Voice et bot modif

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

@@ -1,4 +1,4 @@
-from typing import Literal, Optional, Union
+from typing import Literal, Union
 
 from huggingface_hub.inference._providers.featherless_ai import (
     FeatherlessConversationalTask,
@@ -11,6 +11,7 @@ from .black_forest_labs import BlackForestLabsTextToImageTask
 from .cerebras import CerebrasConversationalTask
 from .clarifai import ClarifaiConversationalTask
 from .cohere import CohereConversationalTask
+from .deepinfra import DeepInfraConversationalTask, DeepInfraTextGenerationTask
 from .fal_ai import (
     FalAIAutomaticSpeechRecognitionTask,
     FalAIImageSegmentationTask,
@@ -37,6 +38,7 @@ from .nebius import (
 )
 from .novita import NovitaConversationalTask, NovitaTextGenerationTask, NovitaTextToVideoTask
 from .nscale import NscaleConversationalTask, NscaleTextToImageTask
+from .nvidia import NvidiaConversationalTask
 from .openai import OpenAIConversationalTask
 from .ovhcloud import OVHcloudConversationalTask
 from .publicai import PublicAIConversationalTask
@@ -49,7 +51,16 @@ from .replicate import (
 )
 from .sambanova import SambanovaConversationalTask, SambanovaFeatureExtractionTask
 from .scaleway import ScalewayConversationalTask, ScalewayFeatureExtractionTask
-from .together import TogetherConversationalTask, TogetherTextGenerationTask, TogetherTextToImageTask
+from .together import (
+    TogetherConversationalTask,
+    TogetherFeatureExtractionTask,
+    TogetherImageToImageTask,
+    TogetherImageToVideoTask,
+    TogetherTextGenerationTask,
+    TogetherTextToImageTask,
+    TogetherTextToSpeechTask,
+    TogetherTextToVideoTask,
+)
 from .wavespeed import (
     WavespeedAIImageToImageTask,
     WavespeedAIImageToVideoTask,
@@ -67,6 +78,7 @@ PROVIDER_T = Literal[
     "cerebras",
     "clarifai",
     "cohere",
+    "deepinfra",
     "fal-ai",
     "featherless-ai",
     "fireworks-ai",
@@ -76,6 +88,7 @@ PROVIDER_T = Literal[
     "nebius",
     "novita",
     "nscale",
+    "nvidia",
     "openai",
     "ovhcloud",
     "publicai",
@@ -104,6 +117,10 @@ PROVIDERS: dict[PROVIDER_T, dict[str, TaskProviderHelper]] = {
     "cohere": {
         "conversational": CohereConversationalTask(),
     },
+    "deepinfra": {
+        "conversational": DeepInfraConversationalTask(),
+        "text-generation": DeepInfraTextGenerationTask(),
+    },
     "fal-ai": {
         "automatic-speech-recognition": FalAIAutomaticSpeechRecognitionTask(),
         "text-to-image": FalAITextToImageTask(),
@@ -171,6 +188,9 @@ PROVIDERS: dict[PROVIDER_T, dict[str, TaskProviderHelper]] = {
         "conversational": NscaleConversationalTask(),
         "text-to-image": NscaleTextToImageTask(),
     },
+    "nvidia": {
+        "conversational": NvidiaConversationalTask(),
+    },
     "openai": {
         "conversational": OpenAIConversationalTask(),
     },
@@ -196,9 +216,14 @@ PROVIDERS: dict[PROVIDER_T, dict[str, TaskProviderHelper]] = {
         "feature-extraction": ScalewayFeatureExtractionTask(),
     },
     "together": {
-        "text-to-image": TogetherTextToImageTask(),
         "conversational": TogetherConversationalTask(),
+        "feature-extraction": TogetherFeatureExtractionTask(),
+        "image-to-image": TogetherImageToImageTask(),
+        "image-to-video": TogetherImageToVideoTask(),
         "text-generation": TogetherTextGenerationTask(),
+        "text-to-image": TogetherTextToImageTask(),
+        "text-to-speech": TogetherTextToSpeechTask(),
+        "text-to-video": TogetherTextToVideoTask(),
     },
     "wavespeed": {
         "text-to-image": WavespeedAITextToImageTask(),
@@ -213,9 +238,7 @@ PROVIDERS: dict[PROVIDER_T, dict[str, TaskProviderHelper]] = {
 }
 
 
-def get_provider_helper(
-    provider: Optional[PROVIDER_OR_POLICY_T], task: str, model: Optional[str]
-) -> TaskProviderHelper:
+def get_provider_helper(provider: PROVIDER_OR_POLICY_T | None, task: str, model: str | None) -> TaskProviderHelper:
     """Get provider helper instance by name and task.
 
