B buffers stay resident on the
+accelerator while DLO streams only the base blocks, so budget for that
+additional fixed HBM usage. The native artifact is rank 64 over 259 target
+modules, and its packed `qkv_proj` and `fc1` layers reuse the full-input A
+tensor per slice while B carries slice-local output rows, so the resident
+footprint exceeds the on-disk payload; measure it for your parallel layout
+rather than assuming the checkpoint size. Pure Ulysses replicates the adapter
+on every rank, while DiT tensor parallelism shards the B buffers. Model-level
+and standard layerwise offload remain unsupported, and `--step-execution`
+cannot be combined with `--enable-distributed-layerwise-offload`.
+
+To validate a deployment, post the same fixed-seed T2VA request twice with the
+adapter and twice without it, then compare the four output digests. The adapter
+is bound and deterministic when each pair matches internally and the two pairs
+differ from each other.
+
 ## Key parameters

 | Parameter | Recommended value | Notes |
@@ -18,11 +18,11 @@ examples in this repository.

 Qwen3-TTS supports three task types, each backed by a dedicated model checkpoint:

-| Task Type     | Model                                    | Description                                                  |
-| ------------- | ---------------------------------------- | ------------------------------------------------------------ |
-| `CustomVoice` | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice`  | Predefined speaker voices with optional style/emotion control |
-| `VoiceDesign` | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`  | Generate speech from a natural language voice description     |
-| `Base`        | `Qwen/Qwen3-TTS-12Hz-1.7B-Base`         | Voice cloning from reference audio + transcript              |
+| Task Type     | Model                                    | Description                                                   |
+| ------------- | ---------------------------------------- | ------------------------------------------------------------- |
+| `CustomVoice` | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice`   | Predefined speaker voices with optional style/emotion control |
+| `VoiceDesign` | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`   | Generate speech from a natural language voice description     |
+| `Base`        | `Qwen/Qwen3-TTS-12Hz-1.7B-Base`          | Voice cloning from reference audio + transcript               |

 Smaller 0.6B variants are also available for `CustomVoice` and `Base`.

@@ -166,6 +166,15 @@ python examples/offline_inference/text_to_speech/qwen3_tts/end2end.py --query-ty
 - Key flags: `--omni` is required. `--deploy-config` points to the bundled two-stage pipeline config.
 - Async chunking: Enabled by default in `qwen3_tts.yaml` for streaming-friendly first-audio latency. Raw audio streaming requires `stream=true`, `stream_format="audio"`, and `response_format="pcm"`.
 - Task/model matching: Each task type requires its matching model checkpoint. Using a CustomVoice model for a Base (voice clone) request will fail.
+- Base codec termination: Base requests without an explicit `max_new_tokens`
+  use a text-scaled safety ceiling (at least 192 codec frames and no more than
+  the configured model limit). If the Talker reaches that ceiling without
+  codec EOS, non-streaming serving discards the incomplete audio and retries
+  once with a fresh seed; an explicit `seed` or `max_new_tokens` disables the
+  retry. SSE and WebSocket clients receive structured errors and must discard
+  previously emitted audio when the error contains `"action":"discard"`;
+  raw PCM streams terminate with a connection error after any partial bytes
+  already sent.
 - Known limitations: The server serves one model variant at a time. To switch task types (e.g., CustomVoice to Base), restart the server with the corresponding model.

 ## Hardware Support
@@ -279,17 +279,79 @@ def test_unmeasured_tpot_stays_missing_after_tokenizer_fallback():
    assert metrics.request_goodput == 0.0


-def test_measured_zero_itl_is_not_treated_as_missing():
+def test_zero_itl_does_not_create_zero_tpot():
    output = _make_output(100, output_tokens=3)
    output.itl = [0.0, 0.0]
    output.duplex_session_metrics = {"mean_ttft_ms": 100.0}

    metrics = _calculate_test_metrics([output], {"tpot": 1.0})

-    assert (metrics.num_tpot_samples, metrics.num_itl_samples) == (1, 2)
-    assert metrics.mean_tpot_ms == 0.0
+    assert (metrics.num_tpot_samples, metrics.num_itl_samples) == (0, 2)
+    assert math.isnan(metrics.mean_tpot_ms)
    assert metrics.mean_itl_ms == 0.0
-    assert metrics.request_goodput == 1.0
+    assert metrics.request_goodput == 0.0
+
+
+def test_stage_output_tokens_without_client_timing_omit_tpot(capsys):
+    output = _make_output(100, output_tokens=5)
+    output.itl = []
+    output.text_latency = 0.0
+    output.ttft = 0.1
+
+    metrics, _ = calculate_metrics(
+        input_requests=[],
+        outputs=[output],
+        dur_s=1.0,
+        tokenizer=None,
+        selected_percentiles=[50.0, 99.0],
+        goodput_config_dict={},
+        task_type=TaskType.GENERATION,
+        selected_percentile_metrics=["tpot"],
+        max_concurrency=None,
+        request_rate=float("inf"),
+        benchmark_duration=1.0,
+    )
+
+    printed = capsys.readouterr().out
+    assert metrics.num_tpot_samples == 0
+    assert math.isnan(metrics.mean_tpot_ms)
+    assert "Time per Output Token" not in printed
+    assert "Mean TPOT" not in printed
+
+
+def test_consistent_client_latency_can_supply_tpot_fallback():
+    output = _make_output(100, output_tokens=10)
+    output.itl = []
+    output.ttft = 0.1
+    output.text_latency = 1.0
+
+    metrics = _calculate_test_metrics([output])
+
+    assert metrics.num_tpot_samples == 1
+    assert metrics.mean_tpot_ms == pytest.approx(100.0)
+
+
+def test_single_token_responses_do_not_report_zero_tpot(capsys):
+    output = _make_output(100, output_tokens=1)
+
+    metrics, _ = calculate_metrics(
+        input_requests=[],
+        outputs=[output],
+        dur_s=1.0,
+        tokenizer=None,
+        selected_percentiles=[50.0],
+        goodput_config_dict={},
+        task_type=TaskType.GENERATION,
+        selected_percentile_metrics=["tpot"],
+        max_concurrency=None,
+        request_rate=float("inf"),
+        benchmark_duration=1.0,
+    )
+
+    printed = capsys.readouterr().out
+    assert metrics.num_tpot_samples == 0
+    assert math.isnan(metrics.mean_tpot_ms)
+    assert "Time per Output Token" not in printed


 def test_duplex_goodput_does_not_pair_measurements_from_different_requests():
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """
 Unit tests for patch.py
@@ -17,6 +17,7 @@ from vllm.benchmarks.lib.endpoint_request_func import RequestFuncInput

 from vllm_omni.benchmarks.patch.patch import (
    MixRequestFuncOutput,
+    _apply_stage0_token_timings,
    async_request_openai_chat_omni_completions,
    async_request_openai_realtime_duplex,
 )
@@ -102,7 +103,24 @@ async def test_seed_tts_realtime_duplex_exports_per_request_metrics(monkeypatch)
                received_at_s=now + 0.03,
            )
            self.events.add(
-                {"type": "response.done", "response": {"id": response_id}},
+                {
+                    "type": "response.done",
+                    "response": {
+                        "id": response_id,
+                        "metadata": {
+                            "vllm_omni": {
+                                "stage_metrics": {
+                                    "0": {
+                                        "num_tokens_out": 3,
+                                        "vllm_ttft_ms": 20.0,
+                                        "vllm_tpot_ms": 10.0,
+                                        "vllm_itls_ms": [10.0, 10.0],
+                                    }
+                                }
+                            }
+                        },
+                    },
+                },
                received_at_s=now + 0.04,
            )

@@ -143,6 +161,7 @@ async def test_seed_tts_realtime_duplex_exports_per_request_metrics(monkeypatch)
    assert client.configure_kwargs["native_duplex"] is False
    assert client.configure_kwargs["extra_body"] == {
        "ref_audio": "data:audio/wav;base64,AAAA",
+        "return_stage_metrics": True,
    }
    assert client.sent == [
        event
@@ -167,6 +186,10 @@ async def test_seed_tts_realtime_duplex_exports_per_request_metrics(monkeypatch)
    assert output.audio_ttfp > output.ttft
    assert output.audio_rtf > 0
    assert output.latency > 0
+    assert output.output_tokens == 12
+    assert output.itl == [0.01, 0.01] * 4
+    assert output.text_latency == pytest.approx(output.ttft + 0.08)
+    assert output.tpot_measured is True
    assert output.tts_turn_pcm_bytes == [b"\x00\x00" * 2400] * 4
    assert output.tts_output_pcm_bytes == b"\x00\x00" * 9600
    session_id = output.duplex_request_metrics[0]["session_id"]
@@ -198,6 +221,54 @@ async def test_seed_tts_realtime_duplex_exports_per_request_metrics(monkeypatch)
    }


+def test_stage0_token_timings_use_weighted_tpot_when_itls_are_incomplete():
+    output = MixRequestFuncOutput(ttft=0.1)
+
+    measured = _apply_stage0_token_timings(
+        output,
+        [
+            {"output_token_count": 3, "itls_ms": [], "tpot_ms": 10.0},
+            {"output_token_count": 2, "itls_ms": [], "tpot_ms": 40.0},
+        ],
+        expected_output_tokens=5,
+    )
+
+    assert measured is True
+    assert output.itl == []
+    assert output.tpot_measured is True
+    assert (output.text_latency - output.ttft) / 4 == pytest.approx(0.02)
+
+
+def test_stage0_zero_itls_fall_back_to_positive_tpot():
+    output = MixRequestFuncOutput(ttft=0.1)
+
+    measured = _apply_stage0_token_timings(
+        output,
+        [{"output_token_count": 3, "itls_ms": [0.0, 0.0], "tpot_ms": 25.0}],
+        expected_output_tokens=3,
+    )
+
+    assert measured is True
+    assert output.itl == []
+    assert output.tpot_measured is True
+    assert (output.text_latency - output.ttft) / 2 == pytest.approx(0.025)
+
+
+def test_stage0_token_count_without_timing_is_not_measured():
+    output = MixRequestFuncOutput(ttft=0.1)
+
+    measured = _apply_stage0_token_timings(
+        output,
+        [{"output_token_count": 3, "itls_ms": [], "tpot_ms": None}],
+        expected_output_tokens=3,
+    )
+
+    assert measured is False
+    assert output.itl == []
+    assert output.text_latency == output.ttft
+    assert output.tpot_measured is False
+
+
 def create_sse_chunk(data_dict):
    """Helper to create SSE formatted chunk"""
    return f"data: {json.dumps(data_dict)}\n\n".encode()
@@ -585,7 +585,7 @@ async def test_adapter_requires_and_forwards_exact_prepared_payload(tmp_path: Pa

@pytest.mark.asyncio
@pytest.mark.parametrize(
-    ("request_metrics", "expected_itl", "expected_text_latency"),
+    ("request_metrics", "expected_itl", "expected_text_latency", "expected_tpot_measured"),
    [
        (
            [
@@ -594,6 +594,7 @@ async def test_adapter_requires_and_forwards_exact_prepared_payload(tmp_path: Pa
            ],
            [0.01, 0.02, 0.03],
            0.16,
+            True,
        ),
        (
            [
@@ -602,6 +603,7 @@ async def test_adapter_requires_and_forwards_exact_prepared_payload(tmp_path: Pa
            ],
            [],
            0.18,
+            True,
        ),
        (
            [
@@ -610,6 +612,7 @@ async def test_adapter_requires_and_forwards_exact_prepared_payload(tmp_path: Pa
            ],
            [],
            0.1,
+            False,
        ),
    ],
 )
@@ -619,6 +622,7 @@ async def test_adapter_reports_exact_or_weighted_token_timing(
    request_metrics: list[dict[str, object]],
    expected_itl: list[float],
    expected_text_latency: float,
+    expected_tpot_measured: bool,
 ):
    case, options = _case(tmp_path), _session_options(tmp_path)
    request = _request(case, options)
@@ -643,7 +647,7 @@ async def test_adapter_reports_exact_or_weighted_token_timing(
    assert output.success
    assert output.itl == expected_itl
    assert output.text_latency == pytest.approx(expected_text_latency)
-    assert output.tpot_measured is True
+    assert output.tpot_measured is expected_tpot_measured


 def test_batch_finalization_publishes_only_measured_results(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
@@ -1159,7 +1159,7 @@ class TestDeployConfigLoading:
        stages = merge_pipeline_deploy(pipeline, deploy)

        assert deploy.session_mode == "duplex"
-        assert deploy.active_stream_window == 1
+        assert deploy.active_stream_window == max_sessions
        assert deploy.duplex_session.max_sessions == max_sessions
        assert [stage.session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
        assert [stage.to_omegaconf().session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
@@ -2399,6 +2399,20 @@ class TestPlatformOverrides:
        assert rocm.stages[0].enforce_eager is None
        assert rocm.stages[1].enforce_eager is True

+    def test_higgs_audio_v3_rocm_uses_triton_attention(self):
+        deploy_path = Path(get_deploy_config_path("higgs_multimodal_qwen3.yaml"))
+
+        base = load_deploy_config(deploy_path)
+        assert base.stages[0].engine_extras["attention_backend"] == "FLASHINFER"
+
+        rocm = _apply_platform_overrides(base, platform="rocm")
+        assert rocm.stages[0].engine_extras["attention_backend"] == "TRITON_ATTN"
+
+        pipeline = resolve_pipeline_config("higgs_multimodal_qwen3")
+        assert isinstance(pipeline, PipelineConfig)
+        stages = merge_pipeline_deploy(pipeline, rocm)
+        assert stages[0].yaml_engine_args["attention_backend"] == "TRITON_ATTN"
+
    def test_qwen3_omni_cuda_uses_thinker_rotary_custom_op(self):
        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
        pipeline = resolve_pipeline_config(
@@ -653,6 +653,7 @@ def test_sub_config_fields_match_structured_scopes():
        "ring_degree",
        "allgather_degree",
        "ulysses_mode",
+        "ulysses_a2a_permute",
        "cfg_parallel_size",
        "vae_patch_parallel_size",
        "vae_parallel_mode",
@@ -858,6 +859,23 @@ def test_from_pipeline_config_derives_sequence_parallel_size_from_allgather_degr
    assert stage.parallel_config.world_size == 2


+def test_from_pipeline_config_preserves_deployed_ulysses_a2a_permute() -> None:
+    pipeline = _resolve_pipeline_or_skip("dreamzero")
+    deploy = DeployConfig(
+        stages=[
+            StageDeployConfig(
+                stage_id=0,
+                ulysses_a2a_permute=True,
+            )
+        ]
+    )
+
+    stage = VllmOmniConfig.from_pipeline_config(pipeline, user_deploy_config=deploy).stage_by_id(0)
+
+    assert isinstance(stage, VllmOmniDiffusionStageConfig)
+    assert stage.parallel_config.ulysses_a2a_permute is True
+
+
 def test_diffusion_parallel_config_accepts_four_way_guidance_parallelism():
    cfg = OmniStageDiffusionParallelConfig(cfg_parallel_size=4)

@@ -25,9 +25,9 @@ pytest_plugins = (
    "tests.helpers.fixtures.clean",
    "tests.helpers.fixtures.log",
    "tests.helpers.fixtures.media",
-    "tests.helpers.fixtures.run_args",
+    "tests.helpers.fixtures.pytest_collection",
+    "tests.helpers.fixtures.pytest_run_args",
    "tests.helpers.fixtures.runtime",
-    "tests.helpers.fixtures.xdist",
 )


@@ -0,0 +1,89 @@
+"""Tests for ``OmniARScheduler._drop_aborted_queued_requests``.
+
+``schedule()`` sweeps ``FINISHED_ABORTED`` requests out of the queues before
+handing control to upstream, because upstream ``Scheduler.schedule()`` raises
+``RuntimeError: Invalid request status: FINISHED_ABORTED`` for any request it
+admits in a finished state -- which kills the stage's engine core, not just
+the request.
+
+Regression for the sweep missing ``skipped_waiting``, the third queue
+upstream admits from. An aborted duplex session parked there was re-selected
+by ``_select_waiting_queue_for_scheduling`` on a later tick and crashed the
+stage.
+"""
+
+from __future__ import annotations
+
+import pytest
+from vllm.v1.core.sched.request_queue import SchedulingPolicy, create_request_queue
+from vllm.v1.request import RequestStatus
+
+import vllm_omni.core.sched.omni_ar_scheduler as ar_sched_mod
+from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+_POLICIES = [
+    pytest.param(SchedulingPolicy.FCFS, id="fcfs"),
+    pytest.param(SchedulingPolicy.PRIORITY, id="priority"),
+]
+
+
+class _StubRequest:
+    """Minimal ``Request`` stub with the surface the sweep exercises.
+
+    Carries the ``(priority, arrival_time)`` ordering ``PriorityRequestQueue``
+    relies on so the same stub works under both scheduling policies.
+    """
+
+    def __init__(self, request_id: str, status: RequestStatus, arrival_time: float = 0.0) -> None:
+        self.request_id = request_id
+        self.status = status
+        self.priority = 0
+        self.arrival_time = arrival_time
+
+    def __lt__(self, other: _StubRequest) -> bool:
+        return (self.priority, self.arrival_time) < (other.priority, other.arrival_time)
+
+
+def _make_scheduler(policy: SchedulingPolicy, *, waiting=(), skipped_waiting=(), running=()):
+    """Bare scheduler with just the surface the sweep reads."""
+    scheduler = OmniARScheduler.__new__(OmniARScheduler)
+    scheduler.waiting = create_request_queue(policy)
+    for req in waiting:
+        scheduler.waiting.add_request(req)
+    scheduler.skipped_waiting = create_request_queue(policy)
+    for req in skipped_waiting:
+        scheduler.skipped_waiting.add_request(req)
+    scheduler.running = list(running)
+    return scheduler
+
+
+@pytest.mark.parametrize("policy", _POLICIES)
+def test_schedule_sweeps_skipped_waiting_before_upstream_selection(
+    monkeypatch: pytest.MonkeyPatch,
+    policy: SchedulingPolicy,
+) -> None:
+    """The public schedule path must remove aborted skipped requests first."""
+    aborted = _StubRequest("req-aborted", RequestStatus.FINISHED_ABORTED)
+    scheduler = _make_scheduler(policy, skipped_waiting=[aborted])
+    scheduler.chunk_transfer_adapter = None
+    scheduler.input_coordinator = None
+    scheduler.num_waiting_for_streaming_input = 0
+    scheduler._process_pending_omni_inputs = lambda model_mode: None
+    scheduler._resync_streaming_input_counter = lambda: None
+    scheduler._should_defer_waiting_admission = lambda: False
+    scheduler._restore_omni_wait_queues = lambda: None
+    scheduler._postprocess_omni_schedule_output = lambda *args, **kwargs: None
+    scheduler.get_finished_requests_needing_kv_transfer = lambda: {}
+    scheduler._wrap_omni_scheduler_output = lambda output, **kwargs: output
+
+    sentinel = object()
+
+    def upstream_schedule(_self, _throttle_prefills: bool = False):
+        assert list(scheduler.skipped_waiting) == []
+        return sentinel
+
+    monkeypatch.setattr(ar_sched_mod.VLLMScheduler, "schedule", upstream_schedule)
+
+    assert OmniARScheduler.schedule(scheduler) is sentinel
@@ -142,13 +142,16 @@ def test_exact_drain_delivers_new_segment_frame() -> None:
    session.num_computed_tokens = 6
    session.num_output_placeholders = 1
    session.num_in_flight_tokens = 1
+    session.async_tokens_to_discard = 1

    _replace_streaming_session(session)
    assert session.num_stale_output_tokens == 1
+    assert session.async_tokens_to_discard == 1

    sched = _make_drain_sched(session)
    assert _run_step(sched, session, num_scheduled=1, token=42) is False  # late frame dropped
    assert session.num_stale_output_tokens == 0
+    assert session.async_tokens_to_discard == 0
    assert _run_step(sched, session, num_scheduled=1, token=43) is True  # new segment survives


@@ -18,7 +18,7 @@ from vllm.sampling_params import SamplingParams
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.outputs import ModelRunnerOutput
 from vllm.v1.request import Request, RequestStatus, StreamingUpdate
-from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
+from vllm_omni.core.sched.omni_ar_scheduler import OmniARAsyncScheduler, OmniARScheduler

 # isort: on

@@ -28,7 +28,9 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 def _make_scheduler(*, stage_id: int = 0, session_mode: str = "turn") -> OmniARScheduler:
    sched = OmniARScheduler.__new__(OmniARScheduler)
    sched._new_prompt_len_snapshot = {}
-    sched.vllm_config = SimpleNamespace(model_config=SimpleNamespace(stage_id=stage_id, session_mode=session_mode))
+    sched.vllm_config = SimpleNamespace(
+        model_config=SimpleNamespace(stage_id=stage_id, session_mode=session_mode),
+    )
    sched.num_waiting_for_streaming_input = 0
    sched.log_stats = False
    sched.chunk_transfer_adapter = None
@@ -185,6 +187,7 @@ def test_resumable_segment_boundary_keeps_pre_transition_send_watermark() -> Non

    def replace_with_next_segment(request: Request) -> bool:
        request.num_computed_tokens = 0
+        request._omni_segment_generation = 1
        request.status = RequestStatus.WAITING
        return False

@@ -201,6 +204,7 @@ def test_resumable_segment_boundary_keeps_pre_transition_send_watermark() -> Non
        True,
        new_token_ids=[42],
        confirmed_num_computed_tokens=26,
+        segment_generation=0,
    )


@@ -286,6 +290,140 @@ def test_queued_streaming_update_on_async_stop_fences_in_flight_once() -> None:
    assert session.num_stale_output_tokens == session.num_in_flight_tokens == 1


+def test_stale_async_frame_is_dropped_before_output_processing() -> None:
+    session = _make_request()
+    session.status = RequestStatus.RUNNING
+    session.num_in_flight_tokens = 2
+    session.num_computed_tokens = session.num_tokens + 1
+    session.num_output_placeholders = 1
+    session.async_tokens_to_discard = 1
+    session.sampling_params = SimpleNamespace(num_logprobs=1)
+    num_computed_tokens = session.num_computed_tokens
+    num_output_placeholders = session.num_output_placeholders
+
+    sched = MagicMock()
+    sched.requests = {session.request_id: session}
+    sched.perf_metrics = None
+    sched.structured_output_manager.should_advance.return_value = False
+
+    def discard_stale_output(request: Request, token_ids: list[int]) -> tuple[list[int], bool]:
+        request.async_tokens_to_discard = 0
+        return token_ids, False
+
+    sched._update_request_with_output.side_effect = discard_stale_output
+    sched._process_kv_transfer_trigger.return_value = False
+    sched.chunk_transfer_adapter = MagicMock()
+    sched.running = [session]
+    sched.waiting_for_transfer_free = set()
+    sched.transfer_triggered_requests = set()
+    sched.active_kv_transfers = set()
+    sched.pending_stop_after_extraction = set()
+    sched.connector = None
+    sched.kv_cache_manager.take_events.return_value = None
+    sched.kv_cache_manager.estimate_cached_tokens.return_value = 0
+    sched.finished_req_ids_dict = {}
+    sched.make_stats.return_value = None
+
+    scheduler_output = MagicMock(spec=SchedulerOutput)
+    scheduler_output.num_scheduled_tokens = {session.request_id: 1}
+    scheduler_output.scheduled_spec_decode_tokens = {}
+    scheduler_output.num_invalid_spec_tokens = 0
+
+    model_runner_output = MagicMock(spec=ModelRunnerOutput)
+    model_runner_output.sampled_token_ids = [[42]]
+    model_runner_output.logprobs = None
+    model_runner_output.prompt_logprobs_dict = {}
+    model_runner_output.pooler_output = None
+    model_runner_output.multimodal_outputs = None
+    model_runner_output.inter_stage_outputs = [{"hidden": object()}]
+    model_runner_output.num_nans_in_logits = None
+    model_runner_output.kv_connector_output = None
+    model_runner_output.cudagraph_stats = None
+    model_runner_output.req_id_to_index = {session.request_id: 0}
+    model_runner_output.routed_experts = None
+
+    OmniARScheduler.update_from_output(sched, scheduler_output, model_runner_output)
+
+    assert session.async_tokens_to_discard == 0
+    assert session.status == RequestStatus.RUNNING
+    assert session.num_computed_tokens == num_computed_tokens
+    assert session.num_output_placeholders == num_output_placeholders
+    sched.chunk_transfer_adapter.save_async.assert_not_called()
+
+    session.sampling_params = SamplingParams(max_tokens=8)
+    next_payload = {"hidden": object()}
+    sched._update_request_with_output.side_effect = None
+    sched._update_request_with_output.return_value = ([43], False)
+    model_runner_output.sampled_token_ids = [[43]]
+    model_runner_output.inter_stage_outputs = [next_payload]
+
+    OmniARScheduler.update_from_output(sched, scheduler_output, model_runner_output)
+
+    sched.chunk_transfer_adapter.save_async.assert_called_once_with(
+        next_payload,
+        session,
+        False,
+        new_token_ids=[43],
+        confirmed_num_computed_tokens=None,
+    )
+
+
+def test_legacy_stale_async_marker_bypasses_real_placeholder_accounting() -> None:
+    request = _make_request()
+    request.status = RequestStatus.RUNNING
+    request.num_in_flight_tokens = 1
+    request.num_output_placeholders = 0
+    request.async_tokens_to_discard = 1
+    request.num_stale_output_tokens = 0
+
+    sched = OmniARAsyncScheduler.__new__(OmniARAsyncScheduler)
+    sched.requests = {request.request_id: request}
+    sched.perf_metrics = None
+    sched.chunk_transfer_adapter = None
+    sched.connector = None
+    sched.ec_connector = None
+    sched.waiting_for_transfer_free = set()
+    sched.transfer_triggered_requests = set()
+    sched.active_kv_transfers = set()
+    sched.pending_stop_after_extraction = set()
+    sched.finished_req_ids_dict = {}
+    sched._new_prompt_len_snapshot = {}
+    sched.kv_cache_manager = MagicMock()
+    sched.kv_cache_manager.take_events.return_value = None
+    sched._remove_stopped_requests_from_queues = MagicMock()
+    sched._handle_failed_kv_load_outputs = MagicMock(return_value=[])
+    sched._cleanup_kv_tracking = MagicMock()
+    sched._aggregate_kv_connector_stats = MagicMock(return_value=None)
+    sched._publish_kv_cache_events = MagicMock()
+    sched._attach_finished_request_sets = MagicMock()
+    sched._attach_scheduler_stats = MagicMock()
+    sched._capture_omni_connector_output = MagicMock()
+
+    scheduler_output = MagicMock(spec=SchedulerOutput)
+    scheduler_output.num_scheduled_tokens = {request.request_id: 1}
+    scheduler_output.scheduled_spec_decode_tokens = {}
+    scheduler_output.num_invalid_spec_tokens = 0
+
+    model_runner_output = MagicMock(spec=ModelRunnerOutput)
+    model_runner_output.sampled_token_ids = [[42]]
+    model_runner_output.logprobs = None
+    model_runner_output.prompt_logprobs_dict = {}
+    model_runner_output.pooler_output = None
+    model_runner_output.multimodal_outputs = None
+    model_runner_output.inter_stage_outputs = None
+    model_runner_output.num_nans_in_logits = None
+    model_runner_output.kv_connector_output = None
+    model_runner_output.cudagraph_stats = None
+    model_runner_output.req_id_to_index = {request.request_id: 0}
+    model_runner_output.routed_experts = None
+
+    OmniARScheduler.update_from_output(sched, scheduler_output, model_runner_output)
+
+    assert request.async_tokens_to_discard == 0
+    assert request.num_output_placeholders == 0
+    assert request._output_token_ids == []
+
+
 def test_stage0_streaming_update_discards_outstanding_async_placeholder_token() -> None:
    sched = _make_scheduler(stage_id=0)
    session = _make_request()
@@ -76,6 +76,9 @@ def _make_scheduler(
        else:
            hf_config.voxcpm2_runtime_config = runtime_config
    sched.vllm_config = SimpleNamespace(model_config=model_config)
+    # Upstream ``Scheduler.__init__`` always creates ``skipped_waiting``; the
+    # abort sweep at the top of ``schedule()`` reads it.
+    sched.skipped_waiting = _MockQueue()
    return sched


@@ -166,3 +166,15 @@ def test_fast_path_schedule_defer_disabled_leaves_fence() -> None:
    OmniGenerationScheduler.schedule(sched)

    assert sched.sched_step_seq == 0
+
+
+def test_generation_schedule_sweeps_and_resyncs_around_pending_inputs() -> None:
+    sched, request = _make_fast_path_sched(defer_block_free=False)
+    calls: list[str] = []
+    sched._drop_aborted_queued_requests.side_effect = lambda: calls.append("sweep")
+    sched._process_pending_omni_inputs.side_effect = lambda model_mode: calls.append(model_mode)
+    sched._resync_streaming_input_counter.side_effect = lambda: calls.append("resync")
+
+    OmniGenerationScheduler.schedule(sched)
+
+    assert calls == ["sweep", "generation", "sweep", "resync"]
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Unit coverage for the two stage-input timeout hooks.

 ``_process_pending_input_timeouts`` covers the full-payload transport;
@@ -253,6 +253,38 @@ def test_log_failed_chunk_sends_noop_without_adapter():
    scheduler._log_failed_chunk_sends()


+class _FakeReceiveFailureAdapter:
+    def __init__(self, failures):
+        self._failures = dict(failures)
+
+    def collect_failed_receive_request_ids(self):
+        failures, self._failures = self._failures, {}
+        return failures
+
+
+class _FakeReceiveFailureScheduler(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_chunk_receive_failures_delegates_to_finish_requests():
+    request_id = "invalid-chunk"
+    adapter = _FakeReceiveFailureAdapter({request_id: "prompt exceeds limit"})
+    scheduler = _FakeReceiveFailureScheduler({request_id: SimpleNamespace(request_id=request_id)}, adapter)
+
+    scheduler._process_chunk_receive_failures()
+
+    assert len(scheduler.finish_calls) == 1
+    finished_ids, status = scheduler.finish_calls[0]
+    assert finished_ids == {request_id}
+    assert getattr(status, "name", str(status)).endswith("FINISHED_ERROR")
+
+
 # --- 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
@@ -0,0 +1,168 @@
+"""Tests for the ``num_waiting_for_streaming_input`` resync placed in
+``OmniARScheduler.finish_requests`` / ``OmniGenerationScheduler.finish_requests``
+and in ``OmniARScheduler.schedule``.
+
+Upstream keeps that counter incrementally (``+1`` when a resumable request
+parks in a waiting queue as ``WAITING_FOR_STREAMING_REQ``, ``-1`` when the
+next update arrives or the request is finished) and
+``Scheduler.get_num_unfinished_requests`` *subtracts* it from the waiting
+queues. Omni rewrites ``request.status`` outside both hooks -- the
+chunk-transfer adapter's park/restore and
+``_realign_request_status_to_queues`` -- so a counted request can leave
+``WAITING_FOR_STREAMING_REQ`` without the matching decrement and inflate the
+counter for the lifetime of the process.
+
+An inflated counter does not degrade throughput, it hangs the stage:
+``EngineCore.has_work()`` reads false while a live request sits in
+``waiting``, so the engine blocks in ``input_queue.get()`` and never calls
+``schedule()`` again. Regression for the Qwen3-Omni duplex symptom where the
+server answered the first conversation after a boot and was silent from the
+second on while still reporting healthy.
+"""
+
+from __future__ import annotations
+
+import pytest
+from vllm.v1.core.sched.request_queue import SchedulingPolicy, create_request_queue
+from vllm.v1.request import RequestStatus
+
+import vllm_omni.core.sched.omni_ar_scheduler as ar_sched_mod
+import vllm_omni.core.sched.omni_generation_scheduler as gen_sched_mod
+from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
+from vllm_omni.core.sched.omni_generation_scheduler import OmniGenerationScheduler
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+class _StubRequest:
+    def __init__(self, request_id: str, status: RequestStatus) -> None:
+        self.request_id = request_id
+        self.status = status
+
+    def is_finished(self) -> bool:
+        return RequestStatus.is_finished(self.status)
+
+
+_SCHEDULER_PARAMS = [
+    pytest.param(OmniARScheduler, ar_sched_mod, id="ar"),
+    pytest.param(OmniGenerationScheduler, gen_sched_mod, id="generation"),
+]
+
+
+def _make_scheduler(scheduler_cls, *, requests, running, waiting, counter, skipped=None):
+    scheduler = scheduler_cls.__new__(scheduler_cls)
+    scheduler.chunk_transfer_adapter = None
+    scheduler.input_coordinator = None
+    scheduler.requests = requests
+    scheduler.running = running
+    scheduler.waiting = waiting
+    scheduler.skipped_waiting = skipped if skipped is not None else []
+    scheduler.num_waiting_for_streaming_input = counter
+    return scheduler
+
+
+@pytest.mark.parametrize(("scheduler_cls", "scheduler_mod"), _SCHEDULER_PARAMS)
+def test_finish_requests_clears_counter_left_by_a_stomped_status(
+    monkeypatch: pytest.MonkeyPatch,
+    scheduler_cls,
+    scheduler_mod,
+) -> None:
+    """The abort case that wedged the talker.
+
+    The request was counted while parked for streaming input, then had its
+    status rewritten to ``RUNNING`` by omni's queue surgery, so upstream's
+    ``finish_requests`` took the running branch and never decremented. With
+    no parked requests left, the counter must land at 0.
+    """
+    aborted = _StubRequest("req-aborted", RequestStatus.RUNNING)
+    scheduler = _make_scheduler(
+        scheduler_cls,
+        requests={},  # upstream already dropped it
+        running=[],
+        waiting=[],
+        counter=1,
+    )
+
+    monkeypatch.setattr(
+        scheduler_mod.VLLMScheduler,
+        "finish_requests",
+        lambda self, request_ids, finished_status: [aborted],
+    )
+
+    scheduler_cls.finish_requests(scheduler, [aborted.request_id], RequestStatus.FINISHED_ABORTED)
+
+    assert scheduler.num_waiting_for_streaming_input == 0
+
+
+def test_resync_counts_parked_requests_in_skipped_waiting() -> None:
+    """``get_num_unfinished_requests`` sums ``waiting`` and ``skipped_waiting``,
+    so the resync has to look at both or it under-counts and reintroduces the
+    spin the counter exists to prevent."""
+    parked = _StubRequest("req-skipped", RequestStatus.WAITING_FOR_STREAMING_REQ)
+    scheduler = _make_scheduler(
+        OmniARScheduler,
+        requests={parked.request_id: parked},
+        running=[],
+        waiting=[],
+        counter=0,
+        skipped=[parked],
+    )
+
+    scheduler._resync_streaming_input_counter()
+
+    assert scheduler.num_waiting_for_streaming_input == 1
+
+
+def test_ar_schedule_resyncs_before_delegating_upstream(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Pins placement in ``schedule``: the chunk-transfer adapter rewrites
+    status during ``process_pending_chunks``, so the resync has to run after
+    that and before upstream schedules, while ``self.waiting`` is still the
+    real queue."""
+    parked = _StubRequest("req-resumed", RequestStatus.RUNNING)
+    scheduler = OmniARScheduler.__new__(OmniARScheduler)
+    scheduler.chunk_transfer_adapter = None
+    scheduler.input_coordinator = None
+    scheduler.requests = {parked.request_id: parked}
+    scheduler.running = [parked]
+    scheduler.waiting = []
+    scheduler.skipped_waiting = []
+    scheduler.num_waiting_for_streaming_input = 1  # leaked by the status rewrite
+    scheduler.policy = "fcfs"
+
+    seen: dict[str, int] = {}
+
+    monkeypatch.setattr(
+        ar_sched_mod.OmniSchedulerMixin, "_consume_pending_connector_output", lambda self, model_mode: None
+    )
+    monkeypatch.setattr(ar_sched_mod.OmniSchedulerMixin, "_process_pending_input_timeouts", lambda self: None)
+    monkeypatch.setattr(OmniARScheduler, "_should_defer_waiting_admission", lambda self: False)
+
+    def fake_schedule(self, throttle_prefills=False):
+        seen["counter"] = self.num_waiting_for_streaming_input
+        raise _StopScheduleError
+
+    class _StopScheduleError(Exception):
+        pass
+
+    monkeypatch.setattr(ar_sched_mod.VLLMScheduler, "schedule", fake_schedule)
+
+    with pytest.raises(_StopScheduleError):
+        OmniARScheduler.schedule(scheduler)
+
+    assert seen["counter"] == 0
+
+
+def test_generation_scheduler_sweep_removes_aborted_requests_from_all_queues() -> None:
+    aborted = _StubRequest("req-generation-aborted", RequestStatus.FINISHED_ABORTED)
+    scheduler = OmniGenerationScheduler.__new__(OmniGenerationScheduler)
+    scheduler.waiting = create_request_queue(SchedulingPolicy.FCFS)
+    scheduler.waiting.add_request(aborted)
+    scheduler.skipped_waiting = create_request_queue(SchedulingPolicy.FCFS)
+    scheduler.skipped_waiting.add_request(aborted)
+    scheduler.running = [aborted]
+
+    scheduler._drop_aborted_queued_requests()
+
+    assert list(scheduler.waiting) == []
+    assert list(scheduler.skipped_waiting) == []
+    assert scheduler.running == []
@@ -1,4 +1,8 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 import json
+import math
 import os
 import threading
 from collections.abc import Callable
@@ -159,6 +163,18 @@ def benchmark_params(request):

 def assert_result(result, params, num_prompt) -> None:
    assert result["completed"] == num_prompt, "Request failures exist"
+    baseline = params.get("baseline")
+    hardware = result.get("Hardware")
+    hardware_baseline = baseline.get(hardware) if isinstance(baseline, dict) and isinstance(hardware, str) else None
+    if isinstance(hardware_baseline, dict) and "mean_tpot_ms" in hardware_baseline:
+        num_tpot_samples = result.get("num_tpot_samples")
+        mean_tpot_ms = result.get("mean_tpot_ms")
+        assert isinstance(num_tpot_samples, int) and not isinstance(num_tpot_samples, bool) and num_tpot_samples > 0, (
+            "TPOT baseline is configured, but no measurable TPOT samples were produced"
+        )
+        assert (
+            isinstance(mean_tpot_ms, int | float) and not isinstance(mean_tpot_ms, bool) and math.isfinite(mean_tpot_ms)
+        ), "TPOT baseline is configured, but mean_tpot_ms is not finite"
    expected_audio_turns = params.get("expected_duplex_audio_turns_per_session")
    if expected_audio_turns is not None:
        session_metrics = result.get("duplex_session_metrics")
@@ -43,7 +43,7 @@
        "extra_body": {
          "save_duplex_request_metrics": true
        },
-        "percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration",
+        "percentile-metrics": "ttft,tpot,itl,e2el,audio_ttfp,audio_rtf,audio_duration",
        "baseline": {
          "H100": {
            "request_throughput": [
@@ -46,9 +46,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
-            "mean_e2el_ms": [
-              3646.797
-            ],
            "mean_audio_rtf": [
              0.1098
            ]
@@ -82,9 +79,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
-            "mean_e2el_ms": [
-              4059.0878
-            ],
            "mean_audio_rtf": [
              0.1335
            ]
@@ -120,9 +114,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
-            "mean_e2el_ms": [
-              5639.5769
-            ],
            "mean_audio_rtf": [
              0.1798
            ]
@@ -45,13 +45,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
-            "mean_e2el_ms": [
-              15793.1209,
-              30658.6632,
-              37463.4943,
-              47990.5839,
-              64980.2199
-            ],
            "mean_audio_rtf": [
              0.169,
              0.2041,
@@ -87,9 +80,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
-            "mean_e2el_ms": [
-              4556.4347
-            ],
            "mean_audio_rtf": [
              0.138
            ]
@@ -123,9 +113,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
-            "mean_e2el_ms": [
-              4768.0972
-            ],
            "mean_audio_rtf": [
              0.1609
            ]
@@ -161,9 +148,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
-            "mean_e2el_ms": [
-              6395.3517
-            ],
            "mean_audio_rtf": [
              0.2048
            ]
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """Tests for DFX runner metadata field exclusion."""

