huggingface-hub 0.31.3__py3-none-any.whl → 0.32.0__py3-none-any.whl

huggingface_hub/inference/_mcp/mcp_client.py

@@ -0,0 +1,322 @@
+import json
+import logging
+from contextlib import AsyncExitStack
+from datetime import timedelta
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, AsyncIterable, Dict, List, Literal, Optional, Union, overload
+
+from typing_extensions import NotRequired, TypeAlias, TypedDict, Unpack
+
+from ...utils._runtime import get_hf_hub_version
+from .._generated._async_client import AsyncInferenceClient
+from .._generated.types import (
+    ChatCompletionInputMessage,
+    ChatCompletionInputTool,
+    ChatCompletionStreamOutput,
+    ChatCompletionStreamOutputDeltaToolCall,
+)
+from .._providers import PROVIDER_OR_POLICY_T
+from .utils import format_result
+
+
+if TYPE_CHECKING:
+    from mcp import ClientSession
+
+logger = logging.getLogger(__name__)
+
+# Type alias for tool names
+ToolName: TypeAlias = str
+
+ServerType: TypeAlias = Literal["stdio", "sse", "http"]
+
+
+class StdioServerParameters_T(TypedDict):
+    command: str
+    args: NotRequired[List[str]]
+    env: NotRequired[Dict[str, str]]
+    cwd: NotRequired[Union[str, Path, None]]
+
+
+class SSEServerParameters_T(TypedDict):
+    url: str
+    headers: NotRequired[Dict[str, Any]]
+    timeout: NotRequired[float]
+    sse_read_timeout: NotRequired[float]
+
+
+class StreamableHTTPParameters_T(TypedDict):
+    url: str
+    headers: NotRequired[dict[str, Any]]
+    timeout: NotRequired[timedelta]
+    sse_read_timeout: NotRequired[timedelta]
+    terminate_on_close: NotRequired[bool]
+
+
+class MCPClient:
+    """
+    Client for connecting to one or more MCP servers and processing chat completions with tools.
+
+    <Tip warning={true}>
+
+    This class is experimental and might be subject to breaking changes in the future without prior notice.
+
+    </Tip>
+
+    Args:
+        model (`str`, `optional`):
+            The model to run inference with. Can be a model id hosted on the Hugging Face Hub, e.g. `meta-llama/Meta-Llama-3-8B-Instruct`
+            or a URL to a deployed Inference Endpoint or other local or remote endpoint.
+        provider (`str`, *optional*):
+            Name of the provider to use for inference. Defaults to "auto" i.e. the first of the providers available for the model, sorted by the user's order in https://hf.co/settings/inference-providers.
+            If model is a URL or `base_url` is passed, then `provider` is not used.
+        api_key (`str`, `optional`):
+            Token to use for authentication. Will default to the locally Hugging Face saved token if not provided. You can also use your own provider API key to interact directly with the provider's service.
+    """
+
+    def __init__(
+        self,
+        *,
+        model: str,
+        provider: Optional[PROVIDER_OR_POLICY_T] = None,
+        api_key: Optional[str] = None,
+    ):
+        # Initialize MCP sessions as a dictionary of ClientSession objects
+        self.sessions: Dict[ToolName, "ClientSession"] = {}
+        self.exit_stack = AsyncExitStack()
+        self.available_tools: List[ChatCompletionInputTool] = []
+
+        # Initialize the AsyncInferenceClient
+        self.client = AsyncInferenceClient(model=model, provider=provider, api_key=api_key)
+
+    async def __aenter__(self):
+        """Enter the context manager"""
+        await self.client.__aenter__()
+        await self.exit_stack.__aenter__()
+        return self
+
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        """Exit the context manager"""
+        await self.client.__aexit__(exc_type, exc_val, exc_tb)
+        await self.cleanup()
+
+    @overload
+    async def add_mcp_server(self, type: Literal["stdio"], **params: Unpack[StdioServerParameters_T]): ...
+
+    @overload
+    async def add_mcp_server(self, type: Literal["sse"], **params: Unpack[SSEServerParameters_T]): ...
+
+    @overload
+    async def add_mcp_server(self, type: Literal["http"], **params: Unpack[StreamableHTTPParameters_T]): ...
+
+    async def add_mcp_server(self, type: ServerType, **params: Any):
+        """Connect to an MCP server
+
+        Args:
+            type (`str`):
+                Type of the server to connect to. Can be one of:
+                - "stdio": Standard input/output server (local)
+                - "sse": Server-sent events (SSE) server
+                - "http": StreamableHTTP server
+            **params (`Dict[str, Any]`):
+                Server parameters that can be either:
+                    - For stdio servers:
