rebase: port harness-tier support, wheel hardening and bridge host env to main

adapters/vllm_omni/manifest.yaml

@@ -25,11 +25,8 @@ upstream:
   kind: fork_tracking
   repo_path: ${VLLM_UPSTREAM_REPO}
   remote: origin
-  # v0.30.0 alignment campaign (2026-09-17): wheel pick and commit
-  # assignment target the release branch, not upstream main. The branch
-  # is cut but NOT yet tagged, so this tracks the branch head and is
-  # re-pinned to the tag commit once v0.30.0 is released.
-  target_branch: releases/v0.30.0
+  # v0.32.0 alignment tracks upstream main until the release branch is cut.
+  target_branch: main
 
 # Repo-specific slice of the SHARED, human-curated knowledge base (the vendored
 # community docs at the repo-root `knowledge/` = settings.knowledge_dir;
@@ -67,10 +64,10 @@ knowledge:
 # scoping) may only ever gain coverage here, never lose it.
 modules:
   worker_runner:
-    local_paths: [vllm_omni/worker/]
+    local_paths: [vllm_omni/worker/, vllm_omni/worker_v2/]
     upstream_paths: [vllm/v1/worker/]
-    test_paths: [tests/worker/, tests/diffusion/test_diffusion_worker.py, tests/e2e/offline_inference/custom_pipeline/, tests/e2e/online_serving/test_qwen3_tts_base.py, tests/entrypoints/test_omni_sleep_mode.py]
-    import_check: "from vllm_omni.worker.gpu_model_runner import *; from vllm_omni.worker.gpu_generation_model_runner import *; from vllm_omni.worker.gpu_ar_model_runner import *; from vllm_omni.worker.gpu_ar_worker import *; print('Worker/Runner imports OK')"
+    test_paths: [tests/worker/, tests/worker_v2/, tests/diffusion/test_diffusion_worker.py, tests/e2e/features/custom_pipeline/, tests/e2e/online_serving/test_qwen3_tts_base.py, tests/entrypoints/test_omni_sleep_mode.py]
+    import_check: "from vllm_omni.worker.gpu_model_runner import *; from vllm_omni.worker.gpu_generation_model_runner import *; from vllm_omni.worker.gpu_ar_model_runner import *; from vllm_omni.worker.gpu_ar_worker import *; from vllm_omni.worker_v2.omni_ar_model_runner import *; from vllm_omni.worker_v2.omni_generation_model_runner import *; print('Worker/Runner imports OK')"
     wave: 1
     risk: high
   model_executor:
@@ -82,14 +79,14 @@ modules:
     risk: high
   input_output:
     local_paths: [vllm_omni/inputs/, vllm_omni/outputs/, vllm_omni/utils/mm_outputs.py, vllm_omni/data_entry_keys.py, vllm_omni/engine/async_omni_engine.py, vllm_omni/engine/orchestrator.py, vllm_omni/engine/serialization.py, vllm_omni/engine/stage_init_utils.py, vllm_omni/engine/stage_engine_core_client.py, vllm_omni/engine/output_processor.py, vllm_omni/request.py]
-    upstream_paths: [vllm/inputs/, vllm/v1/engine/input_processor.py, vllm/v1/engine/output_processor.py, vllm/v1/request.py]
-    test_paths: [tests/entrypoints/test_omni_new_request_data.py, tests/engine/test_async_omni_engine_input.py, tests/test_outputs.py]
+    upstream_paths: [vllm/inputs/, vllm/renderers/, vllm/v1/engine/input_processor.py, vllm/v1/engine/output_processor.py, vllm/v1/request.py]
+    test_paths: [tests/inputs/, tests/entrypoints/test_omni_new_request_data.py, tests/engine/test_async_omni_engine_input.py, tests/engine/test_omni_request_output.py]