 import json
@@ -105,7 +108,7 @@ def test_resolve_pytest_marks_hardware_dict_with_extra():
    assert "H100" in names
    assert "cuda" in names
    assert "gpu" in names
-    assert "distributed_cuda" in names
+    assert "cards_2" in names
    assert "full_model" in names
    assert "diffusion" in names
    assert "local_model" in names
@@ -512,3 +515,41 @@ def test_omni_duplex_expected_audio_turns_rejects_incomplete_session():
            {"expected_duplex_audio_turns_per_session": 4},
            1,
        )
+
+
+def test_omni_tpot_baseline_accepts_measured_finite_sample():
+    from tests.dfx.perf.scripts.run_benchmark import assert_result
+
+    assert_result(
+        {
+            "completed": 1,
+            "Hardware": "H100",
+            "num_tpot_samples": 1,
+            "mean_tpot_ms": 10.0,
+        },
+        {"baseline": {"H100": {"mean_tpot_ms": 20.0}}},
+        1,
+    )
+
+
+@pytest.mark.parametrize(
+    ("num_tpot_samples", "mean_tpot_ms", "match"),
+    [
+        (0, float("nan"), "no measurable TPOT samples"),
+        (1, float("nan"), "mean_tpot_ms is not finite"),
+    ],
+)
+def test_omni_tpot_baseline_rejects_missing_or_nonfinite_sample(num_tpot_samples, mean_tpot_ms, match):
+    from tests.dfx.perf.scripts.run_benchmark import assert_result
+
+    with pytest.raises(AssertionError, match=match):
+        assert_result(
+            {
+                "completed": 1,
+                "Hardware": "H100",
+                "num_tpot_samples": num_tpot_samples,
+                "mean_tpot_ms": mean_tpot_ms,
+            },
+            {"baseline": {"H100": {"mean_tpot_ms": 20.0}}},
+            1,
+        )
@@ -101,7 +101,12 @@ class TestAttentionSpec:
    def test_block_sparse_defaults_applied_when_backend_selected(self):
        spec = AttentionSpec(backend="RAINFUSION_ATTN")
        assert spec.block_sparse.sparsity == 0.8
-        assert spec.backend_kwargs() == {"sparsity": 0.8, "start_step": 0}
+        assert spec.backend_kwargs() == {
+            "sparsity": 0.8,
+            "start_step": 0,
+            "end_step": 0,
+            "precision": "bf16",
+        }

    def test_block_sparse_skip_layers_selector_expanded(self):
        spec = AttentionSpec(
@@ -112,6 +117,8 @@ class TestAttentionSpec:
        assert spec.backend_kwargs() == {
            "sparsity": 0.9,
            "start_step": 12,
+            "end_step": 0,
+            "precision": "bf16",
            "skip_layers": [0, 1, 2, 38],
        }

@@ -121,12 +128,44 @@ class TestAttentionSpec:

    @pytest.mark.parametrize(
        "block_sparse",
-        [{"sparsity": 1.5}, {"start_step": -1}],
+        [{"sparsity": 1.5}, {"start_step": -1}, {"end_step": -1}, {"precision": "bogus"}],
    )
    def test_block_sparse_invalid_values(self, block_sparse):
        with pytest.raises(ValueError):
            AttentionSpec(backend="RAINFUSION_ATTN", block_sparse=block_sparse)

+    def test_block_sparse_end_step_precision_forwarded(self):
+        spec = AttentionSpec(
+            backend="RAINFUSION_ATTN",
+            block_sparse={"sparsity": 0.9, "start_step": 5, "end_step": 3, "precision": "fp8"},
+        )
+        assert spec.backend_kwargs() == {
+            "sparsity": 0.9,
+            "start_step": 5,
+            "end_step": 3,
+            "precision": "fp8",
+        }
+
+    def test_block_sparse_precision_enum_input_normalized(self):
+        """RainFusionPrecision enum members must normalize to their string value.
+
+        Regression: the str,Enum members passed the value check but str() later
+        produced "RainFusionPrecision.MIX" instead of "mix" (StrEnum fixes it).
+        """
+        from vllm_omni.diffusion.attention.backends.rainfusion_attn import RainFusionConfig
+        from vllm_omni.diffusion.data import RainFusionPrecision
+
+        spec = AttentionSpec(
+            backend="RAINFUSION_ATTN",
+            block_sparse={"sparsity": 0.8, "precision": RainFusionPrecision.MIX},
+        )
+        kwargs = spec.backend_kwargs()
+        assert kwargs["precision"] == "mix"
+        # The config path is where str(value) used to mangle enum members.
+        cfg = RainFusionConfig.from_backend_kwargs(kwargs)
+        assert cfg.precision == "mix"
+        assert type(cfg.precision) is str
+

 class TestAttentionConfig:
    def test_empty_config(self):
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """Geometry contract RAINFUSION_ATTN enforces before handing a forward to rf_v2.

@@ -11,18 +11,25 @@ to the always-kept prefix segment before generating the block mask.
 """

 import dataclasses
+import sys
+import types
+from unittest import mock

 import pytest
 import torch

-from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata, VideoTokenLayout
+from vllm_omni.diffusion.attention.backends import rainfusion_attn
+from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata, VideoTokenLayout, VideoTokenSpan
 from vllm_omni.diffusion.attention.backends.rainfusion_attn import (
    _BLOCK_SIZE,
+    RainFusionAttentionBackend,
    RainFusionAttentionImpl,
    RainFusionPlan,
 )
 from vllm_omni.platforms import current_omni_platform

+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.npu]
+
 PREFIX_ROWS = 710  # 14 text rows + 696 audio rows
 ALIGNED_GRID = (62, 24, 40)  # 1280x768 -> 59520 video rows, 465 blocks
 MISALIGNED_GRID = (62, 24, 42)  # 1344x768 -> 62496 video rows, 488 blocks + 32 rows
@@ -97,6 +104,62 @@ def test_video_segment_must_be_the_tail_of_packed_document_zero():
    assert make_impl()._resolve_plan(metadata) is None


+def test_ref2va_multi_video_spans_resolve():
+    layout = VideoTokenLayout(
+        used_len=12000,
+        video_spans=(
+            VideoTokenSpan(start=128, latent_grid=(4, 16, 64), role="reference"),
+            VideoTokenSpan(start=5000, latent_grid=(4, 16, 64), role="target"),
+        ),
+    )
+    plan = make_impl()._resolve_plan(AttentionMetadata(extra={"max_seqlen_q": 12000}, video_layout=layout))
+
+    assert plan is not None
+    assert plan.used_len == 12000
+    assert plan.video_spans == [
+        {"start": 128, "latent_shape": [4, 16, 64]},
+        {"start": 5000, "latent_shape": [4, 16, 64]},
+    ]
+
+
+def test_invalid_ref2va_video_spans_fall_back_to_dense():
+    layout = VideoTokenLayout(
+        used_len=12000,
+        video_spans=(
+            VideoTokenSpan(start=128, latent_grid=(4, 16, 64), role="reference"),
+            VideoTokenSpan(start=4000, latent_grid=(4, 16, 64), role="target"),
+        ),
+    )
+    assert make_impl()._resolve_plan(AttentionMetadata(extra={"max_seqlen_q": 12000}, video_layout=layout)) is None
+
+
+def test_validate_available_rejects_legacy_mindiesd(monkeypatch):
+    mindiesd = types.ModuleType("mindiesd")
+
+    def sparse_attention(query, key, value, **kwargs):
+        return query
+
+    mindiesd.sparse_attention = sparse_attention
+    monkeypatch.setitem(sys.modules, "mindiesd", mindiesd)
+    monkeypatch.setattr("importlib.util.find_spec", lambda _: object())
+
+    with pytest.raises(ValueError, match="video_spans"):
+        RainFusionAttentionBackend.validate_available()
+
+
+def test_validate_available_accepts_new_mindiesd(monkeypatch):
+    mindiesd = types.ModuleType("mindiesd")
+
+    def sparse_attention(query, key, value, *, video_spans=None, **kwargs):
+        return query
+
+    mindiesd.sparse_attention = sparse_attention
+    monkeypatch.setitem(sys.modules, "mindiesd", mindiesd)
+    monkeypatch.setattr("importlib.util.find_spec", lambda _: object())
+
+    RainFusionAttentionBackend.validate_available()
+
+
@pytest.mark.parametrize("grid", [(4, 24, 40), (1, 24, 40)])
 def test_short_video_stays_dense(grid):
    assert make_impl()._resolve_plan(make_metadata(grid)) is None
@@ -169,7 +232,9 @@ def test_irregular_video_tail_reproduces_dense_attention_with_updated_mindiesd()
    heads, head_dim = 4, 128
    q, k, v = (torch.randn(1, used, heads, head_dim, dtype=torch.bfloat16, device="npu") for _ in range(3))

-    impl = make_impl()
+    # Explicit bf16: the dense comparison exercises the plain sparse kernel,
+    # and older MindIE-SD releases do not accept precision= (gate would raise).
+    impl = make_impl(precision="bf16")
    impl.rainfusion = dataclasses.replace(impl.rainfusion, sparsity=0.0)
    plan = RainFusionPlan(prefix_len=prefix_len, used_len=used, latent_shape=list(grid))
    out = impl._forward_sparse_npu(q, k, v, plan)
@@ -182,3 +247,90 @@ def test_irregular_video_tail_reproduces_dense_attention_with_updated_mindiesd()
    ).transpose(1, 2)
    error = (out.float() - reference.float()).abs().mean() / reference.float().abs().mean()
    assert error < 2e-3, f"mean relative error {error:.4%} against dense attention"
+
+
+# end_step tail fallback: the last ``end_step`` denoise steps must stay dense.
+
+
+def test_end_step_tail_window_stays_dense():
+    """step_idx inside the last end_step window must resolve no sparse plan."""
+    impl = make_impl(end_step=3)
+    fc = mock.Mock()
+    fc.denoise_step_idx = 47
+    fc.total_denoise_steps = 50
+    with (
+        mock.patch.object(rainfusion_attn, "is_forward_context_available", return_value=True),
+        mock.patch.object(rainfusion_attn, "get_forward_context", return_value=fc),
+    ):
+        assert impl._resolve_plan(make_metadata(ALIGNED_GRID)) is None
+
+
+def test_end_step_before_tail_window_still_sparse():
+    """step_idx before the tail window must still resolve a sparse plan."""
+    impl = make_impl(end_step=3)
+    fc = mock.Mock()
+    fc.denoise_step_idx = 46
+    fc.total_denoise_steps = 50
+    with (
+        mock.patch.object(rainfusion_attn, "is_forward_context_available", return_value=True),
+        mock.patch.object(rainfusion_attn, "get_forward_context", return_value=fc),
+    ):
+        assert impl._resolve_plan(make_metadata(ALIGNED_GRID)) is not None
+
+
+def test_end_step_zero_never_triggers_tail_fallback():
+    """end_step=0 must not fall back even on the final denoise step."""
+    impl = make_impl(end_step=0)
+    fc = mock.Mock()
+    fc.denoise_step_idx = 49
+    fc.total_denoise_steps = 50
+    with (
+        mock.patch.object(rainfusion_attn, "is_forward_context_available", return_value=True),
+        mock.patch.object(rainfusion_attn, "get_forward_context", return_value=fc),
+    ):
+        assert impl._resolve_plan(make_metadata(ALIGNED_GRID)) is not None
+
+
+# precision capability gate: a mindiesd without sparse_attention(precision=)
+# must fail fast instead of silently running the BF16 path.
+
+
+def _fake_mindiesd_module():
+    """Install a minimal stand-in so the import inside _forward_sparse_npu succeeds."""
+    import types
+
+    fake = types.ModuleType("mindiesd")
+    fake.sparse_attention = lambda *args, **kwargs: None
+    return fake
+
+
+def test_precision_non_bf16_requires_mindiesd_support():
+    """precision != bf16 against a mindiesd lacking the kwarg must raise RuntimeError."""
+    import sys
+
+    impl = make_impl(precision="mix")
+    sys.modules["mindiesd"] = _fake_mindiesd_module()
+    try:
+        with mock.patch.object(rainfusion_attn, "_mindiesd_supports_precision", return_value=False):
+            with pytest.raises(RuntimeError, match="requires MindIE-SD"):
+                impl._forward_sparse_npu(None, None, None, None)
+    finally:
+        sys.modules.pop("mindiesd", None)
+
+
+def test_precision_non_bf16_passes_gate_when_supported():
+    """precision != bf16 with a capable mindiesd must not raise from the gate."""
+    import sys
+
+    impl = make_impl(precision="mix")
+    sys.modules["mindiesd"] = _fake_mindiesd_module()
+    try:
+        with mock.patch.object(rainfusion_attn, "_mindiesd_supports_precision", return_value=True):
+            # q/k/v shapes: [B, S, N, D]; plan geometry must match S.
+            q = torch.randn(1, 8, 4, 128)
+            plan = RainFusionPlan(prefix_len=0, used_len=8, latent_shape=[2, 2, 2])
+            # The gate must pass; the fake mindiesd returns None so no crash.
+            out = impl._forward_sparse_npu(q, q, q, plan)
+            assert out is None
+    finally:
+        sys.modules.pop("mindiesd", None)
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """
 Unit tests for TeaCache extractor functions.
@@ -18,6 +18,7 @@ Currently implemented:
 - TestMiniMaxH3Extractor: MiniMaxH3DiTModel extractor
 """

+import math
 from abc import ABC, abstractmethod
 from types import SimpleNamespace
 from unittest.mock import MagicMock, Mock, patch
@@ -650,6 +651,29 @@ class TestMiniMaxH3Extractor(BaseExtractorTest):
        torch.testing.assert_close(actual_video, expected_video)
        torch.testing.assert_close(actual_audio, expected_audio)

+    def test_extractor_uses_prepared_rope_table(
+        self,
+        minimax_h3_module,
+        sample_inputs,
+        monkeypatch,
+    ):
+        """TeaCache must consume the denoise-loop RoPE table without rebuilding it."""
+        rope_table = minimax_h3_module.prepare_rope_table(
+            sample_inputs["img_position_ids"],
+            seq_len=sample_inputs["x"].shape[1],
+        )
+
+        def unexpected_rope_build(*args, **kwargs):
+            raise AssertionError("TeaCache extractor rebuilt the prepared RoPE table")
+
+        monkeypatch.setattr(minimax_h3_module, "prepare_rope_table", unexpected_rope_build)
+        monkeypatch.setattr(minimax_h3_module.rope, "forward", unexpected_rope_build)
+        context = extract_minimax_h3_context(
+            minimax_h3_module,
+            **{**sample_inputs, "rope_table": rope_table},
+        )
+        context.run_transformer_blocks()
+
    def test_run_transformer_blocks_forwards_packed_layout(
        self,
        minimax_h3_module,
@@ -705,6 +729,222 @@ class TestMiniMaxH3Extractor(BaseExtractorTest):
        assert video_logits.shape[0] == sample_inputs["img_pos_info"]["position_ids"].shape[0]
        assert audio_logits.shape[0] == sample_inputs["audio_pos_info"]["position_ids"].shape[0]

+    @pytest.mark.sp
+    def test_teacache_strict_sp_keeps_cache_state_rank_local(
+        self,
+        minimax_h3_module,
+        sample_inputs,
+        monkeypatch,
+    ):
+        """Strict SP must not mix local TeaCache state with gathered block rows."""
+        from vllm_omni.diffusion.attention.ops import minimax_h3_modulation
+        from vllm_omni.diffusion.cache.teacache.config import TeaCacheConfig
+        from vllm_omni.diffusion.cache.teacache.hook import TeaCacheHook
+        from vllm_omni.diffusion.distributed import parallel_state
+
+        seq_len = sample_inputs["x"].shape[1]
+        local_len = seq_len // 2
+        hidden_size = minimax_h3_module.hidden_size
+        video_width = minimax_h3_module.arch.latents_dim * math.prod(minimax_h3_module.arch.patch_size)
+        audio_width = minimax_h3_module.arch.audio_latents_dim
+        captured_modulation_indices: list[torch.Tensor] = []
+
+        class _StrictSPPrepare(nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.global_indices: list[torch.Tensor] = []
+
+            def forward(self, hidden, rope_table, combined_indices):
+                assert hidden.shape[0] == local_len
+                assert rope_table.shape[0] == local_len
+                self.global_indices.append(combined_indices.clone())
+                return hidden, rope_table, combined_indices.narrow(0, 0, local_len)
+
+        class _StrictSPGather(nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.inputs: list[torch.Tensor] = []
+                self.outputs: list[torch.Tensor] = []
+
+            def forward(self, tensor):
+                self.inputs.append(tensor.clone())
+                if tensor.shape[-1] == hidden_size:
+                    raise AssertionError("TeaCache must not gather hidden states in strict SP")
+                assert tensor.shape == (local_len, video_width + audio_width)
+                # Model the all-gather of compact logits from every SP rank.
+                # postprocess() indexes the resulting global layout with
+                # infer_out_pos/audio_pos, including rows owned by another rank.
+                gathered = tensor.repeat(seq_len // local_len, 1)
+                assert gathered.shape == (seq_len, video_width + audio_width)
+                self.outputs.append(gathered)
+                return gathered
+
+        class _StrictSPGroup:
+            def __init__(self):
+                super().__init__()
+                self.local_decisions: list[int] = []
+                self.ops = []
+
+            def all_reduce(self, decision, op):
+                self.local_decisions.append(int(decision.item()))
+                self.ops.append(op)
+                # Simulate another rank requiring the second step to execute
+                # the collective block path even though this rank would cache.
+                if len(self.local_decisions) == 2:
+                    decision.fill_(1)
+                return decision
+
+        class _StrictSPBlock(nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.block_indices: list[torch.Tensor] = []
+
+            def adaln_proj(self, t_emb):
+                del t_emb
+                zeros = torch.zeros(3, hidden_size, dtype=torch.float32)
+                return zeros, zeros
+
+            def norm1(self, hidden):
+                return hidden.clone()
+
+            def forward(self, hidden, *, combined_indices, rope_table, **kwargs):
+                del kwargs
+                assert hidden.shape[0] == local_len
+                assert rope_table.shape[0] == local_len
+                self.block_indices.append(combined_indices.clone())
+                return hidden + 1
+
+        class _StrictSPFinalLayer(nn.Module):
+            def forward(self, hidden, *, t_emb, inverse_indices):
+                del t_emb, inverse_indices
+                assert hidden.shape == (local_len, hidden_size)
+                return hidden[:, :video_width].float(), hidden[:, :audio_width].float()
+
+        local_prepare = _StrictSPPrepare()
+        gather = _StrictSPGather()
+        sp_group = _StrictSPGroup()
+        block = _StrictSPBlock()
+        minimax_h3_module.local_sp_prepare = local_prepare
+        minimax_h3_module.sp_gather = gather
+        minimax_h3_module.blocks = nn.ModuleList([block])
+        minimax_h3_module.final_layer = _StrictSPFinalLayer()
+        monkeypatch.setattr(minimax_h3_module, "_rope_local_span", lambda _seq_len: (0, local_len))
+        monkeypatch.setattr(
+            minimax_h3_module,
+            "_embed",
+            lambda **kwargs: (
+                torch.arange(local_len * hidden_size, dtype=torch.float32).reshape(local_len, hidden_size),
+                torch.zeros(3, hidden_size, dtype=torch.float32),
+            ),
+        )
+        monkeypatch.setattr(
+            minimax_h3_module,
+            "prepare_rope_table",
+            lambda *args, **kwargs: torch.zeros(local_len, 6, dtype=torch.float32),
+        )
+        monkeypatch.setattr(parallel_state, "get_sp_group", lambda: sp_group)
+
+        original_indexed_scale_shift = minimax_h3_modulation.indexed_scale_shift_
+
+        def capture_indexed_scale_shift(hidden, shift, scale, indices):
+            captured_modulation_indices.append(indices.clone())
+            return original_indexed_scale_shift(hidden, shift, scale, indices)
+
+        monkeypatch.setattr(minimax_h3_modulation, "indexed_scale_shift_", capture_indexed_scale_shift)
+
+        local_img_pos = torch.tensor([0, 1])
+        local_audio_pos = torch.tensor([2, 3])
+        token_tags = torch.full((seq_len,), -1, dtype=torch.long)
+        token_tags[local_img_pos] = 0
+        token_tags[local_audio_pos] = 2
+        inverse_indices = torch.zeros(seq_len, dtype=torch.long)
+        inputs = {
+            **sample_inputs,
+            "inverse_indices": inverse_indices,
+            "token_tags": token_tags,
+            "update_mask": torch.ones(local_img_pos.numel(), dtype=torch.bool),
+            "prompt_embeds": torch.empty(0, 6),
+            "img_pos_info": {"position_ids": local_img_pos},
+            "audio_pos_info": {"position_ids": local_audio_pos},
+            "text_pos_info": {"position_ids": torch.empty(0, dtype=torch.long)},
+            "img_pos_for_infer_output_info": {"position_ids": local_img_pos},
+            "refiner_packed_seq_params": {
+                "cu_seqlens_q": torch.tensor([0, 0], dtype=torch.int32),
+                "max_seqlen_q": 0,
+            },
+        }
+        expected_local_indices = (inverse_indices * 3 + token_tags.clamp(min=0))[:local_len]
+
+        hook = TeaCacheHook(
+            TeaCacheConfig(
+                transformer_type="MiniMaxH3DiTModel",
+                rel_l1_thresh=0.3,
+            )
+        )
+        hook.initialize_hook(minimax_h3_module)
+
+        hook.new_forward(minimax_h3_module, **inputs)
+        hook.new_forward(minimax_h3_module, **inputs)
+        video_logits, audio_logits = hook.new_forward(minimax_h3_module, **inputs)
+        state = hook.state_manager.get_state()
+
+        assert state.previous_residual is not None
+        assert state.previous_residual.shape == (local_len, hidden_size)
+        assert len(captured_modulation_indices) == 3
+        assert all(torch.equal(indices, expected_local_indices) for indices in captured_modulation_indices)
+        assert sp_group.local_decisions == [1, 0, 0]
+        assert len(sp_group.ops) == 3
+        assert all(op == torch.distributed.ReduceOp.MAX for op in sp_group.ops)
+        assert len(local_prepare.global_indices) == 2
+        assert all(indices.shape == (seq_len,) for indices in local_prepare.global_indices)
+        assert len(block.block_indices) == 2
+        assert all(torch.equal(indices, expected_local_indices) for indices in block.block_indices)
+        assert len(gather.inputs) == 3
+        assert video_logits.shape == (local_img_pos.numel(), video_width)
+        assert audio_logits.shape == (local_audio_pos.numel(), audio_width)
+
+    @pytest.mark.cpu
+    def test_teacache_non_strict_sp_gathers_hidden_once(
+        self,
+        minimax_h3_module,
+        sample_inputs,
+        monkeypatch,
+    ):
+        """The non-strict SP cache path must not gather global hidden rows twice."""
+        from vllm_omni.diffusion.cache.teacache.config import TeaCacheConfig
+        from vllm_omni.diffusion.cache.teacache.hook import TeaCacheHook
+
+        class _Prepare(nn.Module):
+            def forward(self, hidden, rope_table, combined_indices):
+                return hidden, rope_table, combined_indices
+
+        class _Gather(nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.inputs: list[torch.Tensor] = []
+
+            def forward(self, tensor):
+                self.inputs.append(tensor.clone())
+                return tensor
+
+        seq_len = sample_inputs["x"].shape[1]
+        gather = _Gather()
+        minimax_h3_module.sp_prepare = _Prepare()
+        minimax_h3_module.sp_gather = gather
+        monkeypatch.setattr(minimax_h3_module, "_rope_local_span", lambda _seq_len: (0, seq_len))
+
+        hook = TeaCacheHook(
+            TeaCacheConfig(
+                transformer_type="MiniMaxH3DiTModel",
+                rel_l1_thresh=0.3,
+            )
+        )
+        hook.initialize_hook(minimax_h3_module)
+        hook.new_forward(minimax_h3_module, **sample_inputs)
+
+        assert len(gather.inputs) == 1
+        assert gather.inputs[0].shape == (seq_len, minimax_h3_module.hidden_size)
+
    def test_invalid_module_raises_error(self):
        """Test that invalid module without blocks raises ValueError."""
        invalid_module = Mock()
@@ -9,6 +9,7 @@ import torch
 from diffusers import DiffusionPipeline  # pyright: ignore[reportPrivateImportUsage]
 from PIL import Image

+import vllm_omni.diffusion.data as diffusion_data
 from tests.helpers.mark import hardware_test
 from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient, dummy_messages_from_mix_data
 from vllm_omni.diffusion.data import (
@@ -629,6 +630,46 @@ class TestPipelineArgumentsHandling:
        with pytest.raises(ValueError):
            pipeline.forward(DiffusionRequestBatch(requests=[problematic_request]))

+        interpolation_request = _make_request(
+            sampling_params=OmniDiffusionSamplingParams(
+                enable_frame_interpolation=True,
+            ),
+        )
+        with pytest.raises(ValueError, match="Frame interpolation is not supported with the Diffusers backend"):
+            pipeline.forward(DiffusionRequestBatch(requests=[interpolation_request]))
+
+    def test_diffusers_config_preserves_checkpoint_class_for_metadata(self, mocker):
+        class MockWanImageToVideoPipeline:
+            pass
+
+        od_config = _make_od_config(model_class_name=None)
+        mocker.patch(
+            "vllm_omni.diffusion.utils.hf_utils.get_diffusion_model_index",
+            return_value={"_class_name": "WanImageToVideoPipeline"},
+        )
+        mocker.patch.object(
+            diffusion_data,
+            "diffusers",
+            SimpleNamespace(WanImageToVideoPipeline=MockWanImageToVideoPipeline),
+        )
+        mocker.patch.object(OmniDiffusionConfig, "update_multimodal_support")
+
+        od_config.enrich_config()
+
+        assert od_config.model_class_name == "WanImageToVideoPipeline"
+        assert od_config.diffusers_pipeline_cls is MockWanImageToVideoPipeline
+
+    def test_diffusers_config_falls_back_to_adapter_without_pipeline_index(self, mocker):
+        od_config = _make_od_config(model_class_name=None)
+        mocker.patch(
+            "vllm_omni.diffusion.utils.hf_utils.get_diffusion_model_index",
+            return_value=None,
+        )
+
+        od_config.enrich_config()
+
+        assert od_config.model_class_name == "DiffusersAdapterPipeline"
+

@pytest.mark.advanced_model
@hardware_test(res={"cuda": "L4"}, num_cards=1)
@@ -0,0 +1,117 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from types import SimpleNamespace
+
+import pytest
+import torch
+
+from vllm_omni.diffusion.distributed import a2a_permute
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+def test_jit_build_includes_cuda_headers_from_nvidia_wheels(tmp_path, monkeypatch) -> None:
+    cu13_include = tmp_path / "nvidia" / "cu13" / "include"
+    nccl_include = tmp_path / "nvidia" / "nccl" / "include"
+    nccl_lib = tmp_path / "nvidia" / "nccl" / "lib" / "libnccl.so.2"
+    cu13_include.mkdir(parents=True)
+    nccl_include.mkdir(parents=True)
+    nccl_lib.parent.mkdir(parents=True)
+    (cu13_include / "cusparse.h").touch()
+    (nccl_include / "nccl.h").touch()
+    nccl_lib.touch()
+
+    load_kwargs = {}
+    monkeypatch.setattr(a2a_permute.sysconfig, "get_paths", lambda: {"purelib": str(tmp_path)})
+    monkeypatch.setattr(
+        torch.utils.cpp_extension,
+        "load",
+        lambda **kwargs: load_kwargs.update(kwargs),
+    )
+    monkeypatch.setattr(a2a_permute.symm_mem, "set_backend", lambda _backend: None)
+    monkeypatch.setattr(a2a_permute, "_BUILT", False)
+
+    a2a_permute.ensure_a2a_permute_available()
+
+    assert set(load_kwargs["extra_include_paths"]) == {str(cu13_include), str(nccl_include)}
+    assert load_kwargs["extra_ldflags"] == [str(nccl_lib)]
+
+
+@dataclass
+class _FakeAllocation:
+    size: int
+    device: torch.device
+
+
+class _FakeHandle:
+    rank = 0
+
+    def __init__(self, allocation: _FakeAllocation) -> None:
+        self.allocation = allocation
+        self.views: list[tuple[tuple[int, ...], torch.dtype]] = []
+
+    def get_buffer(self, rank, shape, dtype):
+        assert rank == self.rank
+        self.views.append((shape, dtype))
+        return self.allocation, shape, dtype
+
+
+def test_workspace_reuses_peak_capacity_across_shapes(monkeypatch) -> None:
+    allocations: list[_FakeAllocation] = []
+    all_reduce_calls: list[object] = []
+    synchronize_calls: list[torch.device] = []
+    stream_id = [1]
+    a2a_permute._SYMM_WORKSPACES.clear()
+
+    monkeypatch.setattr(a2a_permute.torch.cuda, "is_current_stream_capturing", lambda: False)
+    monkeypatch.setattr(
+        a2a_permute.torch.cuda,
+        "current_stream",
+        lambda _device: SimpleNamespace(cuda_stream=stream_id[0]),
+    )
+    monkeypatch.setattr(a2a_permute, "_resolve_process_group", lambda _name: object())
+    monkeypatch.setattr(a2a_permute.torch, "ones", lambda *_args, **_kwargs: object())
+    monkeypatch.setattr(
+        a2a_permute.dist,
+        "all_reduce",
+        lambda _tensor, group: all_reduce_calls.append(group),
+    )
+    monkeypatch.setattr(
+        a2a_permute.torch.accelerator,
+        "synchronize",
+        lambda device: synchronize_calls.append(device),
+    )
+
+    def empty(size, *, dtype, device):
+        assert dtype == torch.uint8
+        allocation = _FakeAllocation(size=size, device=torch.device(device))
+        allocations.append(allocation)
+        return allocation
+
+    monkeypatch.setattr(a2a_permute.symm_mem, "empty", empty)
+    monkeypatch.setattr(a2a_permute.symm_mem, "rendezvous", lambda allocation, _group: _FakeHandle(allocation))
+
+    device = torch.device("cuda:0")
+    first = a2a_permute._get_symm_buffer((2, 3), torch.float16, device, "group")
+    smaller = a2a_permute._get_symm_buffer((1, 4), torch.float16, device, "group")
+    larger = a2a_permute._get_symm_buffer((4, 3), torch.float16, device, "group")
+
+    assert first[0].size == 12
+    assert smaller[0] is first[0]
+    assert larger[0].size == 24
+    assert [allocation.size for allocation in allocations] == [12, 24]
+    assert len(a2a_permute._SYMM_WORKSPACES) == 1
+    assert len(all_reduce_calls) == 1
+    assert synchronize_calls == [device, device]
+
+    stream_id[0] = 2
+    with pytest.raises(RuntimeError, match="single CUDA stream"):
+        a2a_permute._get_symm_buffer((1, 4), torch.float16, device, "group")
+
+    a2a_permute.clear_a2a_permute_workspaces()
+    assert not a2a_permute._SYMM_WORKSPACES
+    assert synchronize_calls == [device, device, device]
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 from types import SimpleNamespace

@@ -194,6 +194,47 @@ def test_moe_ep_maps_diffusion_sp_cfg_dp_to_vllm_groups(monkeypatch):
    assert ep_groups == [parallel_state.vllm_parallel_state._EP.local_group]


+@pytest.mark.cpu
+@pytest.mark.core_model
+@pytest.mark.parametrize(
+    ("fully_shard_degree", "error_message"),
+    [(0, "fully_shard_degree must be positive"), (3, "must be divisible by fully_shard_degree")],
+)
+def test_invalid_hsdp_shard_size_fails_before_group_creation(monkeypatch, fully_shard_degree: int, error_message: str):
+    """Reject invalid HSDP shard sizes before allocating any process groups."""
+    world_size = 4
+    created_groups: list[object] = []
+
+    def fail_if_group_created(*args, **kwargs):
+        del args, kwargs
+        created_groups.append(object())
+        raise AssertionError("a process group was created before HSDP validation")
+
+    fake_world_group = SimpleNamespace(
+        rank_in_group=0,
+        local_rank=0,
+        device_group=object(),
+    )
+    monkeypatch.setattr(parallel_state.torch.distributed, "is_initialized", lambda: True)
+    monkeypatch.setattr(parallel_state.torch.distributed, "get_world_size", lambda: world_size)
+    monkeypatch.setattr(parallel_state, "get_world_group", lambda: fake_world_group)
+    monkeypatch.setattr(parallel_state, "init_model_parallel_group", fail_if_group_created)
+
+    for name in ("_DP", "_CFG", "_SP", "_PP", "_FS", "_HSDP_REPLICATE", "_EXPERT_PARALLEL_GROUP_RANKS"):
+        monkeypatch.setattr(parallel_state, name, None)
+    for name in ("_TP", "_PCP", "_DP", "_EP", "_PP"):
+        monkeypatch.setattr(parallel_state.vllm_parallel_state, name, None, raising=False)
+
+    with pytest.raises(ValueError, match=error_message):
+        parallel_state.initialize_model_parallel(
+            fully_shard_degree=fully_shard_degree,
+            use_hsdp=True,
+            backend="gloo",
+        )
+
+    assert created_groups == []
+
+
@pytest.mark.cpu
@pytest.mark.core_model
 def test_cfg_parallel_keeps_diffusion_dp_without_ep(monkeypatch):
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Tests for the Sequence Parallelism (SP) framework.