     Args:

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

@@ -1,5 +1,5 @@
 from functools import lru_cache
-from typing import Any, Optional, Union, overload
+from typing import Any, overload
 
 from huggingface_hub import constants
 from huggingface_hub.hf_api import InferenceProviderMapping
@@ -25,6 +25,7 @@ HARDCODED_MODEL_INFERENCE_MAPPING: dict[str, dict[str, InferenceProviderMapping]
     "cerebras": {},
     "cohere": {},
     "clarifai": {},
+    "deepinfra": {},
     "fal-ai": {},
     "fireworks-ai": {},
     "groq": {},
@@ -32,6 +33,7 @@ HARDCODED_MODEL_INFERENCE_MAPPING: dict[str, dict[str, InferenceProviderMapping]
     "hyperbolic": {},
     "nebius": {},
     "nscale": {},
+    "nvidia": {},
     "ovhcloud": {},
     "replicate": {},
     "sambanova": {},
@@ -48,7 +50,7 @@ def filter_none(obj: dict[str, Any]) -> dict[str, Any]: ...
 def filter_none(obj: list[Any]) -> list[Any]: ...
 
 
-def filter_none(obj: Union[dict[str, Any], list[Any]]) -> Union[dict[str, Any], list[Any]]:
+def filter_none(obj: dict[str, Any] | list[Any]) -> dict[str, Any] | list[Any]:
     if isinstance(obj, dict):
         cleaned: dict[str, Any] = {}
         for k, v in obj.items():
@@ -79,9 +81,9 @@ class TaskProviderHelper:
         inputs: Any,
         parameters: dict[str, Any],
         headers: dict,
-        model: Optional[str],
-        api_key: Optional[str],
-        extra_payload: Optional[dict[str, Any]] = None,
+        model: str | None,
+        api_key: str | None,
+        extra_payload: dict[str, Any] | None = None,
     ) -> RequestParameters:
         """
         Prepare the request to be sent to the provider.
@@ -128,8 +130,8 @@ class TaskProviderHelper:
 
     def get_response(
         self,
-        response: Union[bytes, dict],
-        request_params: Optional[RequestParameters] = None,
+        response: bytes | dict,
+        request_params: RequestParameters | None = None,
     ) -> Any:
         """
         Return the response in the expected format.
@@ -137,7 +139,7 @@ class TaskProviderHelper:
         Override this method in subclasses for customized response handling."""
         return response
 
-    def _prepare_api_key(self, api_key: Optional[str]) -> str:
+    def _prepare_api_key(self, api_key: str | None) -> str:
         """Return the API key to use for the request.
 
         Usually not overwritten in subclasses."""
@@ -149,7 +151,7 @@ class TaskProviderHelper:
             )
         return api_key
 
-    def _prepare_mapping_info(self, model: Optional[str]) -> InferenceProviderMapping:
+    def _prepare_mapping_info(self, model: str | None) -> InferenceProviderMapping:
         """Return the mapped model ID to use for the request.
 
         Usually not overwritten in subclasses."""
@@ -187,7 +189,7 @@ class TaskProviderHelper:
         return provider_mapping
 
     def _normalize_headers(
-        self, headers: dict[str, Any], payload: Optional[dict[str, Any]], data: Optional[MimeBytes]
+        self, headers: dict[str, Any], payload: dict[str, Any] | None, data: MimeBytes | None
     ) -> dict[str, Any]:
         """Normalize the headers to use for the request.
 
@@ -237,7 +239,7 @@ class TaskProviderHelper:
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         """Return the payload to use for the request, as a dict.
 
         Override this method in subclasses for customized payloads.
@@ -250,8 +252,8 @@ class TaskProviderHelper:
         inputs: Any,
         parameters: dict,
         provider_mapping_info: InferenceProviderMapping,
-        extra_payload: Optional[dict],
-    ) -> Optional[MimeBytes]:
+        extra_payload: dict | None,
+    ) -> MimeBytes | None:
         """Return the body to use for the request, as bytes.
 