+                        - command (str): The command to run the MCP server
+                        - args (List[str], optional): Arguments for the command
+                        - env (Dict[str, str], optional): Environment variables for the command
+                        - cwd (Union[str, Path, None], optional): Working directory for the command
+                    - For SSE servers:
+                        - url (str): The URL of the SSE server
+                        - headers (Dict[str, Any], optional): Headers for the SSE connection
+                        - timeout (float, optional): Connection timeout
+                        - sse_read_timeout (float, optional): SSE read timeout
+                    - For StreamableHTTP servers:
+                        - url (str): The URL of the StreamableHTTP server
+                        - headers (Dict[str, Any], optional): Headers for the StreamableHTTP connection
+                        - timeout (timedelta, optional): Connection timeout
+                        - sse_read_timeout (timedelta, optional): SSE read timeout
+                        - terminate_on_close (bool, optional): Whether to terminate on close
+        """
+        from mcp import ClientSession, StdioServerParameters
+        from mcp import types as mcp_types
+
+        # Determine server type and create appropriate parameters
+        if type == "stdio":
+            # Handle stdio server
+            from mcp.client.stdio import stdio_client
+
+            logger.info(f"Connecting to stdio MCP server with command: {params['command']} {params.get('args', [])}")
+
+            client_kwargs = {"command": params["command"]}
+            for key in ["args", "env", "cwd"]:
+                if params.get(key) is not None:
+                    client_kwargs[key] = params[key]
+            server_params = StdioServerParameters(**client_kwargs)
+            read, write = await self.exit_stack.enter_async_context(stdio_client(server_params))
+        elif type == "sse":
+            # Handle SSE server
+            from mcp.client.sse import sse_client
+
+            logger.info(f"Connecting to SSE MCP server at: {params['url']}")
+
+            client_kwargs = {"url": params["url"]}
+            for key in ["headers", "timeout", "sse_read_timeout"]:
+                if params.get(key) is not None:
+                    client_kwargs[key] = params[key]
+            read, write = await self.exit_stack.enter_async_context(sse_client(**client_kwargs))
+        elif type == "http":
+            # Handle StreamableHTTP server
+            from mcp.client.streamable_http import streamablehttp_client
+
+            logger.info(f"Connecting to StreamableHTTP MCP server at: {params['url']}")
+
+            client_kwargs = {"url": params["url"]}
+            for key in ["headers", "timeout", "sse_read_timeout", "terminate_on_close"]:
+                if params.get(key) is not None:
+                    client_kwargs[key] = params[key]
+            read, write, _ = await self.exit_stack.enter_async_context(streamablehttp_client(**client_kwargs))
+            # ^ TODO: should be handle `get_session_id_callback`? (function to retrieve the current session ID)
+        else:
+            raise ValueError(f"Unsupported server type: {type}")
+
+        session = await self.exit_stack.enter_async_context(
+            ClientSession(
+                read_stream=read,
+                write_stream=write,
+                client_info=mcp_types.Implementation(
+                    name="huggingface_hub.MCPClient",
+                    version=get_hf_hub_version(),
+                ),
+            )
+        )
+
+        logger.debug("Initializing session...")
+        await session.initialize()
+
+        # List available tools
+        response = await session.list_tools()
+        logger.debug("Connected to server with tools:", [tool.name for tool in response.tools])
+
+        for tool in response.tools:
+            if tool.name in self.sessions:
+                logger.warning(f"Tool '{tool.name}' already defined by another server. Skipping.")
+                continue
+
+            # Map tool names to their server for later lookup
+            self.sessions[tool.name] = session
+
+            # Add tool to the list of available tools (for use in chat completions)
+            self.available_tools.append(
+                ChatCompletionInputTool.parse_obj_as_instance(
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": tool.name,
+                            "description": tool.description,
+                            "parameters": tool.inputSchema,
