Initialisation du repository de Beta

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/__init__.py

@@ -0,0 +1,267 @@
+from typing import Literal, Optional, Union
+
+from huggingface_hub.inference._providers.featherless_ai import (
+    FeatherlessConversationalTask,
+    FeatherlessTextGenerationTask,
+)
+from huggingface_hub.utils import logging
+
+from ._common import AutoRouterConversationalTask, TaskProviderHelper, _fetch_inference_provider_mapping
+from .black_forest_labs import BlackForestLabsTextToImageTask
+from .cerebras import CerebrasConversationalTask
+from .clarifai import ClarifaiConversationalTask
+from .cohere import CohereConversationalTask
+from .fal_ai import (
+    FalAIAutomaticSpeechRecognitionTask,
+    FalAIImageSegmentationTask,
+    FalAIImageToImageTask,
+    FalAIImageToVideoTask,
+    FalAITextToImageTask,
+    FalAITextToSpeechTask,
+    FalAITextToVideoTask,
+)
+from .fireworks_ai import FireworksAIConversationalTask
+from .groq import GroqConversationalTask
+from .hf_inference import (
+    HFInferenceBinaryInputTask,
+    HFInferenceConversational,
+    HFInferenceFeatureExtractionTask,
+    HFInferenceTask,
+)
+from .hyperbolic import HyperbolicTextGenerationTask, HyperbolicTextToImageTask
+from .nebius import (
+    NebiusConversationalTask,
+    NebiusFeatureExtractionTask,
+    NebiusTextGenerationTask,
+    NebiusTextToImageTask,
+)
+from .novita import NovitaConversationalTask, NovitaTextGenerationTask, NovitaTextToVideoTask
+from .nscale import NscaleConversationalTask, NscaleTextToImageTask
+from .openai import OpenAIConversationalTask
+from .ovhcloud import OVHcloudConversationalTask
+from .publicai import PublicAIConversationalTask
+from .replicate import (
+    ReplicateAutomaticSpeechRecognitionTask,
+    ReplicateImageToImageTask,
+    ReplicateTask,
+    ReplicateTextToImageTask,
+    ReplicateTextToSpeechTask,
+)
+from .sambanova import SambanovaConversationalTask, SambanovaFeatureExtractionTask
+from .scaleway import ScalewayConversationalTask, ScalewayFeatureExtractionTask
+from .together import TogetherConversationalTask, TogetherTextGenerationTask, TogetherTextToImageTask
+from .wavespeed import (
+    WavespeedAIImageToImageTask,
+    WavespeedAIImageToVideoTask,
+    WavespeedAITextToImageTask,
+    WavespeedAITextToVideoTask,
+)
+from .zai_org import ZaiConversationalTask, ZaiTextToImageTask
+
+
+logger = logging.get_logger(__name__)
+
+
+PROVIDER_T = Literal[
+    "black-forest-labs",
+    "cerebras",
+    "clarifai",
+    "cohere",
+    "fal-ai",
+    "featherless-ai",
+    "fireworks-ai",
+    "groq",
+    "hf-inference",
+    "hyperbolic",
+    "nebius",
+    "novita",
+    "nscale",
+    "openai",
+    "ovhcloud",
+    "publicai",
+    "replicate",
+    "sambanova",
+    "scaleway",
+    "together",
+    "wavespeed",
+    "zai-org",
+]
+
+PROVIDER_OR_POLICY_T = Union[PROVIDER_T, Literal["auto"]]
+
+CONVERSATIONAL_AUTO_ROUTER = AutoRouterConversationalTask()
+
+PROVIDERS: dict[PROVIDER_T, dict[str, TaskProviderHelper]] = {
+    "black-forest-labs": {
+        "text-to-image": BlackForestLabsTextToImageTask(),
+    },
+    "cerebras": {
+        "conversational": CerebrasConversationalTask(),
+    },
+    "clarifai": {
+        "conversational": ClarifaiConversationalTask(),
+    },
+    "cohere": {
+        "conversational": CohereConversationalTask(),
+    },
+    "fal-ai": {
+        "automatic-speech-recognition": FalAIAutomaticSpeechRecognitionTask(),
+        "text-to-image": FalAITextToImageTask(),
+        "text-to-speech": FalAITextToSpeechTask(),
+        "text-to-video": FalAITextToVideoTask(),
+        "image-to-video": FalAIImageToVideoTask(),
+        "image-to-image": FalAIImageToImageTask(),
+        "image-segmentation": FalAIImageSegmentationTask(),
+    },
+    "featherless-ai": {
+        "conversational": FeatherlessConversationalTask(),
+        "text-generation": FeatherlessTextGenerationTask(),
+    },
+    "fireworks-ai": {
+        "conversational": FireworksAIConversationalTask(),
+    },
+    "groq": {
+        "conversational": GroqConversationalTask(),
+    },
+    "hf-inference": {
+        "text-to-image": HFInferenceTask("text-to-image"),
+        "conversational": HFInferenceConversational(),
+        "text-generation": HFInferenceTask("text-generation"),
+        "text-classification": HFInferenceTask("text-classification"),
+        "question-answering": HFInferenceTask("question-answering"),
+        "audio-classification": HFInferenceBinaryInputTask("audio-classification"),
+        "automatic-speech-recognition": HFInferenceBinaryInputTask("automatic-speech-recognition"),
+        "fill-mask": HFInferenceTask("fill-mask"),
+        "feature-extraction": HFInferenceFeatureExtractionTask(),
+        "image-classification": HFInferenceBinaryInputTask("image-classification"),
+        "image-segmentation": HFInferenceBinaryInputTask("image-segmentation"),
+        "document-question-answering": HFInferenceTask("document-question-answering"),
+        "image-to-text": HFInferenceBinaryInputTask("image-to-text"),
+        "object-detection": HFInferenceBinaryInputTask("object-detection"),
+        "audio-to-audio": HFInferenceBinaryInputTask("audio-to-audio"),
+        "zero-shot-image-classification": HFInferenceBinaryInputTask("zero-shot-image-classification"),
+        "zero-shot-classification": HFInferenceTask("zero-shot-classification"),
+        "image-to-image": HFInferenceBinaryInputTask("image-to-image"),
+        "sentence-similarity": HFInferenceTask("sentence-similarity"),
+        "table-question-answering": HFInferenceTask("table-question-answering"),
+        "tabular-classification": HFInferenceTask("tabular-classification"),
+        "text-to-speech": HFInferenceTask("text-to-speech"),
+        "token-classification": HFInferenceTask("token-classification"),
+        "translation": HFInferenceTask("translation"),
+        "summarization": HFInferenceTask("summarization"),
+        "visual-question-answering": HFInferenceBinaryInputTask("visual-question-answering"),
+    },
+    "hyperbolic": {
+        "text-to-image": HyperbolicTextToImageTask(),
+        "conversational": HyperbolicTextGenerationTask("conversational"),
+        "text-generation": HyperbolicTextGenerationTask("text-generation"),
+    },
+    "nebius": {
+        "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(),
+    },
+    "ovhcloud": {
+        "conversational": OVHcloudConversationalTask(),
+    },
+    "publicai": {
+        "conversational": PublicAIConversationalTask(),
+    },
+    "replicate": {
+        "automatic-speech-recognition": ReplicateAutomaticSpeechRecognitionTask(),
+        "image-to-image": ReplicateImageToImageTask(),
+        "text-to-image": ReplicateTextToImageTask(),
+        "text-to-speech": ReplicateTextToSpeechTask(),
+        "text-to-video": ReplicateTask("text-to-video"),
+    },
+    "sambanova": {
+        "conversational": SambanovaConversationalTask(),
+        "feature-extraction": SambanovaFeatureExtractionTask(),
+    },
+    "scaleway": {
+        "conversational": ScalewayConversationalTask(),
+        "feature-extraction": ScalewayFeatureExtractionTask(),
+    },
+    "together": {
+        "text-to-image": TogetherTextToImageTask(),
+        "conversational": TogetherConversationalTask(),
+        "text-generation": TogetherTextGenerationTask(),
+    },
+    "wavespeed": {
+        "text-to-image": WavespeedAITextToImageTask(),
+        "text-to-video": WavespeedAITextToVideoTask(),
+        "image-to-image": WavespeedAIImageToImageTask(),
+        "image-to-video": WavespeedAIImageToVideoTask(),
+    },
+    "zai-org": {
+        "conversational": ZaiConversationalTask(),
+        "text-to-image": ZaiTextToImageTask(),
+    },
+}
+
+
+def get_provider_helper(
+    provider: Optional[PROVIDER_OR_POLICY_T], task: str, model: Optional[str]
+) -> TaskProviderHelper:
+    """Get provider helper instance by name and task.
+
+    Args:
+        provider (`str`, *optional*): name of the provider, or "auto" to automatically select the provider for the model.
+        task (`str`): Name of the task
+        model (`str`, *optional*): Name of the model
+    Returns:
+        TaskProviderHelper: Helper instance for the specified provider and task
+
+    Raises:
+        ValueError: If provider or task is not supported
+    """
+
+    if (model is None and provider in (None, "auto")) or (
+        model is not None and model.startswith(("http://", "https://"))
+    ):
+        provider = "hf-inference"
+
+    if provider is None:
+        logger.info(
+            "No provider specified for task `conversational`. Defaulting to server-side auto routing."
+            if task == "conversational"
+            else "Defaulting to 'auto' which will select the first provider available for the model, sorted by the user's order in https://hf.co/settings/inference-providers."
+        )
+        provider = "auto"
+
+    if provider == "auto":
+        if model is None:
+            raise ValueError("Specifying a model is required when provider is 'auto'")
+        if task == "conversational":
+            # Special case: we have a dedicated auto-router for conversational models. No need to fetch provider mapping.
+            return CONVERSATIONAL_AUTO_ROUTER
+
+        provider_mapping = _fetch_inference_provider_mapping(model)
+        provider = next(iter(provider_mapping)).provider
+
+    provider_tasks = PROVIDERS.get(provider)  # type: ignore
+    if provider_tasks is None:
+        raise ValueError(
+            f"Provider '{provider}' not supported. Available values: 'auto' or any provider from {list(PROVIDERS.keys())}."
+            "Passing 'auto' (default value) will automatically select the first provider available for the model, sorted "
+            "by the user's order in https://hf.co/settings/inference-providers."
+        )
+
+    if task not in provider_tasks:
+        raise ValueError(
+            f"Task '{task}' not supported for provider '{provider}'. Available tasks: {list(provider_tasks.keys())}"
+        )
+    return provider_tasks[task]