         Override this method in subclasses for customized body data.
@@ -274,10 +276,10 @@ class BaseConversationalTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self,
-        inputs: list[Union[dict, ChatCompletionInputMessage]],
+        inputs: list[dict | ChatCompletionInputMessage],
         parameters: dict,
         provider_mapping_info: InferenceProviderMapping,
-    ) -> Optional[dict]:
+    ) -> dict | None:
         return filter_none({"messages": inputs, **parameters, "model": provider_mapping_info.provider_id})
 
 
@@ -302,7 +304,7 @@ class AutoRouterConversationalTask(BaseConversationalTask):
         else:
             return self.base_url  # No `/auto` suffix in the URL
 
-    def _prepare_mapping_info(self, model: Optional[str]) -> InferenceProviderMapping:
+    def _prepare_mapping_info(self, model: str | None) -> 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.
@@ -333,7 +335,7 @@ class BaseTextGenerationTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         return filter_none({"prompt": inputs, **parameters, "model": provider_mapping_info.provider_id})
 
 

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

@@ -1,5 +1,5 @@
 import time
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._common import RequestParameters, _as_dict
@@ -30,7 +30,7 @@ class BlackForestLabsTextToImageTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         parameters = filter_none(parameters)
         if "num_inference_steps" in parameters:
             parameters["steps"] = parameters.pop("num_inference_steps")
@@ -39,7 +39,7 @@ class BlackForestLabsTextToImageTask(TaskProviderHelper):
 
         return {"prompt": inputs, **parameters}
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         """
         Polling mechanism for Black Forest Labs since the API is asynchronous.
         """

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

@@ -1,4 +1,4 @@
-from typing import Any, Optional
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 
@@ -18,13 +18,13 @@ class CohereConversationalTask(BaseConversationalTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         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]
+                payload["response_format"] = {  # type: ignore
                     "type": "json_object",
                     "schema": json_schema_details["schema"],
                 }

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

@@ -0,0 +1,44 @@
+from typing import Any
+
+from huggingface_hub.hf_api import InferenceProviderMapping
+from huggingface_hub.inference._common import RequestParameters, _as_dict
+
+from ._common import BaseConversationalTask, BaseTextGenerationTask, filter_none
+
+
+_PROVIDER = "deepinfra"
+_BASE_URL = "https://api.deepinfra.com"
+
+
+class DeepInfraTextGenerationTask(BaseTextGenerationTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/v1/openai/completions"
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> dict | None:
+        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: bytes | dict, request_params: RequestParameters | None = 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 DeepInfraConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/v1/openai/chat/completions"

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

@@ -1,7 +1,7 @@
 import base64
 import time
 from abc import ABC
-from typing import Any, Optional, Union
+from typing import Any
 from urllib.parse import urlparse
 
 from huggingface_hub import constants
@@ -50,8 +50,8 @@ class FalAIQueueTask(TaskProviderHelper, ABC):
 
     def get_response(
         self,
-        response: Union[bytes, dict],
-        request_params: Optional[RequestParameters] = None,
+        response: bytes | dict,
+        request_params: RequestParameters | None = None,
     ) -> Any:
         response_dict = _as_dict(response)
 
@@ -92,7 +92,7 @@ class FalAIAutomaticSpeechRecognitionTask(FalAITask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         if isinstance(inputs, str) and inputs.startswith(("http://", "https://")):
             # If input is a URL, pass it directly
             audio_url = inputs
@@ -108,7 +108,7 @@ class FalAIAutomaticSpeechRecognitionTask(FalAITask):
 
         return {"audio_url": audio_url, **filter_none(parameters)}
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = 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)}.")
@@ -121,7 +121,7 @@ class FalAITextToImageTask(FalAITask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         payload: dict[str, Any] = {
             "prompt": inputs,
             **filter_none(parameters),
@@ -145,7 +145,7 @@ class FalAITextToImageTask(FalAITask):
 
         return payload
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         url = _as_dict(response)["images"][0]["url"]
         return get_session().get(url).content
 
@@ -156,10 +156,10 @@ class FalAITextToSpeechTask(FalAITask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         return {"text": inputs, **filter_none(parameters)}
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         url = _as_dict(response)["audio"]["url"]
         return get_session().get(url).content
 