 These tests verify the SP plan mechanism and hooks work correctly without
@@ -17,6 +17,7 @@ import pytest
 import torch
 import torch.nn as nn

+from tests.helpers.mark import hardware_marks
 from vllm_omni.diffusion.distributed.sp_plan import (
    SequenceParallelInput,
    SequenceParallelOutput,
@@ -419,7 +420,6 @@ class TestHookRegistration:
 # =============================================================================


-@pytest.mark.L4
 class TestModelSpPlans:
    """Test that model _sp_plan definitions are valid.

@@ -427,6 +427,8 @@ class TestModelSpPlans:
    May require GPU for model imports.
    """

+    pytestmark = hardware_marks(res={"cuda": "L4"})
+
    def test_zimage_transformer_sp_plan(self):
        """Test ZImageTransformer2DModel _sp_plan structure.

@@ -0,0 +1,32 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+import pytest
+import torch
+import torch.nn.functional as F
+
+from tests.helpers.mark import hardware_test
+from vllm_omni.diffusion.layers.activation import SiluAndMul
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion]
+
+
+@pytest.mark.cpu
+def test_silu_and_mul_native_matches_packed_reference() -> None:
+    packed = torch.randn(257, 256, dtype=torch.float32)
+    gate, up = packed.chunk(2, dim=-1)
+
+    torch.testing.assert_close(SiluAndMul().forward_native(packed), F.silu(gate) * up)
+
+
+@hardware_test(res={"npu": "A3"}, num_cards=1)
+def test_silu_and_mul_npu_matches_packed_reference() -> None:
+    packed = torch.randn(257, 256, device="npu", dtype=torch.bfloat16)
+    gate, up = packed.chunk(2, dim=-1)
+
+    torch.testing.assert_close(
+        SiluAndMul()(packed),
+        F.silu(gate) * up,
+        atol=2e-2,
+        rtol=2e-2,
+    )
@@ -0,0 +1,194 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import sys
+import types
+
+import pytest
+import torch
+import torch.nn.functional as F
+
+from tests.helpers.mark import hardware_test
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion]
+
+_HEAD_DIM = 128
+_ROTARY_DIM = 96
+_EPS = 1e-5
+
+
+def _reference(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    q_weight: torch.Tensor,
+    k_weight: torch.Tensor,
+    rope_table: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    half = _ROTARY_DIM // 2
+
+    def apply(x: torch.Tensor, weight: torch.Tensor) -> torch.Tensor:
+        x = F.rms_norm(x, (_HEAD_DIM,), weight, _EPS)
+        cos = rope_table[..., :half].unsqueeze(1)
+        sin = rope_table[..., half:].unsqueeze(1)
+        return torch.cat(
+            (
+                x[..., :half] * cos - x[..., half:_ROTARY_DIM] * sin,
+                x[..., half:_ROTARY_DIM] * cos + x[..., :half] * sin,
+                x[..., _ROTARY_DIM:],
+            ),
+            dim=-1,
+        )
+
+    return apply(q, q_weight), apply(k, k_weight)
+
+
+def _inputs() -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+    torch.manual_seed(7)
+    tokens, heads = 11, 4
+    q = torch.randn(tokens, heads, _HEAD_DIM, dtype=torch.bfloat16)
+    k = torch.randn(tokens, heads, _HEAD_DIM, dtype=torch.bfloat16)
+    q_weight = torch.randn(_HEAD_DIM, dtype=torch.bfloat16)
+    k_weight = torch.randn(_HEAD_DIM, dtype=torch.bfloat16)
+    freqs = torch.randn(tokens, _ROTARY_DIM // 2)
+    rope_table = torch.cat((torch.cos(freqs), torch.sin(freqs)), dim=-1).to(torch.bfloat16)
+    return q, k, q_weight, k_weight, rope_table
+
+
+@pytest.mark.cpu
+def test_npu_qk_norm_rope_uses_torch_npu_fused_primitives_without_mindiesd(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    from vllm_omni.diffusion.layers import fused_qk_norm_rope as fused
+
+    calls: dict[str, int] = {"rms_norm": 0, "rotary_mul": 0}
+
+    def npu_rms_norm(x: torch.Tensor, gamma: torch.Tensor, epsilon: float):
+        calls["rms_norm"] += 1
+        return (F.rms_norm(x, (_HEAD_DIM,), gamma, epsilon),)
+
+    def npu_rotary_mul(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor):
+        calls["rotary_mul"] += 1
+        assert x.shape[-1] == _ROTARY_DIM
+        assert cos.shape == (1, x.shape[1], 1, _ROTARY_DIM)
+        assert sin.shape == cos.shape
+        half = _ROTARY_DIM // 2
+        return torch.cat(
+            (
+                x[..., :half] * cos[..., :half] - x[..., half:] * sin[..., :half],
+                x[..., half:] * cos[..., half:] + x[..., :half] * sin[..., half:],
+            ),
+            dim=-1,
+        )
+
+    monkeypatch.setitem(
+        sys.modules,
+        "torch_npu",
+        types.SimpleNamespace(npu_rms_norm=npu_rms_norm, npu_rotary_mul=npu_rotary_mul),
+    )
+    monkeypatch.setattr(fused, "find_spec", lambda _: None)
+    q, k, q_weight, k_weight, rope_table = _inputs()
+
+    actual_q, actual_k = fused._npu_qk_norm_rope(
+        q,
+        k,
+        q_weight,
+        k_weight,
+        rope_table,
+        _EPS,
+        _ROTARY_DIM,
+    )
+    expected_q, expected_k = _reference(q, k, q_weight, k_weight, rope_table)
+
+    assert calls == {"rms_norm": 2, "rotary_mul": 2}
+    torch.testing.assert_close(actual_q, expected_q, atol=0, rtol=0)
+    torch.testing.assert_close(actual_k, expected_k, atol=0, rtol=0)
+
+
+@pytest.mark.cpu
+def test_npu_qk_norm_rope_uses_mindiesd_and_normalizes_packed_layout(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    from vllm_omni.diffusion.layers import fused_qk_norm_rope as fused
+
+    calls: dict[str, int] = {"rms_norm": 0, "mindiesd_rope": 0}
+
+    def npu_rms_norm(x: torch.Tensor, gamma: torch.Tensor, epsilon: float):
+        calls["rms_norm"] += 1
+        return (F.rms_norm(x, (_HEAD_DIM,), gamma, epsilon),)
+
+    def rotary_position_embedding(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, **kwargs):
+        calls["mindiesd_rope"] += 1
+        assert x.dim() == 4
+        assert x.shape[0] == 1
+        assert x.shape[-1] == _ROTARY_DIM
+        assert cos.shape == (1, x.shape[1], 1, _ROTARY_DIM)
+        assert sin.shape == cos.shape
+        assert kwargs == {
+            "rotated_mode": "rotated_half",
+            "head_first": False,
+            "fused": True,
+        }
+        half = _ROTARY_DIM // 2
+        return torch.cat(
+            (
+                x[..., :half] * cos[..., :half] - x[..., half:] * sin[..., :half],
+                x[..., half:] * cos[..., half:] + x[..., :half] * sin[..., half:],
+            ),
+            dim=-1,
+        )
+
+    monkeypatch.setitem(sys.modules, "torch_npu", types.SimpleNamespace(npu_rms_norm=npu_rms_norm))
+    monkeypatch.setitem(
+        sys.modules,
+        "mindiesd",
+        types.SimpleNamespace(rotary_position_embedding=rotary_position_embedding),
+    )
+    monkeypatch.setattr(fused, "find_spec", lambda _: object())
+    q, k, q_weight, k_weight, rope_table = _inputs()
+
+    actual_q, actual_k = fused._npu_qk_norm_rope(
+        q,
+        k,
+        q_weight,
+        k_weight,
+        rope_table,
+        _EPS,
+        _ROTARY_DIM,
+    )
+    expected_q, expected_k = _reference(q, k, q_weight, k_weight, rope_table)
+
+    assert calls == {"rms_norm": 2, "mindiesd_rope": 2}
+    torch.testing.assert_close(actual_q, expected_q, atol=0, rtol=0)
+    torch.testing.assert_close(actual_k, expected_k, atol=0, rtol=0)
+
+
+@hardware_test(res={"npu": "A3"}, num_cards=1)
+def test_fused_qk_norm_rope_npu_matches_reference(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Exercise the public Ascend dispatch with real torch_npu/MindIE operators."""
+    from vllm_omni.diffusion.layers import fused_qk_norm_rope as fused
+
+    q, k, q_weight, k_weight, rope_table = (tensor.to("npu") for tensor in _inputs())
+    calls = 0
+    original_npu_impl = fused._npu_qk_norm_rope
+
+    def capture_npu_impl(*args, **kwargs):
+        nonlocal calls
+        calls += 1
+        return original_npu_impl(*args, **kwargs)
+
+    monkeypatch.setattr(fused, "_npu_qk_norm_rope", capture_npu_impl)
+    actual_q, actual_k = fused.fused_qk_norm_rope(
+        q,
+        k,
+        q_weight,
+        k_weight,
+        rope_table,
+        _EPS,
+        head_dim=_HEAD_DIM,
+        rotary_dim=_ROTARY_DIM,
+    )
+    expected_q, expected_k = _reference(q, k, q_weight, k_weight, rope_table)
+
+    assert calls == 1
+    torch.testing.assert_close(actual_q, expected_q, atol=2e-2, rtol=2e-2)
+    torch.testing.assert_close(actual_k, expected_k, atol=2e-2, rtol=2e-2)
@@ -273,12 +273,15 @@ def test_required_hwr_miss_fails_before_ordinary_loading_or_publication(
    assert loader.take_host_weight_plan() is None


-def _make_dlo_online_quant_config() -> OmniDiffusionConfig:
+def _make_dlo_online_quant_config(dp_size: int = 2) -> OmniDiffusionConfig:
    return OmniDiffusionConfig(
        model="",
        dtype=torch.float32,
        quantization_config="fp8",
-        parallel_config=DiffusionParallelConfig(data_parallel_size=2),
+        parallel_config=DiffusionParallelConfig(
+            data_parallel_size=dp_size,
+            sequence_parallel_size=1,
+        ),
        enable_distributed_layerwise_offload=True,
        dlo_use_allgather=True,
    )
@@ -824,6 +827,93 @@ def test_dlo_allgather_online_fp8_uses_ordinary_loader(monkeypatch):
    assert loader.take_host_weight_plan() is None


+def test_dlo_allgather_online_int8_uses_ordinary_loader(monkeypatch):
+    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
+    from vllm_omni.quantization.int8_config import NPUInt8OnlineLinearMethod
+
+    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__(NPUInt8OnlineLinearMethod)
+    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_online_quant_group_size_one_skips_allgather_gate(monkeypatch):
+    """A DLO group of one runs no weight collective, so an otherwise
+    unvalidated online method must not be rejected by the AllGather gate."""
+
+    class UnsupportedOnlineMethod:
+        uses_meta_device = True
+
+    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
+
+    od_config = _make_dlo_online_quant_config(dp_size=1)
+    loader = DiffusersPipelineLoader(LoadConfig(), od_config)
+    model = nn.Module()
+    model.transformer = nn.Linear(2, 2, bias=False)
+    model.transformer.quant_method = UnsupportedOnlineMethod()
+    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 == []
+    assert calls == [("load", True), "process"]
+
+
 def test_dlo_allgather_rejects_unvalidated_online_quant_method(monkeypatch):
    import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod

@@ -847,10 +937,54 @@ def test_dlo_allgather_rejects_unvalidated_online_quant_method(monkeypatch):
        ),
    )

-    with pytest.raises(ValueError, match="per-tensor FP8 linears"):
+    with pytest.raises(ValueError, match="per-tensor FP8, INT8, and MXFP8 linears"):
        loader.load_model(load_device="cpu")


+def test_dlo_allgather_online_mxfp8_uses_ordinary_loader(monkeypatch):
+    mxfp8_config = pytest.importorskip("vllm_omni.quantization.mxfp8_config")
+    NPUMxfp8OnlineLinearMethod = mxfp8_config.NPUMxfp8OnlineLinearMethod
+
+    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__(NPUMxfp8OnlineLinearMethod)
+    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_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
@@ -0,0 +1,62 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+from collections.abc import Iterator
+
+import pytest
+import torch
+import torch.nn as nn
+
+from vllm_omni.diffusion.models.minimax_h3.encoder import MiniMaxH3Qwen3VLEncoder
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
+
+
+@pytest.fixture(autouse=True)
+def restore_cudnn_sdp_state() -> Iterator[None]:
+    original = torch.backends.cuda.cudnn_sdp_enabled()
+    try:
+        yield
+    finally:
+        torch.backends.cuda.enable_cudnn_sdp(original)
+
+
+def _encoder_without_weights() -> MiniMaxH3Qwen3VLEncoder:
+    encoder = MiniMaxH3Qwen3VLEncoder("unused", device=torch.device("cpu"), load_model=False)
+    encoder.text_model = nn.Linear(1, 1)
+    return encoder
+
+
+@pytest.mark.parametrize("initial_state", [False, True])
+def test_encode_ids_restores_cudnn_sdp_after_success(monkeypatch: pytest.MonkeyPatch, initial_state: bool) -> None:
+    encoder = _encoder_without_weights()
+    expected = torch.tensor([1.0])
+
+    def encode(*args, **kwargs):
+        assert torch.backends.cuda.cudnn_sdp_enabled()
+        return expected
+
+    monkeypatch.setattr(encoder, "_encode", encode)
+    torch.backends.cuda.enable_cudnn_sdp(initial_state)
+
+    actual = encoder.encode_ids(torch.tensor([1]))
+
+    torch.testing.assert_close(actual, expected)
+    assert torch.backends.cuda.cudnn_sdp_enabled() is initial_state
+
+
+@pytest.mark.parametrize("initial_state", [False, True])
+def test_encode_ids_restores_cudnn_sdp_after_failure(monkeypatch: pytest.MonkeyPatch, initial_state: bool) -> None:
+    encoder = _encoder_without_weights()
+
+    def fail(*args, **kwargs):
+        assert torch.backends.cuda.cudnn_sdp_enabled()
+        raise RuntimeError("encode failed")
+
+    monkeypatch.setattr(encoder, "_encode", fail)
+    torch.backends.cuda.enable_cudnn_sdp(initial_state)
+
+    with pytest.raises(RuntimeError, match="encode failed"):
+        encoder.encode_ids(torch.tensor([1]))
+
+    assert torch.backends.cuda.cudnn_sdp_enabled() is initial_state
@@ -223,6 +223,8 @@ def test_h3_turbo_allows_distributed_layerwise_offload(monkeypatch):
    torch.nn.Module.__init__(pipeline)
    pipeline.partition = "fl2va"
    pipeline._turbo_lora_adapter_ids = set()
+    pipeline._native_lora_adapter_ids = set()
+    pipeline._lora_sigma_schedules = {}
    pipeline.od_config = SimpleNamespace(
        enable_cpu_offload=False,
        enable_layerwise_offload=False,
@@ -321,6 +323,8 @@ def test_only_an_active_recognized_turbo_adapter_restricts_ref2va(monkeypatch):
    pipeline.partition = "combined"
    pipeline.supported_tasks = frozenset({"t2va", "fl2va", "ref2va"})
    pipeline._turbo_lora_adapter_ids = set()
+    pipeline._native_lora_adapter_ids = set()
+    pipeline._lora_sigma_schedules = {}
    request = _request("generic-peft")