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/_common.py

@@ -0,0 +1,363 @@
+from functools import lru_cache
+from typing import Any, Optional, Union, overload
+
+from huggingface_hub import constants
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import MimeBytes, RequestParameters
+from huggingface_hub.inference._generated.types.chat_completion import ChatCompletionInputMessage
+from huggingface_hub.utils import build_hf_headers, get_token, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+# Dev purposes only.
+# If you want to try to run inference for a new model locally before it's registered on huggingface.co
+# for a given Inference Provider, you can add it to the following dictionary.
+HARDCODED_MODEL_INFERENCE_MAPPING: dict[str, dict[str, InferenceProviderMapping]] = {
+    # "HF model ID" => InferenceProviderMapping object initialized with "Model ID on Inference Provider's side"
+    #
+    # Example:
+    # "Qwen/Qwen2.5-Coder-32B-Instruct": InferenceProviderMapping(hf_model_id="Qwen/Qwen2.5-Coder-32B-Instruct",
+    #                                    provider_id="Qwen2.5-Coder-32B-Instruct",
+    #                                    task="conversational",
+    #                                    status="live")
+    "cerebras": {},
+    "cohere": {},
+    "clarifai": {},
+    "fal-ai": {},
+    "fireworks-ai": {},
+    "groq": {},
+    "hf-inference": {},
+    "hyperbolic": {},
+    "nebius": {},
+    "nscale": {},
+    "ovhcloud": {},
+    "replicate": {},
+    "sambanova": {},
+    "scaleway": {},
+    "together": {},
+    "wavespeed": {},
+    "zai-org": {},
+}
+
+
+@overload
+def filter_none(obj: dict[str, Any]) -> dict[str, Any]: ...
+@overload
+def filter_none(obj: list[Any]) -> list[Any]: ...
+
+
+def filter_none(obj: Union[dict[str, Any], list[Any]]) -> Union[dict[str, Any], list[Any]]:
+    if isinstance(obj, dict):
+        cleaned: dict[str, Any] = {}
+        for k, v in obj.items():
+            if v is None:
+                continue
+            if isinstance(v, (dict, list)):
+                v = filter_none(v)
+            cleaned[k] = v
+        return cleaned
+
+    if isinstance(obj, list):
+        return [filter_none(v) if isinstance(v, (dict, list)) else v for v in obj]
+
+    raise ValueError(f"Expected dict or list, got {type(obj)}")
+
+
+class TaskProviderHelper:
+    """Base class for task-specific provider helpers."""
+
+    def __init__(self, provider: str, base_url: str, task: str) -> None:
+        self.provider = provider
+        self.task = task
+        self.base_url = base_url
+
+    def prepare_request(
+        self,
+        *,
+        inputs: Any,
+        parameters: dict[str, Any],
+        headers: dict,
+        model: Optional[str],
+        api_key: Optional[str],
+        extra_payload: Optional[dict[str, Any]] = None,
+    ) -> RequestParameters:
+        """
+        Prepare the request to be sent to the provider.
+
+        Each step (api_key, model, headers, url, payload) can be customized in subclasses.
+        """
+        # api_key from user, or local token, or raise error
+        api_key = self._prepare_api_key(api_key)
+
+        # mapped model from HF model ID
+        provider_mapping_info = self._prepare_mapping_info(model)
+
+        # default HF headers + user headers (to customize in subclasses)
+        headers = self._prepare_headers(headers, api_key)
+
+        # routed URL if HF token, or direct URL (to customize in '_prepare_route' in subclasses)
+        url = self._prepare_url(api_key, provider_mapping_info.provider_id)
+
+        # prepare payload (to customize in subclasses)
+        payload = self._prepare_payload_as_dict(inputs, parameters, provider_mapping_info=provider_mapping_info)
+        if payload is not None:
+            payload = recursive_merge(payload, filter_none(extra_payload or {}))
+
+        # body data (to customize in subclasses)
+        data = self._prepare_payload_as_bytes(inputs, parameters, provider_mapping_info, extra_payload)
+
+        # check if both payload and data are set and return
+        if payload is not None and data is not None:
+            raise ValueError("Both payload and data cannot be set in the same request.")
+        if payload is None and data is None:
+            raise ValueError("Either payload or data must be set in the request.")
+
+        # normalize headers to lowercase and add content-type if not present
+        normalized_headers = self._normalize_headers(headers, payload, data)
+
+        return RequestParameters(
+            url=url,
+            task=self.task,
+            model=provider_mapping_info.provider_id,
+            json=payload,
+            data=data,
+            headers=normalized_headers,
+        )
+
+    def get_response(
+        self,
+        response: Union[bytes, dict],
+        request_params: Optional[RequestParameters] = None,
+    ) -> Any:
+        """
+        Return the response in the expected format.
+
+        Override this method in subclasses for customized response handling."""
+        return response
+
+    def _prepare_api_key(self, api_key: Optional[str]) -> str:
+        """Return the API key to use for the request.
+
+        Usually not overwritten in subclasses."""
+        if api_key is None:
+            api_key = get_token()
+        if api_key is None:
+            raise ValueError(
+                f"You must provide an api_key to work with {self.provider} API or log in with `hf auth login`."
+            )
+        return api_key
+
+    def _prepare_mapping_info(self, model: Optional[str]) -> InferenceProviderMapping:
+        """Return the mapped model ID to use for the request.
+
+        Usually not overwritten in subclasses."""
+        if model is None:
+            raise ValueError(f"Please provide an HF model ID supported by {self.provider}.")
+
+        # hardcoded mapping for local testing
+        if HARDCODED_MODEL_INFERENCE_MAPPING.get(self.provider, {}).get(model):
+            return HARDCODED_MODEL_INFERENCE_MAPPING[self.provider][model]
+
+        provider_mapping = None
+        for mapping in _fetch_inference_provider_mapping(model):
+            if mapping.provider == self.provider:
+                provider_mapping = mapping
+                break
+
+        if provider_mapping is None:
+            raise ValueError(f"Model {model} is not supported by provider {self.provider}.")
+
+        if provider_mapping.task != self.task:
+            raise ValueError(
+                f"Model {model} is not supported for task {self.task} and provider {self.provider}. "
+                f"Supported task: {provider_mapping.task}."
+            )
+
+        if provider_mapping.status == "staging":
+            logger.warning(
+                f"Model {model} is in staging mode for provider {self.provider}. Meant for test purposes only."
+            )
+        if provider_mapping.status == "error":
+            logger.warning(
+                f"Our latest automated health check on model '{model}' for provider '{self.provider}' did not complete successfully.  "
+                "Inference call might fail."
+            )
+        return provider_mapping
+
+    def _normalize_headers(
+        self, headers: dict[str, Any], payload: Optional[dict[str, Any]], data: Optional[MimeBytes]
+    ) -> dict[str, Any]:
+        """Normalize the headers to use for the request.
+
+        Override this method in subclasses for customized headers.
+        """
+        normalized_headers = {key.lower(): value for key, value in headers.items() if value is not None}
+        if normalized_headers.get("content-type") is None:
+            if data is not None and data.mime_type is not None:
+                normalized_headers["content-type"] = data.mime_type
+            elif payload is not None:
+                normalized_headers["content-type"] = "application/json"
+        return normalized_headers
+
+    def _prepare_headers(self, headers: dict, api_key: str) -> dict[str, Any]:
+        """Return the headers to use for the request.
+
+        Override this method in subclasses for customized headers.
+        """
+        return {**build_hf_headers(token=api_key), **headers}
+
+    def _prepare_url(self, api_key: str, mapped_model: str) -> str:
+        """Return the URL to use for the request.
+
+        Usually not overwritten in subclasses."""
+        base_url = self._prepare_base_url(api_key)
+        route = self._prepare_route(mapped_model, api_key)
+        return f"{base_url.rstrip('/')}/{route.lstrip('/')}"
+
+    def _prepare_base_url(self, api_key: str) -> str:
+        """Return the base URL to use for the request.
+
+        Usually not overwritten in subclasses."""
+        # Route to the proxy if the api_key is a HF TOKEN
+        if api_key.startswith("hf_"):
+            logger.info(f"Calling '{self.provider}' provider through Hugging Face router.")
+            return constants.INFERENCE_PROXY_TEMPLATE.format(provider=self.provider)
+        else:
+            logger.info(f"Calling '{self.provider}' provider directly.")
+            return self.base_url
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        """Return the route to use for the request.
+
+        Override this method in subclasses for customized routes.
+        """
+        return ""
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        """Return the payload to use for the request, as a dict.
+
+        Override this method in subclasses for customized payloads.
+        Only one of `_prepare_payload_as_dict` and `_prepare_payload_as_bytes` should return a value.
+        """
+        return None
+
+    def _prepare_payload_as_bytes(
+        self,
+        inputs: Any,
+        parameters: dict,
+        provider_mapping_info: InferenceProviderMapping,
+        extra_payload: Optional[dict],
+    ) -> Optional[MimeBytes]:
+        """Return the body to use for the request, as bytes.
+
+        Override this method in subclasses for customized body data.
+        Only one of `_prepare_payload_as_dict` and `_prepare_payload_as_bytes` should return a value.
+        """
+        return None
+
+
+class BaseConversationalTask(TaskProviderHelper):
+    """
+    Base class for conversational (chat completion) tasks.
+    The schema follows the OpenAI API format defined here: https://platform.openai.com/docs/api-reference/chat
+    """
+
+    def __init__(self, provider: str, base_url: str):
+        super().__init__(provider=provider, base_url=base_url, task="conversational")
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/v1/chat/completions"
+
+    def _prepare_payload_as_dict(
+        self,
+        inputs: list[Union[dict, ChatCompletionInputMessage]],
+        parameters: dict,
+        provider_mapping_info: InferenceProviderMapping,
+    ) -> Optional[dict]:
+        return filter_none({"messages": inputs, **parameters, "model": provider_mapping_info.provider_id})
+
+
+class AutoRouterConversationalTask(BaseConversationalTask):
+    """
+    Auto-router for conversational tasks.
+
+    We let the Hugging Face router select the best provider for the model, based on availability and user preferences.
+    This is a special case since the selection is done server-side (avoid 1 API call to fetch provider mapping).
+    """
+
+    def __init__(self):
+        super().__init__(provider="auto", base_url="https://router.huggingface.co")
+
+    def _prepare_base_url(self, api_key: str) -> str:
+        """Return the base URL to use for the request.
+
+        Usually not overwritten in subclasses."""
+        # Route to the proxy if the api_key is a HF TOKEN
+        if not api_key.startswith("hf_"):
+            raise ValueError("Cannot select auto-router when using non-Hugging Face API key.")
+        else:
+            return self.base_url  # No `/auto` suffix in the URL
+
+    def _prepare_mapping_info(self, model: Optional[str]) -> InferenceProviderMapping:
+        """
+        In auto-router, we don't need to fetch provider mapping info.
+        We just return a dummy mapping info with provider_id set to the HF model ID.
+        """
+        if model is None:
+            raise ValueError("Please provide an HF model ID.")
+
+        return InferenceProviderMapping(
+            provider="auto",
+            hf_model_id=model,
+            providerId=model,
+            status="live",
+            task="conversational",
+        )
+
+
+class BaseTextGenerationTask(TaskProviderHelper):
+    """
+    Base class for text-generation (completion) tasks.
+    The schema follows the OpenAI API format defined here: https://platform.openai.com/docs/api-reference/completions
+    """
+
+    def __init__(self, provider: str, base_url: str):
+        super().__init__(provider=provider, base_url=base_url, task="text-generation")
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/v1/completions"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        return filter_none({"prompt": inputs, **parameters, "model": provider_mapping_info.provider_id})
+
+
+@lru_cache(maxsize=None)
+def _fetch_inference_provider_mapping(model: str) -> list["InferenceProviderMapping"]:
+    """
+    Fetch provider mappings for a model from the Hub.
+    """
+    from huggingface_hub.hf_api import HfApi
+
+    info = HfApi().model_info(model, expand=["inferenceProviderMapping"])
+    provider_mapping = info.inference_provider_mapping
+    if provider_mapping is None:
+        raise ValueError(f"No provider mapping found for model {model}")
+    return provider_mapping
+
+
+def recursive_merge(dict1: dict, dict2: dict) -> dict:
+    return {
+        **dict1,
+        **{
+            key: recursive_merge(dict1[key], value)
+            if (key in dict1 and isinstance(dict1[key], dict) and isinstance(value, dict))
+            else value
+            for key, value in dict2.items()
+        },
+    }

