chore: import upstream snapshot with attribution

doc/source/llm/doc_code/serve/direct_streaming/direct_streaming_config.yaml

@@ -0,0 +1,20 @@
+# config.yaml
+applications:
+  - name: llm-direct-streaming
+    route_prefix: /
+    import_path: ray.serve.llm:build_openai_app
+    args:
+      llm_configs:
+        - model_loading_config:
+            model_id: qwen3.5-0.8b
+            model_source: Qwen/Qwen3.5-0.8B
+          deployment_config:
+            autoscaling_config:
+              min_replicas: 1
+              max_replicas: 4
+          engine_kwargs:
+            trust_remote_code: true
+            tensor_parallel_size: 1
+            enable_auto_tool_choice: true
+            tool_call_parser: qwen3_coder
+            reasoning_parser: qwen3

doc/source/llm/doc_code/serve/direct_streaming/direct_streaming_custom_router_example.py

@@ -0,0 +1,88 @@
+"""
+This file serves as a documentation example and CI test.
+
+Structure:
+1. Monkeypatch setup: Ensures serve.run is non-blocking and removes accelerator requirements for CI testing.
+2. Docs example (between __direct_streaming_custom_router_example_start/end__): Embedded in Sphinx docs via literalinclude.
+3. Test validation (deployment status polling + cleanup)
+
+Scope: validates the documented request_router_config snippet (config and the
+build_openai_app call signature) for direct streaming.
+"""
+
+import time
+from ray import serve
+from ray.serve.schema import ApplicationStatus
+from ray.serve._private.constants import SERVE_DEFAULT_APP_NAME
+from ray.serve import llm
+
+_original_serve_run = serve.run
+_original_build_openai_app = llm.build_openai_app
+
+
+def _non_blocking_serve_run(app, **kwargs):
+    """Forces blocking=False for testing"""
+    kwargs["blocking"] = False
+    return _original_serve_run(app, **kwargs)
+
+
+def _testing_build_openai_app(llm_serving_args):
+    """Removes accelerator requirements for testing"""
+    for config in llm_serving_args["llm_configs"]:
+        config.accelerator_type = None
+
+    return _original_build_openai_app(llm_serving_args)
+
+
+serve.run = _non_blocking_serve_run
+llm.build_openai_app = _testing_build_openai_app
+
+# __direct_streaming_custom_router_example_start__
+from ray.serve.config import RequestRouterConfig
+from ray.serve.llm import LLMConfig
+
+llm_config = LLMConfig(
+    model_loading_config={
+        "model_id": "qwen3.5-0.8b",
+        "model_source": "Qwen/Qwen3.5-0.8B",
+    },
+    deployment_config={
+        "request_router_config": RequestRouterConfig(
+            request_router_class="ray.serve.experimental.consistent_hash_router.ConsistentHashRouter",
+        ),
+    },
+    engine_kwargs={
+        "trust_remote_code": True,
+        "tensor_parallel_size": 1,
+        "enable_auto_tool_choice": True,
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
+    },
+)
+# __direct_streaming_custom_router_example_end__
+
+from ray.serve.llm import build_openai_app
+
+app = build_openai_app({"llm_configs": [llm_config]})
+serve.run(app)
+
+status = ApplicationStatus.NOT_STARTED
+timeout_seconds = 180
+start_time = time.time()
+
+while (
+    status != ApplicationStatus.RUNNING and time.time() - start_time < timeout_seconds
+):
+    status = serve.status().applications[SERVE_DEFAULT_APP_NAME].status
+
+    if status in [ApplicationStatus.DEPLOY_FAILED, ApplicationStatus.UNHEALTHY]:
+        raise AssertionError(f"Deployment failed with status: {status}")
+
+    time.sleep(1)
+
+if status != ApplicationStatus.RUNNING:
+    raise AssertionError(
+        f"Deployment failed to reach RUNNING status within {timeout_seconds}s. Current status: {status}"
+    )
+
+serve.shutdown()

