star

venv/lib/python3.9/site-packages/mlflow/prompt/constants.py

@@ -0,0 +1,17 @@
+# A special tag in RegisteredModel to indicate that it is a prompt
+import re
+
+IS_PROMPT_TAG_KEY = "mlflow.prompt.is_prompt"
+# A special tag in ModelVersion to store the prompt text
+PROMPT_TEXT_TAG_KEY = "mlflow.prompt.text"
+# TODO: Replace this with model_ids in MLflow 3
+PROMPT_ASSOCIATED_RUN_IDS_TAG_KEY = "mlflow.prompt.run_ids"
+
+PROMPT_TEMPLATE_VARIABLE_PATTERN = re.compile(
+    r"\{\{\s*([a-zA-Z_][a-zA-Z0-9_]*(?:\.[a-zA-Z_][a-zA-Z0-9_]*)*)\s*\}\}"
+)
+
+PROMPT_TEXT_DISPLAY_LIMIT = 30
+
+# Alphanumeric, underscore, hyphen, and dot are allowed in prompt name
+PROMPT_NAME_RULE = re.compile(r"^[a-zA-Z0-9_.-]+$")

venv/lib/python3.9/site-packages/mlflow/prompt/promptlab_model.py

@@ -0,0 +1,193 @@
+import os
+import re
+
+import yaml
+
+from mlflow.exceptions import MlflowException
+from mlflow.version import VERSION as __version__
+
+
+class _PromptlabModel:
+    import pandas as pd
+
+    def __init__(self, prompt_template, prompt_parameters, model_parameters, model_route):
+        self.prompt_parameters = prompt_parameters
+        self.model_parameters = model_parameters
+        self.model_route = model_route
+        self.prompt_template = prompt_template
+
+    def predict(self, inputs: pd.DataFrame) -> list[str]:
+        from mlflow.gateway import query
+
+        results = []
+        for idx in inputs.index:
+            prompt_parameters_as_dict = {
+                param.key: inputs[param.key][idx] for param in self.prompt_parameters
+            }
+
+            # copy replacement logic from PromptEngineering.utils.ts for consistency
+            prompt = self.prompt_template
+            for key, value in prompt_parameters_as_dict.items():
+                prompt = re.sub(r"\{\{\s*" + key + r"\s*\}\}", value, prompt)
+
+            model_parameters_as_dict = {param.key: param.value for param in self.model_parameters}
+            query_data = self._construct_query_data(prompt)
+
+            response = query(
+                route=self.model_route, data={**query_data, **model_parameters_as_dict}
+            )
+            results.append(self._parse_gateway_response(response))
+
+        return results
+
+    def _construct_query_data(self, prompt):
+        from mlflow.gateway import get_route
+
+        route_type = get_route(self.model_route).route_type
+
+        if route_type == "llm/v1/completions":
+            return {"prompt": prompt}
+        elif route_type == "llm/v1/chat":
+            return {"messages": [{"content": prompt, "role": "user"}]}
+        else:
+            raise MlflowException(
+                "Error when constructing gateway query: "
+                f"Unsupported route type for _PromptlabModel: {route_type}"
+            )
+
+    def _parse_gateway_response(self, response):
+        from mlflow.gateway import get_route
+
+        route_type = get_route(self.model_route).route_type
+
+        if route_type == "llm/v1/completions":
+            return response["choices"][0]["text"]
+        elif route_type == "llm/v1/chat":
+            return response["choices"][0]["message"]["content"]
+        else:
+            raise MlflowException(
+                "Error when parsing gateway response: "
+                f"Unsupported route type for _PromptlabModel: {route_type}"
+            )
+
+
+def _load_pyfunc(path):
+    from mlflow import pyfunc
+    from mlflow.entities.param import Param
+    from mlflow.utils.model_utils import (
+        _get_flavor_configuration,
+    )
+
+    pyfunc_flavor_conf = _get_flavor_configuration(model_path=path, flavor_name=pyfunc.FLAVOR_NAME)
+    parameters_path = os.path.join(path, pyfunc_flavor_conf["parameters_path"])
+    with open(parameters_path) as f:
+        parameters = yaml.safe_load(f)
+
+        prompt_parameters_as_params = [