+                        },
+                    }
+                )
+            )
+
+    async def process_single_turn_with_tools(
+        self,
+        messages: List[Union[Dict, ChatCompletionInputMessage]],
+        exit_loop_tools: Optional[List[ChatCompletionInputTool]] = None,
+        exit_if_first_chunk_no_tool: bool = False,
+    ) -> AsyncIterable[Union[ChatCompletionStreamOutput, ChatCompletionInputMessage]]:
+        """Process a query using `self.model` and available tools, yielding chunks and tool outputs.
+
+        Args:
+            messages (`List[Dict]`):
+                List of message objects representing the conversation history
+            exit_loop_tools (`List[ChatCompletionInputTool]`, *optional*):
+                List of tools that should exit the generator when called
+            exit_if_first_chunk_no_tool (`bool`, *optional*):
+                Exit if no tool is present in the first chunks. Default to False.
+
+        Yields:
+            [`ChatCompletionStreamOutput`] chunks or [`ChatCompletionInputMessage`] objects
+        """
+        # Prepare tools list based on options
+        tools = self.available_tools
+        if exit_loop_tools is not None:
+            tools = [*exit_loop_tools, *self.available_tools]
+
+        # Create the streaming request
+        response = await self.client.chat.completions.create(
+            messages=messages,
+            tools=tools,
+            tool_choice="auto",
+            stream=True,
+        )
+
+        message = {"role": "unknown", "content": ""}
+        final_tool_calls: Dict[int, ChatCompletionStreamOutputDeltaToolCall] = {}
+        num_of_chunks = 0
+
+        # Read from stream
+        async for chunk in response:
+            # Yield each chunk to caller
+            yield chunk
+
+            num_of_chunks += 1
+            delta = chunk.choices[0].delta if chunk.choices and len(chunk.choices) > 0 else None
+            if not delta:
+                continue
+
+            # Process message
+            if delta.role:
+                message["role"] = delta.role
+            if delta.content:
+                message["content"] += delta.content
+
+            # Process tool calls
+            if delta.tool_calls:
+                for tool_call in delta.tool_calls:
+                    # Aggregate chunks into tool calls
+                    if tool_call.index not in final_tool_calls:
+                        if tool_call.function.arguments is None:  # Corner case (depends on provider)
+                            tool_call.function.arguments = ""
+                        final_tool_calls[tool_call.index] = tool_call
+
+                    if tool_call.function.arguments:
+                        final_tool_calls[tool_call.index].function.arguments += tool_call.function.arguments
+
+            # Optionally exit early if no tools in first chunks
+            if exit_if_first_chunk_no_tool and num_of_chunks <= 2 and len(final_tool_calls) == 0:
+                return
+
+        if message["content"]:
+            messages.append(message)
+
+        # Process tool calls one by one
+        for tool_call in final_tool_calls.values():
+            function_name = tool_call.function.name
+            function_args = json.loads(tool_call.function.arguments or "{}")
+
+            tool_message = {"role": "tool", "tool_call_id": tool_call.id, "content": "", "name": function_name}
+
+            # Check if this is an exit loop tool
+            if exit_loop_tools and function_name in [t.function.name for t in exit_loop_tools]:
+                tool_message_as_obj = ChatCompletionInputMessage.parse_obj_as_instance(tool_message)
+                messages.append(tool_message_as_obj)
+                yield tool_message_as_obj
+                return
+
+            # Execute tool call with the appropriate session
+            session = self.sessions.get(function_name)
+            if session is not None:
+                result = await session.call_tool(function_name, function_args)
+                tool_message["content"] = format_result(result)
+            else:
+                error_msg = f"Error: No session found for tool: {function_name}"
+                tool_message["content"] = error_msg
+
+            # Yield tool message
+            tool_message_as_obj = ChatCompletionInputMessage.parse_obj_as_instance(tool_message)
+            messages.append(tool_message_as_obj)
+            yield tool_message_as_obj
+
+    async def cleanup(self):
+        """Clean up resources"""
+        await self.exit_stack.aclose()