     import_check: "from vllm_omni.inputs.data import *; from vllm_omni.inputs.preprocess import *; from vllm_omni.engine.async_omni_engine import AsyncOmniEngine; from vllm_omni.engine.orchestrator import Orchestrator; from vllm_omni.engine.serialization import serialize_additional_information; print('Input/Output imports OK')"
     wave: 1
   scheduler:
     local_paths: [vllm_omni/core/]
     upstream_paths: [vllm/v1/core/sched/]
-    test_paths: [tests/distributed/, tests/diffusion/distributed/test_ulysses_uaa_perf.py]
+    test_paths: [tests/core/sched/, tests/distributed/, tests/model_executor/test_cfg_pairing.py, tests/diffusion/distributed/test_ulysses_uaa_perf.py]
     import_check: "from vllm_omni.core.sched.omni_ar_scheduler import *; from vllm_omni.core.sched.omni_generation_scheduler import *; print('Scheduler imports OK')"
     wave: 1
     risk: high
@@ -213,18 +210,10 @@ rebase:
   # adapter is active; this whole `rebase:` section is HIGH-RISK (human-only
   # writes) so an agent can never self-install hooks
   hooks: rebase/hooks.py
-  # READ-ONLY parent-store read-compat (Rev 8 §9 PR4c, delivered by the
-  # completion design D2): v3 knowledge retrieval unions these layers LAST
-  # (runtime > adapter seed > parent) until the PR4d migration executes,
-  # after which the owner REMOVES these two keys (RUNBOOK step) and the
-  # compat layer dies. The v3 prelude fails CLOSED when a declared layer
-  # is unreadable and records open/close logical digests (§8 fairness).
-  knowledge:
-    parent_debug_db: ${REBASE_AGENT_ROOT}/agent/store/debug_memory.db
-    parent_skills_dir: ${REBASE_AGENT_ROOT}/agent/skills
-    # the parent store's own column carrying the upstream commit — repo
-    # data, so the neutral read-compat/migration code stays literal-free
-    parent_upstream_column: vllm_commit
+  # The READ-ONLY parent-store read-compat `knowledge:` block (Rev 8 §9
+  # PR4c) is removed (owner decision 2026-10-01): the parent rebase-agent
+  # checkout no longer exists, so v3 runs on its own runtime + adapter seed
+  # knowledge only. An absent declaration attests nothing and blocks nothing.
   # the shared checkout flock name: v3, the v1 backend, EXT1's external
   # startup guard, and PR7 archival ALL take locks/<lock_name>.lock inside
   # the checkout, so exclusion is mutual (repo data — the core stays neutral)
@@ -244,7 +233,10 @@ rebase:
     import_check_modules: [vllm._C_stable_libtorch, vllm._C]
     ops_namespace: _C
     stale_artifact_globs: ["_C*.so", "*.abi3.so", "_moe_C*.so"]
-    install_env: {VLLM_USE_PRECOMPILED: "1"}
+    install_env:
+      VLLM_USE_PRECOMPILED: "1"
+      VLLM_PRECOMPILED_WHEEL_COMMIT: "{commit}"
+      VLLM_PRECOMPILED_WHEEL_VARIANT: "{variant}"
     reinstall_retries: 3
     pin:
       dockerfile: docker/Dockerfile.ci
@@ -260,7 +252,7 @@ rebase:
     # trigger env (parent create_build env + trigger_and_monitor extra_env,
     # verbatim): opts the FULL suite in — a webhook build without these
     # runs fewer tests, which is why pushed rounds create their own build
-    build_env: {NIGHTLY: "1", RUN_HUNYUAN_IMAGE3_PERF: "1"}
