kiln-ai 0.13.2__py3-none-any.whl → 0.15.0__py3-none-any.whl

kiln_ai/adapters/eval/base_eval.py

@@ -2,11 +2,13 @@ import json
 from abc import abstractmethod
 from typing import Dict
 
+import jsonschema
+
 from kiln_ai.adapters.adapter_registry import adapter_for_task
 from kiln_ai.adapters.ml_model_list import ModelProviderName
 from kiln_ai.adapters.model_adapters.base_adapter import AdapterConfig
 from kiln_ai.datamodel.eval import Eval, EvalConfig, EvalScores
-from kiln_ai.datamodel.json_schema import validate_schema
+from kiln_ai.datamodel.json_schema import validate_schema_with_value_error
 from kiln_ai.datamodel.task import RunConfig, TaskOutputRatingType, TaskRun
 from kiln_ai.utils.exhaustive_error import raise_exhaustive_enum_error
 
@@ -72,7 +74,10 @@ class BaseEval:
         run_output = await run_adapter.invoke(parsed_input)
 
         eval_output, intermediate_outputs = await self.run_eval(run_output)
-        validate_schema(eval_output, self.score_schema)
+
+        validate_schema_with_value_error(
+            eval_output, self.score_schema, "Eval output does not match score schema."
+        )
 
         return run_output, eval_output, intermediate_outputs
 

kiln_ai/adapters/fine_tune/base_finetune.py

@@ -72,8 +72,6 @@ class BaseFinetuneAdapter(ABC):
         Create and start a fine-tune.
         """
 
-        cls.check_valid_provider_model(provider_id, provider_base_model_id)
-
         if not dataset.id:
             raise ValueError("Dataset must have an id")
 
@@ -168,9 +166,12 @@ class BaseFinetuneAdapter(ABC):
 
                 # Strict type checking for numeric types
                 if expected_type is float and not isinstance(value, float):
-                    raise ValueError(
-                        f"Parameter {parameter.name} must be a float, got {type(value)}"
-                    )
+                    if isinstance(value, int):
+                        value = float(value)
+                    else:
+                        raise ValueError(
+                            f"Parameter {parameter.name} must be a float, got {type(value)}"
+                        )
                 elif expected_type is int and not isinstance(value, int):
                     raise ValueError(
                         f"Parameter {parameter.name} must be an integer, got {type(value)}"
@@ -184,21 +185,3 @@ class BaseFinetuneAdapter(ABC):
         for parameter_key in parameters:
             if parameter_key not in allowed_parameters:
                 raise ValueError(f"Parameter {parameter_key} is not available")
-
-    @classmethod
-    def check_valid_provider_model(
-        cls, provider_id: str, provider_base_model_id: str
-    ) -> None:
-        """
-        Check if the provider and base model are valid.
-        """
-        for model in built_in_models:
-            for provider in model.providers:
-                if (
-                    provider.name == provider_id
-                    and provider.provider_finetune_id == provider_base_model_id
-                ):
-                    return
-        raise ValueError(
-            f"Provider {provider_id} with base model {provider_base_model_id} is not available"
-        )

kiln_ai/adapters/fine_tune/dataset_formatter.py

@@ -30,8 +30,8 @@ class DatasetFormat(str, Enum):
         "huggingface_chat_template_toolcall_jsonl"
     )
 
-    """Vertex Gemini 1.5 format (flash and pro)"""
-    VERTEX_GEMINI_1_5 = "vertex_gemini_1_5"
+    """Vertex Gemini format"""
+    VERTEX_GEMINI = "vertex_gemini"
 
 
 @dataclass
@@ -288,7 +288,7 @@ def generate_huggingface_chat_template_toolcall(
     return {"conversations": conversations}
 