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/black_forest_labs.py

@@ -0,0 +1,69 @@
+import time
+from typing import Any, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+from huggingface_hub.inference._providers._common import TaskProviderHelper, filter_none
+from huggingface_hub.utils import logging
+from huggingface_hub.utils._http import get_session
+
+
+logger = logging.get_logger(__name__)
+
+MAX_POLLING_ATTEMPTS = 6
+POLLING_INTERVAL = 1.0
+
+
+class BlackForestLabsTextToImageTask(TaskProviderHelper):
+    def __init__(self):
+        super().__init__(provider="black-forest-labs", base_url="https://api.us1.bfl.ai", task="text-to-image")
+
+    def _prepare_headers(self, headers: dict, api_key: str) -> dict[str, Any]:
+        headers = super()._prepare_headers(headers, api_key)
+        if not api_key.startswith("hf_"):
+            _ = headers.pop("authorization")
+            headers["X-Key"] = api_key
+        return headers
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return f"/v1/{mapped_model}"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        parameters = filter_none(parameters)
+        if "num_inference_steps" in parameters:
+            parameters["steps"] = parameters.pop("num_inference_steps")
+        if "guidance_scale" in parameters:
+            parameters["guidance"] = parameters.pop("guidance_scale")
+
+        return {"prompt": inputs, **parameters}
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        """
+        Polling mechanism for Black Forest Labs since the API is asynchronous.
+        """
+        url = _as_dict(response).get("polling_url")
+        session = get_session()
+        for _ in range(MAX_POLLING_ATTEMPTS):
+            time.sleep(POLLING_INTERVAL)
+
+            response = session.get(url, headers={"Content-Type": "application/json"})  # type: ignore
+            response.raise_for_status()  # type: ignore
+            response_json: dict = response.json()  # type: ignore
+            status = response_json.get("status")
+            logger.info(
+                f"Polling generation result from {url}. Current status: {status}. "
+                f"Will retry after {POLLING_INTERVAL} seconds if not ready."
+            )
+
+            if (
+                status == "Ready"
+                and isinstance(response_json.get("result"), dict)
+                and (sample_url := response_json["result"].get("sample"))
+            ):
+                image_resp = session.get(sample_url)
+                image_resp.raise_for_status()
+                return image_resp.content
+
+        raise TimeoutError(f"Failed to get the image URL after {MAX_POLLING_ATTEMPTS} attempts.")