+    build_env: {NIGHTLY: "1", RUN_HUNYUAN_IMAGE3_PERF: "1", EMAIL_DISTRIBUTION: "0"}
     # monitor classification data (parent monitor.py constants — EXCEPT
     # the pipeline upload/load lanes, moved to structural: their failure
     # suppresses every dynamic test job, so ignoring it was a false-green
@@ -310,9 +302,9 @@ rebase:
     # behavioral routing replay).
     assignment_paths:
       model_config: [tests/entrypoints/test_stage_utils.py, tests/metrics/test_stats.py]
-      input_output: [tests/entrypoints/test_omni_new_request_data.py, tests/engine/, tests/test_outputs.py]
+      input_output: [tests/entrypoints/test_omni_new_request_data.py, tests/engine/, tests/engine/test_omni_request_output.py]
       scheduler: [tests/distributed/]
-      worker_runner: [tests/worker/, tests/diffusion/test_diffusion_worker.py, tests/e2e/offline_inference/custom_pipeline/]
+      worker_runner: [tests/worker/, tests/worker_v2/, tests/diffusion/test_diffusion_worker.py, tests/e2e/features/custom_pipeline/]
       model_executor: [tests/model_executor/, tests/diffusion/, tests/e2e/offline_inference/]
       online_serving: [tests/entrypoints/, tests/e2e/online_serving/]
       benchmarks: [tests/benchmarks/test_serve_cli.py]
@@ -332,6 +324,13 @@ rebase:
     # full offline inference suite takes about four hours on the shared host,
     # so the previous 7200s default could never produce a terminal result.
     module_timeout_sec: 21600
+    # The wrapper refuses a TEST_TIMEOUT_SEC above this cap; it must not be
+    # lower than the planned budget above or every required command is refused.
+    module_max_timeout_sec: 21600
+    # Phase-2 module agents run on the CUDA wheel/runtime above. Broad pytest
+    # paths must retain CUDA-capable and platform-neutral items while leaving
+    # NPU/ROCm/XPU/MUSA-only items to their native CI lanes.
+    module_test_platform: cuda
     # root-level artifacts tests leak between runs (test_artifact_cleanup.sh
     # TEST_ARTIFACT_GLOBS, verbatim) — gitignored upstream but pile up on disk
     artifact_globs: ["audio_chatcmpl-*.wav", "audio_[0-9]*ch_*.wav",
@@ -372,7 +371,7 @@ rebase:
           candidates: [tests/entrypoints/test_stage_utils.py, tests/metrics/test_stats.py, tests/entrypoints/test_omni_llm.py]
           fallback: tests/entrypoints/test_stage_utils.py
         input_output:
-          candidates: [tests/entrypoints/test_omni_new_request_data.py, tests/engine/test_async_omni_engine_input.py, tests/entrypoints/test_omni_input_preprocessor.py, tests/test_outputs.py]
+          candidates: [tests/entrypoints/test_omni_new_request_data.py, tests/engine/test_async_omni_engine_input.py, tests/entrypoints/test_omni_input_preprocessor.py, tests/engine/test_omni_request_output.py]
           fallback: tests/entrypoints/test_omni_new_request_data.py
         online_serving:
           candidates: [tests/entrypoints/test_async_omni.py, tests/entrypoints/test_async_omni_diffusion_config.py, tests/entrypoints/test_omni_diffusion.py, tests/entrypoints/test_omni_stage_diffusion_config.py, tests/entrypoints/openai_api/]

