AI Quick Actions v1.0.3 - llama-cpp container support (#910)

Author: VipulMascarenhas

.pre-commit-config.yaml

@@ -45,7 +45,7 @@ repos:
       rev: v8.18.4
       hooks:
           - id: gitleaks
-            exclude: .github/workflows/reusable-actions/set-dummy-conf.yml
+            exclude: .github/workflows/reusable-actions/set-dummy-conf.yml|./tests/operators/common/test_load_data.py
     # Oracle copyright checker
     - repo: https://github.com/oracle-samples/oci-data-science-ai-samples/
       rev: 1bc5270a443b791c62f634233c0f4966dfcc0dd6

ads/aqua/common/entities.py

@@ -0,0 +1,17 @@
+#!/usr/bin/env python
+# Copyright (c) 2024 Oracle and/or its affiliates.
+# Licensed under the Universal Permissive License v 1.0 as shown at https://oss.oracle.com/licenses/upl/
+
+
+class ContainerSpec:
+    """
+    Class to hold to hold keys within the container spec.
+    """
+
+    CONTAINER_SPEC = "containerSpec"
+    CLI_PARM = "cliParam"
+    SERVER_PORT = "serverPort"
+    HEALTH_CHECK_PORT = "healthCheckPort"
+    ENV_VARS = "envVars"
+    RESTRICTED_PARAMS = "restrictedParams"
+    EVALUATION_CONFIGURATION = "evaluationConfiguration"

ads/aqua/common/enums.py

@@ -1,5 +1,4 @@
 #!/usr/bin/env python
-# -*- coding: utf-8 -*-
 # Copyright (c) 2024 Oracle and/or its affiliates.
 # Licensed under the Universal Permissive License v 1.0 as shown at https://oss.oracle.com/licenses/upl/
 
