Merge branch 'main' of https://github.com/oracle/accelerated-data-science into model_group

Author: lu-ohai

ads/aqua/app.py

@@ -64,6 +64,7 @@ def __init__(self) -> None:
             set_auth("resource_principal")
         self._auth = default_signer({"service_endpoint": OCI_ODSC_SERVICE_ENDPOINT})
         self.ds_client = oc.OCIClientFactory(**self._auth).data_science
+        self.compute_client = oc.OCIClientFactory(**default_signer()).compute
         self.logging_client = oc.OCIClientFactory(**default_signer()).logging_management
         self.identity_client = oc.OCIClientFactory(**default_signer()).identity
         self.region = extract_region(self._auth)

ads/aqua/common/entities.py

@@ -157,6 +157,8 @@ class AquaMultiModelRef(Serializable):
         Optional environment variables to override during deployment.
     artifact_location : Optional[str]
         Artifact path of model in the multimodel group.
+    fine_tune_weights_location : Optional[str]
+        For fine tuned models, the artifact path of the modified model weights
     """
 
     model_id: str = Field(..., description="The model OCID to deploy.")
@@ -171,6 +173,9 @@ class AquaMultiModelRef(Serializable):
     artifact_location: Optional[str] = Field(
         None, description="Artifact path of model in the multimodel group."
     )
+    fine_tune_weights_location: Optional[str] = Field(
+        None, description="For fine tuned models, the artifact path of the modified model weights"
+    )
 
     class Config:
         extra = "ignore"

ads/aqua/common/enums.py

@@ -20,6 +20,12 @@ class Resource(ExtendedEnum):
     MODEL_VERSION_SET = "model-version-sets"
 
 
+class PredictEndpoints(ExtendedEnum):
+    CHAT_COMPLETIONS_ENDPOINT = "/v1/chat/completions"
+    TEXT_COMPLETIONS_ENDPOINT = "/v1/completions"
+    EMBEDDING_ENDPOINT = "/v1/embedding"
+
+
 class Tags(ExtendedEnum):
     TASK = "task"
     LICENSE = "license"
@@ -49,6 +55,7 @@ class InferenceContainerType(ExtendedEnum):
 class InferenceContainerTypeFamily(ExtendedEnum):
     AQUA_VLLM_CONTAINER_FAMILY = "odsc-vllm-serving"
     AQUA_VLLM_V1_CONTAINER_FAMILY = "odsc-vllm-serving-v1"
+    AQUA_VLLM_LLAMA4_CONTAINER_FAMILY = "odsc-vllm-serving-llama4"
     AQUA_TGI_CONTAINER_FAMILY = "odsc-tgi-serving"
     AQUA_LLAMA_CPP_CONTAINER_FAMILY = "odsc-llama-cpp-serving"
 
@@ -119,4 +126,9 @@ class Platform(ExtendedEnum):
         InferenceContainerTypeFamily.AQUA_VLLM_V1_CONTAINER_FAMILY,
         InferenceContainerTypeFamily.AQUA_VLLM_CONTAINER_FAMILY,
     ],
+    InferenceContainerTypeFamily.AQUA_VLLM_LLAMA4_CONTAINER_FAMILY: [
+        InferenceContainerTypeFamily.AQUA_VLLM_LLAMA4_CONTAINER_FAMILY,
+        InferenceContainerTypeFamily.AQUA_VLLM_V1_CONTAINER_FAMILY,
+        InferenceContainerTypeFamily.AQUA_VLLM_CONTAINER_FAMILY,
+    ],
 }

ads/aqua/common/utils.py

@@ -832,7 +832,9 @@ def get_params_dict(params: Union[str, List[str]]) -> dict:
     """
     params_list = get_params_list(params) if isinstance(params, str) else params
     return {
-        split_result[0]: split_result[1] if len(split_result) > 1 else UNKNOWN
+        split_result[0]: " ".join(split_result[1:])
+        if len(split_result) > 1
+        else UNKNOWN
         for split_result in (x.split() for x in params_list)
     }
 
@@ -881,7 +883,9 @@ def build_params_string(params: dict) -> str:
         A params string.
     """
     return (
-        " ".join(f"{name} {value}" for name, value in params.items()).strip()
+        " ".join(
+            f"{name} {value}" if value else f"{name}" for name, value in params.items()
+        ).strip()
         if params
         else UNKNOWN
     )
@@ -1158,9 +1162,11 @@ def validate_cmd_var(cmd_var: List[str], overrides: List[str]) -> List[str]:
 
 
 def build_pydantic_error_message(ex: ValidationError):
-    """Added to handle error messages from pydantic model validator.
+    """
+    Added to handle error messages from pydantic model validator.
     Combine both loc and msg for errors where loc (field) is present in error details, else only build error
-    message using msg field."""
+    message using msg field.
+    """
 
     return {
         ".".join(map(str, e["loc"])): e["msg"]
@@ -1185,67 +1191,71 @@ def is_pydantic_model(obj: object) -> bool:
 