+            Param(key=key, value=value) for key, value in parameters["prompt_parameters"].items()
+        ]
+        model_parameters_as_params = [
+            Param(key=key, value=value) for key, value in parameters["model_parameters"].items()
+        ]
+
+        return _PromptlabModel(
+            prompt_template=parameters["prompt_template"],
+            prompt_parameters=prompt_parameters_as_params,
+            model_parameters=model_parameters_as_params,
+            model_route=parameters["model_route"],
+        )
+
+
+def save_model(
+    path,
+    conda_env=None,
+    code_paths=None,
+    mlflow_model=None,
+    signature=None,
+    input_example=None,
+    pip_requirements=None,
+    prompt_template=None,
+    prompt_parameters=None,
+    model_parameters=None,
+    model_route=None,
+):
+    from mlflow import pyfunc
+    from mlflow.models import Model
+    from mlflow.models.model import MLMODEL_FILE_NAME, Model
+    from mlflow.models.utils import _save_example
+    from mlflow.utils.environment import (
+        _CONDA_ENV_FILE_NAME,
+        _CONSTRAINTS_FILE_NAME,
+        _PYTHON_ENV_FILE_NAME,
+        _REQUIREMENTS_FILE_NAME,
+        _process_conda_env,
+        _process_pip_requirements,
+        _PythonEnv,
+        _validate_env_arguments,
+        infer_pip_requirements,
+    )
+    from mlflow.utils.file_utils import write_to
+    from mlflow.utils.model_utils import (
+        _validate_and_copy_code_paths,
+        _validate_and_prepare_target_save_path,
+    )
+
+    _validate_env_arguments(conda_env, pip_requirements, None)
+
+    _validate_and_prepare_target_save_path(path)
+    code_dir_subpath = _validate_and_copy_code_paths(code_paths, path)
+
+    if mlflow_model is None:
+        mlflow_model = Model()
+    if signature is not None:
+        mlflow_model.signature = signature
+    if input_example is not None:
+        _save_example(mlflow_model, input_example, path)
+
+    parameters_sub_path = "parameters.yaml"
+    parameters_path = os.path.join(path, parameters_sub_path)
+    # dump prompt_template, prompt_parameters, model_parameters, model_route to parameters_path
+
+    parameters = {
+        "prompt_template": prompt_template,
+        "prompt_parameters": {param.key: param.value for param in prompt_parameters},
+        "model_parameters": {param.key: param.value for param in model_parameters},
+        "model_route": model_route,
+    }
+    with open(parameters_path, "w") as f:
+        yaml.safe_dump(parameters, stream=f, default_flow_style=False)
+
+    pyfunc.add_to_model(
+        mlflow_model,
+        loader_module="mlflow.prompt.promptlab_model",
+        parameters_path=parameters_sub_path,
+        conda_env=_CONDA_ENV_FILE_NAME,
+        python_env=_PYTHON_ENV_FILE_NAME,
+        code=code_dir_subpath,
+    )
+
+    mlflow_model.save(os.path.join(path, MLMODEL_FILE_NAME))
+
+    if conda_env is None:
+        if pip_requirements is None:
+            inferred_reqs = infer_pip_requirements(
+                path, "mlflow._promptlab", [f"mlflow[gateway]=={__version__}"]
+            )
+            default_reqs = sorted(inferred_reqs)
+        else:
+            default_reqs = None
+        conda_env, pip_requirements, pip_constraints = _process_pip_requirements(
+            default_reqs, pip_requirements, None
+        )
+    else:
+        conda_env, pip_requirements, pip_constraints = _process_conda_env(conda_env)
+
+    with open(os.path.join(path, _CONDA_ENV_FILE_NAME), "w") as f:
+        yaml.safe_dump(conda_env, stream=f, default_flow_style=False)
+
+    if pip_constraints:
+        write_to(os.path.join(path, _CONSTRAINTS_FILE_NAME), "\n".join(pip_constraints))
+
+    write_to(os.path.join(path, _REQUIREMENTS_FILE_NAME), "\n".join(pip_requirements))
+
+    _PythonEnv.current().to_yaml(os.path.join(path, _PYTHON_ENV_FILE_NAME))