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/cerebras.py

@@ -0,0 +1,6 @@
+from ._common import BaseConversationalTask
+
+
+class CerebrasConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider="cerebras", base_url="https://api.cerebras.ai")

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/clarifai.py

@@ -0,0 +1,13 @@
+from ._common import BaseConversationalTask
+
+
+_PROVIDER = "clarifai"
+_BASE_URL = "https://api.clarifai.com"
+
+
+class ClarifaiConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/v2/ext/openai/v1/chat/completions"

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/cohere.py

@@ -0,0 +1,32 @@
+from typing import Any, Optional
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+
+from ._common import BaseConversationalTask
+
+
+_PROVIDER = "cohere"
+_BASE_URL = "https://api.cohere.com"
+
+
+class CohereConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    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

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/fal_ai.py

@@ -0,0 +1,299 @@
+import base64
+import time
+from abc import ABC
+from typing import Any, Optional, Union
+from urllib.parse import urlparse
+
+from huggingface_hub import constants
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict, _as_url
+from huggingface_hub.inference._providers._common import TaskProviderHelper, filter_none
+from huggingface_hub.utils import get_session, hf_raise_for_status
+from huggingface_hub.utils.logging import get_logger
+
+
+logger = get_logger(__name__)
+
+# Arbitrary polling interval
+_POLLING_INTERVAL = 0.5
+
+
+class FalAITask(TaskProviderHelper, ABC):
+    def __init__(self, task: str):
+        super().__init__(provider="fal-ai", base_url="https://fal.run", task=task)
+
+    def _prepare_headers(self, headers: dict, api_key: str) -> dict[str, Any]:
+        headers = super()._prepare_headers(headers, api_key)
+        if not api_key.startswith("hf_"):
+            headers["authorization"] = f"Key {api_key}"
+        return headers
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return f"/{mapped_model}"
+
+
+class FalAIQueueTask(TaskProviderHelper, ABC):
+    def __init__(self, task: str):
+        super().__init__(provider="fal-ai", base_url="https://queue.fal.run", task=task)
+
+    def _prepare_headers(self, headers: dict, api_key: str) -> dict[str, Any]:
+        headers = super()._prepare_headers(headers, api_key)
+        if not api_key.startswith("hf_"):
+            headers["authorization"] = f"Key {api_key}"
+        return headers
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        if api_key.startswith("hf_"):
+            # Use the queue subdomain for HF routing
+            return f"/{mapped_model}?_subdomain=queue"
+        return f"/{mapped_model}"
+
+    def get_response(
+        self,
+        response: Union[bytes, dict],
+        request_params: Optional[RequestParameters] = None,
+    ) -> Any:
+        response_dict = _as_dict(response)
+
+        request_id = response_dict.get("request_id")
+        if not request_id:
+            raise ValueError("No request ID found in the response")
+        if request_params is None:
+            raise ValueError(
+                f"A `RequestParameters` object should be provided to get {self.task} responses with Fal AI."
+            )
+
+        # extract the base url and query params
+        parsed_url = urlparse(request_params.url)
+        # a bit hacky way to concatenate the provider name without parsing `parsed_url.path`
+        base_url = f"{parsed_url.scheme}://{parsed_url.netloc}{'/fal-ai' if parsed_url.netloc == 'router.huggingface.co' else ''}"
+        query_param = f"?{parsed_url.query}" if parsed_url.query else ""
+
+        # extracting the provider model id for status and result urls
+        # from the response as it might be different from the mapped model in `request_params.url`
+        model_id = urlparse(response_dict.get("response_url")).path
+        status_url = f"{base_url}{str(model_id)}/status{query_param}"
+        result_url = f"{base_url}{str(model_id)}{query_param}"
+
+        status = response_dict.get("status")
+        logger.info("Generating the output.. this can take several minutes.")
+        while status != "COMPLETED":
+            time.sleep(_POLLING_INTERVAL)
+            status_response = get_session().get(status_url, headers=request_params.headers)
+            hf_raise_for_status(status_response)
+            status = status_response.json().get("status")
+
+        return get_session().get(result_url, headers=request_params.headers).json()
+
+
+class FalAIAutomaticSpeechRecognitionTask(FalAITask):
+    def __init__(self):
+        super().__init__("automatic-speech-recognition")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        if isinstance(inputs, str) and inputs.startswith(("http://", "https://")):
+            # If input is a URL, pass it directly
+            audio_url = inputs
+        else:
+            # If input is a file path, read it first
+            if isinstance(inputs, str):
+                with open(inputs, "rb") as f:
+                    inputs = f.read()
+
+            audio_b64 = base64.b64encode(inputs).decode()
+            content_type = "audio/mpeg"
+            audio_url = f"data:{content_type};base64,{audio_b64}"
+
+        return {"audio_url": audio_url, **filter_none(parameters)}
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        text = _as_dict(response)["text"]
+        if not isinstance(text, str):
+            raise ValueError(f"Unexpected output format from FalAI API. Expected string, got {type(text)}.")
+        return {"text": text}
+
+
+class FalAITextToImageTask(FalAITask):
+    def __init__(self):
+        super().__init__("text-to-image")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        payload: dict[str, Any] = {
+            "prompt": inputs,
+            **filter_none(parameters),
+        }
+        if "width" in payload and "height" in payload:
+            payload["image_size"] = {
+                "width": payload.pop("width"),
+                "height": payload.pop("height"),
+            }
+        if provider_mapping_info.adapter_weights_path is not None:
+            lora_path = constants.HUGGINGFACE_CO_URL_TEMPLATE.format(
+                repo_id=provider_mapping_info.hf_model_id,
+                revision="main",
+                filename=provider_mapping_info.adapter_weights_path,
+            )
+            payload["loras"] = [{"path": lora_path, "scale": 1}]
+            if provider_mapping_info.provider_id == "fal-ai/lora":
+                # little hack: fal requires the base model for stable-diffusion-based loras but not for flux-based
+                # See payloads in https://fal.ai/models/fal-ai/lora/api vs https://fal.ai/models/fal-ai/flux-lora/api
+                payload["model_name"] = "stabilityai/stable-diffusion-xl-base-1.0"
+
+        return payload
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        url = _as_dict(response)["images"][0]["url"]
+        return get_session().get(url).content
+
+
+class FalAITextToSpeechTask(FalAITask):
+    def __init__(self):
+        super().__init__("text-to-speech")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        return {"text": inputs, **filter_none(parameters)}
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        url = _as_dict(response)["audio"]["url"]
+        return get_session().get(url).content
+
+
+class FalAITextToVideoTask(FalAIQueueTask):
+    def __init__(self):
+        super().__init__("text-to-video")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        return {"prompt": inputs, **filter_none(parameters)}
+
+    def get_response(
+        self,
+        response: Union[bytes, dict],
+        request_params: Optional[RequestParameters] = None,
+    ) -> Any:
+        output = super().get_response(response, request_params)
+        url = _as_dict(output)["video"]["url"]
+        return get_session().get(url).content
+
+
+class FalAIImageToImageTask(FalAIQueueTask):
+    def __init__(self):
+        super().__init__("image-to-image")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        image_url = _as_url(inputs, default_mime_type="image/jpeg")
+        if "target_size" in parameters:
+            parameters["image_size"] = parameters.pop("target_size")
+        payload: dict[str, Any] = {
+            "image_url": image_url,
+            **filter_none(parameters),
+        }
+        if provider_mapping_info.adapter_weights_path is not None:
+            lora_path = constants.HUGGINGFACE_CO_URL_TEMPLATE.format(
+                repo_id=provider_mapping_info.hf_model_id,
+                revision="main",
+                filename=provider_mapping_info.adapter_weights_path,
+            )
+            payload["loras"] = [{"path": lora_path, "scale": 1}]
+
+        return payload
+
+    def get_response(
+        self,
+        response: Union[bytes, dict],
+        request_params: Optional[RequestParameters] = None,
+    ) -> Any:
+        output = super().get_response(response, request_params)
+        url = _as_dict(output)["images"][0]["url"]
+        return get_session().get(url).content
+
+
+class FalAIImageToVideoTask(FalAIQueueTask):
+    def __init__(self):
+        super().__init__("image-to-video")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        image_url = _as_url(inputs, default_mime_type="image/jpeg")
+        payload: dict[str, Any] = {
+            "image_url": image_url,
+            **filter_none(parameters),
+        }
+        if provider_mapping_info.adapter_weights_path is not None:
+            lora_path = constants.HUGGINGFACE_CO_URL_TEMPLATE.format(
+                repo_id=provider_mapping_info.hf_model_id,
+                revision="main",
+                filename=provider_mapping_info.adapter_weights_path,
+            )
+            payload["loras"] = [{"path": lora_path, "scale": 1}]
+        return payload
+
+    def get_response(
+        self,
+        response: Union[bytes, dict],
+        request_params: Optional[RequestParameters] = None,
+    ) -> Any:
+        output = super().get_response(response, request_params)
+        url = _as_dict(output)["video"]["url"]
+        return get_session().get(url).content
+
+
+class FalAIImageSegmentationTask(FalAIQueueTask):
+    def __init__(self):
+        super().__init__("image-segmentation")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        image_url = _as_url(inputs, default_mime_type="image/png")
+        payload: dict[str, Any] = {
+            "image_url": image_url,
+            **filter_none(parameters),
+            "sync_mode": True,
+        }
+        return payload
+
+    def get_response(
+        self,
+        response: Union[bytes, dict],
+        request_params: Optional[RequestParameters] = None,
+    ) -> Any:
+        result = super().get_response(response, request_params)
+        result_dict = _as_dict(result)
+
+        if "image" not in result_dict:
+            raise ValueError(f"Response from fal ai image-segmentation API does not contain an image: {result_dict}")
+
+        image_data = result_dict["image"]
+        if "url" not in image_data:
+            raise ValueError(f"Image data from fal ai image-segmentation API does not contain a URL: {image_data}")
+
+        image_url = image_data["url"]
+
+        if isinstance(image_url, str) and image_url.startswith("data:"):
+            if "," in image_url:
+                mask_base64 = image_url.split(",", 1)[1]
+            else:
+                raise ValueError(f"Invalid data URL format: {image_url}")
+        else:
+            # or it's a regular URL, fetch it
+            mask_response = get_session().get(image_url)
+            hf_raise_for_status(mask_response)
+            mask_base64 = base64.b64encode(mask_response.content).decode()
+
+        return [
+            {
+                "label": "mask",
+                "mask": mask_base64,
+            }
+        ]