doc/source/llm/doc_code/serve/direct_streaming/direct_streaming_example.py

@@ -0,0 +1,88 @@
+"""
+This file serves as a documentation example and CI test.
+
+Structure:
+1. Monkeypatch setup: Ensures serve.run is non-blocking and removes accelerator requirements for CI testing.
+2. Docs example (between __direct_streaming_example_start/end__): Embedded in Sphinx docs via literalinclude.
+3. Test validation (deployment status polling + cleanup)
+
+Scope: builds and deploys the documented snippet with direct streaming enabled.
+The bazel target sets ``RAY_SERVE_ENABLE_HA_PROXY`` and
+``RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING``, so build_openai_app returns the
+direct streaming deployment and the app reaches RUNNING on the HAProxy ingress.
+The end-to-end streaming data path is covered by the openai-compatibility
+direct-streaming test suite.
+"""
+
+import time
+from ray import serve
+from ray.serve.schema import ApplicationStatus
+from ray.serve._private.constants import SERVE_DEFAULT_APP_NAME
+from ray.serve import llm
+
+_original_serve_run = serve.run
+_original_build_openai_app = llm.build_openai_app
+
+
+def _non_blocking_serve_run(app, **kwargs):
+    """Forces blocking=False for testing"""
+    kwargs["blocking"] = False
+    return _original_serve_run(app, **kwargs)
+
+
+def _testing_build_openai_app(llm_serving_args):
+    """Removes accelerator requirements for testing"""
+    for config in llm_serving_args["llm_configs"]:
+        config.accelerator_type = None
+
+    return _original_build_openai_app(llm_serving_args)
+
+
+serve.run = _non_blocking_serve_run
+llm.build_openai_app = _testing_build_openai_app
+
+# __direct_streaming_example_start__
+from ray import serve
+from ray.serve.llm import LLMConfig, build_openai_app
+
+llm_config = LLMConfig(
+    model_loading_config={
+        "model_id": "qwen3.5-0.8b",
+        "model_source": "Qwen/Qwen3.5-0.8B",
+    },
+    deployment_config={
+        "autoscaling_config": {"min_replicas": 1, "max_replicas": 4},
+    },
+    engine_kwargs={
+        "trust_remote_code": True,
+        "tensor_parallel_size": 1,
+        "enable_auto_tool_choice": True,
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
+    },
+)
+
+app = build_openai_app({"llm_configs": [llm_config]})
+serve.run(app)
+# __direct_streaming_example_end__
+
+status = ApplicationStatus.NOT_STARTED
+timeout_seconds = 180
+start_time = time.time()
+
+while (
+    status != ApplicationStatus.RUNNING and time.time() - start_time < timeout_seconds
+):
+    status = serve.status().applications[SERVE_DEFAULT_APP_NAME].status
+
+    if status in [ApplicationStatus.DEPLOY_FAILED, ApplicationStatus.UNHEALTHY]:
+        raise AssertionError(f"Deployment failed with status: {status}")
+
+    time.sleep(1)
+
+if status != ApplicationStatus.RUNNING:
+    raise AssertionError(
+        f"Deployment failed to reach RUNNING status within {timeout_seconds}s. Current status: {status}"
+    )
+
+serve.shutdown()

doc/source/llm/doc_code/serve/direct_streaming/direct_streaming_yaml_config_example.py

@@ -0,0 +1,62 @@
+"""
+This file serves as a documentation example and CI test for the direct
+streaming YAML config.
+
+Structure:
+1. Load the YAML config shown in the guide and convert it to an app with
+   build_openai_app, removing accelerator requirements for CI testing.
+2. Test validation (deployment status polling + cleanup).
+
+Scope: validates that the documented config.yaml parses into a valid
+LLMServingArgs and deploys with direct streaming enabled. The bazel target sets
+RAY_SERVE_ENABLE_HA_PROXY and RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING, so the app
+reaches RUNNING on the HAProxy ingress. The end-to-end streaming data path is
+covered by the openai-compatibility direct-streaming test suite.
+"""
+
+import time
+import os
+import yaml
+from ray import serve
+from ray.serve.schema import ApplicationStatus
+from ray.serve._private.constants import SERVE_DEFAULT_APP_NAME
+from ray.serve import llm
+
+
+config_path = os.path.join(os.path.dirname(__file__), "direct_streaming_config.yaml")
+with open(config_path, "r") as f:
+    config_dict = yaml.safe_load(f)
+
+llm_configs = config_dict["applications"][0]["args"]["llm_configs"]
+for config in llm_configs:
+    config.pop("accelerator_type", None)
+    # Disable compile cache to avoid cache corruption in CI
+    if "runtime_env" not in config:
+        config["runtime_env"] = {}
+    if "env_vars" not in config["runtime_env"]:
+        config["runtime_env"]["env_vars"] = {}
+    config["runtime_env"]["env_vars"]["VLLM_DISABLE_COMPILE_CACHE"] = "1"
+
+app = llm.build_openai_app({"llm_configs": llm_configs})
+serve.run(app, blocking=False)
+
+status = ApplicationStatus.NOT_STARTED
+timeout_seconds = 180
+start_time = time.time()
+
+while (
+    status != ApplicationStatus.RUNNING and time.time() - start_time < timeout_seconds
+):
+    status = serve.status().applications[SERVE_DEFAULT_APP_NAME].status
+
+    if status in [ApplicationStatus.DEPLOY_FAILED, ApplicationStatus.UNHEALTHY]:
+        raise AssertionError(f"Deployment failed with status: {status}")
+
+    time.sleep(1)
+
+if status != ApplicationStatus.RUNNING:
+    raise AssertionError(
+        f"Deployment failed to reach RUNNING status within {timeout_seconds}s. Current status: {status}"
+    )
+
+serve.shutdown()