    def load():
@@ -0,0 +1,860 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+from safetensors.torch import save_file
+from vllm.lora.lora_weights import PackedLoRALayerWeights
+
+from vllm_omni.diffusion.lora.manager import DiffusionLoRAManager
+from vllm_omni.diffusion.models.minimax_h3.npu import lora as lora_module
+from vllm_omni.diffusion.models.minimax_h3.npu.lora import load_minimax_h3_native_lora
+from vllm_omni.diffusion.sched.sigma_schedule import DMD2SigmaSchedule
+from vllm_omni.errors import OmniClientError
+from vllm_omni.lora.request import LoRARequest
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+_TINY_HIDDEN = 2
+_TINY_FFN = 2
+_TINY_TIME = 2
+_TINY_BLOCK_ADALN = 4
+_TINY_FINAL_ADALN = 3
+_TINY_RANK = 4
+_TINY_NUM_QUERY_GROUPS = 2
+_TINY_HEADS_PER_GROUP = 1
+_TINY_HEAD_DIM = 1
+_TINY_QKV_SLICE = _TINY_NUM_QUERY_GROUPS * _TINY_HEADS_PER_GROUP * _TINY_HEAD_DIM
+# Grouped qkv stores (heads_per_group + 2) * head_dim rows per query group, so
+# the attention inner size follows from the head layout and cannot be picked
+# independently: 3 * inner must equal the grouped row count.
+_TINY_INNER = _TINY_QKV_SLICE
+
+
+@pytest.fixture(autouse=True)
+def _use_tiny_native_dimensions(monkeypatch):
+    monkeypatch.setattr(lora_module, "_NATIVE_RANK", _TINY_RANK)
+    monkeypatch.setattr(lora_module, "_NATIVE_ALPHA", _TINY_RANK)
+    monkeypatch.setattr(lora_module, "_NATIVE_HIDDEN_SIZE", _TINY_HIDDEN)
+    monkeypatch.setattr(lora_module, "_NATIVE_ATTENTION_INNER_SIZE", _TINY_INNER)
+    monkeypatch.setattr(lora_module, "_NATIVE_FFN_HIDDEN_SIZE", _TINY_FFN)
+    monkeypatch.setattr(lora_module, "_NATIVE_TIME_EMBED_DIM", _TINY_TIME)
+    monkeypatch.setattr(lora_module, "_NATIVE_BLOCK_ADALN_OUT", _TINY_BLOCK_ADALN)
+    monkeypatch.setattr(lora_module, "_NATIVE_FINAL_ADALN_OUT", _TINY_FINAL_ADALN)
+    monkeypatch.setattr(lora_module, "_NATIVE_NUM_QUERY_GROUPS", _TINY_NUM_QUERY_GROUPS)
+    monkeypatch.setattr(lora_module, "_NATIVE_HEADS_PER_GROUP", _TINY_HEADS_PER_GROUP)
+    monkeypatch.setattr(lora_module, "_NATIVE_HEAD_DIM", _TINY_HEAD_DIM)
+    monkeypatch.setattr(lora_module, "_NATIVE_QKV_SLICE", _TINY_QKV_SLICE)
+    monkeypatch.setattr(
+        lora_module,
+        "_NATIVE_TARGET_DIMS",
+        {
+            "attn.qkv_proj": (_TINY_HIDDEN, 3 * _TINY_INNER),
+            "attn.out_proj": (_TINY_INNER, _TINY_HIDDEN),
+            "mlp.fc1": (_TINY_HIDDEN, 2 * _TINY_FFN),
+            "mlp.fc2": (_TINY_FFN, _TINY_HIDDEN),
+            "adaln_proj.linear": (_TINY_TIME, _TINY_BLOCK_ADALN),
+        },
+    )
+    monkeypatch.setattr(lora_module, "_NATIVE_FINAL_ADALN_DIMS", (_TINY_TIME, _TINY_FINAL_ADALN))
+    monkeypatch.setattr(
+        lora_module,
+        "_NATIVE_EXPECTED_TARGETS",
+        frozenset(
+            [
+                *(
+                    f"blocks.{block_index}.{suffix}"
+                    for block_index in range(50)
+                    for suffix in lora_module._NATIVE_TARGET_SUFFIXES
+                )
+            ]
+            + [
+                *(
+                    f"token_refiner.blocks.{block_index}.{suffix}"
+                    for block_index in range(2)
+                    for suffix in lora_module._NATIVE_TOKEN_REFINER_SUFFIXES
+                )
+            ]
+            + ["final_layer.adaln_proj.linear"]
+        ),
+    )
+
+
+def _request(path) -> LoRARequest:
+    return LoRARequest(
+        lora_name="flashgen",
+        lora_int_id=7,
+        lora_path=str(path),
+    )
+
+
+def _write_tiny_native(
+    path,
+    *,
+    omit_target: str | None = None,
+    shape_overrides: dict[str, tuple[int, int]] | None = None,
+    metadata: dict[str, str] | None = None,
+) -> None:
+    rank = _TINY_RANK
+    tensors = {}
+    overrides = shape_overrides or {}
+    for target in sorted(lora_module._NATIVE_EXPECTED_TARGETS):
+        if target == omit_target:
+            continue
+        input_dim, output_dim = lora_module._native_target_dims(target)
+        raw_target = f"transformer.{target}"
+        a_name = f"{raw_target}.lora_A.default.weight"
+        b_name = f"{raw_target}.lora_B.default.weight"
+        tensors[a_name] = torch.ones(overrides.get(a_name, (rank, input_dim)))
+        if target.endswith(".attn.qkv_proj"):
+            grouped_rows = []
+            for group in range(_TINY_NUM_QUERY_GROUPS):
+                grouped_rows.extend(
+                    [
+                        torch.full((1, rank), float(group * 10 + 1)),
+                        torch.full((1, rank), float(group * 10 + 2)),
+                        torch.full((1, rank), float(group * 10 + 3)),
+                    ]
+                )
+            tensors[b_name] = torch.cat(grouped_rows, dim=0)
+        elif target.endswith(".mlp.fc1"):
+            tensors[b_name] = torch.cat(
+                (
+                    torch.full((output_dim // 2, rank), 2.0),
+                    torch.full((output_dim // 2, rank), 1.0),
+                ),
+                dim=0,
+            )
+        else:
+            tensors[b_name] = torch.ones(overrides.get(b_name, (output_dim, rank)))
+    save_file(
+        tensors,
+        str(path),
+        metadata=metadata
+        or {
+            "format": "pt",
+            "key_format": "minimax-h3-native",
+            "qkv_layout": "grouped",
+            "lora_rank": str(rank),
+            "lora_alpha": str(rank),
+            "base_schedule": "1.0,0.7,0.4,0.15,0.0",
+            "tasks": "t2va",
+        },
+    )
+
+
+def test_h3_native_loads_and_packs_qkv_and_fc1(tmp_path):
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path)
+
+    loaded = load_minimax_h3_native_lora(
+        partition="fl2va",
+        lora_request=_request(path),
+        lora_path=path,
+        dtype=torch.float32,
+    )
+
+    assert loaded is not None
+    lora_model, peft_helper, schedule = loaded
+    assert peft_helper.r == _TINY_RANK
+    assert schedule.num_inference_steps == 4
+    assert len(lora_model.loras) == 259
+    assert "blocks.0.adaln_proj.linear" in lora_model.loras
+    assert "final_layer.adaln_proj.linear" in lora_model.loras
+
+    # The writer tags group g with q=g*10+1, k=g*10+2, v=g*10+3, so a correct
+    # grouped -> q/k/v permutation yields slices that are uniform per role.
+    qkv = lora_model.get_lora("blocks.0.attn.qkv_proj")
+    assert isinstance(qkv, PackedLoRALayerWeights)
+    torch.testing.assert_close(qkv.lora_b[0], torch.tensor([[1.0] * _TINY_RANK, [11.0] * _TINY_RANK]))
+    torch.testing.assert_close(qkv.lora_b[1], torch.tensor([[2.0] * _TINY_RANK, [12.0] * _TINY_RANK]))
+    torch.testing.assert_close(qkv.lora_b[2], torch.tensor([[3.0] * _TINY_RANK, [13.0] * _TINY_RANK]))
+
+    fc1 = lora_model.get_lora("blocks.0.mlp.fc1")
+    assert isinstance(fc1, PackedLoRALayerWeights)
+    torch.testing.assert_close(fc1.lora_b[0], torch.full((_TINY_FFN, _TINY_RANK), 2.0))
+    torch.testing.assert_close(fc1.lora_b[1], torch.full((_TINY_FFN, _TINY_RANK), 1.0))
+
+
+def test_h3_native_rejects_bad_metadata_and_ref2va(tmp_path):
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path, metadata={"key_format": "minimax-h3-native", "qkv_layout": "runtime"})
+    with pytest.raises(ValueError, match="qkv_layout='grouped'"):
+        load_minimax_h3_native_lora(
+            partition="fl2va",
+            lora_request=_request(path),
+            lora_path=path,
+            dtype=torch.float32,
+        )
+
+    # The exact-filename guard runs before the partition check, so the artifact
+    # has to keep the published name for this to reach the Ref2VA rejection.
+    valid_dir = tmp_path / "valid"
+    valid_dir.mkdir()
+    valid = valid_dir / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(valid)
+    with pytest.raises(ValueError, match="supports T2VA only"):
+        load_minimax_h3_native_lora(
+            partition="ref2va",
+            lora_request=_request(valid),
+            lora_path=valid,
+            dtype=torch.float32,
+        )
+
+    renamed = tmp_path / "renamed.safetensors"
+    _write_tiny_native(renamed)
+    with pytest.raises(ValueError, match="supports only"):
+        load_minimax_h3_native_lora(
+            partition="fl2va",
+            lora_request=_request(renamed),
+            lora_path=renamed,
+            dtype=torch.float32,
+        )
+
+
+def test_h3_native_rejects_truncated_artifact(tmp_path):
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path, omit_target="blocks.49.mlp.fc2")
+    with pytest.raises(ValueError, match="target set does not match"):
+        load_minimax_h3_native_lora(
+            partition="fl2va",
+            lora_request=_request(path),
+            lora_path=path,
+            dtype=torch.float32,
+        )
+
+
+def test_h3_native_rejects_schedule_with_wrong_interval_count(tmp_path):
+    """A mislabeled schedule must fail closed instead of silently changing steps."""
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(
+        path,
+        metadata={
+            "format": "pt",
+            "key_format": "minimax-h3-native",
+            "qkv_layout": "grouped",
+            "lora_rank": str(_TINY_RANK),
+            "lora_alpha": str(_TINY_RANK),
+            "base_schedule": "1.0,0.5,0.0",
+            "tasks": "t2va",
+        },
+    )
+    with pytest.raises(ValueError, match="4-interval base_schedule"):
+        load_minimax_h3_native_lora(
+            partition="fl2va",
+            lora_request=_request(path),
+            lora_path=path,
+            dtype=torch.float32,
+        )
+
+
+@pytest.mark.parametrize(
+    ("raw_schedule", "message"),
+    [
+        (None, "requires safetensors metadata base_schedule"),
+        ("", "must not be empty"),
+        (",,", "must not be empty"),
+        ("1.0,oops,0.15,0.0", "is malformed"),
+        ("0.9,0.7,0.4,0.15,0.0", "is malformed"),
+    ],
+)
+def test_h3_native_rejects_malformed_schedule_metadata(tmp_path, raw_schedule, message):
+    """safetensors metadata is string-only, so the loader owns the parse contract."""
+    metadata = {
+        "format": "pt",
+        "key_format": "minimax-h3-native",
+        "qkv_layout": "grouped",
+        "lora_rank": str(_TINY_RANK),
+        "lora_alpha": str(_TINY_RANK),
+        "tasks": "t2va",
+    }
+    if raw_schedule is not None:
+        metadata["base_schedule"] = raw_schedule
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path, metadata=metadata)
+
+    with pytest.raises(ValueError, match=message):
+        load_minimax_h3_native_lora(
+            partition="fl2va",
+            lora_request=_request(path),
+            lora_path=path,
+            dtype=torch.float32,
+        )
+
+
+def test_h3_native_declared_file_fails_closed_on_invalid_metadata(tmp_path):
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path, metadata={"key_format": "other"})
+    with pytest.raises(ValueError, match="requires safetensors metadata"):
+        load_minimax_h3_native_lora(
+            partition="fl2va",
+            lora_request=_request(path),
+            lora_path=path,
+            dtype=torch.float32,
+        )
+
+
+def test_pipeline_native_schedule_and_task_validation(monkeypatch):
+    from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
+
+    pipeline = object.__new__(MiniMaxH3Pipeline)
+    torch.nn.Module.__init__(pipeline)
+    pipeline.partition = "fl2va"
+    pipeline.supported_tasks = frozenset({"t2va", "fl2va", "ref2va"})
+    pipeline._turbo_lora_adapter_ids = set()
+    pipeline._native_lora_adapter_ids = {7}
+    pipeline._lora_sigma_schedules = {7: DMD2SigmaSchedule.from_positions([1.0, 0.7, 0.4, 0.15, 0.0])}
+    pipeline._base_schedule_by_partition = {"fl2va": None}
+
+    sampling = SimpleNamespace(
+        lora_request=LoRARequest("flashgen", 7, "/tmp/native.safetensors"),
+        lora_scale=1.0,
+        num_inference_steps=4,
+        extra_args={},
+    )
+    assert pipeline._sigma_schedule_for_request(sampling, "t2va").num_inference_steps == 4
+    with pytest.raises(OmniClientError, match="num_inference_steps must be 4"):
+        pipeline._validate_native_sampling(
+            SimpleNamespace(lora_request=sampling.lora_request, num_inference_steps=5),
+            task="t2va",
+        )
+    with pytest.raises(OmniClientError, match="supports T2VA requests only"):
+        pipeline._resolve_task(
+            "fl2va",
+            {},
+            has_turbo_lora=False,
+            has_native_lora=True,
+        )
+
+    pipeline._base_schedule_by_partition = {"fl2va": DMD2SigmaSchedule.from_positions([1.0, 0.5, 0.0])}
+    with pytest.raises(OmniClientError, match="already pins base_schedule"):
+        pipeline._sigma_schedule_for_request(sampling, "t2va")
+
+
+def test_pipeline_native_step_count_follows_adapter_schedule():
+    """The accepted step count is the adapter's interval count, never a literal."""
+    from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
+
+    pipeline = object.__new__(MiniMaxH3Pipeline)
+    torch.nn.Module.__init__(pipeline)
+    pipeline.partition = "fl2va"
+    pipeline._native_lora_adapter_ids = {7}
+    pipeline._lora_sigma_schedules = {7: DMD2SigmaSchedule.from_positions([1.0, 0.5, 0.0])}
+    request = LoRARequest("flashgen", 7, "/tmp/native.safetensors")
+
+    pipeline._validate_native_sampling(
+        SimpleNamespace(lora_request=request, num_inference_steps=2),
+        task="t2va",
+    )
+    with pytest.raises(OmniClientError, match="num_inference_steps must be 2 or omitted, not 3"):
+        pipeline._validate_native_sampling(
+            SimpleNamespace(lora_request=request, num_inference_steps=3),
+            task="t2va",
+        )
+    with pytest.raises(OmniClientError, match="requires num_inference_steps=2"):
+        pipeline._validate_native_sampling(
+            SimpleNamespace(lora_request=request, num_inference_steps=4),
+            task="t2va",
+        )
+
+
+@pytest.mark.parametrize("step_execution", [False, True])
+def test_pipeline_native_omitted_step_count_follows_execution_mode(step_execution):
+    """Only request mode can resolve an omitted count from the adapter schedule.
+
+    ``StepScheduler`` reads ``num_inference_steps`` off the request at admission,
+    before this pipeline sees it, so under step execution an omitted value can
+    only disagree with the denoise loop. The advertised contract has to match.
+    """
+    from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
+
+    pipeline = object.__new__(MiniMaxH3Pipeline)
+    torch.nn.Module.__init__(pipeline)
+    pipeline.partition = "fl2va"
+    pipeline._native_lora_adapter_ids = {7}
+    pipeline._lora_sigma_schedules = {7: DMD2SigmaSchedule.from_positions([1.0, 0.7, 0.4, 0.15, 0.0])}
+    pipeline.od_config = SimpleNamespace(step_execution=step_execution)
+    sampling = SimpleNamespace(
+        lora_request=LoRARequest("flashgen", 7, "/tmp/native.safetensors"),
+        num_inference_steps=None,
+    )
+
+    if not step_execution:
+        # Request mode defaults to the adapter's own interval count.
+        pipeline._validate_native_sampling(sampling, task="t2va")
+        return
+
+    with pytest.raises(OmniClientError, match="explicit num_inference_steps=4 under step execution"):
+        pipeline._validate_native_sampling(sampling, task="t2va")
+    # The step-mode message must not offer omission it cannot accept.
+    with pytest.raises(OmniClientError, match=r"must be 4, not 5$"):
+        pipeline._validate_native_sampling(
+            SimpleNamespace(lora_request=sampling.lora_request, num_inference_steps=5),
+            task="t2va",
+        )
+
+
+def test_pipeline_replaces_native_classification_after_reload(monkeypatch, tmp_path):
+    from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
+    from vllm_omni.diffusion.models.minimax_h3 import pipeline_minimax_h3 as pipeline_module
+
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path)
+    request = _request(path)
+
+    pipeline = object.__new__(MiniMaxH3Pipeline)
+    torch.nn.Module.__init__(pipeline)
+    pipeline.partition = "fl2va"
+    pipeline.od_config = SimpleNamespace(
+        enable_cpu_offload=False,
+        enable_layerwise_offload=False,
+        enable_distributed_layerwise_offload=False,
+    )
+    pipeline._turbo_lora_adapter_ids = set()
+    pipeline._native_lora_adapter_ids = set()
+    pipeline._lora_sigma_schedules = {}
+
+    loaded = pipeline._load_diffusion_lora_adapter(
+        lora_request=request,
+        lora_path=path,
+        dtype=torch.float32,
+    )
+    assert loaded is not None
+    assert request.lora_int_id in pipeline._native_lora_adapter_ids
+    assert request.lora_int_id in pipeline._lora_sigma_schedules
+
+    monkeypatch.setattr(pipeline_module, "load_minimax_h3_turbo_lora", lambda **_: None)
+    monkeypatch.setattr(pipeline_module, "load_minimax_h3_native_lora", lambda **_: None)
+    assert pipeline._load_diffusion_lora_adapter(lora_request=request, lora_path=path, dtype=torch.float32) is None
+    assert request.lora_int_id not in pipeline._native_lora_adapter_ids
+    assert request.lora_int_id not in pipeline._lora_sigma_schedules
+
+
+@pytest.mark.parametrize(
+    "offload_mode",
+    [
+        "model-level CPU offload (--enable-cpu-offload)",
+        "layerwise offload (--enable-layerwise-offload)",
+    ],
+)
+def test_h3_native_rejects_offload_modes(tmp_path, offload_mode):
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path)
+
+    with pytest.raises(ValueError, match="does not support"):
+        load_minimax_h3_native_lora(
+            partition="fl2va",
+            lora_request=_request(path),
+            lora_path=path,
+            dtype=torch.float32,
+            unsupported_offload_mode=offload_mode,
+        )
+
+
+def test_h3_native_allows_distributed_layerwise_offload(monkeypatch):
+    """DLO keeps LoRA A/B buffers resident, so native must not fail closed."""
+    from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
+    from vllm_omni.diffusion.models.minimax_h3 import pipeline_minimax_h3 as pipeline_module
+
+    pipeline = object.__new__(MiniMaxH3Pipeline)
+    torch.nn.Module.__init__(pipeline)
+    pipeline.partition = "fl2va"
+    pipeline.od_config = SimpleNamespace(
+        enable_cpu_offload=False,
+        enable_layerwise_offload=False,
+        enable_distributed_layerwise_offload=True,
+    )
+    pipeline._turbo_lora_adapter_ids = set()
+    pipeline._native_lora_adapter_ids = set()
+    pipeline._lora_sigma_schedules = {}
+    captured: dict[str, object] = {}
+    schedule = DMD2SigmaSchedule.from_positions([1.0, 0.7, 0.4, 0.15, 0.0])
+
+    def load_native(**kwargs):
+        captured.update(kwargs)
+        return object(), object(), schedule
+
+    monkeypatch.setattr(pipeline_module, "load_minimax_h3_turbo_lora", lambda **_: None)
+    monkeypatch.setattr(pipeline_module, "load_minimax_h3_native_lora", load_native)
+
+    request = _request("flashgen")
+    loaded = pipeline._load_diffusion_lora_adapter(
+        lora_request=request,
+        lora_path="flashgen",
+        dtype=torch.bfloat16,
+    )
+
+    assert loaded is not None
+    assert captured["unsupported_offload_mode"] is None
+    assert request.lora_int_id in pipeline._native_lora_adapter_ids
+    assert pipeline._lora_sigma_schedules[request.lora_int_id] is schedule
+
+
+def test_h3_native_qkv_reorder_matches_base_loader_contract():
+    from vllm_omni.diffusion.models.minimax_h3.minimax_h3_transformer import _reorder_grouped_qkv_to_qkv
+
+    num_query_groups = 56
+    heads_per_group = 1
+    head_dim = 128
+    grouped = torch.arange(num_query_groups * (heads_per_group + 2) * head_dim, dtype=torch.float32)
+    grouped = grouped.reshape(num_query_groups, (heads_per_group + 2) * head_dim)
+    reordered = _reorder_grouped_qkv_to_qkv(
+        grouped.reshape(-1, 1),
+        num_query_groups=num_query_groups,
+        heads_per_group=heads_per_group,
+        head_dim=head_dim,
+    ).reshape(-1)
+    q_size = num_query_groups * heads_per_group * head_dim
+    k_size = num_query_groups * head_dim
+    q, k, v = torch.split(reordered, [q_size, k_size, k_size])
+    assert q.reshape(num_query_groups, head_dim)[0, 0] == grouped[0, 0]
+    assert k.reshape(num_query_groups, head_dim)[0, 0] == grouped[0, head_dim]
+    assert v.reshape(num_query_groups, head_dim)[0, 0] == grouped[0, 2 * head_dim]
+
+
+def test_native_target_pattern_matches_component_qualified_names():
+    """The manager matches "<component>.<module>", so the pattern needs the prefix."""
+    import regex as re
+
+    pattern = lora_module._NATIVE_TARGET_PATTERN
+    matched = [
+        target
+        for target in lora_module._NATIVE_EXPECTED_TARGETS
+        if re.search(pattern, f"transformer.{target}") is not None
+    ]
+    assert len(matched) == 259
+    assert re.search(pattern, "transformer.blocks.0.attn.q_proj") is None
+    assert re.search(pattern, "transformer.token_refiner.blocks.0.adaln_proj.linear") is None
+
+
+def test_replace_layers_binds_every_native_target(monkeypatch):
+    """Regression for bound=0/259: layer replacement must reach all 259 modules."""
+    from vllm.lora.layers import BaseLayerWithLoRA
+    from vllm.lora.peft_helper import PEFTHelper
+
+    from vllm_omni.diffusion.lora import manager as manager_module
+
+    class _StubLoRALayer(BaseLayerWithLoRA):
+        def __init__(self, base_layer):
+            super().__init__()
+            self.base_layer = base_layer
+
+    def _linear() -> torch.nn.Module:
+        return torch.nn.Linear(_TINY_HIDDEN, _TINY_HIDDEN)
+
+    def _block(*, with_adaln: bool) -> torch.nn.Module:
+        block = torch.nn.Module()
+        block.attn = torch.nn.Module()
+        block.attn.qkv_proj = _linear()
+        block.attn.out_proj = _linear()
+        block.mlp = torch.nn.Module()
+        block.mlp.fc1 = _linear()
+        block.mlp.fc2 = _linear()
+        if with_adaln:
+            block.adaln_proj = torch.nn.Module()
+            block.adaln_proj.linear = _linear()
+        return block
+
+    transformer = torch.nn.Module()
+    transformer.blocks = torch.nn.ModuleList([_block(with_adaln=True) for _ in range(50)])
+    transformer.token_refiner = torch.nn.Module()
+    transformer.token_refiner.blocks = torch.nn.ModuleList([_block(with_adaln=False) for _ in range(2)])
+    transformer.final_layer = torch.nn.Module()
+    transformer.final_layer.adaln_proj = torch.nn.Module()
+    transformer.final_layer.adaln_proj.linear = _linear()
+    pipeline = torch.nn.Module()
+    pipeline.transformer = transformer
+
+    monkeypatch.setattr(manager_module, "from_layer_diffusion", lambda *, layer, **_: _StubLoRALayer(layer))
+
+    manager = object.__new__(DiffusionLoRAManager)
+    manager.pipeline = pipeline
+    manager.dtype = torch.float32
+    manager.max_cached_adapters = 1
+    manager._lora_modules = {}
+    manager._max_lora_rank = 64
+    manager._resident_lora_device = None
+
+    peft_helper = PEFTHelper.from_dict(
+        {
+            "r": _TINY_RANK,
+            "lora_alpha": _TINY_RANK,
+            "target_modules": lora_module._NATIVE_TARGET_PATTERN,
+        }
+    )
+    manager._replace_layers_with_lora(peft_helper)
+
+    assert set(manager._lora_modules) == {f"transformer.{target}" for target in lora_module._NATIVE_EXPECTED_TARGETS}
+
+
+def test_legacy_manager_uses_native_loader(tmp_path):
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path)
+
+    class _Pipeline:
+        def _load_diffusion_lora_adapter(self, **kwargs):
+            from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import (
+                MiniMaxH3Pipeline,
+            )
+
+            pipeline = object.__new__(MiniMaxH3Pipeline)
+            pipeline.partition = "fl2va"
+            pipeline.od_config = SimpleNamespace(
+                enable_cpu_offload=False,
+                enable_layerwise_offload=False,
+                enable_distributed_layerwise_offload=False,
+            )
+            pipeline._turbo_lora_adapter_ids = set()
+            pipeline._native_lora_adapter_ids = set()
+            pipeline._lora_sigma_schedules = {}
+            return pipeline._load_diffusion_lora_adapter(**kwargs)
+
+    manager = object.__new__(DiffusionLoRAManager)
+    manager.pipeline = _Pipeline()
+    manager.dtype = torch.float32
+    manager._expected_lora_modules = {"qkv_proj", "fc1", "adaln_proj.linear"}
+
+    lora_model, peft_helper = manager._load_adapter(_request(path))
+    assert lora_model.id == 7
+    assert peft_helper.lora_alpha == _TINY_RANK
+    assert len(lora_model.loras) == 259
+
+
+def test_pipeline_schedule_inactive_when_scale_zero():
+    from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
+
+    pipeline = object.__new__(MiniMaxH3Pipeline)
+    torch.nn.Module.__init__(pipeline)
+    pipeline.partition = "fl2va"
+    pipeline._turbo_lora_adapter_ids = set()
+    pipeline._native_lora_adapter_ids = {7}
+    pipeline._lora_sigma_schedules = {7: DMD2SigmaSchedule.from_positions([1.0, 0.7, 0.4, 0.15, 0.0])}
+    pipeline._base_schedule_by_partition = {"fl2va": None}
+
+    sampling = SimpleNamespace(
+        lora_request=LoRARequest("flashgen", 7, "/tmp/native.safetensors"),
+        lora_scale=0.0,
+        num_inference_steps=4,
+    )
+    assert pipeline._sigma_schedule_for_request(sampling, "t2va") is None
+    assert not pipeline._has_active_native_lora(sampling)
+
+
+def test_pipeline_schedule_falls_back_after_eviction(monkeypatch, tmp_path):
+    from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
+    from vllm_omni.diffusion.models.minimax_h3 import pipeline_minimax_h3 as pipeline_module
+
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path)
+    request = _request(path)
+
+    pipeline = object.__new__(MiniMaxH3Pipeline)
+    torch.nn.Module.__init__(pipeline)
+    pipeline.partition = "fl2va"
+    pipeline.od_config = SimpleNamespace(
+        enable_cpu_offload=False,
+        enable_layerwise_offload=False,
+        enable_distributed_layerwise_offload=False,
+    )
+    pipeline._turbo_lora_adapter_ids = set()
+    pipeline._native_lora_adapter_ids = set()
+    pipeline._lora_sigma_schedules = {}
+    pipeline._base_schedule_by_partition = {"fl2va": None}
+
+    loaded = pipeline._load_diffusion_lora_adapter(
+        lora_request=request,
+        lora_path=path,
+        dtype=torch.float32,
+    )
+    assert loaded is not None
+    assert request.lora_int_id in pipeline._lora_sigma_schedules
+
+    monkeypatch.setattr(pipeline_module, "load_minimax_h3_turbo_lora", lambda **_: None)
+    monkeypatch.setattr(pipeline_module, "load_minimax_h3_native_lora", lambda **_: None)
+    assert pipeline._load_diffusion_lora_adapter(lora_request=request, lora_path=path, dtype=torch.float32) is None
+    assert request.lora_int_id not in pipeline._lora_sigma_schedules
+
+    sampling = SimpleNamespace(
+        lora_request=request,
+        lora_scale=1.0,
+        num_inference_steps=4,
+    )
+    assert pipeline._sigma_schedule_for_request(sampling, "t2va") is None
+
+
+def test_native_packed_qkv_slices_are_tp2_divisible():
+    assert lora_module._NATIVE_QKV_SLICE % 2 == 0
+    assert lora_module._NATIVE_FFN_HIDDEN_SIZE % 2 == 0
+
+
+def test_lora_manager_activates_native_packed_qkv(tmp_path):
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path)
+    loaded = load_minimax_h3_native_lora(
+        partition="fl2va",
+        lora_request=_request(path),
+        lora_path=path,
+        dtype=torch.float32,
+    )
+    assert loaded is not None
+    lora_model, _, _ = loaded
+    packed = lora_model.get_lora("blocks.0.attn.qkv_proj")
+    assert isinstance(packed, PackedLoRALayerWeights)
+    assert len(packed.lora_b) == 3
+    assert all(b.shape[0] == _TINY_QKV_SLICE for b in packed.lora_b)
+
+    class _DummyLoRALayer:
+        def __init__(self):
+            self.n_slices = 3
+            self.output_slices = (_TINY_QKV_SLICE, _TINY_QKV_SLICE, _TINY_QKV_SLICE)
+            self.set_calls: list[tuple[list, list]] = []
+            self.reset_calls = 0
+
+        def reset_lora(self, index: int):
+            self.reset_calls += 1
+
+        def set_lora(self, index: int, lora_a, lora_b):
+            self.set_calls.append((lora_a, lora_b))
+
+    layer = _DummyLoRALayer()
+    manager = object.__new__(DiffusionLoRAManager)
+    manager.pipeline = torch.nn.Module()
+    manager._lora_modules = {"blocks.0.attn.qkv_proj": layer}
+    manager._registered_adapters = {
+        lora_model.id: lora_model,
+    }
+    manager._active_adapter_id = None
+    manager._adapter_scales = {}
+    manager._activate_adapter(lora_model.id, scale=0.5)
+
+    assert manager._active_adapter_id == lora_model.id
+
+    assert len(layer.set_calls) == 1
+    lora_a_list, lora_b_list = layer.set_calls[0]
+    assert len(lora_a_list) == 3
+    assert len(lora_b_list) == 3
+    torch.testing.assert_close(lora_b_list[0], packed.lora_b[0] * 0.5)
+    torch.testing.assert_close(lora_b_list[1], packed.lora_b[1] * 0.5)
+    torch.testing.assert_close(lora_b_list[2], packed.lora_b[2] * 0.5)
+
+
+def test_lora_manager_tp2_splits_native_packed_qkv_per_rank():
+    """Each Q/K/V slice must be divisible by TP2 so rank-local layers can shard rows."""
+
+    tp_size = 2
+    slice_rows = lora_module._NATIVE_QKV_SLICE
+    assert slice_rows % tp_size == 0
+    local_rows = slice_rows // tp_size
+
+    class _TpQkvLoRALayer:
+        def __init__(self, tp_rank: int):
+            self.tp_rank = tp_rank
+            self.n_slices = 3
+            self.output_slices = (local_rows, local_rows, local_rows)
+            self.set_calls: list[tuple[list, list]] = []
+
+        def reset_lora(self, index: int):
+            return
+
+        def set_lora(self, index: int, lora_a, lora_b):
+            self.set_calls.append((lora_a, lora_b))
+
+    for tp_rank in range(tp_size):
+        layer = _TpQkvLoRALayer(tp_rank)
+        start = tp_rank * local_rows
+        end = start + local_rows
+        full_packed = PackedLoRALayerWeights(
+            module_name="blocks.0.attn.qkv_proj",
+            rank=lora_module._NATIVE_RANK,
+            lora_alphas=[64, 64, 64],
+            lora_a=[torch.ones(lora_module._NATIVE_RANK, lora_module._NATIVE_HIDDEN_SIZE)] * 3,
+            lora_b=[torch.full((slice_rows, lora_module._NATIVE_RANK), float(i + 1)) for i in range(3)],
+            scaling=[1.0, 1.0, 1.0],
+        )
+        local_b = [b[start:end].contiguous() for b in full_packed.lora_b]
+        layer.set_lora(0, full_packed.lora_a, local_b)
+        assert len(layer.set_calls) == 1
+        _, lora_b_list = layer.set_calls[0]
+        assert all(b.shape[0] == local_rows for b in lora_b_list)
+        assert lora_b_list[0][0, 0].item() == 1.0
+        assert lora_b_list[1][0, 0].item() == 2.0
+        assert lora_b_list[2][0, 0].item() == 3.0
+
+
+# 259 modules, of which 52 qkv expand to 3 slices and 52 fc1 expand to 2.
+_EXPECTED_NATIVE_SLICE_PAIRS = 415
+
+
+def _iter_native_slice_pairs(lora_model):
+    """Yield ``(module_name, lora_a, lora_b)`` per slice, flattening packed weights."""
+    for module_name, weights in lora_model.loras.items():
+        lora_a = weights.lora_a if isinstance(weights.lora_a, list) else [weights.lora_a]
+        lora_b = weights.lora_b if isinstance(weights.lora_b, list) else [weights.lora_b]
+        for a, b in zip(lora_a, lora_b, strict=True):
+            yield module_name, a, b
+
+
+@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
+def test_native_lora_load_is_platform_agnostic(tmp_path, dtype):
+    """The artifact is trained on Ascend, but loading it is not Ascend-only.
+
+    This whole module runs in the CPU job with no ``torch_npu`` present, so a
+    passing run is itself the evidence. The assertions pin the property that
+    makes it portable: tensors land on CPU in exactly the requested dtype and
+    the loader never picks a device, leaving placement to the pipeline.
+    """
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path)
+
+    loaded = load_minimax_h3_native_lora(
+        partition="fl2va",
+        lora_request=_request(path),
+        lora_path=path,
+        dtype=dtype,
+    )
+
+    assert loaded is not None
+    lora_model, _, _ = loaded
+    tensors = {
+        f"{module_name}.{side}": tensor
+        for module_name, a, b in _iter_native_slice_pairs(lora_model)
+        for side, tensor in (("lora_a", a), ("lora_b", b))
+    }
+    assert sum(1 for _ in _iter_native_slice_pairs(lora_model)) == _EXPECTED_NATIVE_SLICE_PAIRS
+    assert {name: tensor.device.type for name, tensor in tensors.items() if tensor.device.type != "cpu"} == {}
+    assert {name: tensor.dtype for name, tensor in tensors.items() if tensor.dtype != dtype} == {}
+
+
+@pytest.mark.cuda
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires a CUDA device")
+def test_native_lora_delta_matches_between_cpu_and_cuda(tmp_path):
+    """The packed adapter produces the same delta weight on CUDA as on CPU.
+
+    Covers the GPU side of the same artifact: every ``lora_b @ lora_a`` product,
+    including the reordered q/k/v slices, is recomputed on CUDA and compared
+    against the CPU result, so a GPU deployment cannot silently diverge.
+    """
+    path = tmp_path / "minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors"
+    _write_tiny_native(path)
+
+    loaded = load_minimax_h3_native_lora(
+        partition="fl2va",
+        lora_request=_request(path),
+        lora_path=path,
+        dtype=torch.float32,
+    )
+
+    assert loaded is not None
+    lora_model, _, _ = loaded
+    checked = 0
+    for module_name, a, b in _iter_native_slice_pairs(lora_model):
+        cpu_delta = b @ a
+        cuda_delta = b.cuda() @ a.cuda()
+        assert cuda_delta.device.type == "cuda"
+        torch.testing.assert_close(cuda_delta.cpu(), cpu_delta, msg=f"delta mismatch for {module_name}")
+        checked += 1
+    assert checked == _EXPECTED_NATIVE_SLICE_PAIRS
@@ -74,6 +74,8 @@ def test_t2va_and_fl2va_packing_keep_update_rows_separate():
    assert fl2va["update_mask"].sum().item() == 12
    assert (~fl2va["update_mask"]).sum().item() == 6
    assert t2va["audio_pos"].numel() == fl2va["audio_pos"].numel() == 6
+    assert t2va["video_spans"] == ({"start": 10, "latent_grid": (2, 2, 3), "role": "target"},)
+    assert fl2va["video_spans"] == ({"start": 16, "latent_grid": (2, 2, 3), "role": "target"},)


 def test_ref2va_packing_tracks_video_and_audio_update_masks():
@@ -101,6 +103,33 @@ def test_ref2va_packing_tracks_video_and_audio_update_masks():
    assert packed["cu_seqlens"].tolist() == [0, 30, 64]


+def test_ref2va_packing_publishes_physical_video_spans_only():
+    from vllm_omni.diffusion.models.minimax_h3.packed_sequence import (
+        minimax_h3_packed_sequence_ref2va_blocks,
+    )
+
+    packed = minimax_h3_packed_sequence_ref2va_blocks(
+        text_len=4,
+        latent_t=2,
+        latent_h=4,
+        latent_w=6,
+        audio_t=3,
+        ref_blocks=[
+            {"kind": "image", "latent_h": 4, "latent_w": 4},
+            {"kind": "video_audio", "ref_audio_t": 2, "latent_t": 2, "latent_h": 4, "latent_w": 6},
+            {"kind": "audio", "ref_audio_t": 1},
+            {"kind": "video", "ref_audio_t": 0, "latent_t": 1, "latent_h": 4, "latent_w": 4},
+        ],
+    )
+
+    # Image and all audio rows intentionally stay out of the sparse spans.
+    assert packed["video_spans"] == (
+        {"start": 12, "latent_grid": (2, 2, 3), "role": "reference"},
+        {"start": 26, "latent_grid": (1, 2, 2), "role": "reference"},
+        {"start": 36, "latent_grid": (2, 2, 3), "role": "target"},
+    )
+
+
 def test_condition_noise_is_seeded_and_keeps_clean_anchor_at_timestep_one():
    from vllm_omni.diffusion.models.minimax_h3.condition_noise import (
        minimax_h3_audio_cond_noise_aug_rows,
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 import pytest
 import torch
@@ -113,6 +113,146 @@ def test_h3_rope_table_materializes_local_rows_in_fused_kernel_layout():
    torch.testing.assert_close(actual, expected, atol=0, rtol=0)


+def test_h3_prepared_rope_table_is_rank_local_and_validated(monkeypatch):
+    from vllm_omni.diffusion.models.minimax_h3 import minimax_h3_transformer as h3
+
+    class Rope(nn.Module):
+        def forward(self, position_ids):
+            freqs_half = position_ids[0].float()
+            return torch.cat((freqs_half, freqs_half), dim=-1)
+
+    model = object.__new__(h3.MiniMaxH3DiTModel)
+    nn.Module.__init__(model)
+    model.arch = h3.MiniMaxH3DiTArchConfig(rope_inv_freq_len=1)
+    model.rope = Rope()
+    model.local_sp_prepare = nn.Identity()
+    monkeypatch.setattr(h3, "_sequence_parallel_local_span", lambda *args, **kwargs: (1, 2))
+
+    position_ids = torch.arange(12, dtype=torch.long).reshape(1, 4, 3)
+    actual = model.prepare_rope_table(position_ids, seq_len=4)
+    expected = h3._build_rope_table(model.rope(position_ids[:, 1:3]))
+
+    torch.testing.assert_close(actual, expected, atol=0, rtol=0)
+    model._validate_prepared_rope_table(actual, local_len=2, device=torch.device("cpu"))
+    with pytest.raises(ValueError, match="local_seq_len"):
+        model._validate_prepared_rope_table(actual[:1], local_len=2, device=torch.device("cpu"))
+
+
+def test_denoise_branch_prepares_rope_table_once_per_npu_run(monkeypatch: pytest.MonkeyPatch):
+    from vllm_omni.diffusion.models.minimax_h3 import denoise_loop
+    from vllm_omni.diffusion.models.minimax_h3.denoise_loop import (
+        MiniMaxH3DenoiseBranch,
+        minimax_h3_denoise_loop,
+    )
+
+    monkeypatch.setattr(denoise_loop.current_omni_platform, "is_npu", lambda: True)
+
+    packed = {
+        "seq_len": torch.tensor(3),
+        "img_pos": torch.tensor([1]),
+        "audio_pos": torch.tensor([2]),
+        "text_pos": torch.tensor([0]),
+        "update_mask": torch.tensor([True]),
+        "cu_seqlens": torch.tensor([0, 3, 3], dtype=torch.int32),
+        "img_position_ids": torch.zeros(3, 3, dtype=torch.long),
+        "latent_grid": torch.tensor([1, 1, 1]),
+        "video_row_start": torch.tensor(1),
+    }
+    branch = MiniMaxH3DenoiseBranch(
+        packed=packed,
+        text_embeddings=torch.zeros(1, 2),
+        token_tags=torch.zeros(3, dtype=torch.long),
+        device=torch.device("cpu"),
+    )
+
+    class Model:
+        def __init__(self):
+            self.rope_table = torch.zeros(3, 6, dtype=torch.bfloat16)
+            self.prepare_calls = 0
+            self.forward_calls = 0
+
+        def prepare_rope_table(self, img_position_ids, *, seq_len):
+            assert img_position_ids is branch.static_kwargs["img_position_ids"]
+            assert seq_len == 3
+            self.prepare_calls += 1
+            return self.rope_table
+
+        def __call__(self, **kwargs):
+            assert kwargs["rope_table"] is self.rope_table
+            self.forward_calls += 1
+            return torch.zeros(1, 96), torch.zeros(1, 32)
+
+    model = Model()
+    minimax_h3_denoise_loop(
+        model=model,
+        positive=branch,
+        initial_video_rows=torch.zeros(1, 96),
+        initial_audio_rows=torch.zeros(1, 32),
+        keyframe_cond_rows=None,
+        sigmas_video=[1.0, 0.5, 0.0],
+        sigmas_audio=[1.0, 0.5, 0.0],
+        device=torch.device("cpu"),
+    )
+
+    assert model.prepare_calls == 1
+    assert model.forward_calls == 2
+
+
+def test_denoise_branch_keeps_rope_construction_on_the_reference_path_off_npu(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    from vllm_omni.diffusion.models.minimax_h3 import denoise_loop
+    from vllm_omni.diffusion.models.minimax_h3.denoise_loop import (
+        MiniMaxH3DenoiseBranch,
+        minimax_h3_denoise_loop,
+    )
+
+    monkeypatch.setattr(denoise_loop.current_omni_platform, "is_npu", lambda: False)
+    packed = {
+        "seq_len": torch.tensor(3),
+        "img_pos": torch.tensor([1]),
+        "audio_pos": torch.tensor([2]),
+        "text_pos": torch.tensor([0]),
+        "update_mask": torch.tensor([True]),
+        "cu_seqlens": torch.tensor([0, 3, 3], dtype=torch.int32),
+        "img_position_ids": torch.zeros(3, 3, dtype=torch.long),
+        "latent_grid": torch.tensor([1, 1, 1]),
+        "video_row_start": torch.tensor(1),
+    }
+    branch = MiniMaxH3DenoiseBranch(
+        packed=packed,
+        text_embeddings=torch.zeros(1, 2),
+        token_tags=torch.zeros(3, dtype=torch.long),
+        device=torch.device("cpu"),
+    )
+
+    class Model:
+        def __init__(self):
+            self.forward_calls = 0
+
+        def prepare_rope_table(self, *_args, **_kwargs):
+            raise AssertionError("the reference path must not pre-build a rope_table")
+
+        def __call__(self, **kwargs):
+            assert "rope_table" not in kwargs
+            self.forward_calls += 1
+            return torch.zeros(1, 96), torch.zeros(1, 32)
+
+    model = Model()
+    minimax_h3_denoise_loop(
+        model=model,
+        positive=branch,
+        initial_video_rows=torch.zeros(1, 96),
+        initial_audio_rows=torch.zeros(1, 32),
+        keyframe_cond_rows=None,
+        sigmas_video=[1.0, 0.5, 0.0],
+        sigmas_audio=[1.0, 0.5, 0.0],
+        device=torch.device("cpu"),
+    )
+
+    assert model.forward_calls == 2
+
+
 def test_strict_sp_local_span_uses_rank_owned_rows(monkeypatch):
    from vllm_omni.diffusion import forward_context
    from vllm_omni.diffusion.distributed import parallel_state
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 from pathlib import Path

@@ -187,6 +187,8 @@ def test_minimax_h3_quantization_quality(config: QualityTestConfig):
        "model": model_ref,
        "enforce_eager": True,
        "tensor_parallel_size": 2,
+        # The fused BF16 baseline only fits on H100-80GB with encoder TP.
+        "text_encoder_tp_size": 2,
        "vae_use_tiling": True,
    }

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 import json
 from types import SimpleNamespace

@@ -245,6 +245,160 @@ class TestDistributedLayerwiseOffloadHook:
        )
        torch.testing.assert_close(rank0_block.weight_scale, expected_scale)

+    def test_allgather_reconstructs_online_int8_weight_and_scale(
+        self,
+        patched_offload_runtime,
+        monkeypatch,
+    ):
+        """DLO must reconstruct finalized online-INT8 tensors byte-for-byte.
+
+        One block carries both finalized online-INT8 layouts: the NPU method
+        leaves a contiguous pre-transposed (K, N) int8 weight with a 1-D fp32
+        per-channel scale, while the CUDA method leaves a transposed-view
+        weight (stride (1, K)) with a scalar fp32 scale.  The int8 and fp32
+        dtype groups must survive the flat-shard AllGather independently.
+        """
+
+        class OnlineInt8Block(nn.Module):
+            def __init__(self):
+                super().__init__()
+                logical = torch.arange(1.0, 13.0).reshape(3, 4)
+                # NPU online int8: qweight.t().contiguous() -> (K, N), 1-D scale
+                self.npu_weight = nn.Parameter(logical.t().contiguous().to(torch.int8), requires_grad=False)
+                self.npu_weight_scale = nn.Parameter(torch.tensor([2.0, 3.0, 4.0]), requires_grad=False)
+                # CUDA online int8: Parameter(qweight.t().data) -> transposed
+                # view (stride (1, K)) over the (N, K) storage, scalar scale
+                self.cuda_weight = nn.Parameter((logical + 100).to(torch.int8).t(), requires_grad=False)
+                self.cuda_weight_scale = nn.Parameter(torch.tensor([0.5]), requires_grad=False)
+
+        current_block = OnlineInt8Block()
+        rank0_block = OnlineInt8Block()
+        rank1_block = OnlineInt8Block()
+        expected = {name: tensor.detach().clone() for name, tensor in rank0_block.named_parameters()}
+        expected_strides = {name: tensor.stride() for name, tensor in rank0_block.named_parameters()}
+
+        def make_hook(block: OnlineInt8Block, rank: int) -> DistributedLayerwiseOffloadHook:
+            return DistributedLayerwiseOffloadHook(
+                next_block=block,
+                device=torch.device("cpu"),
+                dp_group=object(),
+                dp_size=2,
+                rank=rank,
+                copy_stream=DummyStream(),
+                comm_stream=DummyStream(),
+                pin_memory=False,
+            )
+
+        rank0_hook = make_hook(rank0_block, 0)
+        rank1_hook = make_hook(rank1_block, 1)
+        rank0_hook.initialize_hook(current_block)
+        rank1_hook.initialize_hook(OnlineInt8Block())
+        rank0_hook.gpu_shard_buffers = [
+            {dtype: torch.empty_like(shard) for dtype, shard in rank0_hook.cpu_shards.items()} for _ in range(2)
+        ]
+
+        def fake_allgather(output, local_shard, *, group):
+            del group
+            remote_shard = rank1_hook.cpu_shards[local_shard.dtype]
+            shard_size = local_shard.numel()
+            output[:shard_size].copy_(local_shard)
+            output[shard_size : 2 * shard_size].copy_(remote_shard)
+
+        monkeypatch.setattr(torch.distributed, "all_gather_into_tensor", fake_allgather)
+        rank0_hook.prefetch_layer(slot=0, non_blocking=False)
+
+        for name, tensor in rank0_block.named_parameters():
+            assert tensor.stride() == expected_strides[name]
+            torch.testing.assert_close(
+                tensor.float(),
+                expected[name].float(),
+                rtol=0,
+                atol=0,
+            )
+
+    def test_allgather_reconstructs_online_mxfp8_weight_and_scale(
+        self,
+        patched_offload_runtime,
+        monkeypatch,
+    ):
+        """DLO must reconstruct finalized online-MXFP8 tensors byte-for-byte.
+
+        One block carries both finalized online-MXFP8 layouts: the NPU method
+        leaves a contiguous pre-transposed (K, N) fp8 weight with a contiguous
+        (K_groups/2, N, 2) e8m0 block scale, while the vLLM-kernel method
+        (XPU) leaves an (N, K) fp8 weight whose e8m0 scale is a transposed
+        view over a contiguous (K/32, N) buffer.  The fp8 and e8m0 dtype
+        groups must survive the flat-shard AllGather independently.
+        """
+
+        class OnlineMxfp8Block(nn.Module):
+            def __init__(self):
+                super().__init__()
+                logical = torch.arange(1.0, 13.0).reshape(3, 4)
+                # Build scales from raw e8m0 bit patterns (1 byte/elem).
+                npu_scale_bits = torch.arange(1, 13, dtype=torch.uint8).reshape(2, 3, 2)
+                kernel_scale_bits = torch.arange(1, 7, dtype=torch.uint8).reshape(2, 3)
+                # NPU online mxfp8: (N, K).t().contiguous() -> (K, N) fp8
+                # weight, contiguous (K_groups/2, N, 2) e8m0 scale
+                self.npu_weight = nn.Parameter(logical.t().contiguous().to(torch.float8_e4m3fn), requires_grad=False)
+                self.npu_weight_scale = nn.Parameter(npu_scale_bits.view(torch.float8_e8m0fnu), requires_grad=False)
+                # vLLM-kernel online mxfp8 (XPU): (N, K) fp8 weight, e8m0
+                # scale kept as a transposed view (stride (1, K/32)) over the
+                # contiguous (K/32, N) storage
+                self.kernel_weight = nn.Parameter((logical + 100).to(torch.float8_e4m3fn), requires_grad=False)
+                self.kernel_weight_scale = nn.Parameter(
+                    kernel_scale_bits.view(torch.float8_e8m0fnu).t(), requires_grad=False
+                )
+
+        current_block = OnlineMxfp8Block()
+        rank0_block = OnlineMxfp8Block()
+        rank1_block = OnlineMxfp8Block()
+        expected = {name: tensor.detach().clone() for name, tensor in rank0_block.named_parameters()}
+        expected_strides = {name: tensor.stride() for name, tensor in rank0_block.named_parameters()}
+
+        def make_hook(block: OnlineMxfp8Block, rank: int) -> DistributedLayerwiseOffloadHook:
+            return DistributedLayerwiseOffloadHook(
+                next_block=block,
+                device=torch.device("cpu"),
+                dp_group=object(),
+                dp_size=2,
+                rank=rank,
+                copy_stream=DummyStream(),
+                comm_stream=DummyStream(),
+                pin_memory=False,
+            )
+
+        rank0_hook = make_hook(rank0_block, 0)
+        rank1_hook = make_hook(rank1_block, 1)
+        rank0_hook.initialize_hook(current_block)
+        rank1_hook.initialize_hook(OnlineMxfp8Block())
+        rank0_hook.gpu_shard_buffers = [
+            {dtype: torch.empty_like(shard) for dtype, shard in rank0_hook.cpu_shards.items()} for _ in range(2)
+        ]
+
+        def fake_allgather(output, local_shard, *, group):
+            del group
+            remote_shard = rank1_hook.cpu_shards[local_shard.dtype]
+            shard_size = local_shard.numel()
+            output[:shard_size].copy_(local_shard)
+            output[shard_size : 2 * shard_size].copy_(remote_shard)
+
+        monkeypatch.setattr(torch.distributed, "all_gather_into_tensor", fake_allgather)
+        rank0_hook.prefetch_layer(slot=0, non_blocking=False)
+
+        for name, tensor in rank0_block.named_parameters():
+            assert tensor.stride() == expected_strides[name]
+            # Compare raw bytes in logical order: float8/e8m0 CPU upcasts are
+            # version-sensitive, while contiguous().view(uint8) is an exact
+            # byte-level comparison for both contiguous and transposed-view
+            # parameters.
+            torch.testing.assert_close(
+                tensor.detach().contiguous().view(torch.uint8),
+                expected[name].contiguous().view(torch.uint8),
+                rtol=0,
+                atol=0,
+            )
+
    def test_dtensor_wrapper_preserved_across_prefetch_and_offload(self, dist_group, patched_offload_runtime):
        """DTensor wrapper should be preserved through prefetch/offload cycle."""
        current_block = TinyBlock(_make_values(1.0))
@@ -81,7 +81,7 @@ class QualityTestConfig:
    num_frames: int = 5  # only for t2v
    seed: int = 42
    gpu: str = "H100"  # minimum GPU requirement
-    negative_prompt: str = ""
+    negative_prompt: str | None = ""
    guidance_scale: float | None = None
    sigmas: list[float] | None = None
    enable_cpu_offload: bool = False
@@ -212,24 +212,14 @@ QUALITY_CONFIGS = [
        quantization="fp8",
        task="t2v",
        prompt="A serene lakeside sunrise with mist over the water",
-        max_lpips=0.10,
-        height=256,
-        width=256,
-        num_frames=25,
-        num_inference_steps=8,
-        # Preserve the final scheduler trajectory used when this FP8 quality
-        # gate was established. Explicit LTX sigmas bypass dynamic shifting.
-        sigmas=[
-            1.0,
-            0.92185378074646,
-            0.8327768445014954,
-            0.7303033471107483,
-            0.6111654043197632,
-            0.4709382653236389,
-            0.30347853899002075,
-            0.10000002384185791,
-            0.0,
-        ],
+        max_lpips=0.20,
+        height=384,
+        width=512,
+        num_frames=73,
+        num_inference_steps=10,
+        # Do not override the recipe's negative conditioning, guidance, or
+        # scheduler trajectory: this gate follows the supported LTX default.
+        negative_prompt=None,
    ),
 ]