huggingface_hub/inference/_mcp/utils.py

@@ -0,0 +1,123 @@
+"""
+Utility functions for MCPClient and Tiny Agents.
+
+Formatting utilities taken from the JS SDK: https://github.com/huggingface/huggingface.js/blob/main/packages/mcp-client/src/ResultFormatter.ts.
+"""
+
+import json
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
+
+from huggingface_hub import snapshot_download
+from huggingface_hub.errors import EntryNotFoundError
+
+from .constants import DEFAULT_AGENT, DEFAULT_REPO_ID, FILENAME_CONFIG, FILENAME_PROMPT
+
+
+if TYPE_CHECKING:
+    from mcp import types as mcp_types
+
+
+def format_result(result: "mcp_types.CallToolResult") -> str:
+    """
+    Formats a mcp.types.CallToolResult content into a human-readable string.
+
+    Args:
+        result (CallToolResult)
+            Object returned by mcp.ClientSession.call_tool.
+
+    Returns:
+        str
+            A formatted string representing the content of the result.
+    """
+    content = result.content
+
+    if len(content) == 0:
+        return "[No content]"
+
+    formatted_parts: List[str] = []
+
+    for item in content:
+        if item.type == "text":
+            formatted_parts.append(item.text)
+
+        elif item.type == "image":
+            formatted_parts.append(
+                f"[Binary Content: Image {item.mimeType}, {_get_base64_size(item.data)} bytes]\n"
+                f"The task is complete and the content accessible to the User"
+            )
+
+        elif item.type == "audio":
+            formatted_parts.append(
+                f"[Binary Content: Audio {item.mimeType}, {_get_base64_size(item.data)} bytes]\n"
+                f"The task is complete and the content accessible to the User"
+            )
+
+        elif item.type == "resource":
+            resource = item.resource
+
+            if hasattr(resource, "text"):
+                formatted_parts.append(resource.text)
+
+            elif hasattr(resource, "blob"):
+                formatted_parts.append(
+                    f"[Binary Content ({resource.uri}): {resource.mimeType}, {_get_base64_size(resource.blob)} bytes]\n"
+                    f"The task is complete and the content accessible to the User"
+                )
+
+    return "\n".join(formatted_parts)
+
+
+def _get_base64_size(base64_str: str) -> int:
+    """Estimate the byte size of a base64-encoded string."""
+    # Remove any prefix like "data:image/png;base64,"
+    if "," in base64_str:
+        base64_str = base64_str.split(",")[1]
+
+    padding = 0
+    if base64_str.endswith("=="):
+        padding = 2
+    elif base64_str.endswith("="):
+        padding = 1
+
+    return (len(base64_str) * 3) // 4 - padding
+
+
+def _load_agent_config(agent_path: Optional[str]) -> Tuple[Dict[str, Any], Optional[str]]:
+    """Load server config and prompt."""
+
+    def _read_dir(directory: Path) -> Tuple[Dict[str, Any], Optional[str]]:
+        cfg_file = directory / FILENAME_CONFIG
+        if not cfg_file.exists():
+            raise FileNotFoundError(f" Config file not found in {directory}! Please make sure it exists locally")
+
+        config: Dict[str, Any] = json.loads(cfg_file.read_text(encoding="utf-8"))
+        prompt_file = directory / FILENAME_PROMPT
+        prompt: Optional[str] = prompt_file.read_text(encoding="utf-8") if prompt_file.exists() else None
+        return config, prompt
+
+    if agent_path is None:
+        return DEFAULT_AGENT, None
+
+    path = Path(agent_path).expanduser()
+
+    if path.is_file():
+        return json.loads(path.read_text(encoding="utf-8")), None
+
+    if path.is_dir():
+        return _read_dir(path)
+
+    # fetch from the Hub
+    try:
+        repo_dir = Path(
+            snapshot_download(
+                repo_id=DEFAULT_REPO_ID,
+                allow_patterns=f"{agent_path}/*",
+                repo_type="dataset",
+            )
+        )
+        return _read_dir(repo_dir / agent_path)
+    except Exception as err:
+        raise EntryNotFoundError(
+            f" Agent {agent_path} not found in tiny-agents/tiny-agents! Please make sure it exists in https://huggingface.co/datasets/tiny-agents/tiny-agents."
+        ) from err