venv/lib/python3.9/site-packages/mlflow/prompt/registry_utils.py

@@ -0,0 +1,146 @@
+import functools
+import re
+from textwrap import dedent
+from typing import Any, Optional, Union
+
+import mlflow
+from mlflow.entities.model_registry.registered_model_tag import RegisteredModelTag
+from mlflow.exceptions import MlflowException
+from mlflow.prompt.constants import IS_PROMPT_TAG_KEY, PROMPT_NAME_RULE
+from mlflow.protos.databricks_pb2 import RESOURCE_ALREADY_EXISTS
+
+
+def add_prompt_filter_string(
+    filter_string: Optional[str], is_prompt: bool = False
+) -> Optional[str]:
+    """
+    Additional filter string to include/exclude prompts from the result.
+    By default, exclude prompts from the result.
+    """
+    if IS_PROMPT_TAG_KEY not in (filter_string or ""):
+        prompt_filter_query = (
+            f"tag.`{IS_PROMPT_TAG_KEY}` = 'true'"
+            if is_prompt
+            else f"tag.`{IS_PROMPT_TAG_KEY}` != 'true'"
+        )
+        if filter_string:
+            filter_string = f"{filter_string} AND {prompt_filter_query}"
+        else:
+            filter_string = prompt_filter_query
+    return filter_string
+
+
+def has_prompt_tag(tags: Optional[Union[list[RegisteredModelTag], dict[str, str]]]) -> bool:
+    """Check if the given tags contain the prompt tag."""
+    if isinstance(tags, dict):
+        return IS_PROMPT_TAG_KEY in tags if tags else False
+    if not tags:
+        return
+    return any(tag.key == IS_PROMPT_TAG_KEY for tag in tags)
+
+
+def is_prompt_supported_registry(registry_uri: Optional[str] = None) -> bool:
+    """
+    Check if the current registry supports prompts.
+
+    Prompts registration is supported only in the OSS MLflow Tracking Server,
+    not in Databricks or OSS Unity Catalog.
+    """
+    registry_uri = registry_uri or mlflow.get_registry_uri()
+    return not registry_uri.startswith("databricks") and not registry_uri.startswith("uc:")
+
+
+def require_prompt_registry(func):
+    """Ensure that the current registry supports prompts."""
+
+    @functools.wraps(func)
+    def wrapper(*args, **kwargs):
+        if args and isinstance(args[0], mlflow.MlflowClient):
+            registry_uri = args[0]._registry_uri
+        else:
+            registry_uri = mlflow.get_registry_uri()
+
+        if not is_prompt_supported_registry(registry_uri):
+            raise MlflowException(
+                f"The '{func.__name__}' API is only available with the OSS MLflow Tracking Server."
+            )
+        return func(*args, **kwargs)
+
+    # Add note about prompt support to the docstring
+    func.__doc__ = dedent(f"""\
+        {func.__doc__}
+
+        .. note::
+
+            This API is supported only when using the OSS MLflow Model Registry. Prompts are not
+            supported in Databricks or the OSS Unity Catalog model registry.
+    """)
+    return wrapper
+
+
+def translate_prompt_exception(func):
+    """
+    Translate MlflowException message related to RegisteredModel / ModelVersion into
+    prompt-specific message.
+    """
+    MODEL_PATTERN = re.compile(r"(registered model|model version)", re.IGNORECASE)
+
+    @functools.wraps(func)
+    def wrapper(*args, **kwargs):
+        try:
+            return func(*args, **kwargs)
+        except MlflowException as e:
+            original_message = e.message
+            # Preserve the case of the first letter
+            new_message = MODEL_PATTERN.sub(
+                lambda m: "Prompt" if m.group(0)[0].isupper() else "prompt", e.message
+            )
+
+            if new_message != original_message:
+                raise MlflowException(new_message) from e
+            else:
+                raise e
+
+    return wrapper
+
+
+def validate_prompt_name(name: Any):
+    """Validate the prompt name against the prompt specific rule"""
+    if not isinstance(name, str) or not name:
+        raise MlflowException.invalid_parameter_value(
+            "Prompt name must be a non-empty string.",
+        )
+
+    if PROMPT_NAME_RULE.match(name) is None:
+        raise MlflowException.invalid_parameter_value(
+            "Prompt name can only contain alphanumeric characters, hyphens, underscores, and dots.",
+        )
+
+
+def handle_resource_already_exist_error(
+    name: str,
+    is_existing_entity_prompt: bool,
+    is_new_entity_prompt: bool,
+):
+    """
+    Show a more specific error message for name conflict in Model Registry.
+
+    1. When creating a model with the same name as an existing model, say "model already exists".
+    2. When creating a prompt with the same name as an existing prompt, say "prompt already exists".
+    3. Otherwise, explain that a prompt and a model cannot have the same name.
+    """
+    old_entity = "Prompt" if is_existing_entity_prompt else "Registered Model"
+    new_entity = "Prompt" if is_new_entity_prompt else "Registered Model"
+
+    if old_entity != new_entity:
+        raise MlflowException(
+            f"Tried to create a {new_entity.lower()} with name {name!r}, but the name is "
+            f"already taken by a {old_entity.lower()}. MLflow does not allow creating a "
+            "model and a prompt with the same name.",
+            RESOURCE_ALREADY_EXISTS,
+        )
+
+    raise MlflowException(
+        f"{new_entity} (name={name}) already exists.",
+        RESOURCE_ALREADY_EXISTS,
+    )