adapters/vllm_omni/rebase/api_drift_guard.py

@@ -2,10 +2,8 @@
 # api_drift_guard.py — Inheritance/constructor/unpack drift guard for vllm-omni.
 # Ported from the rebase agent (agent/lib/api_drift_guard.py); runs
 # standalone with the TARGET repo python (cwd or OMNI_PATH = vllm-omni root).
-# Verbatim except two documented divergences: the pooling constructor entry
-# (parent's spelling could never import) and constructor-check import
-# failures counting as mismatches instead of silent SKIPs (see inline
-# comments at both sites).
+# Documented divergences cover the pooling constructor/import checks and
+# current renderer/EngineClient contracts (see doc/DRIFT_TRIAGE.md).
 # Repo-specific maps (INHERITANCE_MAP etc.) are the point of this file living
 # in the adapter tree; the copilot core never imports it.
 # Run with cwd = vllm-omni repo root (same as tasks/41_check_api_drift.sh).
@@ -19,13 +17,10 @@
 from pathlib import Path
 
 # ---------------------------------------------------------------------------
-# Block flash attention C extensions from loading to prevent SIGBUS (signal 7)
-# when the .so was compiled for a different CUDA / GPU architecture.
-#
-# The guard only inspects method signatures — it does not need the real
-# flash attention kernels.  By raising ImportError (instead of letting the
-# process crash with SIGBUS), the existing try/except in the INHERITANCE_MAP
-# loop can skip affected classes gracefully.
+# Use real extensions by default: current vLLM refuses to import its flash
+# attention package when both extensions are blocked, which skips worker and
+# serving signatures. For a known incompatible binary, an operator may opt
+# into the blocker with IMX_API_DRIFT_BLOCK_FLASH_ATTN=1. That run is partial.
 # ---------------------------------------------------------------------------
 class _BlockFlashAttnExtensions:
     @staticmethod
@@ -40,14 +35,20 @@ def find_spec(fullname, path, target=None):
         return None
 
 
-sys.meta_path.insert(0, _BlockFlashAttnExtensions)
+_block_flash_attn = os.environ.get("IMX_API_DRIFT_BLOCK_FLASH_ATTN", "0")
+if _block_flash_attn not in ("0", "1"):
+    raise ValueError("IMX_API_DRIFT_BLOCK_FLASH_ATTN must be 0 or 1")
+if _block_flash_attn == "1":
+    sys.meta_path.insert(0, _BlockFlashAttnExtensions)
+    print("PARTIAL  Flash attention extensions blocked by operator request; "
+          "affected import checks may be skipped.")
 
 # Ensure relative paths (vllm_omni/...) resolve when invoked from any cwd.
 if "OMNI_PATH" in os.environ:
     os.chdir(os.environ["OMNI_PATH"])
 
 INHERITANCE_MAP = [
-    ("vllm.inputs.preprocess", "InputPreprocessor", "vllm_omni.inputs.preprocess", "OmniInputPreprocessor"),
+    ("vllm.renderers", "BaseRenderer", "vllm_omni.inputs.preprocess", "omni_renderer_cls"),
     ("vllm.v1.engine.core_client", "AsyncMPClient", "vllm_omni.engine.stage_engine_core_client", "StageEngineCoreClient"),
     ("vllm.engine.protocol", "EngineClient", "vllm_omni.entrypoints.async_omni", "AsyncOmni"),
     ("vllm.v1.worker.gpu_model_runner", "GPUModelRunner", "vllm_omni.worker.gpu_model_runner", "OmniGPUModelRunner"),
@@ -219,12 +220,17 @@ def _forbidden_test_imports() -> list[str]:
     return findings
 
 
-def main() -> int:
+def _check_inheritance() -> list[str]:
     mismatches: list[str] = []
     for up_mod, up_cls_name, omni_mod, omni_cls_name in INHERITANCE_MAP:
         try:
             UpCls = import_class(up_mod, up_cls_name)
             OmniCls = import_class(omni_mod, omni_cls_name)
+            # OmniRenderer is a mixin; the actual upstream subclass is
+            # constructed by this factory rather than declared statically.
+            if omni_cls_name == "omni_renderer_cls":
+                OmniCls = OmniCls(UpCls)
+                omni_cls_name = OmniCls.__name__
         except Exception as e:
             print(f"SKIP  {omni_cls_name}: import error — {e}")
             continue
@@ -233,6 +239,14 @@ def main() -> int:
             print(f"SKIP  {omni_cls_name}: not a subclass of {up_cls_name}")
             continue
 
+        if up_cls_name == "EngineClient" and inspect.isabstract(OmniCls):
+            msg = (
+                f"MISMATCH {omni_cls_name}: missing EngineClient implementations "
+                f"{sorted(OmniCls.__abstractmethods__)}; the engine cannot be instantiated"
+            )
+            print(msg)
+            mismatches.append(msg)
+
         for name in dir(OmniCls):
             if name.startswith("__") and name != "__init__":
                 continue
@@ -273,6 +287,11 @@ def main() -> int:
             else:
                 print(f"  OK   {omni_cls_name}.{name}")
 