 @cached(cache=TTLCache(maxsize=1, ttl=timedelta(minutes=5), timer=datetime.now))
 def load_gpu_shapes_index(
-    auth: Optional[Dict] = None,
+    auth: Optional[Dict[str, Any]] = None,
 ) -> GPUShapesIndex:
     """
-    Loads the GPU shapes index from Object Storage or a local resource folder.
+    Load the GPU shapes index, preferring the OS bucket copy over the local one.
 
-    The function first attempts to load the file from an Object Storage bucket using fsspec.
-    If the loading fails (due to connection issues, missing file, etc.), it falls back to
-    loading the index from a local file.
+    Attempts to read `gpu_shapes_index.json` from OCI Object Storage first;
+    if that succeeds, those entries will override the local defaults.
 
     Parameters
     ----------
-    auth: (Dict, optional). Defaults to None.
-        The default authentication is set using `ads.set_auth` API. If you need to override the
-        default, use the `ads.common.auth.api_keys` or `ads.common.auth.resource_principal` to create appropriate
-        authentication signer and kwargs required to instantiate IdentityClient object.
+    auth
+        Optional auth dict (as returned by `ads.common.auth.default_signer()`)
+        to pass through to `fsspec.open()`.
 
     Returns
     -------
-    GPUShapesIndex: The parsed GPU shapes index.
+    GPUShapesIndex
+        Merged index where any shape present remotely supersedes the local entry.
 
     Raises
     ------
-    FileNotFoundError: If the GPU shapes index cannot be found in either Object Storage or locally.
-    json.JSONDecodeError: If the JSON is malformed.
+    json.JSONDecodeError
+        If any of the JSON is malformed.
     """
     file_name = "gpu_shapes_index.json"
-    data: Dict[str, Any] = {}
 
-    # Check if the CONDA_BUCKET_NS environment variable is set.
+    # Try remote load
+    remote_data: Dict[str, Any] = {}
     if CONDA_BUCKET_NS:
         try:
             auth = auth or authutil.default_signer()
-            # Construct the object storage path. Adjust bucket name and path as needed.
             storage_path = (
                 f"oci://{CONDA_BUCKET_NAME}@{CONDA_BUCKET_NS}/service_pack/{file_name}"
             )
-            logger.debug("Loading GPU shapes index from Object Storage")
-            with fsspec.open(storage_path, mode="r", **auth) as file_obj:
-                data = json.load(file_obj)
-            logger.debug("Successfully loaded GPU shapes index.")
-        except Exception as ex:
             logger.debug(
-                f"Failed to load GPU shapes index from Object Storage. Details: {ex}"
-            )
-
-    # If loading from Object Storage failed, load from the local resource folder.
-    if not data:
-        try:
-            local_path = os.path.join(
-                os.path.dirname(__file__), "../resources", file_name
+                "Loading GPU shapes index from Object Storage: %s", storage_path
             )
-            logger.debug(f"Loading GPU shapes index from {local_path}.")
-            with open(local_path) as file_obj:
-                data = json.load(file_obj)
-            logger.debug("Successfully loaded GPU shapes index.")
-        except Exception as e:
+            with fsspec.open(storage_path, mode="r", **auth) as f:
+                remote_data = json.load(f)
             logger.debug(
-                f"Failed to load GPU shapes index from {local_path}. Details: {e}"
+                "Loaded %d shapes from Object Storage",
+                len(remote_data.get("shapes", {})),
             )
+        except Exception as ex:
+            logger.debug("Remote load failed (%s); falling back to local", ex)
+
+    # Load local copy
+    local_data: Dict[str, Any] = {}
+    local_path = os.path.join(os.path.dirname(__file__), "../resources", file_name)
+    try:
+        logger.debug("Loading GPU shapes index from local file: %s", local_path)
+        with open(local_path) as f:
+            local_data = json.load(f)
+        logger.debug(
+            "Loaded %d shapes from local file", len(local_data.get("shapes", {}))
+        )
+    except Exception as ex:
+        logger.debug("Local load GPU shapes index failed (%s)", ex)
+
+    # Merge: remote shapes override local
+    local_shapes = local_data.get("shapes", {})
+    remote_shapes = remote_data.get("shapes", {})
+    merged_shapes = {**local_shapes, **remote_shapes}
 