 
-def generate_vertex_gemini_1_5(
+def generate_vertex_gemini(
     training_data: ModelTrainingData,
 ) -> Dict[str, Any]:
     """Generate Vertex Gemini 1.5 format (flash and pro)"""
@@ -346,7 +346,7 @@ FORMAT_GENERATORS: Dict[DatasetFormat, FormatGenerator] = {
     DatasetFormat.OPENAI_CHAT_TOOLCALL_JSONL: generate_chat_message_toolcall,
     DatasetFormat.HUGGINGFACE_CHAT_TEMPLATE_JSONL: generate_huggingface_chat_template,
     DatasetFormat.HUGGINGFACE_CHAT_TEMPLATE_TOOLCALL_JSONL: generate_huggingface_chat_template_toolcall,
-    DatasetFormat.VERTEX_GEMINI_1_5: generate_vertex_gemini_1_5,
+    DatasetFormat.VERTEX_GEMINI: generate_vertex_gemini,
 }
 
 

kiln_ai/adapters/fine_tune/finetune_registry.py

@@ -4,10 +4,12 @@ from kiln_ai.adapters.fine_tune.base_finetune import BaseFinetuneAdapter
 from kiln_ai.adapters.fine_tune.fireworks_finetune import FireworksFinetune
 from kiln_ai.adapters.fine_tune.openai_finetune import OpenAIFinetune
 from kiln_ai.adapters.fine_tune.together_finetune import TogetherFinetune
+from kiln_ai.adapters.fine_tune.vertex_finetune import VertexFinetune
 from kiln_ai.adapters.ml_model_list import ModelProviderName
 
 finetune_registry: dict[ModelProviderName, Type[BaseFinetuneAdapter]] = {
     ModelProviderName.openai: OpenAIFinetune,
     ModelProviderName.fireworks_ai: FireworksFinetune,
     ModelProviderName.together_ai: TogetherFinetune,
+    ModelProviderName.vertex: VertexFinetune,
 }

kiln_ai/adapters/fine_tune/fireworks_finetune.py

@@ -1,4 +1,5 @@
-from typing import Tuple
+import logging
+from typing import List, Tuple
 from uuid import uuid4
 
 import httpx
@@ -13,6 +14,14 @@ from kiln_ai.adapters.fine_tune.dataset_formatter import DatasetFormat, DatasetF
 from kiln_ai.datamodel import DatasetSplit, StructuredOutputMode, Task
 from kiln_ai.utils.config import Config
 
+logger = logging.getLogger(__name__)
+
+# https://docs.fireworks.ai/fine-tuning/fine-tuning-models#supported-base-models-loras-on-serverless
+serverless_models = [
+    "accounts/fireworks/models/llama-v3p1-8b-instruct",
+    "accounts/fireworks/models/llama-v3p1-70b-instruct",
+]
+
 
 class FireworksFinetune(BaseFinetuneAdapter):
     """
@@ -189,7 +198,8 @@ class FireworksFinetune(BaseFinetuneAdapter):
         if not api_key or not account_id:
             raise ValueError("Fireworks API key or account ID not set")
         url = f"https://api.fireworks.ai/v1/accounts/{account_id}/datasets"
-        dataset_id = str(uuid4())
+        # First char can't be a digit: https://discord.com/channels/1137072072808472616/1363214412395184350/1363214412395184350
+        dataset_id = "kiln-" + str(uuid4())
         payload = {
             "datasetId": dataset_id,
             "dataset": {
@@ -283,32 +293,54 @@ class FireworksFinetune(BaseFinetuneAdapter):
         return {k: v for k, v in payload.items() if v is not None}
 
     async def _deploy(self) -> bool:
-        # Now we "deploy" the model using PEFT serverless.
-        # A bit complicated: most fireworks deploys are server based.
-        # However, a Lora can be serverless (PEFT).
-        # By calling the deploy endpoint WITHOUT first creating a deployment ID, it will only deploy if it can be done serverless.
-        # https://docs.fireworks.ai/models/deploying#deploying-to-serverless
-        # This endpoint will return 400 if already deployed with code 9, so we consider that a success.
+        if self.datamodel.base_model_id in serverless_models:
+            return await self._deploy_serverless()
+        else:
+            return await self._check_or_deploy_server()
 
+    def api_key_and_account_id(self) -> Tuple[str, str]:
         api_key = Config.shared().fireworks_api_key
         account_id = Config.shared().fireworks_account_id
         if not api_key or not account_id:
             raise ValueError("Fireworks API key or account ID not set")
+        return api_key, account_id
 