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/featherless_ai.py

@@ -0,0 +1,38 @@
+from typing import Any, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+
+from ._common import BaseConversationalTask, BaseTextGenerationTask, filter_none
+
+
+_PROVIDER = "featherless-ai"
+_BASE_URL = "https://api.featherless.ai"
+
+
+class FeatherlessTextGenerationTask(BaseTextGenerationTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        params = filter_none(parameters.copy())
+        params["max_tokens"] = params.pop("max_new_tokens", None)
+
+        return {"prompt": inputs, **params, "model": provider_mapping_info.provider_id}
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        output = _as_dict(response)["choices"][0]
+        return {
+            "generated_text": output["text"],
+            "details": {
+                "finish_reason": output.get("finish_reason"),
+                "seed": output.get("seed"),
+            },
+        }
+
+
+class FeatherlessConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/fireworks_ai.py

@@ -0,0 +1,27 @@
+from typing import Any, Optional
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+
+from ._common import BaseConversationalTask
+
+
+class FireworksAIConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider="fireworks-ai", base_url="https://api.fireworks.ai")
+
+    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

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/groq.py

@@ -0,0 +1,9 @@
+from ._common import BaseConversationalTask
+
+
+class GroqConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider="groq", base_url="https://api.groq.com")
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/openai/v1/chat/completions"

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/hf_inference.py

@@ -0,0 +1,228 @@
+import json
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Optional, Union
+from urllib.parse import urlparse, urlunparse
+
+from huggingface_hub import constants
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import (
+    MimeBytes,
+    RequestParameters,
+    _b64_encode,
+    _bytes_to_dict,
+    _open_as_mime_bytes,
+)
+from huggingface_hub.inference._providers._common import TaskProviderHelper, filter_none
+from huggingface_hub.utils import build_hf_headers, get_session, get_token, hf_raise_for_status
+
+
+class HFInferenceTask(TaskProviderHelper):
+    """Base class for HF Inference API tasks."""
+
+    def __init__(self, task: str):
+        super().__init__(
+            provider="hf-inference",
+            base_url=constants.INFERENCE_PROXY_TEMPLATE.format(provider="hf-inference"),
+            task=task,
+        )
+
+    def _prepare_api_key(self, api_key: Optional[str]) -> str:
+        # special case: for HF Inference we allow not providing an API key
+        return api_key or get_token()  # type: ignore[return-value]
+
+    def _prepare_mapping_info(self, model: Optional[str]) -> InferenceProviderMapping:
+        if model is not None and model.startswith(("http://", "https://")):
+            return InferenceProviderMapping(
+                provider="hf-inference", providerId=model, hf_model_id=model, task=self.task, status="live"
+            )
+        model_id = model if model is not None else _fetch_recommended_models().get(self.task)
+        if model_id is None:
+            raise ValueError(
+                f"Task {self.task} has no recommended model for HF Inference. Please specify a model"
+                " explicitly. Visit https://huggingface.co/tasks for more info."
+            )
+        _check_supported_task(model_id, self.task)
+        return InferenceProviderMapping(
+            provider="hf-inference", providerId=model_id, hf_model_id=model_id, task=self.task, status="live"
+        )
+
+    def _prepare_url(self, api_key: str, mapped_model: str) -> str:
+        # hf-inference provider can handle URLs (e.g. Inference Endpoints or TGI deployment)
+        if mapped_model.startswith(("http://", "https://")):
+            return mapped_model
+        return (
+            # Feature-extraction and sentence-similarity are the only cases where we handle models with several tasks.
+            f"{self.base_url}/models/{mapped_model}/pipeline/{self.task}"
+            if self.task in ("feature-extraction", "sentence-similarity")
+            # Otherwise, we use the default endpoint
+            else f"{self.base_url}/models/{mapped_model}"
+        )
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        if isinstance(inputs, bytes):
+            raise ValueError(f"Unexpected binary input for task {self.task}.")
+        if isinstance(inputs, Path):
+            raise ValueError(f"Unexpected path input for task {self.task} (got {inputs})")
+        return filter_none({"inputs": inputs, "parameters": parameters})
+
+
+class HFInferenceBinaryInputTask(HFInferenceTask):
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        return None
+
+    def _prepare_payload_as_bytes(
+        self,
+        inputs: Any,
+        parameters: dict,
+        provider_mapping_info: InferenceProviderMapping,
+        extra_payload: Optional[dict],
+    ) -> Optional[MimeBytes]:
+        parameters = filter_none(parameters)
+        extra_payload = extra_payload or {}
+        has_parameters = len(parameters) > 0 or len(extra_payload) > 0
+
+        # Raise if not a binary object or a local path or a URL.
+        if not isinstance(inputs, (bytes, Path)) and not isinstance(inputs, str):
+            raise ValueError(f"Expected binary inputs or a local path or a URL. Got {inputs}")
+
+        # Send inputs as raw content when no parameters are provided
+        if not has_parameters:
+            return _open_as_mime_bytes(inputs)
+
+        # Otherwise encode as b64
+        return MimeBytes(
+            json.dumps({"inputs": _b64_encode(inputs), "parameters": parameters, **extra_payload}).encode("utf-8"),
+            mime_type="application/json",
+        )
+
+
+class HFInferenceConversational(HFInferenceTask):
+    def __init__(self):
+        super().__init__("conversational")
+
+    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"
+
+        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 = (
+            mapped_model
+            if mapped_model.startswith(("http://", "https://"))
+            else f"{constants.INFERENCE_PROXY_TEMPLATE.format(provider='hf-inference')}/models/{mapped_model}"
+        )
+        return _build_chat_completion_url(base_url)
+
+
+def _build_chat_completion_url(model_url: str) -> str:
+    parsed = urlparse(model_url)
+    path = parsed.path.rstrip("/")
+
+    # If the path already ends with /chat/completions, we're done!
+    if path.endswith("/chat/completions"):
+        return model_url
+
+    # Append /chat/completions if not already present
+    if path.endswith("/v1"):
+        new_path = path + "/chat/completions"
+    # If path was empty or just "/", set the full path
+    elif not path:
+        new_path = "/v1/chat/completions"
+    # Append /v1/chat/completions if not already present
+    else:
+        new_path = path + "/v1/chat/completions"
+
+    # Reconstruct the URL with the new path and original query parameters.
+    new_parsed = parsed._replace(path=new_path)
+    return str(urlunparse(new_parsed))
+
+
+@lru_cache(maxsize=1)
+def _fetch_recommended_models() -> dict[str, Optional[str]]:
+    response = get_session().get(f"{constants.ENDPOINT}/api/tasks", headers=build_hf_headers())
+    hf_raise_for_status(response)
+    return {task: next(iter(details["widgetModels"]), None) for task, details in response.json().items()}
+
+
+@lru_cache(maxsize=None)
+def _check_supported_task(model: str, task: str) -> None:
+    from huggingface_hub.hf_api import HfApi
+
+    model_info = HfApi().model_info(model)
+    pipeline_tag = model_info.pipeline_tag
+    tags = model_info.tags or []
+    is_conversational = "conversational" in tags
+    if task in ("text-generation", "conversational"):
+        if pipeline_tag == "text-generation":
+            # text-generation + conversational tag -> both tasks allowed
+            if is_conversational:
+                return
+            # text-generation without conversational tag -> only text-generation allowed
+            if task == "text-generation":
+                return
+            raise ValueError(f"Model '{model}' doesn't support task '{task}'.")
+
+    if pipeline_tag == "text2text-generation":
+        if task == "text-generation":
+            return
+        raise ValueError(f"Model '{model}' doesn't support task '{task}'.")
+
+    if pipeline_tag == "image-text-to-text":
+        if is_conversational and task == "conversational":
+            return  # Only conversational allowed if tagged as conversational
+        raise ValueError("Non-conversational image-text-to-text task is not supported.")
+
+    if (
+        task in ("feature-extraction", "sentence-similarity")
+        and pipeline_tag in ("feature-extraction", "sentence-similarity")
+        and task in tags
+    ):
+        # feature-extraction and sentence-similarity are interchangeable for HF Inference
+        return
+
+    # For all other tasks, just check pipeline tag
+    if pipeline_tag != task:
+        raise ValueError(
+            f"Model '{model}' doesn't support task '{task}'. Supported tasks: '{pipeline_tag}', got: '{task}'"
+        )
+    return
+
+
+class HFInferenceFeatureExtractionTask(HFInferenceTask):
+    def __init__(self):
+        super().__init__("feature-extraction")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        if isinstance(inputs, bytes):
+            raise ValueError(f"Unexpected binary input for task {self.task}.")
+        if isinstance(inputs, Path):
+            raise ValueError(f"Unexpected path input for task {self.task} (got {inputs})")
+
+        # Parameters are sent at root-level for feature-extraction task
+        # See specs: https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/src/tasks/feature-extraction/spec/input.json
+        return {"inputs": inputs, **filter_none(parameters)}
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        if isinstance(response, bytes):
+            return _bytes_to_dict(response)
+        return response

