mirascope 2.0.0a1__py3-none-any.whl → 2.0.0a3__py3-none-any.whl

mirascope/llm/providers/mlx/provider.py

@@ -0,0 +1,411 @@
+from collections.abc import Sequence
+from functools import cache, lru_cache
+from typing import cast
+from typing_extensions import Unpack
+
+import mlx.nn as nn
+from mlx_lm import load as mlx_load
+from transformers import PreTrainedTokenizer
+
+from ...context import Context, DepsT
+from ...formatting import Format, FormattableT
+from ...messages import Message
+from ...responses import (
+    AsyncContextResponse,
+    AsyncContextStreamResponse,
+    AsyncResponse,
+    AsyncStreamResponse,
+    ContextResponse,
+    ContextStreamResponse,
+    Response,
+    StreamResponse,
+)
+from ...tools import (
+    AsyncContextTool,
+    AsyncContextToolkit,
+    AsyncTool,
+    AsyncToolkit,
+    ContextTool,
+    ContextToolkit,
+    Tool,
+    Toolkit,
+)
+from ..base import BaseProvider, Params
+from . import _utils
+from .encoding import TransformersEncoder
+from .mlx import MLX
+from .model_id import MLXModelId
+
+
+@cache
+def _mlx_client_singleton() -> "MLXProvider":
+    """Get or create the singleton MLX client instance."""
+    return MLXProvider()
+
+
+def client() -> "MLXProvider":
+    """Get the MLX client singleton instance."""
+    return _mlx_client_singleton()
+
+
+@lru_cache(maxsize=16)
+def _get_mlx(model_id: MLXModelId) -> MLX:
+    model, tokenizer = cast(tuple[nn.Module, PreTrainedTokenizer], mlx_load(model_id))
+    encoder = TransformersEncoder(tokenizer)
+    return MLX(
+        model_id,
+        model,
+        tokenizer,
+        encoder,
+    )
+
+
+class MLXProvider(BaseProvider[None]):
+    """Client for interacting with MLX language models.
+
+    This client provides methods for generating responses from MLX models,
+    supporting both synchronous and asynchronous operations, as well as
+    streaming responses.
+    """
+
+    id = "mlx"
+    default_scope = "mlx-community/"
+
+    def _call(
+        self,
+        *,
+        model_id: MLXModelId,
+        messages: Sequence[Message],
+        tools: Sequence[Tool] | Toolkit | None = None,
+        format: type[FormattableT] | Format[FormattableT] | None = None,
+        **params: Unpack[Params],
+    ) -> Response | Response[FormattableT]:
+        """Generate an `llm.Response` using MLX model.
+
+        Args:
+            model_id: Model identifier to use.
+            messages: Messages to send to the LLM.
+            tools: Optional tools that the model may invoke.
+            format: Optional response format specifier.
+            **params: Additional parameters to configure output (e.g. temperature). See `llm.Params`.
+
+        Returns:
+            An `llm.Response` object containing the LLM-generated content.
+        """
+        mlx = _get_mlx(model_id)
+
+        input_messages, format, assistant_message, response = mlx.generate(
+            messages, tools, format, params
+        )
+
+        return Response(
+            raw=response,
+            provider_id="mlx",
+            model_id=model_id,
+            provider_model_name=model_id,
+            params=params,
+            tools=tools,
+            input_messages=input_messages,
+            assistant_message=assistant_message,
+            finish_reason=_utils.extract_finish_reason(response),
+            format=format,
+        )
+
+    def _context_call(
+        self,
+        *,
+        ctx: Context[DepsT],
+        model_id: MLXModelId,
+        messages: Sequence[Message],
+        tools: Sequence[Tool | ContextTool[DepsT]]
+        | ContextToolkit[DepsT]
+        | None = None,
+        format: type[FormattableT] | Format[FormattableT] | None = None,
+        **params: Unpack[Params],
+    ) -> ContextResponse[DepsT, None] | ContextResponse[DepsT, FormattableT]:
+        """Generate an `llm.ContextResponse` using MLX model.
+
+        Args:
+            ctx: Context object with dependencies for tools.
+            model_id: Model identifier to use.
+            messages: Messages to send to the LLM.
+            tools: Optional tools that the model may invoke.
+            format: Optional response format specifier.
+            **params: Additional parameters to configure output (e.g. temperature). See `llm.Params`.
+
+        Returns:
+            An `llm.ContextResponse` object containing the LLM-generated content.
+        """
+        mlx = _get_mlx(model_id)
+
+        input_messages, format, assistant_message, response = mlx.generate(
+            messages, tools, format, params
+        )
+
+        return ContextResponse(
+            raw=response,
+            provider_id="mlx",
+            model_id=model_id,
+            provider_model_name=model_id,
+            params=params,
+            tools=tools,