@@ -8,6 +7,7 @@
 ~~~~~~~~~~~~~~
 This module contains the set of enums used in AQUA.
 """
+
 from ads.common.extended_enum import ExtendedEnumMeta
 
 
@@ -38,21 +38,25 @@ class Tags(str, metaclass=ExtendedEnumMeta):
     READY_TO_IMPORT = "ready_to_import"
     BASE_MODEL_CUSTOM = "aqua_custom_base_model"
     AQUA_EVALUATION_MODEL_ID = "evaluation_model_id"
+    MODEL_FORMAT = "model_format"
 
 
 class InferenceContainerType(str, metaclass=ExtendedEnumMeta):
     CONTAINER_TYPE_VLLM = "vllm"
     CONTAINER_TYPE_TGI = "tgi"
+    CONTAINER_TYPE_LLAMA_CPP = "llama-cpp"
 
 
 class InferenceContainerTypeFamily(str, metaclass=ExtendedEnumMeta):
     AQUA_VLLM_CONTAINER_FAMILY = "odsc-vllm-serving"
     AQUA_TGI_CONTAINER_FAMILY = "odsc-tgi-serving"
+    AQUA_LLAMA_CPP_CONTAINER_FAMILY = "odsc-llama-cpp-serving"
 
 
 class InferenceContainerParamType(str, metaclass=ExtendedEnumMeta):
     PARAM_TYPE_VLLM = "VLLM_PARAMS"
     PARAM_TYPE_TGI = "TGI_PARAMS"
+    PARAM_TYPE_LLAMA_CPP = "LLAMA_CPP_PARAMS"
 
 
 class HuggingFaceTags(str, metaclass=ExtendedEnumMeta):

ads/aqua/common/utils.py

@@ -10,14 +10,17 @@
 import os
 import random
 import re
+from datetime import datetime, timedelta
 from functools import wraps
 from pathlib import Path
 from string import Template
 from typing import List, Union
 
 import fsspec
 import oci
+from cachetools import TTLCache, cached
 from oci.data_science.models import JobRun, Model
+from oci.object_storage.models import ObjectSummary
 
 from ads.aqua.common.enums import (
     InferenceContainerParamType,
@@ -45,7 +48,6 @@
 )
 from ads.aqua.data import AquaResourceIdentifier
 from ads.common.auth import default_signer
-from ads.common.decorator.threaded import threaded
 from ads.common.extended_enum import ExtendedEnumMeta
 from ads.common.object_storage_details import ObjectStorageDetails
 from ads.common.oci_resource import SEARCH_TYPE, OCIResource
@@ -213,7 +215,6 @@ def read_file(file_path: str, **kwargs) -> str:
         return UNKNOWN
 
 
-@threaded()
 def load_config(file_path: str, config_file_name: str, **kwargs) -> dict:
     artifact_path = f"{file_path.rstrip('/')}/{config_file_name}"
     signer = default_signer() if artifact_path.startswith("oci://") else {}
@@ -228,6 +229,32 @@ def load_config(file_path: str, config_file_name: str, **kwargs) -> dict:
     return config
 
 
+def list_os_files_with_extension(oss_path: str, extension: str) -> [str]:
+    """
+    List files in the specified directory with the given extension.
+
+    Parameters:
+    - oss_path: The path to the directory where files are located.
+    - extension: The file extension to filter by (e.g., 'txt' for text files).
+
+    Returns:
+    - A list of file paths matching the specified extension.
+    """
+
+    oss_client = ObjectStorageDetails.from_path(oss_path)
+
+    # Ensure the extension is prefixed with a dot if not already
+    if not extension.startswith("."):
+        extension = "." + extension
+    files: List[ObjectSummary] = oss_client.list_objects().objects
+
+    return [
+        file.name[len(oss_client.filepath) :].lstrip("/")
+        for file in files
+        if file.name.endswith(extension)
+    ]
+
+
 def is_valid_ocid(ocid: str) -> bool:
     """Checks if the given ocid is valid.
 
@@ -503,6 +530,7 @@ def container_config_path():
     return f"oci://{AQUA_SERVICE_MODELS_BUCKET}@{CONDA_BUCKET_NS}/service_models/config"
 
 
+@cached(cache=TTLCache(maxsize=1, ttl=timedelta(hours=5), timer=datetime.now))
 def get_container_config():
     config = load_config(
         file_path=container_config_path(),
@@ -881,6 +909,8 @@ def get_container_params_type(container_type_name: str) -> str:
         return InferenceContainerParamType.PARAM_TYPE_VLLM
     elif InferenceContainerType.CONTAINER_TYPE_TGI in container_type_name.lower():
         return InferenceContainerParamType.PARAM_TYPE_TGI
+    elif InferenceContainerType.CONTAINER_TYPE_LLAMA_CPP in container_type_name.lower():
+        return InferenceContainerParamType.PARAM_TYPE_LLAMA_CPP
     else:
         return UNKNOWN
 

ads/aqua/config/config.py

@@ -1,5 +1,4 @@
 #!/usr/bin/env python
-# -*- coding: utf-8 -*-
 # Copyright (c) 2024 Oracle and/or its affiliates.
 # Licensed under the Universal Permissive License v 1.0 as shown at https://oss.oracle.com/licenses/upl/
 
@@ -14,5 +13,6 @@ def get_finetuning_config_defaults():