@@ -296,8 +286,11 @@ def _generate_image(omni, config: QualityTestConfig):
    ).manual_seed(config.seed)
    torch.accelerator.reset_peak_memory_stats()

+    prompt = {"prompt": config.prompt}
+    if config.negative_prompt is not None:
+        prompt["negative_prompt"] = config.negative_prompt
    outputs = omni.generate(
-        {"prompt": config.prompt, "negative_prompt": config.negative_prompt},
+        prompt,
        OmniDiffusionSamplingParams(
            height=config.height,
            width=config.width,
@@ -328,8 +321,11 @@ def _generate_video(omni, config: QualityTestConfig):
    ).manual_seed(config.seed)
    torch.accelerator.reset_peak_memory_stats()

+    prompt = {"prompt": config.prompt}
+    if config.negative_prompt is not None:
+        prompt["negative_prompt"] = config.negative_prompt
    outputs = omni.generate(
-        {"prompt": config.prompt, "negative_prompt": config.negative_prompt},
+        prompt,
        OmniDiffusionSamplingParams(
            height=config.height,
            width=config.width,
@@ -465,6 +461,7 @@ def test_benchmark_generate_image_unwraps_nested_omni_request_output(monkeypatch


 def test_generate_video_forwards_sigmas(monkeypatch):
+    from vllm_omni.outputs import OmniRequestOutput
    from vllm_omni.platforms import current_omni_platform

    monkeypatch.setattr(current_omni_platform, "device_type", "cpu", raising=False)
@@ -475,7 +472,12 @@ def test_generate_video_forwards_sigmas(monkeypatch):
    class DummyOmni:
        def generate(self, _prompt, sampling):
            captured.sampling = sampling
-            return [SimpleNamespace(images=[np.zeros((1, 2, 2, 3), dtype=np.float32)])]
+            return [
+                OmniRequestOutput.from_diffusion(
+                    request_id="req",
+                    images=[np.zeros((1, 2, 2, 3), dtype=np.float32)],
+                )
+            ]

    config = QualityTestConfig(
        id="ltx-sigmas",
@@ -491,6 +493,41 @@ def test_generate_video_forwards_sigmas(monkeypatch):
    assert captured.sampling.sigmas == [1.0, 0.5]


+@pytest.mark.core_model
+@pytest.mark.cpu
+def test_ltx_quality_gate_uses_official_eager_defaults(monkeypatch):
+    from vllm_omni.outputs import OmniRequestOutput
+    from vllm_omni.platforms import current_omni_platform
+
+    monkeypatch.setattr(current_omni_platform, "device_type", "cpu", raising=False)
+    monkeypatch.setattr(torch.accelerator, "reset_peak_memory_stats", lambda: None, raising=False)
+    monkeypatch.setattr(torch.accelerator, "max_memory_allocated", lambda: 0, raising=False)
+    captured = SimpleNamespace(prompt=None, sampling=None)
+
+    class DummyOmni:
+        def generate(self, prompt, sampling):
+            captured.prompt = prompt
+            captured.sampling = sampling
+            return [
+                OmniRequestOutput.from_diffusion(
+                    request_id="req",
+                    images=[np.zeros((1, 2, 2, 3), dtype=np.float32)],
+                )
+            ]
+
+    config = next(config for config in QUALITY_CONFIGS if config.id == "fp8_ltx2")
+    _generate_video(DummyOmni(), config)
+
+    assert (config.width, config.height, config.num_frames, config.num_inference_steps) == (512, 384, 73, 10)
+    assert config.max_lpips == 0.20
+    assert config.sigmas is None
+    assert config.guidance_scale is None
+    assert config.negative_prompt is None
+    assert "negative_prompt" not in captured.prompt
+    assert captured.sampling.sigmas is None
+    assert captured.sampling.guidance_scale is None
+
+
 _marks = hardware_marks(res={"cuda": "H100"})
 _OUTPUT_DIR = Path(os.environ["VLLM_OMNI_QUALITY_OUTPUT_DIR"]) if "VLLM_OMNI_QUALITY_OUTPUT_DIR" in os.environ else None

@@ -501,8 +538,6 @@ def _quality_param(c: QualityTestConfig):
        marks.append(
            pytest.mark.skip(reason="Qwen-Image FP8 quality gate temporarily disabled (see CI / issue tracker).")
        )
-    if c.id == "fp8_ltx2":
-        marks.append(pytest.mark.skip(reason="https://github.com/vllm-project/vllm-omni/issues/6245"))
    return pytest.param(c, id=c.id, marks=marks)


@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 import asyncio
 import queue
@@ -197,6 +197,59 @@ class TestRequestBatchCapability:

        assert diffusion_engine_module.supports_request_batch(od_config) is True

+    def test_engine_rejects_multi_seq_diffusers_using_adapter_capability(
+        self,
+        monkeypatch: pytest.MonkeyPatch,
+        mocker: MockerFixture,
+    ) -> None:
+        od_config = SimpleNamespace(
+            model_class_name="QwenImagePipeline",
+            custom_pipeline_args=None,
+            diffusion_load_format="diffusers",
+            streaming_output=False,
+            max_num_seqs=2,
+        )
+        registry_load = mocker.Mock(
+            side_effect=lambda model_class_name: (
+                _SingleRequestPipeline if model_class_name == "DiffusersAdapterPipeline" else _BatchCapablePipeline
+            )
+        )
+        monkeypatch.setattr(
+            diffusion_engine_module.DiffusionModelRegistry,
+            "_try_load_model_cls",
+            registry_load,
+        )
+
+        engine = DiffusionEngine.__new__(DiffusionEngine)
+        with pytest.raises(ValueError, match="max_num_seqs=1"):
+            engine._resolve_execution_mode(od_config)
+
+        registry_load.assert_called_once_with("DiffusersAdapterPipeline")
+
+    def test_engine_prefers_batch_capable_custom_pipeline_over_diffusers_adapter(
+        self,
+        monkeypatch: pytest.MonkeyPatch,
+        mocker: MockerFixture,
+    ) -> None:
+        od_config = SimpleNamespace(
+            model_class_name="QwenImagePipeline",
+            custom_pipeline_args={"pipeline_class": _BatchCapablePipeline},
+            diffusion_load_format="diffusers",
+            streaming_output=False,
+            max_num_seqs=2,
+        )
+        registry_load = mocker.Mock(return_value=_SingleRequestPipeline)
+        monkeypatch.setattr(
+            diffusion_engine_module.DiffusionModelRegistry,
+            "_try_load_model_cls",
+            registry_load,
+        )
+
+        engine = DiffusionEngine.__new__(DiffusionEngine)
+
+        assert engine._resolve_execution_mode(od_config) is DiffusionExecutionMode.REQUEST_BATCH
+        registry_load.assert_not_called()
+
    def test_supports_request_batch_uses_custom_pipeline_class(self, monkeypatch: pytest.MonkeyPatch) -> None:
        od_config = SimpleNamespace(
            model_class_name="SinglePipeline",
@@ -328,6 +328,25 @@ def test_formatter_preserves_audio_model_video_audio_and_actions(
    }


+def test_formatter_uses_native_model_identity_for_diffusers_video_frames(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(output_formatter, "supports_audio_output", lambda _: False)
+    frames = [["frame-0", "frame-1"]]
+    postprocess_output = normalize_diffusion_postprocess_output(frames)
+
+    [result] = format_diffusion_outputs(
+        request=_request("make a video"),
+        od_config=_config("WanImageToVideoPipeline"),
+        diffusion_output=DiffusionOutput(output=frames),
+        output_data=frames,
+        postprocess_output=postprocess_output,
+    )
+
+    assert result.images == frames
+    assert result.final_output_type == "video"
+
+
 def test_formatter_preserves_audio_only_postprocess_dict(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """
 Unit tests for diffusion engine plugin extensibility hooks.
@@ -20,6 +20,8 @@ from vllm_omni.diffusion.registry import (
    _DIFFUSION_MODELS,
    _DIFFUSION_POST_PROCESS_FUNCS,
    _DIFFUSION_PRE_PROCESS_FUNCS,
+    get_diffusion_post_process_func,
+    get_diffusion_pre_process_func,
    register_diffusion_model,
 )
 from vllm_omni.platforms.interface import OmniPlatform, OmniPlatformEnum
@@ -112,6 +114,49 @@ class TestRegisterDiffusionModel:
        assert _DIFFUSION_POST_PROCESS_FUNCS["LegacyActionPipeline"] == "test_post_process"


+@pytest.mark.parametrize(
+    ("model_class_name", "get_process_func"),
+    [
+        ("QwenImagePipeline", get_diffusion_post_process_func),
+        ("WanImageToVideoPipeline", get_diffusion_post_process_func),
+        ("WanImageToVideoPipeline", get_diffusion_pre_process_func),
+    ],
+)
+def test_diffusers_backend_skips_native_process_hooks(model_class_name, get_process_func):
+    od_config = SimpleNamespace(
+        model_class_name=model_class_name,
+        custom_pipeline_args=None,
+        diffusion_load_format="diffusers",
+    )
+
+    with patch("vllm_omni.diffusion.registry._load_process_func") as load_process_func:
+        assert get_process_func(od_config) is None
+
+    load_process_func.assert_not_called()
+
+
+@pytest.mark.parametrize(
+    ("model_class_name", "get_process_func"),
+    [
+        ("QwenImagePipeline", get_diffusion_post_process_func),
+        ("WanImageToVideoPipeline", get_diffusion_post_process_func),
+        ("WanImageToVideoPipeline", get_diffusion_pre_process_func),
+    ],
+)
+def test_custom_pipeline_preserves_native_process_hooks(model_class_name, get_process_func):
+    process_func = Mock()
+    od_config = SimpleNamespace(
+        model_class_name=model_class_name,
+        custom_pipeline_args={"pipeline_class": "custom.Pipeline"},
+        diffusion_load_format="diffusers",
+    )
+
+    with patch("vllm_omni.diffusion.registry._load_process_func", return_value=process_func) as load_process_func:
+        assert get_process_func(od_config) is process_func
+
+    load_process_func.assert_called_once()
+
+
 class TestWorkerUsesHook:
    """Test that DiffusionWorker resolves model runner via platform hook."""

@@ -0,0 +1,255 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+import importlib.util
+import shutil
+import subprocess
+import sys
+import textwrap
+import time
+from dataclasses import dataclass
+from unittest.mock import Mock
+
+import pytest
+from pytest_mock import MockerFixture
+
+import vllm_omni.diffusion.worker.diffusion_worker as worker_module
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+
+@dataclass(frozen=True)
+class _FakeGroup:
+    world_size: int = 1
+    rank_in_group: int = 0
+
+
+_GROUP_GETTERS = (
+    "get_dp_group",
+    "get_pp_group",
+    "get_sp_group",
+    "get_cfg_group",
+    "get_tp_group",
+    "get_fs_group",
+    "get_hsdp_replicate_group",
+    "get_ep_group",
+)
+
+
+def _patch_groups(
+    mocker: MockerFixture,
+    groups: dict[str, _FakeGroup],
+) -> dict[str, Mock]:
+    patched_getters = {}
+    for getter_name in _GROUP_GETTERS:
+        patched_getters[getter_name] = mocker.patch.object(
+            worker_module,
+            getter_name,
+            return_value=groups.get(getter_name, _FakeGroup()),
+        )
+    return patched_getters
+
+
+def test_setup_uses_default_name_before_model_parallel_init(
+    mocker: MockerFixture,
+) -> None:
+    mocker.patch.object(
+        worker_module,
+        "model_parallel_is_initialized",
+        return_value=False,
+    )
+    group_getters = _patch_groups(mocker, {})
+    set_process_title = mocker.patch.object(worker_module, "set_process_title")
+    decorate_logs = mocker.patch.object(worker_module, "decorate_logs")
+
+    worker_module._setup_diffusion_worker_proc_title_and_log_prefix(enable_ep=True, use_hsdp=True)
+
+    set_process_title.assert_called_once_with(
+        name="DiffusionWorker",
+        prefix="vLLM-Omni",
+    )
+    decorate_logs.assert_called_once_with("DiffusionWorker")
+    for group_getter in group_getters.values():
+        group_getter.assert_not_called()
+
+
+@pytest.mark.parametrize(
+    ("groups", "enable_ep", "use_hsdp", "hsdp_replicate_size", "expected_name"),
+    [
+        ({}, False, False, 1, "DiffusionWorker"),
+        ({"get_tp_group": _FakeGroup(2, 1)}, False, False, 1, "DiffusionWorker_TP1"),
+        ({"get_dp_group": _FakeGroup(4, 2)}, False, False, 1, "DiffusionWorker_DP2"),
+        ({"get_pp_group": _FakeGroup(2, 1)}, False, False, 1, "DiffusionWorker_PP1"),
+        (
+            {
+                "get_cfg_group": _FakeGroup(2, 1),
+                "get_tp_group": _FakeGroup(2, 1),
+            },
+            False,
+            False,
+            1,
+            "DiffusionWorker_CFG1_TP1",
+        ),
+        (
+            {
+                "get_dp_group": _FakeGroup(2, 1),
+                "get_sp_group": _FakeGroup(4, 2),
+                "get_tp_group": _FakeGroup(2, 1),
+            },
+            False,
+            False,
+            1,
+            "DiffusionWorker_DP1_SP2_TP1",
+        ),
+        (
+            {"get_fs_group": _FakeGroup(4, 3)},
+            False,
+            True,
+            1,
+            "DiffusionWorker_FS3",
+        ),
+        (
+            {
+                "get_sp_group": _FakeGroup(2, 1),
+                "get_fs_group": _FakeGroup(4, 3),
+            },
+            False,
+            True,
+            1,
+            "DiffusionWorker_SP1_FS3",
+        ),
+        (
+            {
+                "get_fs_group": _FakeGroup(1, 0),
+                "get_hsdp_replicate_group": _FakeGroup(2, 1),
+            },
+            False,
+            True,
+            2,
+            "DiffusionWorker_RP1",
+        ),
+        ({"get_ep_group": _FakeGroup(4, 2)}, True, False, 1, "DiffusionWorker_EP2"),
+        ({}, True, False, 1, "DiffusionWorker"),
+    ],
+    ids=[
+        "all-singleton",
+        "tp-only",
+        "dp-only",
+        "pp-only",
+        "tp-cfg",
+        "dp-sp-tp",
+        "hsdp-only",
+        "hsdp-sp",
+        "hsdp-replicated",
+        "expert-parallel",
+        "singleton-ep",
+    ],
+)
+def test_setup_uses_initialized_parallel_groups(
+    mocker: MockerFixture,
+    groups: dict[str, _FakeGroup],
+    enable_ep: bool,
+    use_hsdp: bool,
+    hsdp_replicate_size: int,
+    expected_name: str,
+) -> None:
+    mocker.patch.object(
+        worker_module,
+        "model_parallel_is_initialized",
+        return_value=True,
+    )
+    group_getters = _patch_groups(mocker, groups)
+    set_process_title = mocker.patch.object(worker_module, "set_process_title")
+    decorate_logs = mocker.patch.object(worker_module, "decorate_logs")
+
+    worker_module._setup_diffusion_worker_proc_title_and_log_prefix(
+        enable_ep=enable_ep,
+        use_hsdp=use_hsdp,
+        hsdp_replicate_size=hsdp_replicate_size,
+    )
+
+    set_process_title.assert_called_once_with(
+        name=expected_name,
+        prefix="vLLM-Omni",
+    )
+    decorate_logs.assert_called_once_with(expected_name)
+    if use_hsdp:
+        group_getters["get_fs_group"].assert_called_once_with()
+        if hsdp_replicate_size > 1:
+            group_getters["get_hsdp_replicate_group"].assert_called_once_with()
+        else:
+            group_getters["get_hsdp_replicate_group"].assert_not_called()
+    else:
+        group_getters["get_fs_group"].assert_not_called()
+        group_getters["get_hsdp_replicate_group"].assert_not_called()
+    if enable_ep:
+        group_getters["get_ep_group"].assert_called_once_with()
+    else:
+        group_getters["get_ep_group"].assert_not_called()
+
+
+def test_missing_setproctitle_is_non_fatal(mocker: MockerFixture) -> None:
+    mocker.patch.object(
+        worker_module,
+        "model_parallel_is_initialized",
+        return_value=False,
+    )
+    mocker.patch.dict(sys.modules, {"setproctitle": None})
+    decorate_logs = mocker.patch.object(worker_module, "decorate_logs")
+
+    worker_module._setup_diffusion_worker_proc_title_and_log_prefix(enable_ep=False, use_hsdp=False)
+
+    decorate_logs.assert_called_once_with("DiffusionWorker")
+
+
+@pytest.mark.skipif(
+    sys.platform != "linux" or shutil.which("ps") is None or importlib.util.find_spec("setproctitle") is None,
+    reason="requires Linux ps and setproctitle",
+)
+def test_process_title_is_visible_through_ps() -> None:
+    expected_name = "vLLM-Omni::DiffusionWorker_TP1"
+    child_script = textwrap.dedent(
+        """
+        import time
+
+        from vllm.utils.system_utils import set_process_title
+
+
+        set_process_title(
+            name="DiffusionWorker_TP1",
+            prefix="vLLM-Omni",
+        )
+        time.sleep(30)
+        """
+    )
+    process = subprocess.Popen(
+        [sys.executable, "-c", child_script],
+        stderr=subprocess.PIPE,
+        text=True,
+    )
+
+    observed_title = ""
+    try:
+        for _ in range(100):
+            if process.poll() is not None:
+                stderr = process.stderr.read() if process.stderr is not None else ""
+                pytest.fail(f"child process exited before title check: {stderr}")
+            result = subprocess.run(
+                ["ps", "-o", "args=", "-p", str(process.pid)],
+                check=False,
+                capture_output=True,
+                text=True,
+            )
+            observed_title = result.stdout.strip()
+            if expected_name in observed_title:
+                break
+            time.sleep(0.1)
+        else:
+            pytest.fail(f"expected {expected_name!r} in process title, got {observed_title!r}")
+    finally:
+        process.terminate()
+        try:
+            process.wait(timeout=5)
+        except subprocess.TimeoutExpired:
+            process.kill()
+            process.wait(timeout=5)
@@ -57,6 +57,19 @@ def _req(req_id: str, status: RequestStatus, external_req_id: str | None = None)
    return request


+def _streaming_request(mocker: MockerFixture, num_computed_tokens: int) -> SimpleNamespace:
+    prompt_token_ids = [0] * num_computed_tokens
+    return SimpleNamespace(
+        _all_token_ids=prompt_token_ids.copy(),
+        _output_token_ids=[],
+        prompt_token_ids=prompt_token_ids,
+        num_computed_tokens=num_computed_tokens,
+        num_prompt_tokens=num_computed_tokens,
+        num_output_placeholders=0,
+        update_block_hashes=mocker.Mock(),
+    )
+
+
 def test_streaming_payload_can_replace_placeholder_prompt(mocker: MockerFixture) -> None:
    request = SimpleNamespace(
        _all_token_ids=[0, 0, 7, 8],
@@ -84,30 +97,219 @@ def test_streaming_payload_can_replace_placeholder_prompt(mocker: MockerFixture)
    request.update_block_hashes.assert_called_once_with()


-def test_streaming_payload_can_append_exact_prompt_length(mocker: MockerFixture) -> None:
+def test_turn_start_replacement_ignores_accumulated_prompt_capacity(mocker: MockerFixture) -> None:
+    request = _streaming_request(mocker, num_computed_tokens=4064)
+    payload = {
+        "ids": {"prompt": [1]},
+        "meta": {
+            "replace_streaming_prompt": True,
+            "next_stage_prompt_len": 10,
+            "next_stage_generation_tokens": 26,
+        },
+    }
+
+    replaced = construct_next_stage_streaming_input_prompt(
+        payload,
+        request,
+        max_model_len=4096,
+        previous_condition_len=10,
+        previous_condition_seq=7,
+        condition_seq=8,
+        recompute_previous_chunks=1,
+    )
+
+    assert replaced is True
+    assert request.prompt_token_ids == [0] * 10
+    assert payload["meta"]["streaming_prompt_recompute"] is False
+    assert "streaming_prompt_previous_codes" not in payload["ids"]
+
+
+@pytest.mark.parametrize("reserve", [True, 0, -1, 1.5, "26", None])
+def test_streaming_prompt_rejects_invalid_generation_reserve(
+    mocker: MockerFixture,
+    reserve: object,
+) -> None:
+    request = _streaming_request(mocker, num_computed_tokens=10)
+    payload = {
+        "ids": {"prompt": [1]},
+        "meta": {
+            "next_stage_prompt_len": 10,
+            "next_stage_generation_tokens": reserve,
+        },
+    }
+
+    with pytest.raises(ValueError, match="next_stage_generation_tokens must be a positive integer"):
+        construct_next_stage_streaming_input_prompt(payload, request, max_model_len=4096)
+
+
+def test_capacity_managed_streaming_prompt_at_limit_still_appends(mocker: MockerFixture) -> None:
+    request = _streaming_request(mocker, num_computed_tokens=4060)
+    payload = {
+        "ids": {"prompt": [1]},
+        "meta": {
+            "next_stage_prompt_len": 10,
+            "next_stage_generation_tokens": 26,
+        },
+    }
+
+    replaced = construct_next_stage_streaming_input_prompt(payload, request, max_model_len=4096)
+
+    assert replaced is False
+    assert request.num_computed_tokens == 4060
+    assert request.num_prompt_tokens == 4070
+
+
+def test_streaming_window_builds_one_chunk_recompute_recipe(mocker: MockerFixture) -> None:
+    previous_condition_len = 10
+    previous_codes = list(range(25))
    request = SimpleNamespace(
-        _all_token_ids=[0, 0, 7, 8],
-        _output_token_ids=[7, 8],
-        prompt_token_ids=[0, 0],
-        num_computed_tokens=4,
-        num_prompt_tokens=2,
+        _all_token_ids=[0] * 4039 + previous_codes + [999],
+        _output_token_ids=previous_codes + [999],
+        prompt_token_ids=[0] * 4039,
+        num_computed_tokens=4065,
+        num_prompt_tokens=4039,
+        num_output_placeholders=1,
        update_block_hashes=mocker.Mock(),
    )
    payload = {
-        "ids": {"prompt": [1, 2, 3]},
-        "meta": {"next_stage_prompt_len": 3},
+        "ids": {"prompt": [1]},
+        "meta": {
+            "next_stage_prompt_len": 10,
+            "next_stage_generation_tokens": 26,
+        },
    }

-    construct_next_stage_streaming_input_prompt(payload, request)
+    replaced = construct_next_stage_streaming_input_prompt(
+        payload,
+        request,
+        max_model_len=4096,
+        previous_condition_len=previous_condition_len,
+        previous_condition_seq=7,
+        condition_seq=8,
+        recompute_previous_chunks=1,
+    )

-    assert request.prompt_token_ids == [0, 0, 7, 8, 0, 0, 0]
-    assert request._all_token_ids == [0, 0, 7, 8, 0, 0, 0]
-    assert request._output_token_ids == []
-    assert request.num_computed_tokens == 4
-    assert request.num_prompt_tokens == 7
+    assert replaced is True
+    assert request.prompt_token_ids == [0] * 45
+    assert request.num_prompt_tokens == 45
+    assert request.num_computed_tokens == 0
+    assert payload["ids"]["streaming_prompt_previous_codes"] == previous_codes
+    assert payload["meta"] == {
+        "next_stage_prompt_len": 10,
+        "next_stage_generation_tokens": 26,
+        "streaming_prompt_recompute": True,
+        "streaming_condition_seq": 8,
+    }
    request.update_block_hashes.assert_called_once_with()


+def test_streaming_window_recomputes_each_condition_without_accumulating(mocker: MockerFixture) -> None:
+    request = SimpleNamespace(
+        _all_token_ids=[0] * 10 + [101, 102],
+        _output_token_ids=[101, 102],
+        prompt_token_ids=[0] * 10,
+        num_computed_tokens=12,
+        num_prompt_tokens=10,
+        num_output_placeholders=0,
+        update_block_hashes=mocker.Mock(),
+    )
+    second = {
+        "ids": {"prompt": [1]},
+        "meta": {
+            "next_stage_prompt_len": 12,
+            "next_stage_generation_tokens": 26,
+        },
+    }
+
+    assert construct_next_stage_streaming_input_prompt(
+        second,
+        request,
+        max_model_len=4096,
+        previous_condition_len=10,
+        previous_condition_seq=0,
+        condition_seq=1,
+        recompute_previous_chunks=1,
+    )
+    assert request.num_prompt_tokens == 24
+    assert second["ids"]["streaming_prompt_previous_codes"] == [101, 102]
+
+    request._all_token_ids.extend([201, 202, 203])
+    request._output_token_ids.extend([201, 202, 203])
+    request.num_computed_tokens = 27
+    third = {
+        "ids": {"prompt": [2]},
+        "meta": {
+            "next_stage_prompt_len": 8,
+            "next_stage_generation_tokens": 26,
+        },
+    }
+
+    assert construct_next_stage_streaming_input_prompt(
+        third,
+        request,
+        max_model_len=4096,
+        previous_condition_len=12,
+        previous_condition_seq=1,
+        condition_seq=2,
+        recompute_previous_chunks=1,
+    )
+    assert request.num_prompt_tokens == 23
+    assert third["ids"]["streaming_prompt_previous_codes"] == [201, 202, 203]
+    assert third["meta"]["streaming_condition_seq"] == 2
+
+
+def test_capacity_rollover_requires_explicit_window_contract(mocker: MockerFixture) -> None:
+    request = _streaming_request(mocker, num_computed_tokens=4064)
+    payload = {
+        "ids": {"prompt": [1]},
+        "meta": {
+            "next_stage_prompt_len": 10,
+            "next_stage_generation_tokens": 26,
+        },
+    }
+
+    with pytest.raises(ValueError, match="requires window_size=1"):
+        construct_next_stage_streaming_input_prompt(
+            payload,
+            request,
+            max_model_len=4096,
+            previous_condition_len=10,
+            previous_condition_seq=0,
+            condition_seq=1,
+        )
+
+
+@pytest.mark.parametrize("next_stage_prompt_len", [0, -1])
+def test_capacity_managed_streaming_prompt_rejects_nonpositive_length(
+    mocker: MockerFixture, next_stage_prompt_len: int
+) -> None:
+    request = _streaming_request(mocker, num_computed_tokens=0)
+    payload = {
+        "ids": {"prompt": [1]},
+        "meta": {
+            "next_stage_prompt_len": next_stage_prompt_len,
+            "next_stage_generation_tokens": 26,
+        },
+    }
+
+    with pytest.raises(ValueError, match="positive next_stage_prompt_len"):
+        construct_next_stage_streaming_input_prompt(payload, request, max_model_len=4096)
+
+
+def test_capacity_managed_streaming_prompt_rejects_fresh_overflow(mocker: MockerFixture) -> None:
+    request = _streaming_request(mocker, num_computed_tokens=0)
+    payload = {
+        "ids": {"prompt": [1]},
+        "meta": {
+            "next_stage_prompt_len": 4071,
+            "next_stage_generation_tokens": 26,
+        },
+    }
+
+    with pytest.raises(ValueError, match="exceeds max_model_len"):
+        construct_next_stage_streaming_input_prompt(payload, request, max_model_len=4096)
+
+
@pytest.fixture
 def build_adapter(monkeypatch, mocker: MockerFixture):
    def _build(
@@ -115,6 +317,10 @@ def build_adapter(monkeypatch, mocker: MockerFixture):
        stage_id: int = 1,
        model_mode: str = "ar",
        max_num_seqs: int = 2,
+        max_model_len: int = 0,
+        tts_max_model_len: int = 0,
+        tts_attention_type: str = "full_attention",
+        flat_tts_config: bool = False,
        active_stream_window: int = 0,
        connector_extra: dict | None = None,
    ):
@@ -131,6 +337,10 @@ def build_adapter(monkeypatch, mocker: MockerFixture):
            self._cancelled_load_reqs = set()
            self._pending_save_reqs = deque()
            self._finished_save_reqs = set()
+            self._send_failures = {}
+            self._send_failure_lock = threading.Lock()
+            self._receive_failures = {}
+            self._receive_failure_lock = threading.Lock()
            self.stop_event = threading.Event()
            self._recv_cond = threading.Condition()
            self._save_cond = threading.Condition()
@@ -142,9 +352,23 @@ def build_adapter(monkeypatch, mocker: MockerFixture):
            classmethod(lambda cls, _model_config: connector),
        )

+        hf_config = SimpleNamespace(
+            model_type="minicpmtts",
+            max_position_embeddings=tts_max_model_len,
+            attention_type=tts_attention_type,
+        )
+        if not flat_tts_config:
+            hf_config = SimpleNamespace(
+                tts_config=SimpleNamespace(
+                    max_position_embeddings=tts_max_model_len,
+                    attention_type=tts_attention_type,
+                )
+            )
        model_config = SimpleNamespace(
            worker_type=model_mode,
            max_num_seqs=max_num_seqs,
+            max_model_len=max_model_len,
+            hf_config=hf_config,
            active_stream_window=active_stream_window,
            stage_connector_config={
                "name": "SharedMemoryConnector",
@@ -160,6 +384,27 @@ def build_adapter(monkeypatch, mocker: MockerFixture):
    return _build


+def _dequeue_load_entry(adapter, request):
+    """Register and take the entry as the background recv loop would."""
+    adapter.load_async(request)
+    entry = adapter._registered_load_entries[request.request_id]
+    try:
+        adapter._pending_load_reqs.remove(entry)
+    except ValueError:
+        pass
+    return entry
+
+
+@pytest.mark.parametrize(
+    ("attention_type", "expected_previous_chunks"),
+    [("full_attention", 0), ("sliding_recompute", 1)],
+)
+def test_talker_attention_policy_controls_streaming_recompute(build_adapter, attention_type, expected_previous_chunks):
+    adapter, _ = build_adapter(tts_attention_type=attention_type)
+
+    assert adapter._streaming_prompt_previous_chunks == expected_previous_chunks
+
+
@pytest.mark.parametrize(
    ("raw_cfg", "expected_name", "expected_extra"),
    [
@@ -200,15 +445,78 @@ def test_load_poll(build_adapter):
        "meta": {"finished": torch.tensor(True, dtype=torch.bool)},
    }
    connector.get.return_value = (payload, 16)
-    adapter._poll_single_request(request)
+    adapter._poll_single_request(_dequeue_load_entry(adapter, request))

+    assert request.additional_information is None
+    assert "req-1" in adapter._finished_load_reqs
+    request.status = RequestStatus.WAITING_FOR_CHUNK
+    adapter.process_pending_chunks(
+        DummyWaitingQueue(),
+        [request],
+        scheduler_requests={request.request_id: request},
+    )
    assert request.additional_information == payload
    assert adapter.get_req_chunk["req-1"] == 1
-    assert "req-1" in adapter._finished_load_reqs
+    assert "req-1" not in adapter._finished_load_reqs
    assert "req-1" in adapter.upstream_exhausted_requests
    assert "req-1" not in adapter._pending_load_reqs


+def test_load_async_does_not_requeue_registered_inflight_request(build_adapter):
+    adapter, connector = build_adapter(stage_id=1, model_mode="ar")
+    request = _req("req-inflight", RequestStatus.RUNNING, external_req_id="ext-inflight")
+
+    entry = _dequeue_load_entry(adapter, request)
+
+    # The recv loop owns the popped request until polling completes. A
+    # scheduler-side duplicate registration must not append a second copy.
+    adapter.load_async(request)
+    assert adapter._pending_load_reqs == deque()
+
+    connector.get.return_value = (
+        {
+            "ids": {"prompt": [1]},
+            "meta": {"finished": True},
+        },
+        1,
+    )
+    assert adapter._poll_single_request(entry) is True
+
+    # Successful polling releases the registration for the next chunk.
+    adapter.load_async(request)
+    assert [queued.request for queued in adapter._pending_load_reqs] == [request]
+
+
+def test_cleanup_then_reregister_drops_stale_queued_entry(build_adapter):
+    adapter, connector = build_adapter(stage_id=1, model_mode="ar")
+    request = _req("req-resume", RequestStatus.WAITING, external_req_id="ext-resume")
+
+    adapter.load_async(request)
+    stale_entry = adapter._pending_load_reqs[0]
+    adapter.cleanup_receiver(request.request_id)
+    adapter.load_async(request)
+    fresh_entry = adapter._pending_load_reqs[-1]
+    assert len(adapter._pending_load_reqs) == 2
+    assert adapter._pending_load_reqs.popleft() is stale_entry
+    assert adapter._pending_load_reqs.popleft() is fresh_entry
+
+    connector.get.return_value = (
+        {
+            "ids": {"prompt": [1]},
+            "meta": {"finished": True},
+        },
+        1,
+    )
+    assert adapter._poll_single_request(stale_entry) is True
+    connector.get.assert_not_called()
+    assert adapter._registered_load_entries[request.request_id] is fresh_entry
+
+    assert adapter._poll_single_request(fresh_entry) is True
+    connector.get.assert_called_once()
+    assert request.request_id not in adapter._registered_load_entries
+    assert adapter.get_req_chunk[request.request_id] == 1
+
+
 def test_load_poll_ar_requeues_explicitly_replaced_running_prompt(build_adapter):
    adapter, connector = build_adapter(stage_id=1, model_mode="ar")
    request = _req("req-replace", RequestStatus.RUNNING, external_req_id="external-replace")
@@ -234,15 +542,18 @@ def test_load_poll_ar_requeues_explicitly_replaced_running_prompt(build_adapter)
        1,
    )

-    assert adapter._poll_single_request(request) is True
-    assert request.num_computed_tokens == 0
-    assert request.prompt_token_ids == [0] * 10
-    assert request.request_id in adapter.replaced_streaming_prompt_ids
-
+    assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is True
    request.status = RequestStatus.WAITING_FOR_CHUNK
    running_queue = [request]
    waiting_queue = DummyWaitingQueue()
-    adapter.process_pending_chunks(waiting_queue, running_queue)
+    adapter.process_pending_chunks(
+        waiting_queue,
+        running_queue,
+        scheduler_requests={request.request_id: request},
+    )
+    assert request.num_computed_tokens == 0
+    assert request.prompt_token_ids == [0] * 10
+    assert request.request_id in adapter.replaced_streaming_prompt_ids
    assert running_queue == []
    assert waiting_queue == [request]
    assert request.status == RequestStatus.WAITING
@@ -258,6 +569,134 @@ def test_load_poll_ar_requeues_explicitly_replaced_running_prompt(build_adapter)
    assert request.external_req_id not in adapter.requests_num_chunks_sent


