ai-dynamo · dagil-nvidia · Mar 6, 2026 · Mar 6, 2026 · Mar 6, 2026 · Mar 6, 2026
@@ -157,6 +157,21 @@ jobs:
           path: source-checkout
           fetch-depth: 1
 
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+
+      - name: Install API doc dependencies
+        run: pip install -r source-checkout/docs/scripts/requirements-apidocs.txt
+
+      - name: Fernify API reference docs
+        working-directory: source-checkout
+        run: |
+          python3 docs/scripts/fernify_python_api.py
+          python3 docs/scripts/fernify_rust_api.py
+          python3 docs/scripts/fernify_k8s_api.py
+
       - name: Checkout docs-website branch
         uses: actions/checkout@v4
         with:
@@ -437,6 +452,15 @@ jobs:
           echo "Created fern/pages-$TAG/"
           ls -la "fern/pages-$TAG/" | head -20
 
+      - name: Update Rust API docs.rs links to release version
+        run: |
+          TAG="${{ steps.version.outputs.tag }}"
+          VERSION="${{ steps.version.outputs.version }}"
+          # Update docs.rs links from /latest to version-pinned URLs
+          # The Rust page was synced from pages-dev, just sed-replace the version
+          find "fern/pages-$TAG/api/rust" \( -name "*.md" -o -name "*.mdx" \) -exec \
+            sed -i "s|docs.rs/\([^/]*\)/latest|docs.rs/\1/$VERSION|g" {} +
+
       - name: Update GitHub links to 'main' to version tag
         run: |
           TAG="${{ steps.version.outputs.tag }}"

@@ -11,6 +11,24 @@ class ArgGroup(ABC):
     Base interface for configuration groups.
 
     Each ArgGroup represents a domain of configuration parameters with clear ownership.
+
+    Examples:
+        >>> import argparse
+        >>> from dynamo.common.configuration.arg_group import ArgGroup
+        >>> from dynamo.common.configuration.utils import add_argument
+        >>>
+        >>> class MyArgGroup(ArgGroup):
+        ...     def add_arguments(self, parser) -> None:
+        ...         add_argument(
+        ...             parser,
+        ...             flag_name="--my-option",
+        ...             env_var="DYN_MY_OPTION",
+        ...             default="value",
+        ...             help="A custom option.",
+        ...         )
+        >>>
+        >>> parser = argparse.ArgumentParser()
+        >>> MyArgGroup().add_arguments(parser)
     """
 
     @abstractmethod

@@ -5,7 +5,24 @@
 
 
 class ConfigBase:
-    """Base configuration class that allows properties with and without defaults in arbitrary order."""
+    """Base configuration class that allows properties with and without defaults in arbitrary order.
+
+    Examples:
+        >>> import argparse
+        >>> from dynamo.common.configuration.config_base import ConfigBase
+        >>>
+        >>> class MyConfig(ConfigBase):
+        ...     model_name: str
+        ...     http_port: int = 8000
+        >>>
+        >>> parser = argparse.ArgumentParser()
+        >>> parser.add_argument("--model-name", default="llama")
+        >>> parser.add_argument("--http-port", type=int, default=8000)
+        >>> args = parser.parse_args([])
+        >>> config = MyConfig.from_cli_args(args)
+        >>> config.http_port
+        8000
+    """
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace) -> Self:

@@ -38,7 +38,22 @@
 
 
 class KvRouterConfigBase(ConfigBase):
-    """Mixin carrying the 16 KvRouterConfig fields."""
+    """Mixin carrying the 16 KvRouterConfig fields.
+
+    Examples:
+        >>> from dynamo.common.configuration.groups.kv_router_args import (
+        ...     KvRouterArgGroup,
+        ...     KvRouterConfigBase,
+        ... )
+        >>>
+        >>> class MyRouterConfig(KvRouterConfigBase):
+        ...     endpoint: str = "http://localhost:8080"
+        >>>
+        >>> config = MyRouterConfig()
+        >>> kwargs = config.kv_router_kwargs()
+        >>> "overlap_score_weight" in kwargs
+        True
+    """
 
     overlap_score_weight: float
     router_temperature: float
@@ -63,7 +78,21 @@ def kv_router_kwargs(self) -> dict:
 
 
 class KvRouterArgGroup(ArgGroup):
-    """CLI arguments for the 16 KvRouterConfig parameters."""
+    """CLI arguments for the 16 KvRouterConfig parameters.
+
+    Examples:
+        >>> import argparse
+        >>> from dynamo.common.configuration.groups.kv_router_args import KvRouterArgGroup
+        >>>
+        >>> parser = argparse.ArgumentParser()
+        >>> KvRouterArgGroup().add_arguments(parser)
+        >>> args = parser.parse_args([
+        ...     "--router-kv-overlap-score-weight", "0.8",
+        ...     "--router-temperature", "0.5",
+        ... ])
+        >>> args.overlap_score_weight
+        0.8
+    """
 
     def add_arguments(self, parser) -> None:
         g = parser.add_argument_group("KV Router Options")

@@ -68,6 +68,23 @@ def add_argument(
     """
     Add a CLI argument with env var default, optional alias and dest, and help message construction.
 