             "BM.GPU.A10.4": {"batch_size": 1, "replica": 1},
             "BM.GPU4.8": {"batch_size": 4, "replica": 1},
             "BM.GPU.A100-v2.8": {"batch_size": 6, "replica": 1},
+            "BM.GPU.H100.8": {"batch_size": 6, "replica": 1},
         }
     }

ads/aqua/config/deployment_config_defaults.json

@@ -1,9 +1,37 @@
 {
+  "configuration": {
+    "VM.Standard.A1.Flex": {
+      "parameters": {},
+      "shape_info": {
+        "configs": [
+          {
+            "memory_in_gbs": 128,
+            "ocpu": 20
+          },
+          {
+            "memory_in_gbs": 256,
+            "ocpu": 40
+          },
+          {
+            "memory_in_gbs": 384,
+            "ocpu": 60
+          },
+          {
+            "memory_in_gbs": 512,
+            "ocpu": 80
+          }
+        ],
+        "type": "CPU"
+      }
+    }
+  },
   "shape": [
     "VM.GPU.A10.1",
     "VM.GPU.A10.2",
     "BM.GPU.A10.4",
     "BM.GPU4.8",
-    "BM.GPU.A100-v2.8"
+    "BM.GPU.A100-v2.8",
+    "BM.GPU.H100.8",
+    "VM.Standard.A1.Flex"
   ]
 }

ads/aqua/config/resource_limit_names.json

@@ -1,6 +1,7 @@
 {
   "BM.GPU.A10.4": "ds-gpu-a10-count",
   "BM.GPU.A100-v2.8": "ds-gpu-a100-v2-count",
+  "BM.GPU.H100.8": "ds-gpu-h100-count",
   "BM.GPU4.8": "ds-gpu4-count",
   "VM.GPU.A10.1": "ds-gpu-a10-count",
   "VM.GPU.A10.2": "ds-gpu-a10-count"

ads/aqua/constants.py

@@ -21,7 +21,6 @@
 DEFAULT_FT_REPLICA = 1
 DEFAULT_FT_BATCH_SIZE = 1
 DEFAULT_FT_VALIDATION_SET_SIZE = 0.1
-
 MAXIMUM_ALLOWED_DATASET_IN_BYTE = 52428800  # 1024 x 1024 x 50 = 50MB
 JOB_INFRASTRUCTURE_TYPE_DEFAULT_NETWORKING = "ME_STANDALONE"
 NB_SESSION_IDENTIFIER = "NB_SESSION_OCID"
@@ -34,6 +33,7 @@
 AQUA_MODEL_ARTIFACT_CONFIG = "config.json"
 AQUA_MODEL_ARTIFACT_CONFIG_MODEL_NAME = "_name_or_path"
 AQUA_MODEL_ARTIFACT_CONFIG_MODEL_TYPE = "model_type"
+AQUA_MODEL_ARTIFACT_FILE = "model_file"
 
 TRAINING_METRICS_FINAL = "training_metrics_final"
 VALIDATION_METRICS_FINAL = "validation_metrics_final"
@@ -74,3 +74,7 @@
     "--sharded",
     "--trust-remote-code",
 }
+LLAMA_CPP_INFERENCE_RESTRICTED_PARAMS = {
+    "--port",
+    "--host",
+}

ads/aqua/evaluation/entities.py

@@ -1,5 +1,4 @@
 #!/usr/bin/env python
-# -*- coding: utf-8 -*-
 # Copyright (c) 2024 Oracle and/or its affiliates.
 # Licensed under the Universal Permissive License v 1.0 as shown at https://oss.oracle.com/licenses/upl/
 

ads/aqua/evaluation/evaluation.py

@@ -7,7 +7,7 @@
 import re
 import tempfile
 from concurrent.futures import ThreadPoolExecutor, as_completed
-from dataclasses import asdict
+from dataclasses import asdict, fields
 from datetime import datetime, timedelta
 from pathlib import Path
 from threading import Lock
@@ -76,6 +76,7 @@
     ModelParams,
 )
 from ads.aqua.evaluation.errors import EVALUATION_JOB_EXIT_CODE_MESSAGE
+from ads.aqua.ui import AquaContainerConfig
 from ads.common.auth import default_signer
 from ads.common.object_storage_details import ObjectStorageDetails
 from ads.common.utils import get_console_link, get_files, get_log_links
@@ -90,7 +91,9 @@
 from ads.jobs.builders.runtimes.base import Runtime
 from ads.jobs.builders.runtimes.container_runtime import ContainerRuntime
 from ads.model.datascience_model import DataScienceModel
+from ads.model.deployment import ModelDeploymentContainerRuntime
 from ads.model.deployment.model_deployment import ModelDeployment
+from ads.model.generic_model import ModelDeploymentRuntimeType
 from ads.model.model_metadata import (
     MetadataTaxonomyKeys,
     ModelCustomMetadata,
@@ -157,24 +160,47 @@ def create(
                 create_aqua_evaluation_details = CreateAquaEvaluationDetails(**kwargs)
             except Exception as ex:
                 raise AquaValueError(
-                    "Invalid create evaluation parameters. Allowable parameters are: "
-                    f"{', '.join(list(asdict(CreateAquaEvaluationDetails).keys()))}."
+                    "Invalid create evaluation parameters. "
+                    "Allowable parameters are: "
+                    f"{', '.join([field.name for field in fields(CreateAquaEvaluationDetails)])}."
                 ) from ex
 