-    return GPUShapesIndex(**data)
+    return GPUShapesIndex(shapes=merged_shapes)
 
 
 def get_preferred_compatible_family(selected_families: set[str]) -> str:

ads/aqua/config/container_config.py

@@ -7,6 +7,7 @@
 from oci.data_science.models import ContainerSummary
 from pydantic import Field
 
+from ads.aqua import logger
 from ads.aqua.config.utils.serializer import Serializable
 from ads.aqua.constants import (
     SERVICE_MANAGED_CONTAINER_URI_SCHEME,
@@ -168,50 +169,47 @@ def from_service_config(
             container_type = container.family_name
             usages = [x.upper() for x in container.usages]
             if "INFERENCE" in usages or "MULTI_MODEL" in usages:
+                # Extract additional configurations
+                additional_configurations = {}
+                try:
+                    additional_configurations = (
+                        container.workload_configuration_details_list[
+                            0
+                        ].additional_configurations
+                    )
+                except (AttributeError, IndexError) as ex:
+                    logger.debug(
+                        "Failed to extract `additional_configurations` for container '%s': %s",
+                        getattr(container, "container_name", "<unknown>"),
+                        ex,
+                    )
+
                 container_item.platforms.append(
-                    container.workload_configuration_details_list[
-                        0
-                    ].additional_configurations.get("platforms")
+                    additional_configurations.get("platforms")
                 )
                 container_item.model_formats.append(
-                    container.workload_configuration_details_list[
-                        0
-                    ].additional_configurations.get("modelFormats")
+                    additional_configurations.get("modelFormats")
                 )
+
+                # Parse environment variables from `additional_configurations`.
+                # Only keys present in the configuration will be added to the result.
+                config_keys = {
+                    "MODEL_DEPLOY_PREDICT_ENDPOINT": UNKNOWN,
+                    "MODEL_DEPLOY_HEALTH_ENDPOINT": UNKNOWN,
+                    "MODEL_DEPLOY_ENABLE_STREAMING": UNKNOWN,
+                    "PORT": UNKNOWN,
+                    "HEALTH_CHECK_PORT": UNKNOWN,
+                    "VLLM_USE_V1": UNKNOWN,
+                }
+
                 env_vars = [
-                    {
-                        "MODEL_DEPLOY_PREDICT_ENDPOINT": container.workload_configuration_details_list[
-                            0
-                        ].additional_configurations.get(
-                            "MODEL_DEPLOY_PREDICT_ENDPOINT", UNKNOWN
-                        )
-                    },
-                    {
-                        "MODEL_DEPLOY_HEALTH_ENDPOINT": container.workload_configuration_details_list[
-                            0
-                        ].additional_configurations.get(
-                            "MODEL_DEPLOY_HEALTH_ENDPOINT", UNKNOWN
-                        )
-                    },
-                    {
-                        "MODEL_DEPLOY_ENABLE_STREAMING": container.workload_configuration_details_list[
-                            0
-                        ].additional_configurations.get(
-                            "MODEL_DEPLOY_ENABLE_STREAMING", UNKNOWN
-                        )
-                    },
-                    {
-                        "PORT": container.workload_configuration_details_list[
-                            0
-                        ].additional_configurations.get("PORT", "")
-                    },
-                    {
-                        "HEALTH_CHECK_PORT": container.workload_configuration_details_list[
-                            0
-                        ].additional_configurations.get("HEALTH_CHECK_PORT", UNKNOWN),
-                    },
+                    {key: additional_configurations.get(key, default)}
+                    for key, default in config_keys.items()
+                    if key in additional_configurations
                 ]
-                container_spec = AquaContainerConfigSpec(
+
+                # Build container spec
+                container_item.spec = AquaContainerConfigSpec(
                     cli_param=container.workload_configuration_details_list[0].cmd,
                     server_port=str(
                         container.workload_configuration_details_list[0].server_port
@@ -236,13 +234,14 @@ def from_service_config(
                         )
                     ),
                 )
-                container_item.spec = container_spec
+
             if "INFERENCE" in usages or "MULTI_MODEL" in usages:
                 inference_items[container_type] = container_item
             if "FINE_TUNE" in usages:
                 finetune_items[container_type] = container_item
             if "EVALUATION" in usages:
                 evaluate_items[container_type] = container_item
+
         return cls(
             inference=inference_items, finetune=finetune_items, evaluate=evaluate_items
         )