+@pytest.mark.parametrize("flat_tts_config", [False, True])
+def test_load_poll_ar_recomputes_native_duplex_prompt_each_condition(build_adapter, flat_tts_config):
+    adapter, connector = build_adapter(
+        stage_id=1,
+        model_mode="ar",
+        max_model_len=8192,
+        tts_max_model_len=4096,
+        tts_attention_type="sliding_recompute",
+        flat_tts_config=flat_tts_config,
+    )
+    request = _req("req-rollover", RequestStatus.RUNNING, external_req_id="external-rollover")
+    request.resumable = True
+    previous_codes = list(range(25))
+    request.prompt_token_ids = [0] * 10
+    request._all_token_ids = [0] * 10 + previous_codes
+    request._output_token_ids = previous_codes.copy()
+    request.num_prompt_tokens = 10
+    request.num_computed_tokens = 35
+    request.num_output_placeholders = 0
+    request.update_block_hashes = Mock()
+    adapter.get_req_chunk[request.request_id] = 1
+    adapter.request_ids_mapping[request.request_id] = request.external_req_id
+    adapter._streaming_condition_lengths[request.request_id] = 10
+    adapter._streaming_condition_seqs[request.request_id] = 0
+    connector.get.return_value = (
+        {
+            "ids": {"prompt": [1]},
+            "meta": {
+                "next_stage_prompt_len": 10,
+                "next_stage_generation_tokens": 26,
+                "finished": True,
+            },
+        },
+        1,
+    )
+
+    assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is True
+    request.status = RequestStatus.WAITING_FOR_CHUNK
+    running_queue = [request]
+    waiting_queue = DummyWaitingQueue()
+    adapter.process_pending_chunks(
+        waiting_queue,
+        running_queue,
+        scheduler_requests={request.request_id: request},
+    )
+
+    assert request.num_computed_tokens == 0
+    assert request.prompt_token_ids == [0] * 45
+    assert request.additional_information["ids"]["streaming_prompt_previous_codes"] == previous_codes
+    assert request.additional_information["meta"]["streaming_prompt_recompute"] is True
+    assert request.resumable is False
+    assert request.request_id in adapter.replaced_streaming_prompt_ids
+    assert adapter._max_model_len == 4096
+    assert running_queue == []
+    assert waiting_queue == [request]
+
+
+def test_window_condition_sequence_ignores_control_only_chunks(build_adapter) -> None:
+    adapter, connector = build_adapter(
+        stage_id=1,
+        model_mode="ar",
+        tts_attention_type="sliding_recompute",
+    )
+    request = _req("req-condition-seq", RequestStatus.RUNNING)
+
+    def receive(payload):
+        connector.get.return_value = (payload, 1)
+        assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is True
+        adapter.requests_with_ready_chunks.add(request.request_id)
+        adapter._apply_pending_ar_prompt_updates({request.request_id: request})
+        adapter.requests_with_ready_chunks.discard(request.request_id)
+        return request.additional_information
+
+    first = receive(
+        {
+            "ids": {"prompt": [1]},
+            "meta": {"next_stage_prompt_len": 10},
+        }
+    )
+    assert first["meta"]["streaming_condition_seq"] == 0
+    assert adapter._streaming_condition_seqs[request.request_id] == 0
+
+    control = receive({"meta": {"is_segment_finished": True}})
+    assert "streaming_condition_seq" not in control["meta"]
+    assert adapter._streaming_condition_seqs[request.request_id] == 0
+
+    second = receive(
+        {
+            "ids": {"prompt": [2]},
+            "meta": {"next_stage_prompt_len": 12},
+        }
+    )
+    assert second["meta"]["streaming_condition_seq"] == 1
+    assert adapter._streaming_condition_lengths[request.request_id] == 12
+
+
+def test_load_poll_ar_reports_invalid_capacity_metadata(build_adapter):
+    adapter, connector = build_adapter(stage_id=1, model_mode="ar", max_model_len=4096)
+    request = _req("req-invalid-capacity", RequestStatus.RUNNING)
+    request.resumable = True
+    adapter.get_req_chunk[request.request_id] = 1
+    connector.get.return_value = (
+        {
+            "ids": {"prompt": [1]},
+            "meta": {
+                "next_stage_prompt_len": 4096,
+                "next_stage_generation_tokens": 26,
+                "finished": False,
+            },
+        },
+        1,
+    )
+
+    assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is True
+    request.status = RequestStatus.WAITING_FOR_CHUNK
+    running_queue = [request]
+    adapter.process_pending_chunks(
+        DummyWaitingQueue(),
+        running_queue,
+        scheduler_requests={request.request_id: request},
+    )
+    assert adapter.collect_failed_receive_request_ids() == {
+        request.request_id: "fresh streaming prompt plus generation reserve exceeds max_model_len: "
+        "prompt=4096, reserve=26, limit=4096"
+    }
+    assert request.request_id not in adapter._finished_load_reqs
+
+
 def test_load_poll_generation_tensor_codes_use_placeholder_prompt(build_adapter):
    adapter, connector = build_adapter(stage_id=1, model_mode="generation")
    request = _req("req-tensor", RequestStatus.WAITING, external_req_id="external-tensor")
@@ -272,7 +711,7 @@ def test_load_poll_generation_tensor_codes_use_placeholder_prompt(build_adapter)
    }
    connector.get.return_value = (payload, 16)

-    adapter._poll_single_request(request)
+    adapter._poll_single_request(_dequeue_load_entry(adapter, request))

    assert request.prompt_token_ids == [0]
    assert request.num_computed_tokens == 0
@@ -302,12 +741,13 @@ def test_load_poll_generation_empty_nonterminal_chunk_keeps_polling(build_adapte
    }
    connector.get.side_effect = [(empty_payload, 16), (ready_payload, 16)]

-    assert adapter._poll_single_request(request) is False
+    entry = _dequeue_load_entry(adapter, request)
+    assert adapter._poll_single_request(entry) is False
    assert request.request_id not in adapter._finished_load_reqs
    assert request.request_id not in adapter.requests_with_ready_chunks
    assert adapter.get_req_chunk[request.request_id] == 1

-    assert adapter._poll_single_request(request) is True
+    assert adapter._poll_single_request(entry) is True
    assert request.request_id in adapter._finished_load_reqs
    assert torch.equal(request.additional_information["codes"]["audio"], ready_payload["codes"]["audio"])
    assert adapter.get_req_chunk[request.request_id] == 2
@@ -347,6 +787,7 @@ def test_segment_boundary_starts_new_send_watermark_before_background_flush(buil
    adapter, _ = build_adapter(stage_id=1)
    request = _req("req-stream", RequestStatus.WAITING, external_req_id="ext-stream")
    request.resumable = True
+    request._omni_segment_generation = 1
    request.num_computed_tokens = 0
    adapter.requests_num_chunks_sent["ext-stream"] = 26

@@ -355,13 +796,16 @@ def test_segment_boundary_starts_new_send_watermark_before_background_flush(buil
        request=request,
        is_segment_finished=True,
        confirmed_num_computed_tokens=26,
+        segment_generation=0,
    )

    assert len(adapter._pending_save_reqs) == 1
    boundary_task = adapter._pending_save_reqs.popleft()
+    assert boundary_task["segment_generation"] == 0
    assert "ext-stream" not in adapter.requests_num_chunks_sent

    request.num_computed_tokens = 3
+    request._omni_segment_generation = 1
    adapter.save_async(
        multimodal_output=None,
        request=request,
@@ -414,6 +858,7 @@ def test_background_send_uses_enqueued_request_snapshot(build_adapter):
    request.additional_information["meta"]["segment"] = "next"
    request.prompt_token_ids.append(3)
    request.append_output_token_ids([8])
+    request._omni_segment_generation = 1
    adapter.save_async(
        multimodal_output=None,
        request=request,
@@ -662,7 +1107,7 @@ def test_personaplex_sender_cleanup_drops_delayed_frame_state(build_adapter):
    assert second.codes is None


-def test_save_async_skips_stale_resumable_chunk_until_dedup_is_reset(build_adapter):
+def test_save_async_skips_stale_resumable_chunk_within_segment(build_adapter):
    adapter, _ = build_adapter(stage_id=1)
    request = _req("req-stream", RequestStatus.WAITING, external_req_id="ext-stream")
    request.resumable = True
@@ -681,6 +1126,51 @@ def test_save_async_skips_stale_resumable_chunk_until_dedup_is_reset(build_adapt
    assert adapter.requests_num_chunks_sent["ext-stream"] == 0


+def test_save_async_drops_late_previous_segment_after_boundary_reset(build_adapter):
+    adapter, _ = build_adapter(stage_id=1)
+    request = _req("req-stream", RequestStatus.WAITING, external_req_id="ext-stream")
+    request.resumable = True
+    request._omni_segment_generation = 0
+    adapter.requests_num_chunks_sent["ext-stream"] = 111
+
+    adapter.save_async(multimodal_output=None, request=request, is_segment_finished=True)
+
+    late = _req("req-stream-late", RequestStatus.WAITING, external_req_id="ext-stream")
+    late.resumable = True
+    late.num_computed_tokens = 111
+    late._omni_segment_generation = 0
+    adapter.save_async(multimodal_output=None, request=late, is_segment_finished=False)
+
+    next_segment = _req("req-stream-next", RequestStatus.WAITING, external_req_id="ext-stream")
+    next_segment.resumable = True
+    next_segment.num_computed_tokens = 3
+    next_segment._omni_segment_generation = 1
+    adapter.save_async(multimodal_output=None, request=next_segment, is_segment_finished=False)
+
+    assert len(adapter._pending_save_reqs) == 2
+    assert adapter.requests_num_chunks_sent["ext-stream"] == 3
+    assert adapter._pending_save_reqs[-1]["request"] is next_segment
+
+
+def test_send_single_request_preserves_queued_preboundary_chunk(build_adapter):
+    adapter, connector = build_adapter(stage_id=1)
+    adapter.custom_process_next_stage_input_func = lambda **kwargs: OmniPayloadStruct(meta=MetaStruct())
+    request = _req("req-queued-old", RequestStatus.WAITING, external_req_id="ext-queued-old")
+    request.resumable = True
+    request._omni_segment_generation = 0
+
+    adapter.save_async(multimodal_output=None, request=request, is_segment_finished=False)
+    adapter.save_async(multimodal_output=None, request=request, is_segment_finished=True)
+
+    old_task = adapter._pending_save_reqs.popleft()
+    adapter._send_single_request(old_task)
+    connector.put.assert_called_once()
+
+    boundary_task = adapter._pending_save_reqs.popleft()
+    adapter._send_single_request(boundary_task)
+    assert connector.put.call_count == 2
+
+
 def test_send_single_request_cleans_up_after_finished_payload(build_adapter, monkeypatch):
    adapter, _ = build_adapter(stage_id=1)
    request = _req("req-finished", RequestStatus.FINISHED_STOPPED, external_req_id="ext-finished")
@@ -701,6 +1191,77 @@ def test_send_single_request_cleans_up_after_finished_payload(build_adapter, mon
    assert args[1] == "ext-finished"


+def test_old_turn_terminal_send_preserves_new_turn_state_and_sender(build_adapter):
+    adapter, connector = build_adapter(
+        stage_id=1,
+        model_mode="ar",
+        max_model_len=4096,
+        tts_attention_type="sliding_recompute",
+    )
+    request = _req("req-live", RequestStatus.RUNNING, external_req_id="ext-live")
+    request.resumable = True
+    request._all_token_ids = []
+    request._output_token_ids = []
+    request.update_block_hashes = Mock()
+    adapter.custom_process_next_stage_input_func = lambda **kwargs: OmniPayloadStruct(
+        meta=MetaStruct(
+            finished=torch.tensor(True, dtype=torch.bool),
+            codec_streaming=True,
+        )
+    )
+
+    put_started = threading.Event()
+    release_put = threading.Event()
+    put_keys = []
+
+    def blocking_first_put(*_args, **kwargs):
+        put_keys.append(kwargs["put_key"])
+        if len(put_keys) == 1:
+            put_started.set()
+            assert release_put.wait(timeout=5)
+        return True, 1, {}
+
+    connector.put.side_effect = blocking_first_put
+    adapter.save_async(multimodal_output=None, request=request)
+    first_task = adapter._pending_save_reqs.popleft()
+    first_sender_token = first_task["sender_token"]
+    send_thread = threading.Thread(target=lambda: adapter._send_single_request(first_task))
+    send_thread.start()
+    assert put_started.wait(timeout=5)
+
+    entry = _dequeue_load_entry(adapter, request)
+    connector.get.return_value = (
+        {
+            "ids": {"prompt": [1]},
+            "meta": {
+                "replace_streaming_prompt": True,
+                "next_stage_prompt_len": 10,
+                "next_stage_generation_tokens": 26,
+            },
+        },
+        1,
+    )
+    assert adapter._poll_single_request(entry) is True
+    adapter.requests_with_ready_chunks.add(request.request_id)
+    adapter._apply_pending_ar_prompt_updates({request.request_id: request})
+    assert adapter._streaming_condition_seqs[request.request_id] == 0
+
+    request.num_computed_tokens = 1
+    adapter.save_async(multimodal_output=None, request=request)
+    second_task = adapter._pending_save_reqs.popleft()
+    assert second_task["sender_token"] is first_sender_token
+
+    release_put.set()
+    send_thread.join(timeout=5)
+    assert not send_thread.is_alive()
+    assert adapter._streaming_condition_seqs[request.request_id] == 0
+    assert adapter._sender_tokens[request.external_req_id] is first_sender_token
+    assert first_sender_token.cancelled is False
+
+    adapter._send_single_request(second_task)
+    assert put_keys == ["ext-live_1_0", "ext-live_1_1"]
+
+
 def test_load_poll_non_ar_merges_into_existing_additional_information(build_adapter):
    adapter, connector = build_adapter(stage_id=2, model_mode="diffusion")
    request = _req("req-non-ar", RequestStatus.WAITING, external_req_id="ext-non-ar")
@@ -720,7 +1281,7 @@ def test_load_poll_non_ar_merges_into_existing_additional_information(build_adap
    }
    connector.get.return_value = (payload, 8)

-    assert adapter._poll_single_request(request) is True
+    assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is True

    assert request.prompt_token_ids == [7, 8]
    assert request.num_computed_tokens == 0
@@ -761,7 +1322,7 @@ def test_load_poll_generation_segment_marker_replaces_previous_chunk(build_adapt
        1,
    )

-    assert adapter._poll_single_request(request) is True
+    assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is True

    assert request.prompt_token_ids == [7, 8]
    assert "audio" not in request.additional_information["codes"]
@@ -790,7 +1351,7 @@ def test_load_poll_generation_empty_replacement_snapshot_is_ready(build_adapter)
        1,
    )

-    assert adapter._poll_single_request(request) is True
+    assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is True

    assert request.prompt_token_ids == [0]
    assert "codes" not in request.additional_information
@@ -816,7 +1377,7 @@ def test_load_poll_generation_without_snapshot_marker_keeps_incremental_state(bu
        1,
    )

-    assert adapter._poll_single_request(request) is False
+    assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is False

    assert torch.equal(request.additional_information["codes"]["audio"], torch.tensor([1, 2]))
    assert request.additional_information["meta"]["cache_epoch"] == 3
@@ -840,9 +1401,17 @@ def test_load_poll_ar_request_additional_information_concats_tensors(build_adapt
    }
    connector.get.return_value = (payload, 8)

-    adapter._poll_single_request(request)
+    adapter._poll_single_request(_dequeue_load_entry(adapter, request))

-    # AR mode now forwards the latest payload directly.
+    # The recv thread records the latest payload; scheduler-side queue
+    # restoration publishes it onto the Request.
+    assert torch.equal(request.additional_information["hidden_states"]["output"], torch.tensor([[1.0]]))
+    request.status = RequestStatus.WAITING_FOR_CHUNK
+    adapter.process_pending_chunks(
+        DummyWaitingQueue(),
+        [request],
+        scheduler_requests={request.request_id: request},
+    )
    assert request.additional_information == payload
    assert request.additional_information["meta"]["finished"].item() is True

@@ -860,7 +1429,7 @@ def test_non_ar_poll_reinitializes_prefill_stats_for_later_chunks(build_adapter)
        },
        8,
    )
-    assert adapter._poll_single_request(request) is True
+    assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is True
    OmniGenerationScheduler._record_prefill_stats(request)
    first_chunk_stats = request.prefill_stats
    request.prefill_stats = None
@@ -872,7 +1441,7 @@ def test_non_ar_poll_reinitializes_prefill_stats_for_later_chunks(build_adapter)
        },
        8,
    )
-    assert adapter._poll_single_request(request) is True
+    assert adapter._poll_single_request(_dequeue_load_entry(adapter, request)) is True
    assert isinstance(request.prefill_stats, PrefillStats)
    OmniGenerationScheduler._record_prefill_stats(request)
    second_chunk_stats = request.prefill_stats
@@ -971,7 +1540,7 @@ def test_non_active_waiting_request_is_held_off_scheduler(build_adapter):
    assert waiting_queue == []
    assert active.status == RequestStatus.WAITING_FOR_CHUNK
    assert non_active.status == RequestStatus.WAITING
-    assert list(adapter._pending_load_reqs) == [active]
+    assert [entry.request for entry in adapter._pending_load_reqs] == [active]
    assert list(adapter.waiting_for_chunk_waiting_requests) == [active, non_active]


@@ -1122,8 +1691,19 @@ def _populate_adapter_state(adapter, req_id="req-1", ext_id="ext-1"):
    adapter.get_req_chunk[req_id] = 3
    adapter.requests_with_ready_chunks.add(req_id)
    adapter.request_ids_mapping[req_id] = ext_id
-    adapter._pending_load_reqs.append(SimpleNamespace(request_id=req_id))
+    entry = SimpleNamespace(request_id=req_id)
+    adapter._pending_load_reqs.append(entry)
+    adapter._registered_load_entries[req_id] = entry
    adapter._finished_load_reqs.add(req_id)
+    adapter._pending_ar_prompt_updates[req_id] = (
+        SimpleNamespace(request_id=req_id),
+        {},
+        False,
+        False,
+        False,
+    )
+    adapter._streaming_condition_lengths[req_id] = 10
+    adapter._streaming_condition_seqs[req_id] = 2

    adapter.put_req_chunk[ext_id] = 5
    adapter.request_payload[ext_id] = {"hidden": [1, 2]}
@@ -1145,12 +1725,71 @@ def test_cleanup_clears_all_state(build_adapter):
    assert req_id not in adapter.request_ids_mapping
    assert req_id in adapter._cancelled_load_reqs
    assert req_id not in adapter._finished_load_reqs
+    assert req_id not in adapter._pending_ar_prompt_updates
+    assert req_id not in adapter._streaming_condition_lengths
+    assert req_id not in adapter._streaming_condition_seqs
+    assert req_id not in adapter._registered_load_entries

    assert ext_id not in adapter.put_req_chunk
    assert ext_id not in adapter.request_payload
    assert ext_id not in adapter.code_prompt_token_ids


+def test_cleanup_and_reregister_discards_inflight_old_segment_chunk(build_adapter):
+    """A late old-segment chunk cannot overwrite a resumed registration."""
+    adapter, connector = build_adapter(stage_id=1, model_mode="ar")
+    request = _req("req-late", RequestStatus.RUNNING, external_req_id="ext-late")
+    request.resumable = True
+    entry = _dequeue_load_entry(adapter, request)
+
+    get_started = threading.Event()
+    release_get = threading.Event()
+
+    def blocking_get(*_args):
+        get_started.set()
+        assert release_get.wait(timeout=5)
+        return (
+            {
+                "ids": {"prompt": [1]},
+                "meta": {
+                    "next_stage_prompt_len": 10,
+                },
+            },
+            1,
+        )
+
+    connector.get.side_effect = blocking_get
+    poll_results = []
+    poll_thread = threading.Thread(target=lambda: poll_results.append(adapter._poll_single_request(entry)))
+    poll_thread.start()
+    assert get_started.wait(timeout=5)
+
+    adapter.cleanup_receiver(request.request_id)
+    fresh_entry = _dequeue_load_entry(adapter, request)
+    release_get.set()
+    poll_thread.join(timeout=5)
+
+    assert not poll_thread.is_alive()
+    assert poll_results == [True]
+    assert request.additional_information is None
+    assert request.request_id not in adapter._pending_ar_prompt_updates
+    assert request.request_id not in adapter._streaming_condition_seqs
+    assert adapter.get_req_chunk[request.request_id] == 0
+    assert adapter._registered_load_entries[request.request_id] is fresh_entry
+
+    connector.get.side_effect = None
+    connector.get.return_value = (
+        {
+            "ids": {"prompt": [2]},
+            "meta": {"finished": True},
+        },
+        1,
+    )
+    assert adapter._poll_single_request(fresh_entry) is True
+    assert adapter.get_req_chunk[request.request_id] == 1
+    assert request.request_id not in adapter._registered_load_entries
+
+
 def test_cleanup_infers_external_id(build_adapter):
    """When external_req_id is None, cleanup should look it up from the mapping."""
    adapter, _ = build_adapter(stage_id=1)
@@ -1299,7 +1938,7 @@ def test_cleanup_after_poll_flow(build_adapter):
        "meta": {"finished": torch.tensor(True, dtype=torch.bool)},
    }
    connector.get.return_value = (payload, 8)
-    adapter._poll_single_request(request)
+    adapter._poll_single_request(_dequeue_load_entry(adapter, request))

    assert "req-flow" in adapter.upstream_exhausted_requests
    assert adapter.get_req_chunk["req-flow"] == 1
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """GPU e2e for PersonaPlex elastic batching (bit-exact isolation + recycle).

 Greedy decoding makes every cross-slot leak observable as a bit difference:
@@ -24,12 +24,17 @@ from __future__ import annotations
 import numpy as np
 import pytest

+from tests.helpers.mark import hardware_marks
+
 torch = pytest.importorskip("torch")

-# slow + H100 + omni routes this module into the weekly Buildkite E2E group
-# ("Omni · H100" in .buildkite/cuda/test-weekly.yml: -m "slow and H100 and
-# omni" --run-level full_model); the ready/merge CPU sweeps exclude it by mark.
-pytestmark = [pytest.mark.full_model, pytest.mark.omni, pytest.mark.slow, pytest.mark.H100]
+# slow + H100 + omni + cards_1 routes this module into weekly
+# "Omni · H100 · Single-GPU" (test-weekly.yml).
+pytestmark = [
+    pytest.mark.omni,
+    pytest.mark.slow,
+    *hardware_marks(res={"cuda": "H100"}),
+]

 if not torch.cuda.is_available():  # noqa: E402
    pytest.skip("needs CUDA", allow_module_level=True)
@@ -0,0 +1,129 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""E2E tests for dots.tts native AR offline inference."""
+
+from collections.abc import Mapping
+
+import pytest
+import torch
+
+from tests.helpers.mark import hardware_test
+from tests.helpers.runtime import OmniRunner
+from tests.helpers.stage_config import get_deploy_config_path
+
+DOTS_TTS_MODEL = "rednote-hilab/dots.tts-soar"
+DEPLOY_CONFIG = get_deploy_config_path("dots_tts.yaml")
+SAMPLE_RATE = 48000
+
+# dots.tts ships a custom tokenizer carrying the added ``<|audio_gen_start|>``
+# token, so remote code must be explicitly enabled.
+_OMNI_RUNNER_PARAM = (DOTS_TTS_MODEL, DEPLOY_CONFIG, {"trust_remote_code": True})
+
+pytestmark = pytest.mark.parametrize("omni_runner", [_OMNI_RUNNER_PARAM], indirect=True)
+
+
+def _build_prompt(omni_runner: OmniRunner, text: str) -> dict:
+    """Wrap ``text`` in the prefix scaffold dots.tts prefill expects.
+
+    Unlike most TTS models here, a bare ``{"prompt": ...}`` string does not
+    work: prefill needs ``[文本]<text>[文本对应语音]<|audio_gen_start|>`` as
+    real Qwen2 token ids.
+    """
+    from transformers import AutoTokenizer
+
+    from vllm_omni.model_executor.models.dots_tts.dots_tts_prompt import (
+        build_dots_tts_prompt,
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(omni_runner.model_name, trust_remote_code=True)
+    return build_dots_tts_prompt(tokenizer=tokenizer, text=text)
+
+
+def _extract_audio(multimodal_output: dict) -> torch.Tensor:
+    """Extract the final complete audio tensor from multimodal output."""
+    assert isinstance(multimodal_output, (dict, Mapping)), f"Expected dict/Mapping, got {type(multimodal_output)}"
+
+    # The talker emits delta chunks; the output processor accumulates them
+    # under "model_outputs" ("audio" kept for backwards compat).
+    audio = multimodal_output.get("model_outputs")
+    if audio is None:
+        audio = multimodal_output.get("audio")
+    assert audio is not None, f"No audio key, got {list(multimodal_output.keys())}"
+
+    if isinstance(audio, list):
+        valid = [torch.as_tensor(x).float().cpu().reshape(-1) for x in audio if x is not None]
+        assert valid, "No valid audio tensors in output list"
+        audio = torch.cat(valid, dim=0) if len(valid) > 1 else valid[0]
+
+    assert isinstance(audio, torch.Tensor), f"Expected Tensor, got {type(audio)}"
+    return audio.float().cpu().reshape(-1)
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+def test_dots_tts_zero_shot_001(omni_runner: OmniRunner) -> None:
+    """Zero-shot synthesis produces audio of a plausible duration."""
+    prompt = _build_prompt(omni_runner, "Hello, this is a test of dots TTS running on vLLM Omni.")
+
+    outputs = omni_runner.omni.generate([prompt])
+    assert len(outputs) == 1
+
+    audio = _extract_audio(outputs[0].outputs[0].multimodal_output)
+    duration_s = audio.shape[0] / SAMPLE_RATE
+    assert 0.5 < duration_s < 30.0, f"Audio duration out of range: {duration_s:.2f}s"
+    assert torch.isfinite(audio).all(), "Audio contains NaN/Inf"
+    assert audio.abs().max() > 1e-3, "Audio is silent"
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+def test_dots_tts_prefill_decode_mixed_batch_002(omni_runner: OmniRunner) -> None:
+    """Mixed prefill+decode batches must keep per-request rows aligned.
+
+    ``compute_logits`` fills ``logits[i]`` positionally from
+    ``_results_queue``, so every request in the batch has to push exactly one
+    entry per step, prefills included. When the prefill branch pushed nothing,
+    a shorter request's stop signal could be read into another request's row.
+    Regression guard for that pairing (PR #4765 review).
+    """
+    long_text = (
+        "This is a deliberately long sentence that stays in the decode phase "
+        "for many steps, so the shorter requests behind it keep entering "
+        "prefill alongside it and reproduce the mixed batch pattern."
+    )
+    short_texts = ["Hello one.", "Hello two.", "Hello three.", "Hello four."]
+    prompts = [_build_prompt(omni_runner, long_text)] + [_build_prompt(omni_runner, t) for t in short_texts]
+
+    outputs = omni_runner.omni.generate(prompts)
+    assert len(outputs) == len(prompts)
+
+    for i, out in enumerate(outputs):
+        audio = _extract_audio(out.outputs[0].multimodal_output)
+        duration_s = audio.shape[0] / SAMPLE_RATE
+        assert 0.1 < duration_s < 30.0, f"Request {i} audio duration out of range: {duration_s:.2f}s"
+
+
+@pytest.mark.slow
+@pytest.mark.tts
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+def test_dots_tts_deterministic_flow_matching_noise_003(omni_runner: OmniRunner) -> None:
+    """The same text twice must yield the same audio.
+
+    The DiT flow-matching noise is drawn from a per-request generator seeded
+    from ``blake2b(seed:request_key:noise_step)`` rather than the global CUDA
+    RNG. Drawing from the global RNG made output non-reproducible and let
+    concurrent requests perturb each other's noise streams (PR #4765 review).
+    """
+    text = "Reading aloud every evening helps children build a lasting love of language."
+
+    first = _extract_audio(
+        omni_runner.omni.generate([_build_prompt(omni_runner, text)])[0].outputs[0].multimodal_output
+    )
+    second = _extract_audio(
+        omni_runner.omni.generate([_build_prompt(omni_runner, text)])[0].outputs[0].multimodal_output
+    )
+
+    assert first.shape == second.shape, f"Length differs across runs: {first.shape} vs {second.shape}"
+    assert torch.allclose(first, second, atol=1e-4), "Audio differs across runs; flow-matching noise is not seeded"
@@ -1,13 +1,13 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Smoke test for HunyuanImage-3.0 Image-to-Text (I2T) pipeline."""

 from collections.abc import Generator

 import pytest
-import torch
 from PIL import Image

+from tests.helpers.mark import hardware_test
 from tests.helpers.runtime import OmniRunner
 from tests.helpers.stage_config import get_deploy_config_path
 from vllm_omni import Omni
@@ -44,7 +44,7 @@ EXPECTED_PREFIX_TOKEN_IDS: list[int] = [
 # Decoded form, kept only for human-readable assertion messages.
 EXPECTED_PREFIX_TEXT = "The image is a solid, uniform green color with no variations, objects, or details present. It"

-pytestmark = [pytest.mark.full_model, pytest.mark.diffusion, pytest.mark.cuda]
+pytestmark = [pytest.mark.full_model, pytest.mark.diffusion]


@pytest.fixture(scope="module")
@@ -57,7 +57,7 @@ def omni() -> Generator[Omni, None, None]:
        yield runner.omni


-@pytest.mark.skipif(torch.accelerator.device_count() < 4, reason="Need at least 4 CUDA GPUs.")
+@hardware_test(res={"cuda": "H100"}, num_cards=4)
 def test_i2t_generates_text(omni: Omni) -> None:
    """Verify I2T output's first 20 token IDs match the HF greedy baseline."""
    # Solid-color image keeps the input self-contained and reproducible.
@@ -91,8 +91,8 @@ def test_krea2_cache_dit(offline_client: OfflineOmniClient) -> None:
 def test_krea2_hsdp(offline_client: OfflineOmniClient) -> None:
    """Exercise Hybrid Sharded Data Parallel (HSDP) weight sharding on the DiT across 2 GPUs.

-    Validates the docs' HSDP-support claim for Krea 2. Scheduled on the multi-GPU (distributed_cuda)
-    shard by ``num_cards=2``; skipped automatically on boxes with fewer than 2 CUDA devices.
+    Validates the docs' HSDP-support claim for Krea 2. Scheduled on the multi-GPU
+    shard by ``num_cards=2`` (``not cards_1``); skipped automatically on boxes with fewer than 2 CUDA devices.
    """
    offline_client.send_diffusion_request({"model": MODEL, "prompt": PROMPT, "sampling_params": _sampling()})

@@ -111,7 +111,7 @@ def test_krea2_vae_patch_parallel(offline_client: OfflineOmniClient) -> None:
    """Exercise VAE patch-parallel (decode) across 2 GPUs.

    Validates the docs' VAE-patch-parallel (decode) support claim for Krea 2. Scheduled on the
-    multi-GPU (distributed_cuda) shard by ``num_cards=2``; skipped automatically on boxes with
+    multi-GPU shard by ``num_cards=2`` (``not cards_1``); skipped automatically on boxes with
    fewer than 2 CUDA devices.
    """
    offline_client.send_diffusion_request({"model": MODEL, "prompt": PROMPT, "sampling_params": _sampling()})
@@ -3,6 +3,9 @@

 """
 E2E offline tests for Omni model with video input and audio output.
+
+Abort / sleep-admission lives in ``test_qwen3_omni_colocate_async.py`` so it
+does not overlap this module's OmniRunners.
 """

 import os
@@ -0,0 +1,182 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""Isolated AsyncOmni abort / sleep-admission regression for Qwen3-Omni.
+
+This lives in its own module so the L3 merge job can run it *after*
+``test_qwen3_omni.py`` (offline) and the online serving file. Those modules keep
+module-scoped OmniRunner / OmniServer instances until teardown; overlapping a
+CuMem sleep-mode engine with them kills StageEngineCoreProc at handshake
+(``Failed core proc(s): {}``).
+"""
+
+import asyncio
+import os
+from contextlib import ExitStack
+
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+import pytest
+from vllm import SamplingParams
+from vllm.inputs import TokensPrompt
+
+from tests.helpers.clean import wait_for_gpu_memory_to_clear
+from tests.helpers.mark import hardware_test
+from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
+from vllm_omni.entrypoints.async_omni import AsyncOmni
+from vllm_omni.platforms import current_omni_platform
+
+MODEL = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
+
+
+def _output_token_ids(output) -> list[int]:
+    if not getattr(output, "outputs", None):
+        return []
+    completion = output.outputs[0]
+    token_ids = list(completion.token_ids or [])
+    if token_ids:
+        return token_ids
+    return list(getattr(completion, "cumulative_token_ids", None) or [])
+
+
+def _finish_reason(output) -> str | None:
+    if not getattr(output, "outputs", None):
+        return None
+    return getattr(output.outputs[0], "finish_reason", None)
+
+
+def _colocate_async_deploy() -> str:
+    """Thinker-only Instruct deploy with sleep mode on a single GPU.
+
+    Abort / sleep admission is a thinker control-plane contract. The 3-stage
+    CI yaml plus CuMem on every stage fails engine-core startup in the L3
+    merge job: Stage 1 resolves ``devices: "1"`` while only GPU 0 is visible.
+    """
+    return modify_stage_config(
+        get_deploy_config_path("qwen3_omni_moe_thinking.yaml"),
+        updates={
+            "stages": {
+                0: {
+                    "enable_sleep_mode": True,
+                    "enforce_eager": True,
+                    "trust_remote_code": True,
+                    "tensor_parallel_size": 1,
+                    "devices": "0",
+                    "gpu_memory_utilization": 0.9,
+                    "max_num_seqs": 1,
+                    "max_model_len": 8192,
+                    "max_num_batched_tokens": 8192,
+                    "enable_prefix_caching": False,
+                    "skip_mm_profiling": True,
+                },
+            },
+        },
+    )
+
+
+@pytest.mark.advanced_model
+@pytest.mark.omni
+@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=1)
+@pytest.mark.usefixtures("clean_gpu_memory_between_tests")
+@pytest.mark.asyncio
+async def test_colocate_async_abort_tokens_and_sleep_admission() -> None:
+    """**required** regression. Control-plane APIs (``abort`` / ``sleep`` /
+    ``wake_up`` / ``resume_generation``) are not in ``send_omni_request``.
+
+    On ``origin/main`` this test fails:
+    - abort: ``generate()`` never yields ``finish_reason="abort"`` with the
+      tokens produced so far, so resume has an empty prefix (or the stream
+      hangs after frontend state is dropped).
+    - sleep: ``generate()`` is admitted into EngineCore while weights/KV are
+      offloaded, so the task errors (asleep / corrupted ADD frame) instead of
+      waiting for ``resume_generation()``.
+
+    On this PR both legs pass: abort returns a resumeable prefix, and sleep
+    holds ``generate()`` until wake + resume.
+    """
+    prompt = "What color is the sky? Write a long, detailed explanation."
+    device_count = current_omni_platform.device_count()
+    if device_count > 0:
+        # Fail closed if a prior module-scoped runner is still on the GPU.
+        # ``clean_gpu_memory_between_tests`` only logs a note on timeout.
+        wait_for_gpu_memory_to_clear(
+            devices=list(range(device_count)),
+            threshold_ratio=0.15,
+            timeout_s=120,
+        )
+
+    with ExitStack() as after:
+        engine = AsyncOmni(
+            model=MODEL,
+            deploy_config=_colocate_async_deploy(),
+            enable_sleep_mode=True,
+        )
+        after.callback(engine.shutdown)
+
+        request_id = "qwen3-abort-partial"
+        outputs: list = []
+
+        async def _generate(max_tokens: int, req_id: str, gen_prompt) -> None:
+            async for output in engine.generate(
+                prompt=gen_prompt,
+                request_id=req_id,
+                sampling_params=SamplingParams(temperature=0.0, max_tokens=max_tokens),
+                output_modalities=["text"],
+            ):
+                outputs.append(output)
+
+        abort_task = asyncio.create_task(_generate(256, request_id, prompt))
+        prefix: list[int] = []
+        prompt_token_ids: list[int] = []
+        for _ in range(600):
+            if abort_task.done():
+                break
+            if outputs:
+                latest = outputs[-1]
+                prefix = _output_token_ids(latest)
+                prompt_token_ids = list(getattr(latest, "prompt_token_ids", None) or [])
+                if prefix:
+                    break
+            await asyncio.sleep(0.1)
+
+        assert prefix, "generate produced no tokens before abort"
+        assert not abort_task.done(), "generate finished before abort; raise max_tokens"
+        await engine.abort(request_id)
+        await asyncio.wait_for(abort_task, timeout=90)
+
+        final = outputs[-1]
+        assert final.finished
+        assert _finish_reason(final) == "abort"
+        abort_prefix = _output_token_ids(final)
+        assert abort_prefix, "abort dropped the generated prefix (main-branch behavior)"
+
+        resume_prompt: str | TokensPrompt = (
+            TokensPrompt(prompt_token_ids=prompt_token_ids + abort_prefix) if prompt_token_ids else prompt
+        )
+        outputs.clear()
+        async for output in engine.generate(
+            prompt=resume_prompt,
+            request_id="qwen3-abort-resume",
+            sampling_params=SamplingParams(temperature=0.0, max_tokens=8),
+            output_modalities=["text"],
+        ):
+            outputs.append(output)
+        assert outputs
+        assert any(_output_token_ids(out) for out in outputs)
+
+        # Trainer order is pause → abort → sleep. Pause here after the
+        # resume generate so EngineCore is idle before CuMem offload.
+        await engine.pause_generation(mode="abort", clear_cache=True)
+        await engine.sleep(level=1)
+        outputs.clear()
+        sleep_task = asyncio.create_task(_generate(8, "qwen3-sleep-admission", prompt))
+        await asyncio.sleep(1.0)
+        assert not sleep_task.done(), "generate() ran while EngineCore was sleeping (main-branch admission race)"
+        await engine.wake_up()
+        await asyncio.sleep(0.5)
+        assert not sleep_task.done(), "generate() resumed before resume_generation()"
+        await engine.resume_generation()
+        await asyncio.wait_for(sleep_task, timeout=180)
+        assert outputs
+        assert _finish_reason(outputs[-1]) != "abort"
+    await asyncio.sleep(5)
@@ -101,7 +101,8 @@ def duplex_camera_frames(*, seconds: int, cache_dir: Path) -> list[str]:
    from tests.helpers.media import generate_synthetic_video

    video = generate_synthetic_video(448, 448, 30 * seconds, cache_dir=cache_dir)
-    return _video_frames_from_file(Path(video["file_path"]))
+    frames, _ = _video_frames_from_file(Path(video["file_path"]))
+    return frames


 def realtime_url(omni_server) -> str:
@@ -252,6 +252,19 @@ def summarize_artifacts(
        (summary["done_indices"][0] for summary in reversed(response_summaries) if summary.get("done_indices")),
        None,
    )
+    # A soft interrupt's follow-up response can drain past the commit (residual
+    # model unit), so its response.done lands after input_audio_buffer.committed.
+    # Anchor the "listen after last done, before commit" sandwich on the last
+    # turn that closed its floor strictly before the commit, not the
+    # globally-last done (which may sit after the commit and leave the interval empty).
+    last_done_before_commit_index = next(
+        (
+            summary["done_indices"][0]
+            for summary in reversed(response_summaries)
+            if summary.get("done_indices") and commit_index is not None and summary["done_indices"][0] < commit_index
+        ),
+        None,
+    )
    listen_indices = [index for index, event in enumerate(events) if event.get("type") == "response.listen"]
    effective_min_responses = min_responses if validation_mode == "response-required" else 1
    enough_responses = len(response_summaries) >= effective_min_responses
@@ -279,9 +292,9 @@ def summarize_artifacts(
    )
    listen_after_last_done = last_done_index is not None and any(index > last_done_index for index in listen_indices)
    listen_after_response_before_commit = (
-        last_done_index is not None
+        last_done_before_commit_index is not None
        and commit_index is not None
-        and any(last_done_index < index < commit_index for index in listen_indices)
+        and any(last_done_before_commit_index < index < commit_index for index in listen_indices)
    )
    final_listen_after_commit = commit_index is not None and any(index > commit_index for index in listen_indices)
    transcript = "".join(str(summary.get("transcript") or "") for summary in response_summaries)
@@ -48,10 +48,6 @@ DEFAULT_SPEECH_TIMEOUT_S = 180.0
 # silence-only outputs without flagging short legitimate clips.
 _MIN_AUDIO_BYTES = 20_000