         if not is_valid_ocid(create_aqua_evaluation_details.evaluation_source_id):
             raise AquaValueError(
                 f"Invalid evaluation source {create_aqua_evaluation_details.evaluation_source_id}. "
                 "Specify either a model or model deployment id."
             )
-
         evaluation_source = None
+        eval_inference_configuration = None
         if (
             DataScienceResource.MODEL_DEPLOYMENT
             in create_aqua_evaluation_details.evaluation_source_id
         ):
             evaluation_source = ModelDeployment.from_id(
                 create_aqua_evaluation_details.evaluation_source_id
             )
+            try:
+                if (
+                    evaluation_source.runtime.type
+                    == ModelDeploymentRuntimeType.CONTAINER
+                ):
+                    runtime = ModelDeploymentContainerRuntime.from_dict(
+                        evaluation_source.runtime.to_dict()
+                    )
+                    inference_config = AquaContainerConfig.from_container_index_json(
+                        enable_spec=True
+                    ).inference
+                    for container in inference_config.values():
+                        if container.name == runtime.image.split(":")[0]:
+                            eval_inference_configuration = (
+                                container.spec.evaluation_configuration
+                            )
+            except Exception:
+                logger.debug(
+                    f"Could not load inference config details for the evaluation id: "
+                    f"{create_aqua_evaluation_details.evaluation_source_id}. Please check if the container"
+                    f" runtime has the correct SMC image information."
+                )
         elif (
             DataScienceResource.MODEL
             in create_aqua_evaluation_details.evaluation_source_id
@@ -390,6 +416,9 @@ def create(
                 report_path=create_aqua_evaluation_details.report_path,
                 model_parameters=create_aqua_evaluation_details.model_parameters,
                 metrics=create_aqua_evaluation_details.metrics,
+                inference_configuration=eval_inference_configuration.to_filtered_dict()
+                if eval_inference_configuration
+                else {},
             )
         ).create(**kwargs)  ## TODO: decide what parameters will be needed
         logger.debug(
@@ -511,6 +540,7 @@ def _build_evaluation_runtime(
         report_path: str,
         model_parameters: dict,
         metrics: List = None,
+        inference_configuration: dict = None,
     ) -> Runtime:
         """Builds evaluation runtime for Job."""
         # TODO the image name needs to be extracted from the mapping index.json file.
@@ -520,16 +550,19 @@ def _build_evaluation_runtime(
             .with_environment_variable(
                 **{
                     "AIP_SMC_EVALUATION_ARGUMENTS": json.dumps(
-                        asdict(
-                            self._build_launch_cmd(
-                                evaluation_id=evaluation_id,
-                                evaluation_source_id=evaluation_source_id,
-                                dataset_path=dataset_path,
-                                report_path=report_path,
-                                model_parameters=model_parameters,
-                                metrics=metrics,
-                            )
-                        )
+                        {
+                            **asdict(
+                                self._build_launch_cmd(
+                                    evaluation_id=evaluation_id,
+                                    evaluation_source_id=evaluation_source_id,
+                                    dataset_path=dataset_path,
+                                    report_path=report_path,
+                                    model_parameters=model_parameters,
+                                    metrics=metrics,
+                                ),
+                            ),
+                            **(inference_configuration or {}),
+                        },
                     ),
                     "CONDA_BUCKET_NS": CONDA_BUCKET_NS,
                 },