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/hyperbolic.py

@@ -0,0 +1,47 @@
+import base64
+from typing import Any, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+from huggingface_hub.inference._providers._common import BaseConversationalTask, TaskProviderHelper, filter_none
+
+
+class HyperbolicTextToImageTask(TaskProviderHelper):
+    def __init__(self):
+        super().__init__(provider="hyperbolic", base_url="https://api.hyperbolic.xyz", 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
+        parameters = filter_none(parameters)
+        if "num_inference_steps" in parameters:
+            parameters["steps"] = parameters.pop("num_inference_steps")
+        if "guidance_scale" in parameters:
+            parameters["cfg_scale"] = parameters.pop("guidance_scale")
+        # For Hyperbolic, the width and height are required parameters
+        if "width" not in parameters:
+            parameters["width"] = 512
+        if "height" not in parameters:
+            parameters["height"] = 512
+        return {"prompt": inputs, "model_name": mapped_model, **parameters}
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        response_dict = _as_dict(response)
+        return base64.b64decode(response_dict["images"][0]["image"])
+
+
+class HyperbolicTextGenerationTask(BaseConversationalTask):
+    """
+    Special case for Hyperbolic, where text-generation task is handled as a conversational task.
+    """
+
+    def __init__(self, task: str):
+        super().__init__(
+            provider="hyperbolic",
+            base_url="https://api.hyperbolic.xyz",
+        )
+        self.task = task

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/nebius.py

@@ -0,0 +1,83 @@
+import base64
+from typing import Any, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+from huggingface_hub.inference._providers._common import (
+    BaseConversationalTask,
+    BaseTextGenerationTask,
+    TaskProviderHelper,
+    filter_none,
+)
+
+
+class NebiusTextGenerationTask(BaseTextGenerationTask):
+    def __init__(self):
+        super().__init__(provider="nebius", base_url="https://api.studio.nebius.ai")
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        output = _as_dict(response)["choices"][0]
+        return {
+            "generated_text": output["text"],
+            "details": {
+                "finish_reason": output.get("finish_reason"),
+                "seed": output.get("seed"),
+            },
+        }
+
+
+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):
+        super().__init__(task="text-to-image", provider="nebius", base_url="https://api.studio.nebius.ai")
+
+    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
+        parameters = filter_none(parameters)
+        if "guidance_scale" in parameters:
+            parameters.pop("guidance_scale")
+        if parameters.get("response_format") not in ("b64_json", "url"):
+            parameters["response_format"] = "b64_json"
+
+        return {"prompt": inputs, **parameters, "model": mapped_model}
+
+    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]

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/novita.py

@@ -0,0 +1,69 @@
+from typing import Any, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+from huggingface_hub.inference._providers._common import (
+    BaseConversationalTask,
+    BaseTextGenerationTask,
+    TaskProviderHelper,
+    filter_none,
+)
+from huggingface_hub.utils import get_session
+
+
+_PROVIDER = "novita"
+_BASE_URL = "https://api.novita.ai"
+
+
+class NovitaTextGenerationTask(BaseTextGenerationTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        # there is no v1/ route for novita
+        return "/v3/openai/completions"
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        output = _as_dict(response)["choices"][0]
+        return {
+            "generated_text": output["text"],
+            "details": {
+                "finish_reason": output.get("finish_reason"),
+                "seed": output.get("seed"),
+            },
+        }
+
+
+class NovitaConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        # there is no v1/ route for novita
+        return "/v3/openai/chat/completions"
+
+
+class NovitaTextToVideoTask(TaskProviderHelper):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL, task="text-to-video")
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return f"/v3/hf/{mapped_model}"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        return {"prompt": inputs, **filter_none(parameters)}
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        response_dict = _as_dict(response)
+        if not (
+            isinstance(response_dict, dict)
+            and "video" in response_dict
+            and isinstance(response_dict["video"], dict)
+            and "video_url" in response_dict["video"]
+        ):
+            raise ValueError("Expected response format: { 'video': { 'video_url': string } }")
+
+        video_url = response_dict["video"]["video_url"]
+        return get_session().get(video_url).content

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/nscale.py

@@ -0,0 +1,44 @@
+import base64
+from typing import Any, 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"])

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/openai.py

@@ -0,0 +1,25 @@
+from typing import Optional
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._providers._common import BaseConversationalTask
+
+
+class OpenAIConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider="openai", base_url="https://api.openai.com")
+
+    def _prepare_api_key(self, api_key: Optional[str]) -> str:
+        if api_key is None:
+            raise ValueError("You must provide an api_key to work with OpenAI API.")
+        if api_key.startswith("hf_"):
+            raise ValueError(
+                "OpenAI provider is not available through Hugging Face routing, please use your own OpenAI API key."
+            )
+        return api_key
+
+    def _prepare_mapping_info(self, model: Optional[str]) -> InferenceProviderMapping:
+        if model is None:
+            raise ValueError("Please provide an OpenAI model ID, e.g. `gpt-4o` or `o1`.")
+        return InferenceProviderMapping(
+            provider="openai", providerId=model, task="conversational", status="live", hf_model_id=model
+        )

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/ovhcloud.py

@@ -0,0 +1,10 @@
+from huggingface_hub.inference._providers._common import BaseConversationalTask
+
+
+_PROVIDER = "ovhcloud"
+_BASE_URL = "https://oai.endpoints.kepler.ai.cloud.ovh.net"
+
+
+class OVHcloudConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/publicai.py

@@ -0,0 +1,6 @@
+from ._common import BaseConversationalTask
+
+
+class PublicAIConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider="publicai", base_url="https://api.publicai.co")