@@ -170,13 +170,13 @@ class FalAITextToVideoTask(FalAIQueueTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         return {"prompt": inputs, **filter_none(parameters)}
 
     def get_response(
         self,
-        response: Union[bytes, dict],
-        request_params: Optional[RequestParameters] = None,
+        response: bytes | dict,
+        request_params: RequestParameters | None = None,
     ) -> Any:
         output = super().get_response(response, request_params)
         url = _as_dict(output)["video"]["url"]
@@ -189,12 +189,13 @@ class FalAIImageToImageTask(FalAIQueueTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         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,
+            "image_urls": [image_url],
             **filter_none(parameters),
         }
         if provider_mapping_info.adapter_weights_path is not None:
@@ -209,8 +210,8 @@ class FalAIImageToImageTask(FalAIQueueTask):
 
     def get_response(
         self,
-        response: Union[bytes, dict],
-        request_params: Optional[RequestParameters] = None,
+        response: bytes | dict,
+        request_params: RequestParameters | None = None,
     ) -> Any:
         output = super().get_response(response, request_params)
         url = _as_dict(output)["images"][0]["url"]
@@ -223,7 +224,7 @@ class FalAIImageToVideoTask(FalAIQueueTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         image_url = _as_url(inputs, default_mime_type="image/jpeg")
         payload: dict[str, Any] = {
             "image_url": image_url,
@@ -240,8 +241,8 @@ class FalAIImageToVideoTask(FalAIQueueTask):
 
     def get_response(
         self,
-        response: Union[bytes, dict],
-        request_params: Optional[RequestParameters] = None,
+        response: bytes | dict,
+        request_params: RequestParameters | None = None,
     ) -> Any:
         output = super().get_response(response, request_params)
         url = _as_dict(output)["video"]["url"]
@@ -254,7 +255,7 @@ class FalAIImageSegmentationTask(FalAIQueueTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         image_url = _as_url(inputs, default_mime_type="image/png")
         payload: dict[str, Any] = {
             "image_url": image_url,
@@ -265,8 +266,8 @@ class FalAIImageSegmentationTask(FalAIQueueTask):
 
     def get_response(
         self,
-        response: Union[bytes, dict],
-        request_params: Optional[RequestParameters] = None,
+        response: bytes | dict,
+        request_params: RequestParameters | None = None,
     ) -> Any:
         result = super().get_response(response, request_params)
         result_dict = _as_dict(result)

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

@@ -1,4 +1,4 @@
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._common import RequestParameters, _as_dict
@@ -16,13 +16,13 @@ class FeatherlessTextGenerationTask(BaseTextGenerationTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         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:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         output = _as_dict(response)["choices"][0]
         return {
             "generated_text": output["text"],

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

@@ -1,4 +1,4 @@
-from typing import Any, Optional
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 
@@ -14,13 +14,13 @@ class FireworksAIConversationalTask(BaseConversationalTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         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]
+                payload["response_format"] = {  # type: ignore
                     "type": "json_object",
                     "schema": json_schema_details["schema"],
                 }

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

@@ -1,7 +1,7 @@
 import json
 from functools import lru_cache
 from pathlib import Path
-from typing import Any, Optional, Union
+from typing import Any
 from urllib.parse import urlparse, urlunparse
 
 from huggingface_hub import constants
@@ -27,11 +27,11 @@ class HFInferenceTask(TaskProviderHelper):
             task=task,
         )
 
-    def _prepare_api_key(self, api_key: Optional[str]) -> str:
+    def _prepare_api_key(self, api_key: str | None) -> str:
         # special case: for HF Inference we allow not providing an API key
-        return api_key or get_token()  # type: ignore[return-value]
+        return api_key or get_token()  # type: ignore
 
-    def _prepare_mapping_info(self, model: Optional[str]) -> InferenceProviderMapping:
+    def _prepare_mapping_info(self, model: str | None) -> 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"
@@ -61,7 +61,7 @@ class HFInferenceTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         if isinstance(inputs, bytes):
             raise ValueError(f"Unexpected binary input for task {self.task}.")
         if isinstance(inputs, Path):
@@ -72,7 +72,7 @@ class HFInferenceTask(TaskProviderHelper):
 class HFInferenceBinaryInputTask(HFInferenceTask):
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         return None
 
     def _prepare_payload_as_bytes(
@@ -80,8 +80,8 @@ class HFInferenceBinaryInputTask(HFInferenceTask):
         inputs: Any,
         parameters: dict,
         provider_mapping_info: InferenceProviderMapping,
-        extra_payload: Optional[dict],
-    ) -> Optional[MimeBytes]:
+        extra_payload: dict | None,
+    ) -> MimeBytes | None:
         parameters = filter_none(parameters)
         extra_payload = extra_payload or {}
         has_parameters = len(parameters) > 0 or len(extra_payload) > 0
@@ -107,7 +107,7 @@ class HFInferenceConversational(HFInferenceTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         payload = filter_none(parameters)
         mapped_model = provider_mapping_info.provider_id
         payload_model = parameters.get("model") or mapped_model
@@ -156,7 +156,7 @@ def _build_chat_completion_url(model_url: str) -> str:
 