huggingface_hub/inference/_providers/__init__.py

@@ -20,8 +20,14 @@ from .hf_inference import (
     HFInferenceTask,
 )
 from .hyperbolic import HyperbolicTextGenerationTask, HyperbolicTextToImageTask
-from .nebius import NebiusConversationalTask, NebiusTextGenerationTask, NebiusTextToImageTask
+from .nebius import (
+    NebiusConversationalTask,
+    NebiusFeatureExtractionTask,
+    NebiusTextGenerationTask,
+    NebiusTextToImageTask,
+)
 from .novita import NovitaConversationalTask, NovitaTextGenerationTask, NovitaTextToVideoTask
+from .nscale import NscaleConversationalTask, NscaleTextToImageTask
 from .openai import OpenAIConversationalTask
 from .replicate import ReplicateTask, ReplicateTextToImageTask, ReplicateTextToSpeechTask
 from .sambanova import SambanovaConversationalTask, SambanovaFeatureExtractionTask
@@ -41,6 +47,7 @@ PROVIDER_T = Literal[
     "hyperbolic",
     "nebius",
     "novita",
+    "nscale",
     "openai",
     "replicate",
     "sambanova",
@@ -105,12 +112,17 @@ PROVIDERS: Dict[PROVIDER_T, Dict[str, TaskProviderHelper]] = {
         "text-to-image": NebiusTextToImageTask(),
         "conversational": NebiusConversationalTask(),
         "text-generation": NebiusTextGenerationTask(),
+        "feature-extraction": NebiusFeatureExtractionTask(),
     },
     "novita": {
         "text-generation": NovitaTextGenerationTask(),
         "conversational": NovitaConversationalTask(),
         "text-to-video": NovitaTextToVideoTask(),
     },
+    "nscale": {
+        "conversational": NscaleConversationalTask(),
+        "text-to-image": NscaleTextToImageTask(),
+    },
     "openai": {
         "conversational": OpenAIConversationalTask(),
     },

huggingface_hub/inference/_providers/_common.py

@@ -27,6 +27,7 @@ HARDCODED_MODEL_INFERENCE_MAPPING: Dict[str, Dict[str, InferenceProviderMapping]
     "hf-inference": {},
     "hyperbolic": {},
     "nebius": {},
+    "nscale": {},
     "replicate": {},
     "sambanova": {},
     "together": {},

huggingface_hub/inference/_providers/cerebras.py

@@ -1,4 +1,4 @@
-from huggingface_hub.inference._providers._common import BaseConversationalTask
+from ._common import BaseConversationalTask
 
 
 class CerebrasConversationalTask(BaseConversationalTask):

huggingface_hub/inference/_providers/cohere.py

@@ -1,6 +1,8 @@
-from huggingface_hub.inference._providers._common import (
-    BaseConversationalTask,
-)
+from typing import Any, Dict, Optional
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+
+from ._common import BaseConversationalTask
 
 
 _PROVIDER = "cohere"
@@ -13,3 +15,18 @@ class CohereConversationalTask(BaseConversationalTask):
 
     def _prepare_route(self, mapped_model: str, api_key: str) -> str:
         return "/compatibility/v1/chat/completions"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: Dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[Dict]:
+        payload = super()._prepare_payload_as_dict(inputs, parameters, provider_mapping_info)
+        response_format = parameters.get("response_format")
+        if isinstance(response_format, dict) and response_format.get("type") == "json_schema":
+            json_schema_details = response_format.get("json_schema")
+            if isinstance(json_schema_details, dict) and "schema" in json_schema_details:
+                payload["response_format"] = {  # type: ignore [index]
+                    "type": "json_object",
+                    "schema": json_schema_details["schema"],
+                }
+
+        return payload

huggingface_hub/inference/_providers/fireworks_ai.py

@@ -1,3 +1,7 @@
+from typing import Any, Dict, Optional
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+
 from ._common import BaseConversationalTask
 
 
@@ -7,3 +11,17 @@ class FireworksAIConversationalTask(BaseConversationalTask):
 
     def _prepare_route(self, mapped_model: str, api_key: str) -> str:
         return "/inference/v1/chat/completions"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: Dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[Dict]:
+        payload = super()._prepare_payload_as_dict(inputs, parameters, provider_mapping_info)
+        response_format = parameters.get("response_format")
+        if isinstance(response_format, dict) and response_format.get("type") == "json_schema":
+            json_schema_details = response_format.get("json_schema")
+            if isinstance(json_schema_details, dict) and "schema" in json_schema_details:
+                payload["response_format"] = {  # type: ignore [index]
+                    "type": "json_object",
+                    "schema": json_schema_details["schema"],
+                }
+        return payload

huggingface_hub/inference/_providers/hf_inference.py

@@ -96,13 +96,20 @@ class HFInferenceConversational(HFInferenceTask):
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: Dict, provider_mapping_info: InferenceProviderMapping
     ) -> Optional[Dict]:
+        payload = filter_none(parameters)
         mapped_model = provider_mapping_info.provider_id
         payload_model = parameters.get("model") or mapped_model
 
         if payload_model is None or payload_model.startswith(("http://", "https://")):
             payload_model = "dummy"
 