+    Examples:
+        >>> import argparse
+        >>> from dynamo.common.configuration.utils import add_argument
+        >>>
+        >>> parser = argparse.ArgumentParser()
+        >>> add_argument(
+        ...     parser,
+        ...     flag_name="--http-port",
+        ...     env_var="DYN_HTTP_PORT",
+        ...     default=8000,
+        ...     help="HTTP port for the engine.",
+        ...     arg_type=int,
+        ... )
+        >>> args = parser.parse_args([])
+        >>> args.http_port
+        8000
+
     Args:
         parser: ArgumentParser or argument group
         flag_name: Primary flag (must start with '--', e.g., "--foo")

@@ -7,15 +7,31 @@
 
 
 class DisaggregationMode(Enum):
-    """Disaggregation mode for LLM workers."""
+    """Disaggregation mode for LLM workers.
+
+    Examples:
+        >>> from dynamo.common.constants import DisaggregationMode
+        >>> DisaggregationMode.PREFILL.value
+        'prefill'
+        >>> DisaggregationMode("agg") == DisaggregationMode.AGGREGATED
+        True
+    """
 
     AGGREGATED = "agg"
     PREFILL = "prefill"
     DECODE = "decode"
 
 
 class EmbeddingTransferMode(Enum):
-    """Embedding transfer mode for LLM workers."""
+    """Embedding transfer mode for LLM workers.
+
+    Examples:
+        >>> from dynamo.common.constants import EmbeddingTransferMode
+        >>> EmbeddingTransferMode.NIXL_WRITE.value
+        'nixl-write'
+        >>> EmbeddingTransferMode("local") == EmbeddingTransferMode.LOCAL
+        True
+    """
 
     LOCAL = "local"
     NIXL_WRITE = "nixl-write"

@@ -33,6 +33,16 @@ class LoRAManager:
 
     The manager uses the Rust-based LoRADownloader for S3 and local file sources,
     and allows registering custom Python sources for other protocols.
+
+    Examples:
+        >>> import asyncio
+        >>> from dynamo.common.lora import LoRAManager
+        >>> manager = LoRAManager()
+        >>> result = asyncio.run(
+        ...     manager.download_lora("s3://bucket/adapters/lora-v1")
+        ... )  # doctest: +SKIP
+        >>> manager.is_cached("s3://bucket/adapters/lora-v1")  # doctest: +SKIP
+        True
     """
 
     def __init__(self, cache_path: Optional[Path] = None):

@@ -46,6 +46,17 @@ def get_fs(fs_url: str) -> DirFileSystem:
     Args:
         fs_url: The URL of the filesystem to initialize. e.g. s3://bucket, gs://bucket, file:///local/path
 
+    Examples:
+        >>> from dynamo.common.storage import get_fs
+        >>>
+        >>> fs = get_fs("file:///data/media")
+        >>> protocol = fs.fs.protocol
+        >>> protocol if isinstance(protocol, str) else protocol[0]
+        'file'
+        >>> fs = get_fs("s3://my-bucket")
+        >>> fs.path
+        'my-bucket'
+
     Returns:
         The initialized DirFileSystem wrapper for the filesystem.
 
@@ -81,6 +92,15 @@ def get_media_url(
             ``{base_url}/{storage_path}``.  When *None*, the URL is constructed
             from the filesystem's protocol and root path.
 
+    Examples:
+        >>> from dynamo.common.storage import get_fs, get_media_url
+        >>>
+        >>> fs = get_fs("file:///data/media")
+        >>> get_media_url(fs, "videos/req-123.mp4")
+        'file:///data/media/videos/req-123.mp4'
+        >>> get_media_url(fs, "img.png", base_url="https://cdn.example.com/media")
+        'https://cdn.example.com/media/img.png'
+
     Returns:
         Public URL string for the uploaded file.
     """
@@ -110,6 +130,16 @@ async def upload_to_fs(
         data: Raw bytes to upload.
         base_url: Optional CDN / proxy base URL for URL rewriting.
 