+    def deployment_display_name(self) -> str:
+        # Limit the display name to 60 characters
+        display_name = f"Kiln AI fine-tuned model [ID:{self.datamodel.id}][name:{self.datamodel.name}]"[
+            :60
+        ]
+        return display_name
+
+    async def model_id_checking_status(self) -> str | None:
         # Model ID != fine tune ID on Fireworks. Model is the result of the tune job. Call status to get it.
         status, model_id = await self._status()
         if status.status != FineTuneStatusType.completed:
-            return False
+            return None
         if not model_id or not isinstance(model_id, str):
-            return False
+            return None
+        return model_id
+
+    async def _deploy_serverless(self) -> bool:
+        # Now we "deploy" the model using PEFT serverless.
+        # A bit complicated: most fireworks deploys are server based.
+        # However, a Lora can be serverless (PEFT).
+        # By calling the deploy endpoint WITHOUT first creating a deployment ID, it will only deploy if it can be done serverless.
+        # https://docs.fireworks.ai/models/deploying#deploying-to-serverless
+        # This endpoint will return 400 if already deployed with code 9, so we consider that a success.
+
+        api_key, account_id = self.api_key_and_account_id()
 
         url = f"https://api.fireworks.ai/v1/accounts/{account_id}/deployedModels"
-        # Limit the display name to 60 characters
-        display_name = f"Kiln AI fine-tuned model [ID:{self.datamodel.id}][name:{self.datamodel.name}]"[
-            :60
-        ]
+        model_id = await self.model_id_checking_status()
+        if not model_id:
+            logger.error(
+                "Model ID not found - can't deploy model to Fireworks serverless"
+            )
+            return False
+
         payload = {
-            "displayName": display_name,
+            "displayName": self.deployment_display_name(),
             "model": model_id,
         }
         headers = {
@@ -327,4 +359,120 @@ class FireworksFinetune(BaseFinetuneAdapter):
                     self.datamodel.save_to_file()
             return True
 
+        logger.error(
+            f"Failed to deploy model to Fireworks serverless: [{response.status_code}] {response.text}"
+        )
         return False
+
+    async def _check_or_deploy_server(self) -> bool:
+        """
+        Check if the model is already deployed. If not, deploy it to a dedicated server.
+        """
+
+        # Check if the model is already deployed
+        # If it's fine_tune_model_id is set, it might be deployed. However, Fireworks deletes them over time so we need to check.
+        if self.datamodel.fine_tune_model_id:
+            deployments = await self._fetch_all_deployments()
+            for deployment in deployments:
+                if deployment[
+                    "baseModel"
+                ] == self.datamodel.fine_tune_model_id and deployment["state"] in [
+                    "READY",
+                    "CREATING",
+                ]:
+                    return True
+
+        # If the model is not deployed, deploy it
+        return await self._deploy_server()
+
+    async def _deploy_server(self) -> bool:
+        # For models that are not serverless, we just need to deploy the model to a server.
+        # We use a scale-to-zero on-demand deployment. If you stop using it, it
+        # will scale to zero and charges will stop.
+        model_id = await self.model_id_checking_status()
+        if not model_id:
+            logger.error("Model ID not found - can't deploy model to Fireworks server")
+            return False
+
+        api_key, account_id = self.api_key_and_account_id()
+        url = f"https://api.fireworks.ai/v1/accounts/{account_id}/deployments"
+
+        payload = {
+            "displayName": self.deployment_display_name(),
+            "description": "Deployed by Kiln AI",
+            # Allow scale to zero
+            "minReplicaCount": 0,
+            "autoscalingPolicy": {
+                "scaleUpWindow": "30s",
+                "scaleDownWindow": "300s",
+                # Scale to zero after 5 minutes of inactivity - this is the minimum allowed
+                "scaleToZeroWindow": "300s",
+            },
+            "baseModel": model_id,
+        }
+        headers = {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json",
+        }
+
+        async with httpx.AsyncClient() as client:
+            response = await client.post(url, json=payload, headers=headers)
+
+        if response.status_code == 200:
+            basemodel = response.json().get("baseModel")
+            if basemodel is not None and isinstance(basemodel, str):
+                self.datamodel.fine_tune_model_id = basemodel
+                if self.datamodel.path:
+                    self.datamodel.save_to_file()
+                return True
+
+        logger.error(
+            f"Failed to deploy model to Fireworks server: [{response.status_code}] {response.text}"
+        )
+        return False
+
+    async def _fetch_all_deployments(self) -> List[dict]:
+        """
+        Fetch all deployments for an account.
+        """
+        api_key, account_id = self.api_key_and_account_id()
+
+        url = f"https://api.fireworks.ai/v1/accounts/{account_id}/deployments"
+
+        params = {
+            # Note: filter param does not work for baseModel, which would have been ideal, and ideally would have been documented. Instead we'll fetch all and filter.
+            # Max page size
+            "pageSize": 200,
+        }
+        headers = {
+            "Authorization": f"Bearer {api_key}",
+        }
+
+        deployments = []
+
+        # Paginate through all deployments
+        async with httpx.AsyncClient() as client:
+            while True:
+                response = await client.get(url, params=params, headers=headers)
+                json = response.json()
+                if "deployments" not in json or not isinstance(
+                    json["deployments"], list
+                ):
+                    raise ValueError(
+                        f"Invalid response from Fireworks. Expected list of deployments in 'deployments' key: [{response.status_code}] {response.text}"
+                    )
+                deployments.extend(json["deployments"])
+                next_page_token = json.get("nextPageToken")
+                if (
+                    next_page_token
+                    and isinstance(next_page_token, str)
+                    and len(next_page_token) > 0
+                ):
+                    params = {
+                        "pageSize": 200,
+                        "pageToken": next_page_token,
+                    }
+                else:
+                    break
+
+        return deployments