-_SKIP_ISSUE_6418 = pytest.mark.skip(
-    reason="https://github.com/vllm-project/vllm-omni/issues/6418",
-)
-
 # Reuse the shared TTS reference clip (clean ~5 s 24 kHz mono human speech)
 # vendored under tests/assets/qwen3_tts/. Keeps a single WAV across TTS
 # tests rather than duplicating asset bytes.
@@ -65,7 +61,7 @@ class TestHiggsAudioV3OnlineHappyPath:

    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_plain_text_wav(self, omni_server, online_client) -> None:
        """Single non-streaming WAV request - canonical TTS happy path."""
        online_client.send_audio_speech_request(
@@ -82,10 +78,9 @@ class TestHiggsAudioV3OnlineHappyPath:
            }
        )

-    @_SKIP_ISSUE_6418
    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_plain_text_with_max_new_tokens(self, omni_server, online_client) -> None:
        """``max_new_tokens`` is one of the few extra fields the v3 validator accepts."""
        online_client.send_audio_speech_request(
@@ -95,15 +90,14 @@ class TestHiggsAudioV3OnlineHappyPath:
                "stream": False,
                "response_format": "wav",
                "timeout": DEFAULT_SPEECH_TIMEOUT_S,
-                "max_new_tokens": 32,
+                "max_new_tokens": 1024,
                "min_audio_bytes": _MIN_AUDIO_BYTES,
            }
        )

-    @_SKIP_ISSUE_6418
    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_concurrent_plain_text(self, omni_server, online_client) -> None:
        """Three concurrent non-streaming requests - guards per-slot audio state and Stage-0 PC under batching."""
        online_client.send_audio_speech_request(
@@ -119,7 +113,7 @@ class TestHiggsAudioV3OnlineHappyPath:

    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_plain_text_pcm_streaming(self, omni_server, online_client) -> None:
        """Streaming PCM via talker2code2wav_async_chunk + the connector's codec_streaming path.

@@ -150,7 +144,7 @@ class TestHiggsAudioV3OnlineHappyPath:

    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_concurrent_pcm_streaming(self, omni_server, online_client) -> None:
        """Three concurrent streaming requests - guards per-request frame cursors
        in ``talker2code2wav_async_chunk`` and per-slot delay-pattern state under batched AR.
@@ -185,15 +179,13 @@ class TestHiggsAudioV3OnlineInlineControlTokens:
    tokens (pause, sfx) go inline. SFX tokens must pair with their written
    onomatopoeia.

-    These tests check the *serving surface* - the validator accepts the
-    payload, the engine produces audio, and the WAV has non-trivial size.
-    They do not assert audio quality (out of scope for CI).
+    These tests check the serving surface and spoken text. They do not assert
+    whether the requested control changes audio quality.
    """

-    @_SKIP_ISSUE_6418
    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_inline_emotion_and_expressive(self, omni_server, online_client) -> None:
        """Delivery tokens (emotion + expressive_high) at the start of input."""
        online_client.send_audio_speech_request(
@@ -204,13 +196,13 @@ class TestHiggsAudioV3OnlineInlineControlTokens:
                "response_format": "wav",
                "timeout": DEFAULT_SPEECH_TIMEOUT_S,
                "min_audio_bytes": _MIN_AUDIO_BYTES,
+                "transcript_expected_text": "Wait, that was actually hilarious.",
            }
        )

-    @_SKIP_ISSUE_6418
    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_inline_style_whispering(self, omni_server, online_client) -> None:
        """Style token at the start - ``whispering``."""
        online_client.send_audio_speech_request(
@@ -221,13 +213,13 @@ class TestHiggsAudioV3OnlineInlineControlTokens:
                "response_format": "wav",
                "timeout": DEFAULT_SPEECH_TIMEOUT_S,
                "min_audio_bytes": _MIN_AUDIO_BYTES,
+                "transcript_expected_text": "It is just between you and me, alright?",
            }
        )

-    @_SKIP_ISSUE_6418
    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_inline_prosody_speed_and_pitch(self, omni_server, online_client) -> None:
        """Two prosody tokens at the start - slow speed plus low pitch."""
        online_client.send_audio_speech_request(
@@ -240,13 +232,13 @@ class TestHiggsAudioV3OnlineInlineControlTokens:
                "response_format": "wav",
                "timeout": DEFAULT_SPEECH_TIMEOUT_S,
                "min_audio_bytes": _MIN_AUDIO_BYTES,
+                "transcript_expected_text": "The radar shows a storm approaching from the east.",
            }
        )

-    @_SKIP_ISSUE_6418
    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_inline_pause_mid_text(self, omni_server, online_client) -> None:
        """Positional pause token placed inline between two clauses."""
        online_client.send_audio_speech_request(
@@ -257,13 +249,13 @@ class TestHiggsAudioV3OnlineInlineControlTokens:
                "response_format": "wav",
                "timeout": DEFAULT_SPEECH_TIMEOUT_S,
                "min_audio_bytes": _MIN_AUDIO_BYTES,
+                "transcript_expected_text": "Hold on a moment. Let me think about it.",
            }
        )

-    @_SKIP_ISSUE_6418
    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_inline_sfx_with_onomatopoeia(self, omni_server, online_client) -> None:
        """SFX token paired with its written onomatopoeia (``<|sfx:laughter|>Hehe``)."""
        online_client.send_audio_speech_request(
@@ -277,6 +269,7 @@ class TestHiggsAudioV3OnlineInlineControlTokens:
                "response_format": "wav",
                "timeout": DEFAULT_SPEECH_TIMEOUT_S,
                "min_audio_bytes": _MIN_AUDIO_BYTES,
+                "transcript_expected_text": "I cannot believe that just happened. I am still recovering from it.",
            }
        )

@@ -287,7 +280,7 @@ class TestHiggsAudioV3OnlineVoiceClone:

    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_voice_clone_ref_audio_ref_text(self, omni_server, online_client) -> None:
        """Canonical vllm-omni voice clone via ``ref_audio`` + ``ref_text``."""
        online_client.send_audio_speech_request(
@@ -304,7 +297,7 @@ class TestHiggsAudioV3OnlineVoiceClone:

    @pytest.mark.slow
    @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
    def test_voice_clone_references_alias(self, omni_server, online_client) -> None:
        """BosonAI cookbook payload: ``references=[{audio_path, text}]``.

@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """
-E2E expansion tests for MiMo-Audio online serving (nightly Omni · Function Test with H100).
+E2E expansion tests for MiMo-Audio online serving (nightly Omni · Function Test with H100 · Single-GPU).
 """

 import os
@@ -540,6 +540,111 @@ def test_realtime_duplex_soft_interrupt_accepts_multi_delta_handoff_sequence(tmp
    assert summary["response_summaries"][1]["transcript"] == "一加一等于二。"


+def test_realtime_duplex_soft_interrupt_accepts_followup_done_after_commit(tmp_path):
+    """A follow-up response can drain a residual model unit past the final commit.
+
+    Sequence: ``r1.done -> listen -> r2.created -> ... -> commit -> r2.done -> listen``.
+
+    The globally-last ``response.done`` (r2) lands *after* the commit, so anchoring
+    ``listen_after_response_before_commit`` on it leaves the ``(last_done, commit)``
+    interval empty and fails the contract even though the turn yielded the floor
+    correctly. The check must anchor on the last done strictly before the commit
+    (r1.done), with the listen at 3.1 s satisfying the sandwich.
+    """
+    demo = _load_soft_interrupt_demo_module()
+    output = tmp_path / "soft_interrupt_residual_drain"
+    output.mkdir()
+    first_response_id = "resp-first"
+    second_response_id = "resp-second"
+    events = [
+        {"type": "response.listen", "_client_received_at_s": 1.0},
+        {"type": "response.created", "response": {"id": first_response_id}, "_client_received_at_s": 2.0},
+        {"type": "response.speak", "response_id": first_response_id, "_client_received_at_s": 2.0},
+        {
+            "type": "response.audio.delta",
+            "response_id": first_response_id,
+            "delta": "AAAA",
+            "_client_received_at_s": 2.1,
+        },
+        {
+            "type": "response.audio_transcript.delta",
+            "response_id": first_response_id,
+            "delta": "中国古代四大发明",
+            "_client_received_at_s": 2.1,
+        },
+        {
+            "type": "response.audio.delta",
+            "response_id": first_response_id,
+            "delta": "AAAA",
+            "_client_received_at_s": 2.6,
+        },
+        {
+            "type": "response.audio_transcript.delta",
+            "response_id": first_response_id,
+            "delta": "是造纸术。",
+            "_client_received_at_s": 2.6,
+        },
+        {"type": "response.done", "response_id": first_response_id, "_client_received_at_s": 3.0},
+        {"type": "response.listen", "_client_received_at_s": 3.1},
+        {"type": "response.created", "response": {"id": second_response_id}, "_client_received_at_s": 3.5},
+        {"type": "response.speak", "response_id": second_response_id, "_client_received_at_s": 3.5},
+        {
+            "type": "response.audio.delta",
+            "response_id": second_response_id,
+            "delta": "AAAA",
+            "_client_received_at_s": 3.6,
+        },
+        {
+            "type": "response.audio_transcript.delta",
+            "response_id": second_response_id,
+            "delta": "一加一等于",
+            "_client_received_at_s": 3.6,
+        },
+        {
+            "type": "response.audio.delta",
+            "response_id": second_response_id,
+            "delta": "AAAA",
+            "_client_received_at_s": 3.7,
+        },
+        {
+            "type": "response.audio_transcript.delta",
+            "response_id": second_response_id,
+            "delta": "二。",
+            "_client_received_at_s": 3.7,
+        },
+        # The final commit lands while the follow-up response is still draining.
+        {"type": "input_audio_buffer.committed", "_client_received_at_s": 4.0},
+        {"type": "response.done", "response_id": second_response_id, "_client_received_at_s": 4.5},
+        {"type": "response.listen", "_client_received_at_s": 4.6},
+    ]
+    (output / "events.jsonl").write_text(
+        "".join(demo.json.dumps(event) + "\n" for event in events),
+        encoding="utf-8",
+    )
+    (output / "result.json").write_text(
+        demo.json.dumps({"ok": True, "response_ids": [first_response_id, second_response_id]}),
+        encoding="utf-8",
+    )
+
+    summary = demo.summarize_artifacts(
+        output_dir=output,
+        validation_mode="response-required",
+        min_responses=2,
+        min_audio_deltas_per_response=2,
+        expect_followup_response_substring="一加一等于二",
+    )
+
+    assert summary["ok"] is True
+    # Anchored on r1.done (the last done before the commit), not the late r2.done.
+    assert summary["listen_after_response_before_commit"] is True
+    # listen_after_last_done still keys off the globally-last done, hence the final listen.
+    assert summary["listen_after_last_done"] is True
+    assert summary["final_listen_after_commit"] is True
+    assert summary["second_response_before_final_commit"] is True
+    assert summary["followup_response_transcript_ok"] is True
+    assert summary["response_summaries"][1]["transcript"] == "一加一等于二。"
+
+
 def test_realtime_duplex_soft_interrupt_model_policy_accepts_single_response(tmp_path):
    demo = _load_soft_interrupt_demo_module()
    output = tmp_path / "model_policy"
@@ -1926,6 +2031,28 @@ def test_realtime_duplex_demo_sends_each_units_stacked_composite_next_to_its_bas
    assert _sent_video_frames(messages) == [["f0", "s0"], ["f1", "s1"], ["f2"]]


+def test_minicpmo_duplex_camera_fixture_returns_flat_base_frame_track(tmp_path, monkeypatch):
+    from tests.e2e.online_serving.helpers import minicpmo_4_5_duplex as fixtures
+    from tests.e2e.online_serving.helpers import minicpmo_realtime_duplex_scenarios as demo
+    from tests.helpers import media
+
+    video_path = tmp_path / "camera.mp4"
+    monkeypatch.setattr(
+        media,
+        "generate_synthetic_video",
+        lambda *args, **kwargs: {"file_path": str(video_path)},
+    )
+    monkeypatch.setattr(
+        demo,
+        "_video_frames_from_file",
+        lambda path: (["frame-0", "frame-1"], [None, None]),
+    )
+
+    frames = fixtures.duplex_camera_frames(seconds=2, cache_dir=tmp_path)
+
+    assert frames == ["frame-0", "frame-1"]
+
+
 def test_realtime_duplex_demo_decodes_a_video_file_into_one_frame_per_second(tmp_path):
    pytest.importorskip("cv2")
    pytest.importorskip("imageio")
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """Unit tests for acknowledged abort on AsyncOmniEngine."""

 from __future__ import annotations
@@ -19,6 +22,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]

 def _make_engine(request_q: queue.Queue, rpc_q: queue.Queue) -> AsyncOmniEngine:
    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = False
    engine.request_queue = SimpleNamespace(sync_q=request_q)
    engine.rpc_output_queue = SimpleNamespace(sync_q=rpc_q)
    engine._correlated_rpc_client = CorrelatedRpcClient(request_q, rpc_q)
@@ -63,7 +67,8 @@ def test_abort_async_waits_for_ack(mocker: MockerFixture):
        assert not task.done()

        rpc_q.put(AbortResultMessage(rpc_id="abort-rpc-1", success=True))
-        await task
+        result = await task
+        assert result == []

    try:
        asyncio.run(_run())
@@ -154,3 +159,110 @@ def test_abort_async_preserves_request_queue_backpressure(mocker: MockerFixture)
        asyncio.run(_run())
    finally:
        engine._correlated_rpc_client.close()
+
+
+def test_abort_async_is_noop_after_shutdown_starts():
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = True
+    engine.request_queue = None
+    engine._correlated_rpc_client = None
+
+    asyncio.run(engine.abort_async(["req-late"]))
+
+
+def test_abort_async_is_noop_for_empty_request_ids():
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = False
+    engine.request_queue = None
+    engine._correlated_rpc_client = None
+
+    asyncio.run(engine.abort_async([]))
+
+
+def test_abort_async_tolerates_request_queue_close_during_shutdown(
+    mocker: MockerFixture,
+):
+    class SyntheticSyncQueueShutDownError(Exception):
+        pass
+
+    rpc_q: queue.Queue = queue.Queue()
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = False
+
+    class ClosingRequestQueue:
+        def put(self, item) -> None:
+            del item
+            engine._shutdown_called = True
+            raise SyntheticSyncQueueShutDownError
+
+    request_q = ClosingRequestQueue()
+    engine.request_queue = SimpleNamespace(sync_q=request_q)
+    engine._correlated_rpc_client = CorrelatedRpcClient(request_q, rpc_q)
+    mocker.patch(
+        "vllm_omni.engine.async_engine_utils._JANUS_SYNC_QUEUE_SHUTDOWN",
+        SyntheticSyncQueueShutDownError,
+    )
+
+    try:
+        asyncio.run(engine.abort_async(["req-submit-race"], timeout=1))
+    finally:
+        engine._correlated_rpc_client.close()
+
+
+def test_abort_async_surfaces_request_queue_close_while_engine_is_live(
+    mocker: MockerFixture,
+):
+    class SyntheticSyncQueueShutDownError(Exception):
+        pass
+
+    class ClosedRequestQueue:
+        def put(self, item) -> None:
+            del item
+            raise SyntheticSyncQueueShutDownError
+
+    request_q = ClosedRequestQueue()
+    rpc_q: queue.Queue = queue.Queue()
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = False
+    engine.request_queue = SimpleNamespace(sync_q=request_q)
+    engine._correlated_rpc_client = CorrelatedRpcClient(request_q, rpc_q)
+    mocker.patch(
+        "vllm_omni.engine.async_engine_utils._JANUS_SYNC_QUEUE_SHUTDOWN",
+        SyntheticSyncQueueShutDownError,
+    )
+
+    try:
+        with pytest.raises(SyntheticSyncQueueShutDownError):
+            asyncio.run(engine.abort_async(["req-live-submit"], timeout=1))
+    finally:
+        engine._correlated_rpc_client.close()
+
+
+def test_abort_async_tolerates_rpc_router_close_during_shutdown():
+    request_q: queue.Queue = queue.Queue()
+    rpc_q: queue.Queue = queue.Queue()
+    engine = _make_engine(request_q, rpc_q)
+
+    async def _run() -> None:
+        task = asyncio.create_task(engine.abort_async(["req-ack-race"], timeout=1))
+        await asyncio.to_thread(request_q.get, True, 1)
+        engine._shutdown_called = True
+        engine._correlated_rpc_client.close()
+        await task
+
+    asyncio.run(_run())
+
+
+def test_abort_async_surfaces_rpc_router_close_while_engine_is_live():
+    request_q: queue.Queue = queue.Queue()
+    rpc_q: queue.Queue = queue.Queue()
+    engine = _make_engine(request_q, rpc_q)
+
+    async def _run() -> None:
+        task = asyncio.create_task(engine.abort_async(["req-live"], timeout=1))
+        await asyncio.to_thread(request_q.get, True, 1)
+        engine._correlated_rpc_client.close()
+        with pytest.raises(RuntimeError, match="RPC result router closed"):
+            await task
+
+    asyncio.run(_run())
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 import pytest
 import torch
 from pytest_mock import MockerFixture
@@ -14,6 +17,104 @@ from vllm_omni.model_executor.stage_input_processors.bagel import ExpandedPrompt
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]


+class _SyntheticSyncQueueShutDownError(Exception):
+    pass
+
+
+def test_abort_is_noop_for_empty_request_ids(mocker: MockerFixture):
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = False
+    engine.request_queue = mocker.Mock()
+
+    engine.abort([])
+
+    engine.request_queue.sync_q.put.assert_not_called()
+
+
+def test_abort_is_noop_after_shutdown_starts(mocker: MockerFixture):
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = True
+    engine.request_queue = mocker.Mock()
+
+    engine.abort(["req-1"])
+
+    engine.request_queue.sync_q.put.assert_not_called()
+
+
+def test_abort_tolerates_queue_close_race_during_shutdown(mocker: MockerFixture):
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = False
+    engine.request_queue = mocker.Mock()
+    mocker.patch(
+        "vllm_omni.engine.async_engine_utils._JANUS_SYNC_QUEUE_SHUTDOWN",
+        _SyntheticSyncQueueShutDownError,
+    )
+
+    def close_queue_during_abort(*args, **kwargs):
+        del args, kwargs
+        engine._shutdown_called = True
+        raise _SyntheticSyncQueueShutDownError
+
+    engine.request_queue.sync_q.put.side_effect = close_queue_during_abort
+
+    engine.abort(["req-1"])
+
+
+def test_abort_tolerates_legacy_janus_close_error_during_shutdown(mocker: MockerFixture):
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = False
+    engine.request_queue = mocker.Mock()
+    mocker.patch(
+        "vllm_omni.engine.async_engine_utils._JANUS_SYNC_QUEUE_SHUTDOWN",
+        None,
+    )
+
+    def close_queue_during_abort(*args, **kwargs):
+        del args, kwargs
+        engine._shutdown_called = True
+        raise RuntimeError("Operation on the closed queue is forbidden")
+
+    engine.request_queue.sync_q.put.side_effect = close_queue_during_abort
+
+    engine.abort(["req-1"])
+
+
+def test_abort_surfaces_unrelated_legacy_runtime_error_during_shutdown(
+    mocker: MockerFixture,
+):
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = False
+    engine.request_queue = mocker.Mock()
+    mocker.patch(
+        "vllm_omni.engine.async_engine_utils._JANUS_SYNC_QUEUE_SHUTDOWN",
+        None,
+    )
+
+    def fail_during_abort(*args, **kwargs):
+        del args, kwargs
+        engine._shutdown_called = True
+        raise RuntimeError("unrelated queue failure")
+
+    engine.request_queue.sync_q.put.side_effect = fail_during_abort
+
+    with pytest.raises(RuntimeError, match="unrelated queue failure"):
+        engine.abort(["req-1"])
+
+
+def test_abort_surfaces_unexpected_closed_queue(mocker: MockerFixture):
+    engine = object.__new__(AsyncOmniEngine)
+    engine._shutdown_called = False
+    engine.request_queue = mocker.Mock()
+    mocker.patch(
+        "vllm_omni.engine.async_engine_utils._JANUS_SYNC_QUEUE_SHUTDOWN",
+        _SyntheticSyncQueueShutDownError,
+    )
+    engine.request_queue.sync_q.put.side_effect = _SyntheticSyncQueueShutDownError
+
+    with pytest.raises(_SyntheticSyncQueueShutDownError):
+        engine.abort(["req-1"])
+
+
 def _make_engine_core_request(request_id: str = "req-1") -> EngineCoreRequest:
    return EngineCoreRequest(
        request_id=request_id,
@@ -4,6 +4,7 @@
 import concurrent.futures
 import contextlib
 import importlib
+import json
 import os
 import time
 import types
@@ -1093,6 +1094,8 @@ def test_build_engine_args_cli_tokenizer_overrides_inferred_base_tokenizer(tmp_p
        default_sampling_params={},
    )

+    (tmp_path / "llm").mkdir()
+
    engine_args = build_engine_args_dict(
        stage_cfg,
        str(tmp_path),
@@ -1154,6 +1157,9 @@ def test_build_engine_args_keeps_stage_owned_tokenizer_subdir(tmp_path):
        default_sampling_params={},
    )

+    (tmp_path / "llm").mkdir()
+    (tmp_path / "tokenizer").mkdir()
+
    engine_args = build_engine_args_dict(
        stage_cfg,
        str(tmp_path),
@@ -1164,6 +1170,430 @@ def test_build_engine_args_keeps_stage_owned_tokenizer_subdir(tmp_path):
    assert engine_args["tokenizer"] == os.path.join(str(tmp_path), "tokenizer")


+def test_build_engine_args_rejects_missing_subdir_in_local_model_dir(tmp_path):
+    """A local model directory without the stage subfolder must fail closed.
+
+    Forwarding the joined path as ``EngineArgs.model`` sends a non-directory
+    string to HuggingFace, which reports it as a malformed repo id.
+    """
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={"model_subdir": "language_model", "tokenizer_subdir": "tokenizer"},
+        default_sampling_params={},
+    )
+    (tmp_path / "tokenizer").mkdir()
+
+    with pytest.raises(RuntimeError, match="language_model"):
+        build_engine_args_dict(stage_cfg, str(tmp_path))
+
+
+# A real snapshot subfolder always carries artifacts; empty directories would
+# assert the broken exists-means-complete predicate this suite regresses.
+_SUBDIR_ARTIFACT = {"language_model": "model.safetensors", "tokenizer": "tokenizer.json"}
+
+
+def _make_snapshot(root, subdirs):
+    for subdir in subdirs:
+        folder = root / subdir
+        folder.mkdir(parents=True, exist_ok=True)
+        (folder / _SUBDIR_ARTIFACT.get(subdir, "data.bin")).write_text("x")
+    return str(root)
+
+
+def test_build_engine_args_pulls_stage_subdirs_missing_from_cached_snapshot(monkeypatch, tmp_path):
+    """Regression for issue #6638.
+
+    ``snapshot_download(local_files_only=True)`` hands back a snapshot root as
+    soon as the repo is partly cached, even when this stage's subfolders were
+    never materialized. Stage init must fetch exactly those subfolders instead
+    of joining onto a path that exists nowhere.
+    """
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    snapshot = tmp_path / "snapshots" / "deadbeef"
+    _make_snapshot(snapshot, ["tokenizer"])
+    calls = []
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        calls.append(kwargs)
+        if kwargs.get("local_files_only"):
+            return str(snapshot)
+        _make_snapshot(snapshot, ["language_model"])
+        return str(snapshot)
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={"model_subdir": "language_model", "tokenizer_subdir": "tokenizer"},
+        default_sampling_params={},
+    )
+
+    engine_args = build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+    assert engine_args["model"] == os.path.join(str(snapshot), "language_model")
+    assert engine_args["tokenizer"] == os.path.join(str(snapshot), "tokenizer")
+    assert os.path.isdir(engine_args["model"])
+    # Only the stage's own subfolders are pulled, not the whole repo.
+    assert calls[-1]["allow_patterns"] == ["language_model/*", "tokenizer/*"]
+
+
+def test_build_engine_args_skips_hub_call_when_cached_snapshot_is_complete(monkeypatch, tmp_path):
+    """A warm cache stays offline-friendly: no outgoing Hub request."""
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    snapshot = tmp_path / "snapshots" / "deadbeef"
+    _make_snapshot(snapshot, ["language_model", "tokenizer"])
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        assert kwargs.get("local_files_only"), "warm cache must not reach the Hub"
+        return str(snapshot)
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={"model_subdir": "language_model", "tokenizer_subdir": "tokenizer"},
+        default_sampling_params={},
+    )
+
+    engine_args = build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+    assert engine_args["model"] == os.path.join(str(snapshot), "language_model")
+
+
+def test_build_engine_args_redownloads_a_partial_subdir(monkeypatch, tmp_path):
+    """A subfolder holding only config.json is an interrupted download.
+
+    The warm-cache early return must not accept it: converting the Hub ID into
+    that local path leaves vLLM's loader without any fallback for the missing
+    weights.
+    """
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    snapshot = tmp_path / "snapshots" / "deadbeef"
+    partial = snapshot / "language_model"
+    partial.mkdir(parents=True)
+    (partial / "config.json").write_text("{}")
+    _make_snapshot(snapshot, ["tokenizer"])
+    calls = []
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        calls.append(kwargs)
+        if kwargs.get("local_files_only"):
+            return str(snapshot)
+        _make_snapshot(snapshot, ["language_model"])
+        return str(snapshot)
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={"model_subdir": "language_model", "tokenizer_subdir": "tokenizer"},
+        default_sampling_params={},
+    )
+
+    engine_args = build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+    assert any(not call.get("local_files_only") for call in calls), "a partial subdir must fall through to the Hub"
+    assert engine_args["model"] == os.path.join(str(snapshot), "language_model")
+
+
+def test_build_engine_args_redownloads_when_index_lists_missing_shards(monkeypatch, tmp_path):
+    """An index plus a subset of its shards is an interrupted download.
+
+    The index downloads early; accepting it as proof of weights converts the
+    Hub ID into a local path vLLM cannot fetch the remaining shards for.
+    """
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    snapshot = tmp_path / "snapshots" / "deadbeef"
+    partial = snapshot / "language_model"
+    partial.mkdir(parents=True)
+    (partial / "model.safetensors.index.json").write_text(
+        json.dumps(
+            {
+                "weight_map": {
+                    "a.weight": "model-00001-of-00002.safetensors",
+                    "b.weight": "model-00002-of-00002.safetensors",
+                }
+            }
+        )
+    )
+    (partial / "model-00001-of-00002.safetensors").write_text("x")
+    _make_snapshot(snapshot, ["tokenizer"])
+    calls = []
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        calls.append(kwargs)
+        if not kwargs.get("local_files_only"):
+            (partial / "model-00002-of-00002.safetensors").write_text("x")
+        return str(snapshot)
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={"model_subdir": "language_model", "tokenizer_subdir": "tokenizer"},
+        default_sampling_params={},
+    )
+
+    engine_args = build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+    assert any(not call.get("local_files_only") for call in calls), "missing shards must fall through to the Hub"
+    assert engine_args["model"] == os.path.join(str(snapshot), "language_model")
+
+
+def test_build_engine_args_redownloads_shards_without_their_index(monkeypatch, tmp_path):
+    """Shard-named weights always ship an index; one without it is partial."""
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    snapshot = tmp_path / "snapshots" / "deadbeef"
+    partial = snapshot / "language_model"
+    partial.mkdir(parents=True)
+    (partial / "model-00001-of-00004.safetensors").write_text("x")
+    _make_snapshot(snapshot, ["tokenizer"])
+    calls = []
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        calls.append(kwargs)
+        if not kwargs.get("local_files_only"):
+            shards = {f"w{i}.weight": f"model-0000{i}-of-00004.safetensors" for i in range(1, 5)}
+            (partial / "model.safetensors.index.json").write_text(json.dumps({"weight_map": shards}))
+            for shard in shards.values():
+                (partial / shard).write_text("x")
+        return str(snapshot)
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={"model_subdir": "language_model", "tokenizer_subdir": "tokenizer"},
+        default_sampling_params={},
+    )
+
+    build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+    assert any(not call.get("local_files_only") for call in calls)
+
+
+def test_build_engine_args_redownloads_a_tokenizer_folder_without_vocabulary(monkeypatch, tmp_path):
+    """Templates and configs download first; alone they are not a tokenizer."""
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    snapshot = tmp_path / "snapshots" / "deadbeef"
+    _make_snapshot(snapshot, ["language_model"])
+    partial = snapshot / "tokenizer"
+    partial.mkdir(parents=True)
+    (partial / "chat_template.jinja").write_text("{{ messages }}")
+    calls = []
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        calls.append(kwargs)
+        if not kwargs.get("local_files_only"):
+            _make_snapshot(snapshot, ["tokenizer"])
+        return str(snapshot)
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={"model_subdir": "language_model", "tokenizer_subdir": "tokenizer"},
+        default_sampling_params={},
+    )
+
+    engine_args = build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+    assert any(not call.get("local_files_only") for call in calls)
+    assert engine_args["tokenizer"] == os.path.join(str(snapshot), "tokenizer")
+
+
+def test_build_engine_args_forwards_revision_and_download_dir(monkeypatch, tmp_path):
+    """revision/download_dir must shape snapshot selection (they cannot be
+    corrected downstream once the repo ID is a local path)."""
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    snapshot = tmp_path / "snapshots" / "pinned"
+    _make_snapshot(snapshot, ["language_model", "tokenizer"])
+    seen = []
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        seen.append(kwargs)
+        return str(snapshot)
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={
+            "model_subdir": "language_model",
+            "tokenizer_subdir": "tokenizer",
+            "revision": "v2.0",
+            "download_dir": str(tmp_path / "cache"),
+        },
+        default_sampling_params={},
+    )
+
+    engine_args = build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+    assert engine_args["model"] == os.path.join(str(snapshot), "language_model")
+    assert seen, "snapshot resolution must have been consulted"
+    for call in seen:
+        assert call.get("revision") == "v2.0"
+        assert call.get("cache_dir") == str(tmp_path / "cache")
+
+
+def test_build_engine_args_resolves_tokenizer_revision_separately(monkeypatch, tmp_path):
+    """A tokenizer pinned to a different revision resolves against its own snapshot."""
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    model_snapshot = tmp_path / "snapshots" / "model-rev"
+    _make_snapshot(model_snapshot, ["language_model"])
+    tokenizer_snapshot = tmp_path / "snapshots" / "tok-rev"
+    _make_snapshot(tokenizer_snapshot, ["tokenizer"])
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        if kwargs.get("revision") == "tok-rev":
+            return str(tokenizer_snapshot)
+        return str(model_snapshot)
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={
+            "model_subdir": "language_model",
+            "tokenizer_subdir": "tokenizer",
+            "revision": "model-rev",
+            "tokenizer_revision": "tok-rev",
+        },
+        default_sampling_params={},
+    )
+
+    engine_args = build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+    assert engine_args["model"] == os.path.join(str(model_snapshot), "language_model")
+    assert engine_args["tokenizer"] == os.path.join(str(tokenizer_snapshot), "tokenizer")
+
+
+def test_build_engine_args_resolves_root_tokenizer_revision_separately(monkeypatch, tmp_path):
+    """An empty tokenizer_subdir targets the snapshot root and still honors
+    its own revision."""
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    model_snapshot = tmp_path / "snapshots" / "model-rev"
+    _make_snapshot(model_snapshot, ["language_model"])
+    # An empty tokenizer_subdir means the tokenizer IS the snapshot root, so
+    # the vocabulary artifact must sit at the root. Building it under
+    # ``tokenizer/`` returned a root that holds no loadable tokenizer, which
+    # the root-tokenizer validation now correctly rejects.
+    tokenizer_snapshot = tmp_path / "snapshots" / "tok-rev"
+    tokenizer_snapshot.mkdir(parents=True, exist_ok=True)
+    (tokenizer_snapshot / _SUBDIR_ARTIFACT["tokenizer"]).write_text("x")
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        if kwargs.get("revision") == "tok-rev":
+            return str(tokenizer_snapshot)
+        return str(model_snapshot)
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={
+            "model_subdir": "language_model",
+            "tokenizer_subdir": "",
+            "revision": "model-rev",
+            "tokenizer_revision": "tok-rev",
+        },
+        default_sampling_params={},
+    )
+
+    engine_args = build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+    assert engine_args["model"] == os.path.join(str(model_snapshot), "language_model")
+    assert engine_args["tokenizer"] == str(tokenizer_snapshot)
+
+
+def test_build_engine_args_fails_closed_when_subdir_cannot_be_downloaded(monkeypatch, tmp_path):
+    """An undownloadable subfolder raises here instead of reaching HuggingFace."""
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    snapshot = tmp_path / "snapshots" / "deadbeef"
+    _make_snapshot(snapshot, ["tokenizer"])
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        if kwargs.get("local_files_only"):
+            return str(snapshot)
+        raise OSError("offline")
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={"model_subdir": "language_model", "tokenizer_subdir": "tokenizer"},
+        default_sampling_params={},
+    )
+
+    with pytest.raises(RuntimeError, match="language_model"):
+        build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+
+def test_build_engine_args_fails_closed_on_cold_cache_instead_of_joining_repo_id(monkeypatch):
+    """With nothing cached, the stage must not join a subdir onto the repo id."""
+    from huggingface_hub import HfApi
+
+    from vllm_omni.engine.stage_init_utils import build_engine_args_dict
+
+    def fake_snapshot_download(self, repo_id, **kwargs):
+        raise OSError("offline")
+
+    monkeypatch.setattr(HfApi, "snapshot_download", fake_snapshot_download)
+
+    stage_cfg = types.SimpleNamespace(
+        stage_id=0,
+        stage_type="llm",
+        engine_args={"model_subdir": "language_model", "tokenizer_subdir": "tokenizer"},
+        default_sampling_params={},
+    )
+
+    with pytest.raises(RuntimeError, match="nothing is cached locally"):
+        build_engine_args_dict(stage_cfg, "MiniMaxAI/MiniMax-Music3")
+
+
 def test_model_path_resolver_is_generic_and_model_owned(tmp_path):
    from vllm_omni.engine.stage_init_utils import _resolve_model_path