+    return mismatches
+
+
+def main() -> int:
+    mismatches = _check_inheritance()
     constructor_call_checks = [
         (
             "vllm.entrypoints.openai.responses.serving",

adapters/vllm_omni/rebase/prompt_data.yaml

@@ -32,14 +32,28 @@ debug_prompt_template_live: |
      rebase regression. Separate assertion failures, latency-only failures,
      cancellation, and broken reporting steps. Record unknown attribution
      when evidence is insufficient; do not weaken thresholds or skip tests.
+     A local gated-model download denial is an access failure. Preserve the
+     original model source, tokenizer, configuration, geometry, selected
+     tests and timeout. Do not replace them with a synthetic local builder
+     or a different public tokenizer to make CI pass. If authorized access
+     is unavailable, finish with the exact unresolved failure so the parent
+     can continue the other jobs. A prototype is not the original job's proof.
+     Read run-directory operator evidence when present, including
+     codex_krea_scope_rejection.json and the remote main baseline. Do not
+     claim CI has no model access when its exact job has actually passed.
   3. Read the target vLLM API and `git show origin/main:<path>` for Omni intent.
      Fix the smallest confirmed rebase-specific cause. After a startup API
      fix, inspect adjacent constructor, override, result-field, dummy-run,
      and async-output contracts that the first crash could have masked.
   4. Reproduce with a targeted test using the real upstream API when possible,
      then run the owning suite. Report the exact validated commit and whether
      a new remote CI build has reached a terminal result.
-  5. Write MODULE_DONE {module} to the signal file only after verification.
+  5. Use the tool environment's IMX_LOG_DIR for this run's artifact directory;
+     form the absolute signal path as IMX_LOG_DIR/module.{module}.done.
+     Search only this run directory for signals, never / or /data. Write
+     MODULE_DONE {module} there only after verification. If the original
+     test remains failed, report the exact unresolved result and finish
+     without a success signal so the parent can continue the other jobs.
 
   MODULE={module}
 
@@ -51,6 +65,20 @@ pytest_wrapper: "bash '{script_dir}/lib/run_module_pytest.sh'"
 # imx-omni-pytest instead of the retired shell script, and fix its quoting)
 pytest_wrapper_live: "imx-omni-pytest"
 
+# Operational coverage may grow without changing the historical prompt maps.
+# worker_v2 imports upstream worker/output APIs and must be ported with worker.
+live_module_overrides:
+  scheduler:
+    omni_files: "vllm_omni/core/sched/"
+    test_paths: ["tests/core/sched/", "tests/distributed/", "tests/model_executor/test_cfg_pairing.py"]
+  input_output:
+    vllm_paths: "vllm/inputs/ vllm/renderers/ vllm/v1/engine/input_processor.py vllm/v1/engine/output_processor.py vllm/v1/request.py"
+    test_paths: ["tests/inputs/", "tests/entrypoints/test_omni_new_request_data.py", "tests/engine/", "tests/engine/test_omni_request_output.py"]
+  worker_runner:
+    omni_files: "vllm_omni/worker/ vllm_omni/worker_v2/"
+    test_paths: ["tests/worker/", "tests/worker_v2/", "tests/diffusion/test_diffusion_worker.py", "tests/e2e/features/custom_pipeline/"]
+    import_check: "from vllm_omni.worker.gpu_model_runner import *; from vllm_omni.worker.gpu_generation_model_runner import *; from vllm_omni.worker.gpu_ar_model_runner import *; from vllm_omni.worker.gpu_ar_worker import *; from vllm_omni.worker_v2.omni_ar_model_runner import *; from vllm_omni.worker_v2.omni_generation_model_runner import *; print('OK')"
+
 module_vllm_paths:
   model_config: "vllm/config/ vllm/engine/arg_utils.py"
   input_output: "vllm/inputs/ vllm/v1/engine/input_processor.py vllm/v1/engine/output_processor.py vllm/v1/request.py"

adapters/vllm_omni/rebase/shell_golden.json

@@ -601,7 +601,7 @@
    "Cannot re-initialize CUDA in forked subprocess",
    "cuDNN error",
    "CUBLAS_STATUS",
-   "(?<!non-)FATAL"
+   "(?<![\\w-])FATAL(?!\\w)"
   ],
   "review": [
    "Traceback \\(most recent call last\\)",