+            input_messages=input_messages,
+            assistant_message=assistant_message,
+            finish_reason=_utils.extract_finish_reason(response),
+            format=format,
+        )
+
+    async def _call_async(
+        self,
+        *,
+        model_id: MLXModelId,
+        messages: Sequence[Message],
+        tools: Sequence[AsyncTool] | AsyncToolkit | None = None,
+        format: type[FormattableT] | Format[FormattableT] | None = None,
+        **params: Unpack[Params],
+    ) -> AsyncResponse | AsyncResponse[FormattableT]:
+        """Generate an `llm.AsyncResponse` using MLX model by asynchronously calloing
+        `asycio.to_thread`.
+
+        Args:
+            model_id: Model identifier to use.
+            messages: Messages to send to the LLM.
+            tools: Optional tools that the model may invoke.
+            format: Optional response format specifier.
+            **params: Additional parameters to configure output (e.g. temperature). See `llm.Params`.
+
+        Returns:
+            An `llm.AsyncResponse` object containing the LLM-generated content.
+        """
+        mlx = _get_mlx(model_id)
+
+        (
+            input_messages,
+            format,
+            assistant_message,
+            response,
+        ) = await mlx.generate_async(messages, tools, format, params)
+
+        return AsyncResponse(
+            raw=response,
+            provider_id="mlx",
+            model_id=model_id,
+            provider_model_name=model_id,
+            params=params,
+            tools=tools,
+            input_messages=input_messages,
+            assistant_message=assistant_message,
+            finish_reason=_utils.extract_finish_reason(response),
+            format=format,
+        )
+
+    async def _context_call_async(
+        self,
+        *,
+        ctx: Context[DepsT],
+        model_id: MLXModelId,
+        messages: Sequence[Message],
+        tools: Sequence[AsyncTool | AsyncContextTool[DepsT]]
+        | AsyncContextToolkit[DepsT]
+        | None = None,
+        format: type[FormattableT] | Format[FormattableT] | None = None,
+        **params: Unpack[Params],
+    ) -> AsyncContextResponse[DepsT, None] | AsyncContextResponse[DepsT, FormattableT]:
+        """Generate an `llm.AsyncResponse` using MLX model by asynchronously calloing
+        `asycio.to_thread`.
+
+        Args:
+            ctx: Context object with dependencies for tools.
+            model_id: Model identifier to use.
+            messages: Messages to send to the LLM.
+            tools: Optional tools that the model may invoke.
+            format: Optional response format specifier.
+            **params: Additional parameters to configure output (e.g. temperature). See `llm.Params`.
+
+        Returns:
+            An `llm.AsyncContextResponse` object containing the LLM-generated content.
+        """
+        mlx = _get_mlx(model_id)
+
+        (
+            input_messages,
+            format,
+            assistant_message,
+            response,
+        ) = await mlx.generate_async(messages, tools, format, params)
+
+        return AsyncContextResponse(
+            raw=response,
+            provider_id="mlx",
+            model_id=model_id,
+            provider_model_name=model_id,
+            params=params,
+            tools=tools,
+            input_messages=input_messages,
+            assistant_message=assistant_message,
+            finish_reason=_utils.extract_finish_reason(response),
+            format=format,
+        )
+
+    def _stream(
+        self,
+        *,
+        model_id: MLXModelId,
+        messages: Sequence[Message],
+        tools: Sequence[Tool] | Toolkit | None = None,
+        format: type[FormattableT] | Format[FormattableT] | None = None,
+        **params: Unpack[Params],
+    ) -> StreamResponse | StreamResponse[FormattableT]:
+        """Generate an `llm.StreamResponse` by synchronously streaming from MLX model output.
+
+        Args:
+            model_id: Model identifier to use.
+            messages: Messages to send to the LLM.
+            tools: Optional tools that the model may invoke.
+            format: Optional response format specifier.
+            **params: Additional parameters to configure output (e.g. temperature). See `llm.Params`.
+
+        Returns:
+            An `llm.StreamResponse` object for iterating over the LLM-generated content.
+        """
+        mlx = _get_mlx(model_id)
+
+        input_messages, format, chunk_iterator = mlx.stream(
+            messages, tools, format, params
+        )
+
+        return StreamResponse(
+            provider_id="mlx",
+            model_id=model_id,
+            provider_model_name=model_id,
+            params=params,
+            tools=tools,