-        return {**filter_none(parameters), "model": payload_model, "messages": inputs}
+        response_format = parameters.get("response_format")
+        if isinstance(response_format, dict) and response_format.get("type") == "json_schema":
+            payload["response_format"] = {
+                "type": "json_object",
+                "value": response_format["json_schema"]["schema"],
+            }
+        return {**payload, "model": payload_model, "messages": inputs}
 
     def _prepare_url(self, api_key: str, mapped_model: str) -> str:
         base_url = (

huggingface_hub/inference/_providers/nebius.py

@@ -30,6 +30,17 @@ class NebiusConversationalTask(BaseConversationalTask):
     def __init__(self):
         super().__init__(provider="nebius", base_url="https://api.studio.nebius.ai")
 
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: Dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[Dict]:
+        payload = super()._prepare_payload_as_dict(inputs, parameters, provider_mapping_info)
+        response_format = parameters.get("response_format")
+        if isinstance(response_format, dict) and response_format.get("type") == "json_schema":
+            json_schema_details = response_format.get("json_schema")
+            if isinstance(json_schema_details, dict) and "schema" in json_schema_details:
+                payload["guided_json"] = json_schema_details["schema"]  # type: ignore [index]
+        return payload
+
 
 class NebiusTextToImageTask(TaskProviderHelper):
     def __init__(self):
@@ -53,3 +64,20 @@ class NebiusTextToImageTask(TaskProviderHelper):
     def get_response(self, response: Union[bytes, Dict], request_params: Optional[RequestParameters] = None) -> Any:
         response_dict = _as_dict(response)
         return base64.b64decode(response_dict["data"][0]["b64_json"])
+
+
+class NebiusFeatureExtractionTask(TaskProviderHelper):
+    def __init__(self):
+        super().__init__(task="feature-extraction", provider="nebius", base_url="https://api.studio.nebius.ai")
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/v1/embeddings"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: Dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[Dict]:
+        return {"input": inputs, "model": provider_mapping_info.provider_id}
+
+    def get_response(self, response: Union[bytes, Dict], request_params: Optional[RequestParameters] = None) -> Any:
+        embeddings = _as_dict(response)["data"]
+        return [embedding["embedding"] for embedding in embeddings]

huggingface_hub/inference/_providers/nscale.py

@@ -0,0 +1,44 @@
+import base64
+from typing import Any, Dict, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+
+from ._common import BaseConversationalTask, TaskProviderHelper, filter_none
+
+
+class NscaleConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider="nscale", base_url="https://inference.api.nscale.com")
+
+
+class NscaleTextToImageTask(TaskProviderHelper):
+    def __init__(self):
+        super().__init__(provider="nscale", base_url="https://inference.api.nscale.com", task="text-to-image")
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/v1/images/generations"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: Dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[Dict]:
+        mapped_model = provider_mapping_info.provider_id
+        # Combine all parameters except inputs and parameters
+        parameters = filter_none(parameters)
+        if "width" in parameters and "height" in parameters:
+            parameters["size"] = f"{parameters.pop('width')}x{parameters.pop('height')}"
+        if "num_inference_steps" in parameters:
+            parameters.pop("num_inference_steps")
+        if "cfg_scale" in parameters:
+            parameters.pop("cfg_scale")
+        payload = {
+            "response_format": "b64_json",
+            "prompt": inputs,
+            "model": mapped_model,
+            **parameters,
+        }
+        return payload
+
+    def get_response(self, response: Union[bytes, Dict], request_params: Optional[RequestParameters] = None) -> Any:
+        response_dict = _as_dict(response)
+        return base64.b64decode(response_dict["data"][0]["b64_json"])