@@ -35,6 +35,7 @@ from vllm_omni.engine.messages import (
 from vllm_omni.engine.orchestrator import (
    Orchestrator,
    OrchestratorRequestState,
+    StreamingSegmentState,
    _build_terminal_empty_output,
 )
 from vllm_omni.engine.stage_pool import StagePool
@@ -256,8 +257,32 @@ class FakeOutputProcessor:
        )

    def abort_requests(self, request_ids, internal: bool = False):
-        self.abort_calls.append(request_ids)
-        return request_ids
+        aborted_ids, _outputs = self.abort_requests_collecting_outputs(request_ids, internal=internal)
+        return aborted_ids
+
+    def abort_requests_collecting_outputs(self, request_ids, *, internal: bool = False, commit_state: bool = True):
+        """Mirror MultimodalOutputProcessor collecting for AR abort-prefix tests."""
+        del internal, commit_state
+        ids = list(request_ids)
+        self.abort_calls.append(ids)
+        outputs: list[RequestOutput] = []
+        for rid in ids:
+            seeded = next(
+                (ro for ro in self.request_outputs if getattr(ro, "request_id", None) == rid),
+                None,
+            )
+            token_ids = list(seeded.outputs[0].token_ids) if seeded is not None and seeded.outputs else [1, 2]
+            # Intentionally stamp an internal-looking id on the RequestOutput so
+            # StagePool must re-key by the orchestrator id it passed in.
+            outputs.append(
+                _build_request_output(
+                    f"engine-internal-{rid}",
+                    token_ids=token_ids,
+                    finished=True,
+                    finish_reason="abort",
+                )
+            )
+        return ids, outputs

    def update_scheduler_stats(self, _scheduler_stats) -> None:
        return None
@@ -320,6 +345,7 @@ def _build_request_output(
    prompt_token_ids: list[int] | None = None,
    finished: bool = True,
    text: str = "test",
+    finish_reason: str | None = None,
 ) -> RequestOutput:
    completion = CompletionOutput(
        index=0,
@@ -327,7 +353,7 @@ def _build_request_output(
        token_ids=list(token_ids or [1, 2]),
        cumulative_logprob=0.0,
        logprobs=None,
-        finish_reason="stop" if finished else None,
+        finish_reason=(finish_reason if finish_reason is not None else ("stop" if finished else None)),
        stop_reason=None,
    )
    return RequestOutput(
@@ -983,7 +1009,11 @@ async def test_run_abort(orchestrator_factory) -> None:
 async def test_run_abort_emits_result_when_rpc_id_set(orchestrator_factory) -> None:
    stages = [
        FakeStageClient(stage_type="llm", final_output=False),
-        FakeStageClient(stage_type="llm", final_output=True),
+        FakeStageClient(
+            stage_type="llm",
+            final_output=True,
+            next_inputs=[{"prompt_token_ids": [7, 8, 9]}],
+        ),
    ]
    processors = [
        FakeOutputProcessor(request_outputs=[_build_request_output("req-ack", token_ids=[1], finished=True)]),
@@ -1002,6 +1032,11 @@ async def test_run_abort_emits_result_when_rpc_id_set(orchestrator_factory) -> N
            final_stage_id=1,
        )
        await _wait_for(lambda: len(stages[0].add_request_calls) == 1)
+        # Abort outputs come from the final AR stage's output processor, and
+        # StagePool.abort_requests only collects for requests with a live
+        # replica binding. Forward off stage-0 first so stage-1 is bound.
+        stages[0].push_engine_core_outputs(_engine_core_outputs("stage0-raw", 1.0))
+        await _wait_for(lambda: len(stages[1].add_request_calls) == 1)

        orchestrator_fixture.request_sync_q.put_nowait(AbortRequestMessage(request_ids=["req-ack"], rpc_id="abort-1"))
        result = await _get_rpc_message(orchestrator_fixture)
@@ -1011,11 +1046,53 @@ async def test_run_abort_emits_result_when_rpc_id_set(orchestrator_factory) -> N
        assert result.error is None
        assert result.rpc_correlation_key == ("abort", "abort-1")
        assert stages[0].abort_calls == [["req-ack"]]
+        assert stages[1].abort_calls == [["req-ack"]]
        assert "req-ack" not in orchestrator_fixture.orchestrator.request_states
+        assert result.abort_outputs is not None
+        assert len(result.abort_outputs) == 1
+        abort_msg = result.abort_outputs[0]
+        assert abort_msg.request_id == "req-ack"
+        assert abort_msg.finished is True
+        assert abort_msg.stage_id == 1
+        assert list(abort_msg.engine_outputs.outputs[0].token_ids) == [2]
+        assert abort_msg.engine_outputs.outputs[0].finish_reason == "abort"
    finally:
        await _shutdown_orchestrator(orchestrator_fixture)


+@pytest.mark.asyncio
+async def test_abort_marks_only_last_final_stage_output_finished() -> None:
+    """A request with two final AR stages must keep earlier abort outputs unfinished."""
+    stages = [
+        FakeStageClient(stage_type="llm", final_output=True),
+        FakeStageClient(stage_type="llm", final_output=True),
+    ]
+    processors = [
+        FakeOutputProcessor(request_outputs=[_build_request_output("req-two", token_ids=[1], finished=True)]),
+        FakeOutputProcessor(request_outputs=[_build_request_output("req-two", token_ids=[2], finished=True)]),
+    ]
+    pools = _build_stage_pools([[stages[0]], [stages[1]]], output_processors=processors)
+    pools[0]._request_bindings["req-two"] = 0
+    pools[1]._request_bindings["req-two"] = 0
+
+    orchestrator = Orchestrator(
+        request_async_queue=asyncio.Queue(),
+        output_async_queue=asyncio.Queue(),
+        rpc_async_queue=asyncio.Queue(),
+        stage_pools=pools,
+    )
+    orchestrator.request_states["req-two"] = OrchestratorRequestState(
+        request_id="req-two",
+        final_stage_id=1,
+        final_output_stage_ids={0, 1},
+    )
+
+    outputs = await orchestrator._abort_request_ids(["req-two"])
+    assert [(msg.stage_id, msg.finished) for msg in outputs] == [(0, False), (1, True)]
+    assert list(outputs[0].engine_outputs.outputs[0].token_ids) == [1]
+    assert list(outputs[1].engine_outputs.outputs[0].token_ids) == [2]
+
+
@pytest.mark.asyncio
 async def test_handle_abort_emits_error_result_on_failure() -> None:
    rpc_queue: asyncio.Queue = asyncio.Queue()
@@ -2183,7 +2260,7 @@ async def test_resumable_segment_boundary_builds_stage_metrics() -> None:
        final_stage_id=0,
    )
    req_state.streaming.enabled = True
-    req_state.streaming.segment_finished = True
+    req_state.streaming.segments[0] = StreamingSegmentState(finished=True)
    req_state.stage_submit_ts[0] = time.time()
    orchestrator.request_states = {"req-stream": req_state}
    orchestrator.stage_pools = [pool]
@@ -2216,7 +2293,12 @@ def test_stage_pool_metrics_use_resumable_segment_token_count() -> None:
    )
    output = SimpleNamespace(
        request_id="req-stream",
-        outputs=[SimpleNamespace(cumulative_token_ids=list(range(11)))],
+        outputs=[
+            SimpleNamespace(
+                cumulative_token_ids=list(range(11)),
+                finish_reason=FinishReason.LENGTH,
+            )
+        ],
    )

    metrics = pool.build_stage_metrics(
@@ -2228,6 +2310,7 @@ def test_stage_pool_metrics_use_resumable_segment_token_count() -> None:

    assert metrics.num_tokens_out == 3
    assert metrics.output_unit_count == 3
+    assert metrics.finish_reason == "length"


 def test_image_ttfo_preserves_request_time_and_tracks_stage_time() -> None:
@@ -0,0 +1,130 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Streaming segment-boundary state must be keyed by stage.
+
+``Orchestrator._orchestration_loop`` polls every stage into the same
+``OrchestratorRequestState``. While the segment-boundary fields lived in a single
+flat slot per request, whichever stage polled most recently overwrote the others,
+and the per-stage consumers then read back a different stage's boundary.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+import pytest
+from vllm.sampling_params import SamplingParams
+
+from vllm_omni.engine.orchestrator import (
+    Orchestrator,
+    OrchestratorRequestState,
+    StreamingSegmentState,
+)
+from vllm_omni.experimental.fullduplex.engine.contracts import DuplexRequestIdentity
+from vllm_omni.experimental.fullduplex.engine.messages import DuplexFence
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+THINKER_STAGE = 0
+TALKER_STAGE = 1
+
+
+@dataclass
+class _StageMetrics:
+    pipeline_timings: dict[str, float] = field(default_factory=dict)
+
+
+@dataclass
+class _StageOutput:
+    """Stand-in for a processed stage output, with the fields the router reads."""
+
+    request_id: str
+    finished: bool = False
+    error: object | None = None
+
+
+class _RecordingPool:
+    """Stage pool that records the requests it was asked to build metrics for."""
+
+    def __init__(self) -> None:
+        self.built_for: list[str] = []
+
+    def build_stage_metrics(self, outputs: list[_StageOutput], **_kwargs: object) -> _StageMetrics:
+        self.built_for.extend(output.request_id for output in outputs)
+        return _StageMetrics()
+
+
+def _duplex_req_state(request_id: str = "req-duplex") -> OrchestratorRequestState:
+    req_state = OrchestratorRequestState(
+        request_id=request_id,
+        sampling_params_list=[SamplingParams(max_tokens=1), SamplingParams(max_tokens=1)],
+        final_stage_id=TALKER_STAGE,
+    )
+    req_state.streaming.enabled = True
+    req_state.duplex_identity = DuplexRequestIdentity(
+        session_id="session-1",
+        fence=DuplexFence(session_id="session-1"),
+    )
+    return req_state
+
+
+def test_duplex_output_context_reads_only_its_own_stage_segment() -> None:
+    req_state = _duplex_req_state()
+    req_state.streaming.segments[THINKER_STAGE] = StreamingSegmentState(
+        finished=True,
+        token_ids=[11, 22],
+        output_metadata={"stage": "thinker"},
+    )
+    # The talker polls after the thinker and is still mid-segment.
+    req_state.streaming.segments[TALKER_STAGE] = StreamingSegmentState(finished=False)
+
+    thinker = Orchestrator._duplex_output_context(req_state, stage_id=THINKER_STAGE)
+    talker = Orchestrator._duplex_output_context(req_state, stage_id=TALKER_STAGE)
+
+    assert thinker is not None
+    assert talker is not None
+
+    # These are the assertions that fail on the flat field: the talker is written
+    # second, so the thinker loses its boundary to whichever stage polled last.
+    assert thinker.segment_finished is True
+    assert thinker.segment_token_ids == (11, 22)
+    assert thinker.segment_output_metadata == {"stage": "thinker"}
+
+    assert talker.segment_finished is False
+    assert talker.segment_token_ids == ()
+    assert talker.segment_output_metadata == {}
+
+
+@pytest.mark.asyncio
+async def test_segment_boundary_on_one_stage_does_not_build_metrics_on_another() -> None:
+    req_state = _duplex_req_state()
+    req_state.streaming.segments[THINKER_STAGE] = StreamingSegmentState(finished=True)
+
+    talker_pool = _RecordingPool()
+    orchestrator = object.__new__(Orchestrator)
+    orchestrator.request_states = {req_state.request_id: req_state}
+    orchestrator.stage_pools = [_RecordingPool(), talker_pool]
+
+    routed: list[_StageMetrics | None] = []
+
+    async def record_route(
+        _stage: int,
+        _replica: int,
+        _output: _StageOutput,
+        _state: OrchestratorRequestState,
+        metrics: _StageMetrics | None,
+    ) -> None:
+        routed.append(metrics)
+
+    orchestrator._route_output = record_route
+
+    # The talker's own output is neither finished nor at a segment boundary, so
+    # it must not be treated as one just because the thinker reached one.
+    await orchestrator._handle_processed_outputs(
+        TALKER_STAGE,
+        0,
+        [_StageOutput(request_id=req_state.request_id)],
+    )
+
+    assert talker_pool.built_for == []
+    assert routed == [None]
@@ -17,6 +17,7 @@ from vllm.sampling_params import SamplingParams
 from vllm_omni.engine.orchestrator import (
    Orchestrator,
    OrchestratorRequestState,
+    StreamingSegmentState,
    _OrchestratorDuplexStagePort,
 )
 from vllm_omni.engine.stage_pool import StagePool
@@ -160,6 +161,26 @@ def _duplex_stage_port_submission():
    return port, stage_pools, request_states, prewarm, submission


+def test_duplex_bridge_state_catches_up_after_serving_turn_boundary() -> None:
+    port, _, request_states, _, submission = _duplex_stage_port_submission()
+    request_state = request_states[submission.context.request_id]
+    bridge_state = request_state.streaming.bridge_states["duplex"]
+    bridge_state["model_turn_id"] = 0
+
+    next_context = DuplexStageRequestContext(
+        request_id=submission.context.request_id,
+        session_id=submission.context.session_id,
+        fence=DuplexFence(submission.context.session_id, turn_id=1),
+        stage_id=submission.context.stage_id,
+        final_stage_id=submission.context.final_stage_id,
+        config_generation=submission.context.config_generation,
+        sampling_params=submission.context.sampling_params,
+    )
+    port.ensure_request(next_context)
+
+    assert bridge_state["model_turn_id"] == 1
+
+
 def _request_output(request_id: str) -> RequestOutput:
    completion = CompletionOutput(
        index=0,
@@ -358,7 +379,7 @@ async def test_streaming_segment_does_not_complete_final_output_stage() -> None:
        final_output_stage_ids={0},
    )
    req_state.streaming.enabled = True
-    req_state.streaming.segment_finished = True
+    req_state.streaming.segments[0] = StreamingSegmentState(finished=True)
    output = SimpleNamespace(
        request_id=req_state.request_id,
        finished=True,
@@ -700,3 +700,62 @@ def test_num_cache_creation_tokens_reaches_direct_request_output():

    assert direct.num_cached_tokens == 3
    assert direct.num_cache_creation_tokens == 2
+
+
+def _abort_processor_with_parent(child_ids: list[str]):
+    processor = object.__new__(MultimodalOutputProcessor)
+    processor.request_states = {}
+    processor.external_req_ids = {}
+    processor.parent_requests = {}
+    processor._native_text_metrics_by_request = {}
+    processor.lora_states = SimpleNamespace(request_finished=lambda *_args, **_kwargs: None)
+    parent = SimpleNamespace(request_id="parent", child_requests=set(child_ids))
+    processor.parent_requests["parent"] = parent
+    for child_id in child_ids:
+        req_state = SimpleNamespace(
+            parent_req=parent,
+            lora_name=None,
+            output_kind=RequestOutputKind.CUMULATIVE,
+            detokenizer=SimpleNamespace(output_token_ids=[7, 8]),
+            queue=None,
+            external_req_id="parent",
+            make_request_output=MagicMock(return_value=SimpleNamespace(request_id=child_id)),
+        )
+        processor.request_states[child_id] = req_state
+    return processor, parent
+
+
+def test_abort_last_parallel_child_drops_parent_request():
+    processor, parent = _abort_processor_with_parent(["0_parent", "1_parent"])
+
+    aborted, _outputs = processor.abort_requests_collecting_outputs(["0_parent"], internal=True)
+    assert aborted == ["0_parent"]
+    assert "parent" in processor.parent_requests
+    assert parent.child_requests == {"1_parent"}
+
+    aborted, _outputs = processor.abort_requests_collecting_outputs(["1_parent"], internal=True)
+    assert aborted == ["1_parent"]
+    assert "parent" not in processor.parent_requests
+    assert not parent.child_requests
+
+
+def test_abort_snapshot_leaves_state_until_commit():
+    processor, parent = _abort_processor_with_parent(["0_parent"])
+    processor.external_req_ids["parent"] = ["0_parent"]
+
+    aborted, outputs = processor.abort_requests_collecting_outputs(
+        ["parent"],
+        internal=False,
+        commit_state=False,
+    )
+    assert aborted == ["0_parent"]
+    assert outputs
+    assert "0_parent" in processor.request_states
+    assert processor.external_req_ids["parent"] == ["0_parent"]
+    assert "parent" in processor.parent_requests
+    assert parent.child_requests == {"0_parent"}
+
+    processor.commit_aborted_request_state(["parent"], internal=False)
+    assert "0_parent" not in processor.request_states
+    assert "parent" not in processor.parent_requests
+    assert "parent" not in processor.external_req_ids
@@ -630,8 +630,24 @@ def test_typed_ming_image_engine_args_defer_diffusion_batch_default():
    assert OmniDiffusionConfig(**typed_backend_args).max_num_seqs == 1


+def _fake_model_root(pipeline, tmp_path):
+    """A model directory carrying every subfolder the pipeline's stages declare.
+
+    Stage init fails closed when a declared ``model_subdir``/``tokenizer_subdir``
+    is not a real directory, because the joined path would otherwise reach
+    HuggingFace as a malformed repo id (issue #6638).
+    """
+    root = tmp_path / "model"
+    root.mkdir(exist_ok=True)
+    for stage in pipeline.stages:
+        for subdir in (stage.model_subdir, stage.tokenizer_subdir):
+            if subdir:
+                (root / subdir).mkdir(parents=True, exist_ok=True)
+    return str(root)
+
+
@pytest.mark.parametrize("model_type", sorted(OMNI_PIPELINES))
-def test_typed_engine_args_match_current_registry_backend_semantics(model_type):
+def test_typed_engine_args_match_current_registry_backend_semantics(model_type, tmp_path):
    pipeline = resolve_pipeline_config(model_type)
    if pipeline is None:
        pytest.skip(f"Pipeline {model_type!r} requires an HF config to resolve")
@@ -641,18 +657,19 @@ def test_typed_engine_args_match_current_registry_backend_semantics(model_type):
        if pipeline.default_deploy_config_name is not None
        else DeployConfig()
    )
+    model = _fake_model_root(pipeline, tmp_path)
    legacy_stages, omni_config = _legacy_and_typed_stages(
        pipeline,
        deploy,
-        model="/tmp",
+        model=model,
    )

    for legacy_stage in legacy_stages:
        stage_id = legacy_stage.stage_id
-        legacy_args = build_legacy_engine_args_dict(legacy_stage, model="/tmp")
+        legacy_args = build_legacy_engine_args_dict(legacy_stage, model=model)
        typed_args = build_engine_args_dict_from_omni_stage_config(
            omni_config.stage_by_id(stage_id),
-            model="/tmp",
+            model=model,
        )
        if legacy_stage.stage_type == StageType.DIFFUSION:
            backend_fields = _DIFFUSION_BACKEND_FIELDS
@@ -64,8 +64,61 @@ def test_collective_rpc_control_helper_honors_timeout():
            return "slept"

        pool, _client = _make_pool(sleep_async=slow_sleep)
-        result = await pool.collective_rpc(0, "sleep", timeout=0.01, args=(1,))
-        assert result != "slept"
+        with pytest.raises(asyncio.TimeoutError):
+            await pool.collective_rpc(0, "sleep", timeout=0.01, args=(1,))
+
+    asyncio.run(run())
+
+
+@pytest.mark.cpu
+def test_collective_rpc_control_method_reraises_worker_error():
+    async def run() -> None:
+        pool, _client = _make_pool(wake_up_async=AsyncMock(side_effect=RuntimeError("worker died")))
+        with pytest.raises(RuntimeError, match="worker died"):
+            await pool.collective_rpc(0, "wake_up")
+
+    asyncio.run(run())
+
+
+@pytest.mark.cpu
+def test_collective_rpc_non_control_method_still_returns_error_dict():
+    async def run() -> None:
+        pool, _client = _make_pool(collective_rpc_async=AsyncMock(side_effect=RuntimeError("probe failed")))
+        result = await pool.collective_rpc(0, "reset_prefix_cache")
        assert result["supported"] is False
+        assert "probe failed" in result["error"]
+
+    asyncio.run(run())
+
+
+@pytest.mark.cpu
+def test_abort_requests_does_not_commit_op_state_when_engine_abort_fails():
+    async def run() -> None:
+        class RecordingOutputProcessor:
+            def __init__(self) -> None:
+                self.collected = False
+                self.committed = False
+
+            def abort_requests_collecting_outputs(self, request_ids, *, internal=False, commit_state=True):
+                del internal
+                self.collected = True
+                assert commit_state is False
+                return list(request_ids), [SimpleNamespace(request_id=request_ids[0])]
+
+            def commit_aborted_request_state(self, request_ids, *, internal=False):
+                del request_ids, internal
+                self.committed = True
+
+        abort = AsyncMock(side_effect=RuntimeError("engine abort failed"))
+        output_processor = RecordingOutputProcessor()
+        client = SimpleNamespace(stage_type="llm", abort_requests_async=abort)
+        pool = StagePool(0, [client], output_processor=output_processor)  # type: ignore[arg-type]
+        pool._request_bindings["req-1"] = 0
+
+        with pytest.raises(RuntimeError, match="engine abort failed"):
+            await pool.abort_requests(["req-1"])
+        assert output_processor.collected is True
+        assert output_processor.committed is False
+        abort.assert_awaited_once()

    asyncio.run(run())
@@ -462,6 +462,15 @@ def test_minicpmo_native_capabilities_separate_model_state_from_core_kv_lease():
    assert caps["stage_handoff_transport"] == "scheduler_data_plane"


+def test_duplex_overlap_policy_auto_falls_back_to_listen_only():
+    from vllm_omni.experimental.fullduplex.openai.protocol import (
+        DuplexOverlapPolicy,
+        DuplexSessionConfig,
+    )
+
+    assert DuplexSessionConfig._normalize_overlap_policy("auto") == DuplexOverlapPolicy.LISTEN_ONLY.value
+
+
 def test_minicpmo_native_capabilities_do_not_overclaim_single_session_deployment():
    caps = DuplexCapabilities.minicpmo45_native(max_sessions=1).as_dict()
Author	SHA1	Message	Date
Yueqian Lin	aef61597d3	Merge remote-tracking branch 'origin/main' into HEAD	2026-08-30 03:36:08 -07:00
akshatvishu	c41925689d	[BugFix][Higgs-Audio-V3] Fix transcript and concurrent sampling tests (#6422 ) Signed-off-by: akshatvishu <akshatnayak197@gmail.com>	2026-08-30 08:52:15 +00:00
Yueqian Lin	a6b559cf54	[Bugfix][MiniMax-Music3] Resolve stage subdirs against a real snapshot (#6640 ) Signed-off-by: Yueqian Lin <linyueqian@outlook.com>	2026-08-30 06:51:24 +00:00
baonudesifeizhai	8be601ed7f	[Diffusion] Add native SymmMem Fast Ulysses transport (#6340 ) Signed-off-by: roG0d <baonudesifeizhai@gmail.com>	2026-08-30 04:14:21 +00:00
y-null	99a2ae1fc8	[Bugfix][MiniCPM-o] Anchor duplex soft-interrupt listen sandwich before commit (#6772 ) Signed-off-by: y_null <y_null@qq.com>	2026-08-30 03:25:43 +00:00
Sy03 Nathan Price Eric	21dd4b33ac	[Bugfix][MiniCPM-o] Stabilize native duplex streaming and Stage-1 handoffs (#6529 ) Signed-off-by: Sy03 <1370724210@qq.com> Co-authored-by: Nathan Price <nathan@abridge.com> Co-authored-by: Eric(Qun) <186020953+0xtoward@users.noreply.github.com>	2026-08-29 13:05:24 -07:00
WeiQing ChenandGao Han	9a88d2d21d	[diffusion][bugfix] Separate Diffusers hooks from model metadata (#6749 ) Signed-off-by: david6666666 <530634352@qq.com> Signed-off-by: Gao Han <hgaoaf@connect.ust.hk> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-29 17:37:14 +00:00
YancyandAsthenia	73893e73ea	[Bugfix] Fix create_error_response import across vLLM versions (#6773 ) Signed-off-by: Asthenia <asthenia0412@gmail.com> Co-authored-by: Asthenia <asthenia0412@gmail.com>	2026-08-29 17:01:51 +00:00
NATURE	87736a3cc5	[Bugfix][MiniCPM-o] Bound full-duplex Talker context with sliding recompute (#6626 ) Signed-off-by: natureofnature <wzliu@connect.hku.hk>	2026-08-30 00:49:42 +08:00
Sy03	3adee400a0	[Bugfix][Qwen3-TTS] Handle codec generations that exhaust their token budget (#6728 ) Signed-off-by: Sy03 <1370724210@qq.com>	2026-08-29 08:59:05 -07:00
Gao Han	676a35fa8b	Fix MiniCPM-o concurrent audio first-packet latency (#6767 ) Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-29 23:27:47 +08:00
Honghan Zhuandamy-why-3459	72a1ce488b	[Bugfix] Avoid invalid TPOT metrics (#6696 ) Signed-off-by: zhuhh97 <zhuhonghan@huawei.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-29 11:56:37 +00:00
wangyu	e338f0ece3	[CI]Split GPU jobs by cards_* and reject hand-written SKU marks (#6650 ) Signed-off-by: wangyu <410167048@qq.com> Signed-off-by: [Your Name] <your.email@example.com>	2026-08-29 18:42:33 +08:00
HAAZZZEEEE	c50b7de167	[NPU][Diffusion] MiniMax-H3 RainFusion end_step tail fallback (#6037 ) Signed-off-by: HAAZZZEEEE <wang-taicheng@qq.com>	2026-08-29 18:36:05 +08:00
psv666	a5b8c82788	[Bugfix][MiniCPM-o] Fix duplex camera frame fixture (#6757 ) Signed-off-by: psv666 <2693925048@qq.com>	2026-08-29 09:57:53 +00:00
ooooooye	5a5bc55c25	[Feature][Diffusion] Allow online INT8 quantization with DLO AllGather (#6573 ) Signed-off-by: brandneway <gyuan4892@gmail.com>	2026-08-29 09:50:30 +00:00
Wenjia LiandWeiQing Chen	d9980c75e8	Perf/minimax h3 dit swiglu rope cache npu (#6410 ) Signed-off-by: Wenjia Li <wjialish@gmail.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>	2026-08-29 09:14:52 +00:00
Nguyen Kha Nhat Long Cursor Samit	bb5bf8fd0a	[Bugfix] Fix colocate-async sleep admission race and deliver AR abort tokens (#6367 ) Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com> Signed-off-by: Cursor <cursoragent@cursor.com> Signed-off-by: Samit <285365963@qq.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Samit <285365963@qq.com>	2026-08-29 09:11:39 +00:00
psv666andGao Han	5c47fd0bc8	[Bugfix][Qwen3-Omni] Handle missing packed modules mapping (#6748 ) Signed-off-by: psv666 <2693925048@qq.com> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-29 15:24:49 +08:00
Zijun HuangandWeiQing Chen	08cdee8dc4	[Feature] Support MiniMax-H3 FlashGen native LoRA with the legacy manager (#6666 ) Signed-off-by: Zijun Huang <zijunhua@usc.edu> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>	2026-08-29 06:52:19 +00:00
abrahamzewoudie Hongsheng Liu amy-why-3459	143ccc5a2f	Fix Qwen3-Omni audio encoder TP when heads are not divisible by TP size (#4322 ) Signed-off-by: abrahamzewoudie <azewoudi@amd.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-29 14:52:02 +08:00
WeiQing Chen	55a226dcf1	[video][bugfix] Accept semantic video output in offline examples (#6747 ) Signed-off-by: david6666666 <530634352@qq.com>	2026-08-29 06:26:28 +00:00
Jiatai Wang	f66f05633c	[Bugfix] Honor request seed in MOSS-TTS adapters (#6543 ) Signed-off-by: JiataiWang <wangjiatai@proton.me>	2026-08-28 22:44:13 -07:00
ZacheryAUandamy-why-3459	457c6e2793	[bugfix][MiniCPM-o] Re-land empty full-payload and async_chunk=false processor tests (#6745 ) Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-29 13:39:54 +08:00
boatman	bfe8967ca2	[Bugfix][StepAudio2] Fix async-chunk metadata mismatch and boundary glitches (#5917 ) Signed-off-by: boatman <109857087+sphinxkkkbc@users.noreply.github.com>	2026-08-28 21:45:24 -07:00
WeiQing ChenandOpenAI Codex	fbc93a0e08	[ci][bugfix] Fix MiniMax-H3 FP8 quality test OOM (#6742 ) Signed-off-by: david6666666 <530634352@qq.com> Co-authored-by: OpenAI Codex <codex@openai.com>	2026-08-29 12:11:31 +08:00
Bob Tian	d934a5f7b8	[Bugfix] Report engine-queued requests as waiting in vllm_omni:num_requests gauges (#6549 ) Signed-off-by: Bob Tian <bobzetian@google.com>	2026-08-29 03:58:26 +00:00
Inesh Reddy Chappidiandamy-why-3459	bbb3ae01a1	[CI] Gate qwen3-omni no-async-chunk perf on mean_audio_rtf (#6743 ) Signed-off-by: Inesh Reddy <ineshreddy249@gmail.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-28 20:28:58 -07:00
Zhang BowenandHongsheng Liu	46cfa8435b	[Model][Frontend] MiniMax-H3: Parallelize MP4 response frame conversion (#6499 ) Signed-off-by: MosCloud <bwzhang1991@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-28 21:22:00 -06:00
fan2956andWeiQing Chen	0dc8f45906	[Feat] Minimax H3 support ref and tgt video sparse attention on NPU (#6518 ) Signed-off-by: fan2956 <zhoufan53@huawei.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>	2026-08-29 11:10:53 +08:00
Allen Wu TRAE CLI TRAE CLI	d94061201d	[Bugfix][Engine] Make abort idempotent during shutdown (#6327 ) Signed-off-by: Allen Wu <allenwu2795@gmail.com> Co-authored-by: TRAE CLI <noreply@bytedance.com> Co-authored-by: TRAE CLI <traecli@bytedance.com>	2026-08-28 19:51:25 -07:00
Mu GuanLinandWeiQing Chen	9558fd2ea8	[Bugfix] Scope MiniMax-H3 cuDNN SDPA state (#6710 ) Signed-off-by: mglyn <1203789601@qq.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>	2026-08-28 19:40:50 -07:00
Mu GuanLin	55a8bd6b80	[CI][Bugfix] Align the LTX-2 FP8 quality gate with the default recipe (#5831 ) Signed-off-by: mglyn <1203789601@qq.com>	2026-08-29 02:24:09 +00:00
Honghan Zhu	f584729f24	[BugFix] Fix local media access for diffusion speech (#6622 ) Signed-off-by: zhuhh97 <zhuhonghan@huawei.com>	2026-08-29 02:01:24 +00:00
Chendi.Xue	034af1f657	[CI][XPU]switch to use vLLM base docker for XPU (#6727 ) Signed-off-by: Chendi Xue <chendi.xue@intel.com>	2026-08-29 09:03:00 +08:00
Hongsheng Liuandhsliuustc0106	8bfad23f98	[Core] Restore topology-aware diffusion worker titles (#6722 ) Signed-off-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com> Co-authored-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com>	2026-08-29 08:45:57 +08:00
Gao Han	6b9ce05feb	Revert "[bugfix][MiniCPM-o] Fix offline_inference/test_minicpmo_4_5.py and online_serving/ one" (#6730 )	2026-08-29 08:14:02 +08:00
ZacheryAUandamy-why-3459	6d84409a68	[bugfix][MiniCPM-o] Fix offline_inference/test_minicpmo_4_5.py and online_serving/ one (#5464 ) Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-28 19:20:09 +00:00
Khairul Kabir khairulkabir1661 Claude Opus 4.6	ebb3b20576	[Diffusion] Make Flux2Pipeline text_encoder_out_layers configurable (#6390 ) Co-authored-by: khairulkabir1661 <khairulkabir1661@users.noreply.github.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>	2026-08-28 09:05:41 -06:00
Wallbreazzz Canlin Guo Canlin Guo	be335a86f9	fix(npu): extend bf16 autocast to NPU in MossAudioTokenizer decode (#6664 ) Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com> Signed-off-by: Canlin Guo <961750412@qq.com> Co-authored-by: Canlin Guo <961750412@qq.com> Co-authored-by: Canlin Guo <canlinguosdu@gmail.com>	2026-08-28 04:01:49 -07:00
Yueqian Lin	1ad90653c5	[CI/Build] Add dots.tts weekly e2e coverage dots.tts landed in #4765 with no GPU coverage: nothing under tests/e2e exercises it, so the DiT sampler, streaming vocoder and AR loopback have no CI signal at all. Add an offline e2e module marked slow + tts + L4, which the existing weekly 'TTS · L4' step collects via -m 'slow and L4 and tts'. No pipeline YAML change is needed; the weekly TTS steps are marker driven. Three cases, the latter two guarding specific review findings from #4765: 1. zero-shot synthesis produces finite, non-silent audio 2. mixed prefill+decode batch keeps per-request logits rows aligned 3. the same text twice yields the same audio (seeded FM noise) Verified selector behaviour: 3 collected under the weekly selector, 0 collected under the nightly one. Signed-off-by: Yueqian Lin <linyueqian@outlook.com>	2026-08-13 11:24:50 -07:00