+            input_messages=input_messages,
+            chunk_iterator=chunk_iterator,
+            format=format,
+        )
+
+    def _context_stream(
+        self,
+        *,
+        ctx: Context[DepsT],
+        model_id: MLXModelId,
+        messages: Sequence[Message],
+        tools: Sequence[Tool | ContextTool[DepsT]]
+        | ContextToolkit[DepsT]
+        | None = None,
+        format: type[FormattableT] | Format[FormattableT] | None = None,
+        **params: Unpack[Params],
+    ) -> ContextStreamResponse[DepsT] | ContextStreamResponse[DepsT, FormattableT]:
+        """Generate an `llm.ContextStreamResponse` by synchronously streaming from MLX model output.
+
+        Args:
+            ctx: Context object with dependencies for tools.
+            model_id: Model identifier to use.
+            messages: Messages to send to the LLM.
+            tools: Optional tools that the model may invoke.
+            format: Optional response format specifier.
+            **params: Additional parameters to configure output (e.g. temperature). See `llm.Params`.
+
+        Returns:
+            An `llm.ContextStreamResponse` object for iterating over the LLM-generated content.
+        """
+        mlx = _get_mlx(model_id)
+
+        input_messages, format, chunk_iterator = mlx.stream(
+            messages, tools, format, params
+        )
+
+        return ContextStreamResponse(
+            provider_id="mlx",
+            model_id=model_id,
+            provider_model_name=model_id,
+            params=params,
+            tools=tools,
+            input_messages=input_messages,
+            chunk_iterator=chunk_iterator,
+            format=format,
+        )
+
+    async def _stream_async(
+        self,
+        *,
+        model_id: MLXModelId,
+        messages: Sequence[Message],
+        tools: Sequence[AsyncTool] | AsyncToolkit | None = None,
+        format: type[FormattableT] | Format[FormattableT] | None = None,
+        **params: Unpack[Params],
+    ) -> AsyncStreamResponse | AsyncStreamResponse[FormattableT]:
+        """Generate an `llm.AsyncStreamResponse` by asynchronously streaming from MLX model output.
+
+        Args:
+            model_id: Model identifier to use.
+            messages: Messages to send to the LLM.
+            tools: Optional tools that the model may invoke.
+            format: Optional response format specifier.
+            **params: Additional parameters to configure output (e.g. temperature). See `llm.Params`.
+
+        Returns:
+            An `llm.AsyncStreamResponse` object for asynchronously iterating over the LLM-generated content.
+        """
+        mlx = _get_mlx(model_id)
+
+        input_messages, format, chunk_iterator = await mlx.stream_async(
+            messages, tools, format, params
+        )
+
+        return AsyncStreamResponse(
+            provider_id="mlx",
+            model_id=model_id,
+            provider_model_name=model_id,
+            params=params,
+            tools=tools,
+            input_messages=input_messages,
+            chunk_iterator=chunk_iterator,
+            format=format,
+        )
+
+    async def _context_stream_async(
+        self,
+        *,
+        ctx: Context[DepsT],
+        model_id: MLXModelId,
+        messages: Sequence[Message],
+        tools: Sequence[AsyncTool | AsyncContextTool[DepsT]]
+        | AsyncContextToolkit[DepsT]
+        | None = None,
+        format: type[FormattableT] | Format[FormattableT] | None = None,
+        **params: Unpack[Params],
+    ) -> (
+        AsyncContextStreamResponse[DepsT]
+        | AsyncContextStreamResponse[DepsT, FormattableT]
+    ):
+        """Generate an `llm.AsyncContextStreamResponse` by asynchronously streaming from MLX model output.
+
+        Args:
+            ctx: Context object with dependencies for tools.
+            model_id: Model identifier to use.
+            messages: Messages to send to the LLM.
+            tools: Optional tools that the model may invoke.
+            format: Optional response format specifier.
+            **params: Additional parameters to configure output (e.g. temperature). See `llm.Params`.
+
+        Returns:
+            An `llm.AsyncContextStreamResponse` object for asynchronously iterating over the LLM-generated content.
+        """
+        mlx = _get_mlx(model_id)
+
+        input_messages, format, chunk_iterator = await mlx.stream_async(
+            messages, tools, format, params
+        )
+
+        return AsyncContextStreamResponse(
+            provider_id="mlx",
+            model_id=model_id,
+            provider_model_name=model_id,
+            params=params,
+            tools=tools,
+            input_messages=input_messages,
+            chunk_iterator=chunk_iterator,
+            format=format,
+        )