kiln_ai/adapters/fine_tune/test_base_finetune.py

@@ -98,6 +98,13 @@ def test_validate_parameters_valid():
     }
     MockFinetune.validate_parameters(valid_params)  # Should not raise
 
+    # Test valid parameters (float as int)
+    valid_params = {
+        "learning_rate": 1,
+        "epochs": 10,
+    }
+    MockFinetune.validate_parameters(valid_params)  # Should not raise
+
 
 def test_validate_parameters_missing_required():
     # Test missing required parameter
@@ -261,15 +268,6 @@ async def test_create_and_start_no_parent_task_path():
         )
 
 
-def test_check_valid_provider_model():
-    MockFinetune.check_valid_provider_model("openai", "gpt-4o-mini-2024-07-18")
-
-    with pytest.raises(
-        ValueError, match="Provider openai with base model gpt-99 is not available"
-    ):
-        MockFinetune.check_valid_provider_model("openai", "gpt-99")
-
-
 async def test_create_and_start_invalid_train_split(mock_dataset):
     # Test with an invalid train split name
     mock_dataset.split_contents = {"valid_train": [], "valid_test": []}

kiln_ai/adapters/fine_tune/test_dataset_formatter.py

@@ -15,7 +15,7 @@ from kiln_ai.adapters.fine_tune.dataset_formatter import (
     generate_chat_message_toolcall,
     generate_huggingface_chat_template,
     generate_huggingface_chat_template_toolcall,
-    generate_vertex_gemini_1_5,
+    generate_vertex_gemini,
 )
 from kiln_ai.adapters.model_adapters.base_adapter import COT_FINAL_ANSWER_PROMPT
 from kiln_ai.datamodel import (
@@ -447,7 +447,7 @@ def test_generate_vertex_template():
         final_output="test output",
     )
 
-    result = generate_vertex_gemini_1_5(training_data)
+    result = generate_vertex_gemini(training_data)
 
     assert result == {
         "systemInstruction": {
@@ -475,7 +475,7 @@ def test_generate_vertex_template_thinking():
         thinking_final_answer_prompt="thinking final answer prompt",
     )
 
-    result = generate_vertex_gemini_1_5(training_data)
+    result = generate_vertex_gemini(training_data)
 
     logger.info(result)