+    Examples:
+        >>> import asyncio
+        >>> from dynamo.common.storage import get_fs, upload_to_fs
+        >>>
+        >>> fs = get_fs("s3://my-media-bucket")
+        >>> image_bytes = b"\\x89PNG..."
+        >>> url = asyncio.run(
+        ...     upload_to_fs(fs, "images/req-123/output.png", image_bytes)
+        ... )  # doctest: +SKIP
+
     Returns:
         Public URL string for the uploaded file.
     """

@@ -40,7 +40,24 @@ def validate_model_path(value: str) -> str:
 
 
 class FrontendConfig(KvRouterConfigBase):
-    """Configuration for the Dynamo frontend."""
+    """Configuration for the Dynamo frontend.
+
+    Examples:
+        >>> import argparse
+        >>> from dynamo.frontend.frontend_args import FrontendArgGroup, FrontendConfig
+        >>>
+        >>> parser = argparse.ArgumentParser()
+        >>> FrontendArgGroup().add_arguments(parser)
+        >>> args = parser.parse_args([
+        ...     "--model-name", "Llama-3.2-1B-Instruct",
+        ...     "--http-port", "8080",
+        ... ])
+        >>> config = FrontendConfig.from_cli_args(args)
+        >>> config.model_name
+        'Llama-3.2-1B-Instruct'
+        >>> config.http_port
+        8080
+    """
 
     interactive: bool
     kv_cache_block_size: Optional[int]
@@ -93,7 +110,23 @@ def _preprocess_for_encode_config(config: FrontendConfig) -> Dict[str, Any]:
 
 
 class FrontendArgGroup(ArgGroup):
-    """Frontend configuration parameters."""
+    """Frontend configuration parameters.
+
+    Examples:
+        >>> import argparse
+        >>> from dynamo.frontend.frontend_args import FrontendArgGroup
+        >>>
+        >>> parser = argparse.ArgumentParser()
+        >>> group = FrontendArgGroup()
+        >>> group.add_arguments(parser)
+        >>> args = parser.parse_args([
+        ...     "--model-name", "Llama-3.2-1B-Instruct",
+        ...     "--router-mode", "kv",
+        ...     "--http-port", "8080",
+        ... ])
+        >>> args.model_name
+        'Llama-3.2-1B-Instruct'
+    """
 
     def add_arguments(self, parser) -> None:
         parser.add_argument(

@@ -70,6 +70,21 @@ async def worker():
 
     Each mocker gets its own DistributedRuntime instance for true isolation,
     while still sharing the same event loop and tokio runtime.
+
+    Examples:
+        Launch via CLI (the standard invocation pattern):
+
+        >>> # python -m dynamo.mocker --model-path /data/models/Qwen3-0.6B
+        >>>
+        >>> import uvloop
+        >>> from dynamo.mocker.main import worker
+        >>> uvloop.run(worker())
+
+        Launch multiple workers with stagger delay:
+
+        >>> # python -m dynamo.mocker \\
+        >>> #     --model-path /data/models/Qwen3-0.6B \\
+        >>> #     --num-workers 4 --stagger-delay 0.1
     """
     args = parse_args()
 

@@ -49,6 +49,23 @@ class BasePlannerDefaults:
 
 
 class SLAPlannerDefaults(BasePlannerDefaults):
+    """SLA-based planner defaults for throughput and latency targets.
+
+    Provides default values for SLA objectives (TTFT, ITL), load prediction,
+    and scaling parameters used by the planner.
+
+    Examples:
+        >>> from dynamo.planner.defaults import SLAPlannerDefaults
+        >>> SLAPlannerDefaults.ttft
+        500.0
+        >>> SLAPlannerDefaults.itl
+        50.0
+        >>> SLAPlannerDefaults.mode
+        'disagg'
+        >>> SLAPlannerDefaults.max_gpu_budget
+        8
+    """
+
     # Prometheus endpoint URL for pulling/querying metrics
     metric_pulling_prometheus_endpoint = os.environ.get(
         "PROMETHEUS_ENDPOINT",
@@ -143,6 +160,16 @@ class MockerComponentName:
 
 
 class SubComponentType(str, Enum):
+    """Type of sub-component in a disaggregated deployment.
+
+    Examples:
+        >>> from dynamo.planner.defaults import SubComponentType
+        >>> SubComponentType.PREFILL.value
+        'prefill'
+        >>> SubComponentType("decode") == SubComponentType.DECODE
+        True
+    """
+
     PREFILL = "prefill"
     DECODE = "decode"
 

@@ -28,6 +28,19 @@ class GlobalPlannerConnector(PlannerConnector):
     This connector wraps RemotePlannerClient and implements the PlannerConnector
     interface, allowing planner_core.py to treat global-planner environment mode
     consistently with kubernetes and virtual modes.
+
+    Examples:
+        >>> from dynamo.planner.global_planner_connector import GlobalPlannerConnector
+        >>>
+        >>> connector = GlobalPlannerConnector(
+        ...     runtime=runtime,
+        ...     dynamo_namespace="test-ns",
+        ...     global_planner_namespace="global-ns",
+        ...     model_name="Llama-3.2-1B-Instruct",
+        ... )
+        >>> await connector._async_init()
+        >>> connector.get_model_name()
+        'Llama-3.2-1B-Instruct'
     """
 
     def __init__(