mirascope/llm/providers/model_id.py

@@ -0,0 +1,16 @@
+from typing import TypeAlias
+
+from .anthropic import (
+    AnthropicModelId,
+)
+from .google import (
+    GoogleModelId,
+)
+from .mlx import (
+    MLXModelId,
+)
+from .openai import (
+    OpenAIModelId,
+)
+
+ModelId: TypeAlias = AnthropicModelId | GoogleModelId | OpenAIModelId | MLXModelId | str

mirascope/llm/providers/openai/__init__.py

@@ -0,0 +1,6 @@
+"""OpenAI client implementation."""
+
+from .model_id import OpenAIModelId
+from .provider import OpenAIProvider
+
+__all__ = ["OpenAIModelId", "OpenAIProvider"]

mirascope/llm/providers/openai/completions/__init__.py

@@ -0,0 +1,20 @@
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from .provider import OpenAICompletionsProvider
+else:
+    try:
+        from .provider import OpenAICompletionsProvider
+    except ImportError:  # pragma: no cover
+        from ..._missing_import_stubs import (
+            create_import_error_stub,
+            create_provider_stub,
+        )
+
+        OpenAICompletionsProvider = create_provider_stub(
+            "openai", "OpenAICompletionsProvider"
+        )
+
+__all__ = [
+    "OpenAICompletionsProvider",
+]

mirascope/llm/{clients/openai/responses → providers/openai/completions}/_utils/__init__.py

@@ -2,6 +2,7 @@ from .decode import (
     decode_async_stream,
     decode_response,
     decode_stream,
+    model_name,
 )
 from .encode import encode_request
 
@@ -10,4 +11,5 @@ __all__ = [
     "decode_response",
     "decode_stream",
     "encode_request",
+    "model_name",
 ]

mirascope/llm/{clients → providers}/openai/completions/_utils/decode.py

@@ -24,7 +24,7 @@ from .....responses import (
     FinishReasonChunk,
     RawStreamEventChunk,
 )
-from ..model_ids import OpenAICompletionsModelId
+from ...model_id import OpenAIModelId, model_name
 