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/replicate.py

@@ -0,0 +1,151 @@
+from typing import Any, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict, _as_url
+from huggingface_hub.inference._providers._common import TaskProviderHelper, filter_none
+from huggingface_hub.utils import get_session
+
+
+_PROVIDER = "replicate"
+_BASE_URL = "https://api.replicate.com"
+
+
+class ReplicateTask(TaskProviderHelper):
+    def __init__(self, task: str):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL, task=task)
+
+    def _prepare_headers(self, headers: dict, api_key: str) -> dict[str, Any]:
+        headers = super()._prepare_headers(headers, api_key)
+        headers["Prefer"] = "wait"
+        return headers
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        if ":" in mapped_model:
+            return "/v1/predictions"
+        return f"/v1/models/{mapped_model}/predictions"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        mapped_model = provider_mapping_info.provider_id
+        payload: dict[str, Any] = {"input": {"prompt": inputs, **filter_none(parameters)}}
+        if ":" in mapped_model:
+            version = mapped_model.split(":", 1)[1]
+            payload["version"] = version
+        return payload
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        response_dict = _as_dict(response)
+        if response_dict.get("output") is None:
+            raise TimeoutError(
+                f"Inference request timed out after 60 seconds. No output generated for model {response_dict.get('model')}"
+                "The model might be in cold state or starting up. Please try again later."
+            )
+        output_url = (
+            response_dict["output"] if isinstance(response_dict["output"], str) else response_dict["output"][0]
+        )
+        return get_session().get(output_url).content
+
+
+class ReplicateTextToImageTask(ReplicateTask):
+    def __init__(self):
+        super().__init__("text-to-image")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        payload: dict = super()._prepare_payload_as_dict(inputs, parameters, provider_mapping_info)  # type: ignore[assignment]
+        if provider_mapping_info.adapter_weights_path is not None:
+            payload["input"]["lora_weights"] = f"https://huggingface.co/{provider_mapping_info.hf_model_id}"
+        return payload
+
+
+class ReplicateTextToSpeechTask(ReplicateTask):
+    def __init__(self):
+        super().__init__("text-to-speech")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        payload: dict = super()._prepare_payload_as_dict(inputs, parameters, provider_mapping_info)  # type: ignore[assignment]
+        payload["input"]["text"] = payload["input"].pop("prompt")  # rename "prompt" to "text" for TTS
+        return payload
+
+
+class ReplicateAutomaticSpeechRecognitionTask(ReplicateTask):
+    def __init__(self) -> None:
+        super().__init__("automatic-speech-recognition")
+
+    def _prepare_payload_as_dict(
+        self,
+        inputs: Any,
+        parameters: dict,
+        provider_mapping_info: InferenceProviderMapping,
+    ) -> Optional[dict]:
+        mapped_model = provider_mapping_info.provider_id
+        audio_url = _as_url(inputs, default_mime_type="audio/wav")
+
+        payload: dict[str, Any] = {
+            "input": {
+                **{"audio": audio_url},
+                **filter_none(parameters),
+            }
+        }
+
+        if ":" in mapped_model:
+            payload["version"] = mapped_model.split(":", 1)[1]
+
+        return payload
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        response_dict = _as_dict(response)
+        output = response_dict.get("output")
+
+        if isinstance(output, str):
+            return {"text": output}
+
+        if isinstance(output, list) and output:
+            first_item = output[0]
+            if isinstance(first_item, str):
+                return {"text": first_item}
+            if isinstance(first_item, dict):
+                output = first_item
+
+        text: Optional[str] = None
+        if isinstance(output, dict):
+            transcription = output.get("transcription")
+            if isinstance(transcription, str):
+                text = transcription
+
+            translation = output.get("translation")
+            if isinstance(translation, str):
+                text = translation
+
+            txt_file = output.get("txt_file")
+            if isinstance(txt_file, str):
+                text_response = get_session().get(txt_file)
+                text_response.raise_for_status()
+                text = text_response.text
+
+        if text is not None:
+            return {"text": text}
+
+        raise ValueError("Received malformed response from Replicate automatic-speech-recognition API")
+
+
+class ReplicateImageToImageTask(ReplicateTask):
+    def __init__(self):
+        super().__init__("image-to-image")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        image_url = _as_url(inputs, default_mime_type="image/jpeg")
+
+        payload: dict[str, Any] = {"input": {"input_image": image_url, **filter_none(parameters)}}
+
+        mapped_model = provider_mapping_info.provider_id
+        if ":" in mapped_model:
+            version = mapped_model.split(":", 1)[1]
+            payload["version"] = version
+        return payload

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/sambanova.py

@@ -0,0 +1,42 @@
+from typing import Any, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+from huggingface_hub.inference._providers._common import BaseConversationalTask, TaskProviderHelper, filter_none
+
+
+class SambanovaConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider="sambanova", base_url="https://api.sambanova.ai")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        response_format_config = parameters.get("response_format")
+        if isinstance(response_format_config, dict):
+            if response_format_config.get("type") == "json_schema":
+                json_schema_config = response_format_config.get("json_schema", {})
+                strict = json_schema_config.get("strict")
+                if isinstance(json_schema_config, dict) and (strict is True or strict is None):
+                    json_schema_config["strict"] = False
+
+        payload = super()._prepare_payload_as_dict(inputs, parameters, provider_mapping_info)
+        return payload
+
+
+class SambanovaFeatureExtractionTask(TaskProviderHelper):
+    def __init__(self):
+        super().__init__(provider="sambanova", base_url="https://api.sambanova.ai", task="feature-extraction")
+
+    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]:
+        parameters = filter_none(parameters)
+        return {"input": inputs, "model": provider_mapping_info.provider_id, **parameters}
+
+    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]

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/scaleway.py

@@ -0,0 +1,28 @@
+from typing import Any, Dict, Optional, Union
+
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+
+from ._common import BaseConversationalTask, InferenceProviderMapping, TaskProviderHelper, filter_none
+
+
+class ScalewayConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider="scaleway", base_url="https://api.scaleway.ai")
+
+
+class ScalewayFeatureExtractionTask(TaskProviderHelper):
+    def __init__(self):
+        super().__init__(provider="scaleway", base_url="https://api.scaleway.ai", task="feature-extraction")
+
+    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]:
+        parameters = filter_none(parameters)
+        return {"input": inputs, "model": provider_mapping_info.provider_id, **parameters}
+
+    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]

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/together.py

@@ -0,0 +1,88 @@
+import base64
+from abc import ABC
+from typing import Any, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+from huggingface_hub.inference._providers._common import (
+    BaseConversationalTask,
+    BaseTextGenerationTask,
+    TaskProviderHelper,
+    filter_none,
+)
+
+
+_PROVIDER = "together"
+_BASE_URL = "https://api.together.xyz"
+
+
+class TogetherTask(TaskProviderHelper, ABC):
+    """Base class for Together API tasks."""
+
+    def __init__(self, task: str):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL, task=task)
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        if self.task == "text-to-image":
+            return "/v1/images/generations"
+        elif self.task == "conversational":
+            return "/v1/chat/completions"
+        elif self.task == "text-generation":
+            return "/v1/completions"
+        raise ValueError(f"Unsupported task '{self.task}' for Together API.")
+
+
+class TogetherTextGenerationTask(BaseTextGenerationTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+        output = _as_dict(response)["choices"][0]
+        return {
+            "generated_text": output["text"],
+            "details": {
+                "finish_reason": output.get("finish_reason"),
+                "seed": output.get("seed"),
+            },
+        }
+
+
+class TogetherConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    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
+
+
+class TogetherTextToImageTask(TogetherTask):
+    def __init__(self):
+        super().__init__("text-to-image")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        mapped_model = provider_mapping_info.provider_id
+        parameters = filter_none(parameters)
+        if "num_inference_steps" in parameters:
+            parameters["steps"] = parameters.pop("num_inference_steps")
+        if "guidance_scale" in parameters:
+            parameters["guidance"] = parameters.pop("guidance_scale")
+
+        return {"prompt": inputs, "response_format": "base64", **parameters, "model": mapped_model}
+
+    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"])

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/wavespeed.py

@@ -0,0 +1,138 @@
+import base64
+import time
+from abc import ABC
+from typing import Any, Optional, Union
+from urllib.parse import urlparse
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+from huggingface_hub.inference._providers._common import TaskProviderHelper, filter_none
+from huggingface_hub.utils import get_session, hf_raise_for_status
+from huggingface_hub.utils.logging import get_logger
+
+
+logger = get_logger(__name__)
+
+# Polling interval (in seconds)
+_POLLING_INTERVAL = 0.5
+
+
+class WavespeedAITask(TaskProviderHelper, ABC):
+    def __init__(self, task: str):
+        super().__init__(provider="wavespeed", base_url="https://api.wavespeed.ai", task=task)
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return f"/api/v3/{mapped_model}"
+
+    def get_response(
+        self,
+        response: Union[bytes, dict],
+        request_params: Optional[RequestParameters] = None,
+    ) -> Any:
+        response_dict = _as_dict(response)
+        data = response_dict.get("data", {})
+        result_path = data.get("urls", {}).get("get")
+
+        if not result_path:
+            raise ValueError("No result URL found in the response")
+        if request_params is None:
+            raise ValueError("A `RequestParameters` object should be provided to get responses with WaveSpeed AI.")
+
+        # Parse the request URL to determine base URL
+        parsed_url = urlparse(request_params.url)
+        # Add /wavespeed to base URL if going through HF router
+        if parsed_url.netloc == "router.huggingface.co":
+            base_url = f"{parsed_url.scheme}://{parsed_url.netloc}/wavespeed"
+        else:
+            base_url = f"{parsed_url.scheme}://{parsed_url.netloc}"
+
+        # Extract path from result_path URL
+        if isinstance(result_path, str):
+            result_url_path = urlparse(result_path).path
+        else:
+            result_url_path = result_path
+
+        result_url = f"{base_url}{result_url_path}"
+
+        logger.info("Processing request, polling for results...")
+
+        # Poll until task is completed
+        while True:
+            time.sleep(_POLLING_INTERVAL)
+            result_response = get_session().get(result_url, headers=request_params.headers)
+            hf_raise_for_status(result_response)
+
+            result = result_response.json()
+            task_result = result.get("data", {})
+            status = task_result.get("status")
+
+            if status == "completed":
+                # Get content from the first output URL
+                if not task_result.get("outputs") or len(task_result["outputs"]) == 0:
+                    raise ValueError("No output URL in completed response")
+
+                output_url = task_result["outputs"][0]
+                return get_session().get(output_url).content
+            elif status == "failed":
+                error_msg = task_result.get("error", "Task failed with no specific error message")
+                raise ValueError(f"WaveSpeed AI task failed: {error_msg}")
+            elif status in ["processing", "created"]:
+                continue
+            else:
+                raise ValueError(f"Unknown status: {status}")
+
+
+class WavespeedAITextToImageTask(WavespeedAITask):
+    def __init__(self):
+        super().__init__("text-to-image")
+
+    def _prepare_payload_as_dict(
+        self,
+        inputs: Any,
+        parameters: dict,
+        provider_mapping_info: InferenceProviderMapping,
+    ) -> Optional[dict]:
+        return {"prompt": inputs, **filter_none(parameters)}
+
+
+class WavespeedAITextToVideoTask(WavespeedAITextToImageTask):
+    def __init__(self):
+        WavespeedAITask.__init__(self, "text-to-video")
+
+
+class WavespeedAIImageToImageTask(WavespeedAITask):
+    def __init__(self):
+        super().__init__("image-to-image")
+
+    def _prepare_payload_as_dict(
+        self,
+        inputs: Any,
+        parameters: dict,
+        provider_mapping_info: InferenceProviderMapping,
+    ) -> Optional[dict]:
+        # Convert inputs to image (URL or base64)
+        if isinstance(inputs, str) and inputs.startswith(("http://", "https://")):
+            image = inputs
+        elif isinstance(inputs, str):
+            # If input is a file path, read it first
+            with open(inputs, "rb") as f:
+                file_content = f.read()
+            image_b64 = base64.b64encode(file_content).decode("utf-8")
+            image = f"data:image/jpeg;base64,{image_b64}"
+        else:
+            # If input is binary data
+            image_b64 = base64.b64encode(inputs).decode("utf-8")
+            image = f"data:image/jpeg;base64,{image_b64}"
+
+        # Extract prompt from parameters if present
+        prompt = parameters.pop("prompt", None)
+        payload = {"image": image, **filter_none(parameters)}
+        if prompt is not None:
+            payload["prompt"] = prompt
+
+        return payload
+
+
+class WavespeedAIImageToVideoTask(WavespeedAIImageToImageTask):
+    def __init__(self):
+        WavespeedAITask.__init__(self, "image-to-video")

venv/lib/python3.12/site-packages/huggingface_hub/inference/_providers/zai_org.py

@@ -0,0 +1,127 @@
+import time
+from abc import ABC
+from typing import Any, Optional, Union
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+from huggingface_hub.inference._providers._common import BaseConversationalTask, TaskProviderHelper, filter_none
+from huggingface_hub.utils import get_session
+
+
+_PROVIDER = "zai-org"
+_BASE_URL = "https://api.z.ai"
+_POLLING_INTERVAL = 5  # seconds
+_MAX_POLL_ATTEMPTS = 60
+
+
+class ZaiTask(TaskProviderHelper, ABC):
+    def __init__(self, task: str):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL, task=task)
+
+    def _prepare_headers(self, headers: dict, api_key: str) -> dict[str, Any]:
+        headers = super()._prepare_headers(headers, api_key)
+        headers["Accept-Language"] = "en-US,en"
+        headers["x-source-channel"] = "hugging_face"
+        return headers
+
+
+class ZaiConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    def _prepare_headers(self, headers: dict, api_key: str) -> dict[str, Any]:
+        headers = super()._prepare_headers(headers, api_key)
+        headers["Accept-Language"] = "en-US,en"
+        headers["x-source-channel"] = "hugging_face"
+        return headers
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/api/paas/v4/chat/completions"
+
+
+class ZaiTextToImageTask(ZaiTask):
+    """Text-to-image task for ZAI provider using async API."""
+
+    def __init__(self):
+        super().__init__("text-to-image")
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/api/paas/v4/async/images/generations"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> Optional[dict]:
+        width = parameters.pop("width", None)
+        height = parameters.pop("height", None)
+        size = None
+        if width is not None and height is not None:
+            size = f"{width}x{height}"
+
+        payload: dict[str, Any] = {
+            "model": provider_mapping_info.provider_id,
+            "prompt": inputs,
+        }
+        if size is not None:
+            payload["size"] = size
+
+        payload.update(filter_none(parameters))
+        return payload
+
+    def get_response(
+        self,
+        response: Union[bytes, dict],
+        request_params: Optional[RequestParameters] = None,
+    ) -> Any:
+        """Handle async response by polling for results."""
+        response_dict = _as_dict(response)
+
+        task_id = response_dict.get("id")
+        if task_id is None:
+            raise ValueError("No task_id in response from ZAI API")
+
+        task_status = response_dict.get("task_status")
+        if task_status == "FAIL":
+            raise ValueError(f"ZAI image generation failed for request {task_id}")
+
+        if task_status == "PROCESSING" and request_params is not None:
+            return self._poll_for_result(task_id, request_params)
+
+        return self._extract_image(response_dict)
+
+    def _poll_for_result(self, task_id: str, request_params: RequestParameters) -> bytes:
+        """Poll the async-result endpoint until completion."""
+        session = get_session()
+        base_url = request_params.url.rsplit("/api/paas/v4/async/images/generations", 1)[0]
+        poll_url = f"{base_url}/api/paas/v4/async-result/{task_id}"
+
+        for _ in range(_MAX_POLL_ATTEMPTS):
+            poll_response = session.get(poll_url, headers=request_params.headers)
+            poll_response.raise_for_status()
+            result = poll_response.json()
+
+            task_status = result.get("task_status")
+            if task_status == "SUCCESS":
+                return self._extract_image(result)
+            elif task_status == "FAIL":
+                raise ValueError(f"Zai text-to-image generation failed for request {task_id}")
+
+            time.sleep(_POLLING_INTERVAL)
+
+        raise ValueError(
+            f"Timed out while waiting for the result from Zai API - aborting after {_MAX_POLL_ATTEMPTS} attempts"
+        )
+
+    def _extract_image(self, result: dict) -> bytes:
+        """Extract and download the image from the result."""
+        image_result = result.get("image_result")
+        if not image_result or not isinstance(image_result, list) or len(image_result) == 0:
+            raise ValueError("No image_result in response from ZAI API")
+
+        image_url = image_result[0].get("url")
+        if not image_url:
+            raise ValueError("No image URL in response from ZAI API")
+
+        session = get_session()
+        image_response = session.get(image_url)
+        image_response.raise_for_status()
+        return image_response.content