 OPENAI_FINISH_REASON_MAP = {
     "length": FinishReason.MAX_TOKENS,
@@ -34,7 +34,8 @@ OPENAI_FINISH_REASON_MAP = {
 
 def decode_response(
     response: openai_types.ChatCompletion,
-    model_id: OpenAICompletionsModelId,
+    model_id: OpenAIModelId,
+    provider_id: Literal["openai", "openai:completions"],
 ) -> tuple[AssistantMessage, FinishReason | None]:
     """Convert OpenAI ChatCompletion to mirascope AssistantMessage."""
     choice = response.choices[0]
@@ -69,8 +70,9 @@ def decode_response(
 
     assistant_message = AssistantMessage(
         content=parts,
-        provider="openai:completions",
+        provider_id=provider_id,
         model_id=model_id,
+        provider_model_name=model_name(model_id, "completions"),
         raw_message=message.model_dump(exclude_none=True),
     )
 

mirascope/llm/{clients → providers}/openai/completions/_utils/encode.py

@@ -19,11 +19,11 @@ from .....formatting import (
     resolve_format,
 )
 from .....messages import AssistantMessage, Message, UserMessage
-from .....tools import FORMAT_TOOL_NAME, BaseToolkit, ToolSchema
+from .....tools import FORMAT_TOOL_NAME, AnyToolSchema, BaseToolkit
 from ....base import Params, _utils as _base_utils
+from ...model_id import OpenAIModelId, model_name
+from ...model_info import MODELS_WITHOUT_AUDIO_SUPPORT
 from ...shared import _utils as _shared_utils
-from ..model_ids import OpenAICompletionsModelId
-from .model_features import MODEL_FEATURES
 
 
 class ChatCompletionCreateKwargs(TypedDict, total=False):
@@ -49,7 +49,7 @@ class ChatCompletionCreateKwargs(TypedDict, total=False):
 
 def _encode_user_message(
     message: UserMessage,
-    model_id: OpenAICompletionsModelId,
+    model_id: OpenAIModelId,
 ) -> list[openai_types.ChatCompletionMessageParam]:
     """Convert Mirascope `UserMessage` to a list of OpenAI `ChatCompletionMessageParam`.
 
@@ -98,11 +98,11 @@ def _encode_user_message(
             )
             current_content.append(content)
         elif part.type == "audio":
-            model_status = MODEL_FEATURES.get(model_id)
-            if model_status == "no_audio_support":
+            base_model_name = model_name(model_id, None)
+            if base_model_name in MODELS_WITHOUT_AUDIO_SUPPORT:
                 raise FeatureNotSupportedError(
                     feature="Audio inputs",
-                    provider="openai:completions",
+                    provider_id="openai",
                     message=f"Model '{model_id}' does not support audio inputs.",
                 )
 
@@ -111,7 +111,7 @@ def _encode_user_message(
                 if audio_format not in ("wav", "mp3"):
                     raise FeatureNotSupportedError(
                         feature=f"Audio format: {audio_format}",
-                        provider="openai:completions",
+                        provider_id="openai",
                         message="OpenAI only supports 'wav' and 'mp3' audio formats.",
                     )  # pragma: no cover
                 audio_content = openai_types.ChatCompletionContentPartInputAudioParam(
@@ -141,13 +141,14 @@ def _encode_user_message(
 
 
 def _encode_assistant_message(
-    message: AssistantMessage, model_id: OpenAICompletionsModelId, encode_thoughts: bool
+    message: AssistantMessage, model_id: OpenAIModelId, encode_thoughts: bool
 ) -> openai_types.ChatCompletionAssistantMessageParam:
     """Convert Mirascope `AssistantMessage` to OpenAI `ChatCompletionAssistantMessageParam`."""
 
     if (
-        message.provider == "openai:completions"
-        and message.model_id == model_id
+        message.provider_id in ("openai", "openai:completions")
+        and message.provider_model_name
+        == model_name(model_id=model_id, api_mode="completions")
         and message.raw_message
         and not encode_thoughts
     ):
@@ -188,7 +189,7 @@ def _encode_assistant_message(
     elif text_params:
         content = text_params
 
-    message_params = {
+    message_params: openai_types.ChatCompletionAssistantMessageParam = {
         "role": "assistant",
         "content": content,
     }
@@ -199,7 +200,7 @@ def _encode_assistant_message(
 
 
 def _encode_message(
-    message: Message, model_id: OpenAICompletionsModelId, encode_thoughts: bool
+    message: Message, model_id: OpenAIModelId, encode_thoughts: bool
 ) -> list[openai_types.ChatCompletionMessageParam]:
     """Convert a Mirascope `Message` to OpenAI `ChatCompletionMessageParam` format.
 
@@ -227,12 +228,12 @@ def _encode_message(
 
 @lru_cache(maxsize=128)
 def _convert_tool_to_tool_param(
-    tool: ToolSchema,
+    tool: AnyToolSchema,
 ) -> openai_types.ChatCompletionToolParam:
     """Convert a single Mirascope `Tool` to OpenAI ChatCompletionToolParam with caching."""
     schema_dict = tool.parameters.model_dump(by_alias=True, exclude_none=True)
     schema_dict["type"] = "object"
-    _shared_utils._ensure_additional_properties_false(schema_dict)
+    _shared_utils.ensure_additional_properties_false(schema_dict)
     return openai_types.ChatCompletionToolParam(
         type="function",
         function={
@@ -257,7 +258,7 @@ def _create_strict_response_format(
     """
     schema = format.schema.copy()
 
-    _shared_utils._ensure_additional_properties_false(schema)
+    _shared_utils.ensure_additional_properties_false(schema)
 
     json_schema = JSONSchema(
         name=format.name,
@@ -274,23 +275,32 @@ def _create_strict_response_format(
 
 def encode_request(
     *,
-    model_id: OpenAICompletionsModelId,
+    model_id: OpenAIModelId,
     messages: Sequence[Message],
-    tools: Sequence[ToolSchema] | BaseToolkit | None,
+    tools: Sequence[AnyToolSchema] | BaseToolkit[AnyToolSchema] | None,
     format: type[FormattableT] | Format[FormattableT] | None,
     params: Params,
 ) -> tuple[Sequence[Message], Format[FormattableT] | None, ChatCompletionCreateKwargs]:
     """Prepares a request for the `OpenAI.chat.completions.create` method."""
+    if model_id.endswith(":responses"):
+        raise FeatureNotSupportedError(
+            feature="responses API",
+            provider_id="openai:completions",
+            model_id=model_id,
+            message=f"Can't use completions client for responses model: {model_id}",
+        )
+    base_model_name = model_name(model_id, None)
+
     kwargs: ChatCompletionCreateKwargs = ChatCompletionCreateKwargs(
         {
-            "model": model_id,
+            "model": base_model_name,
         }
     )
     encode_thoughts = False
 
     with _base_utils.ensure_all_params_accessed(
         params=params,
-        provider="openai:completions",
+        provider_id="openai",
         unsupported_params=["top_k", "thinking"],
     ) as param_accessor:
         if param_accessor.temperature is not None:
@@ -312,7 +322,7 @@ def encode_request(
     openai_tools = [_convert_tool_to_tool_param(tool) for tool in tools]
 
     model_supports_strict = (
-        model_id not in _shared_utils.MODELS_WITHOUT_JSON_SCHEMA_SUPPORT
+        base_model_name not in _shared_utils.MODELS_WITHOUT_JSON_SCHEMA_SUPPORT
     )
     default_mode = "strict" if model_supports_strict else "tool"
     format = resolve_format(format, default_mode=default_mode)
@@ -321,7 +331,7 @@ def encode_request(
             if not model_supports_strict:
                 raise FormattingModeNotSupportedError(
                     formatting_mode="strict",
-                    provider="openai:completions",
+                    provider_id="openai",
                     model_id=model_id,
                 )
             kwargs["response_format"] = _create_strict_response_format(format)
@@ -338,7 +348,7 @@ def encode_request(
             openai_tools.append(_convert_tool_to_tool_param(format_tool_schema))
         elif (
             format.mode == "json"
-            and model_id not in _shared_utils.MODELS_WITHOUT_JSON_OBJECT_SUPPORT
+            and base_model_name not in _shared_utils.MODELS_WITHOUT_JSON_OBJECT_SUPPORT
         ):
             kwargs["response_format"] = {"type": "json_object"}