 
 @lru_cache(maxsize=1)
-def _fetch_recommended_models() -> dict[str, Optional[str]]:
+def _fetch_recommended_models() -> dict[str, str | None]:
     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()}
@@ -212,7 +212,7 @@ class HFInferenceFeatureExtractionTask(HFInferenceTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         if isinstance(inputs, bytes):
             raise ValueError(f"Unexpected binary input for task {self.task}.")
         if isinstance(inputs, Path):
@@ -222,7 +222,7 @@ class HFInferenceFeatureExtractionTask(HFInferenceTask):
         # 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:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = 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

@@ -1,5 +1,5 @@
 import base64
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._common import RequestParameters, _as_dict
@@ -15,7 +15,7 @@ class HyperbolicTextToImageTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         mapped_model = provider_mapping_info.provider_id
         parameters = filter_none(parameters)
         if "num_inference_steps" in parameters:
@@ -29,7 +29,7 @@ class HyperbolicTextToImageTask(TaskProviderHelper):
             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:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         response_dict = _as_dict(response)
         return base64.b64decode(response_dict["images"][0]["image"])
 

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

@@ -1,5 +1,5 @@
 import base64
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._common import RequestParameters, _as_dict
@@ -15,7 +15,7 @@ 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:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         output = _as_dict(response)["choices"][0]
         return {
             "generated_text": output["text"],
@@ -32,13 +32,13 @@ class NebiusConversationalTask(BaseConversationalTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         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]
+                payload["guided_json"] = json_schema_details["schema"]  # type: ignore
         return payload
 
 
@@ -51,7 +51,7 @@ class NebiusTextToImageTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         mapped_model = provider_mapping_info.provider_id
         parameters = filter_none(parameters)
         if "guidance_scale" in parameters:
@@ -61,7 +61,7 @@ class NebiusTextToImageTask(TaskProviderHelper):
 
         return {"prompt": inputs, **parameters, "model": mapped_model}
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         response_dict = _as_dict(response)
         return base64.b64decode(response_dict["data"][0]["b64_json"])
 
@@ -75,9 +75,9 @@ class NebiusFeatureExtractionTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         return {"input": inputs, "model": provider_mapping_info.provider_id}
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = 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

@@ -1,4 +1,4 @@
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._common import RequestParameters, _as_dict
@@ -23,7 +23,7 @@ class NovitaTextGenerationTask(BaseTextGenerationTask):
         # 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:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         output = _as_dict(response)["choices"][0]
         return {
             "generated_text": output["text"],
@@ -52,10 +52,10 @@ class NovitaTextToVideoTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         return {"prompt": inputs, **filter_none(parameters)}
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         response_dict = _as_dict(response)
         if not (
             isinstance(response_dict, dict)

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

@@ -1,5 +1,5 @@
 import base64
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._common import RequestParameters, _as_dict
@@ -21,7 +21,7 @@ class NscaleTextToImageTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         mapped_model = provider_mapping_info.provider_id
         # Combine all parameters except inputs and parameters
         parameters = filter_none(parameters)
@@ -39,6 +39,6 @@ class NscaleTextToImageTask(TaskProviderHelper):
         }
         return payload
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = 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/nvidia.py

@@ -0,0 +1,6 @@
+from huggingface_hub.inference._providers._common import BaseConversationalTask
+
+
+class NvidiaConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider="nvidia", base_url="https://integrate.api.nvidia.com")

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

@@ -1,5 +1,3 @@
-from typing import Optional
-
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._providers._common import BaseConversationalTask
 
@@ -8,7 +6,7 @@ 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:
+    def _prepare_api_key(self, api_key: str | None) -> str:
         if api_key is None:
             raise ValueError("You must provide an api_key to work with OpenAI API.")
         if api_key.startswith("hf_"):
@@ -17,7 +15,7 @@ class OpenAIConversationalTask(BaseConversationalTask):
             )
         return api_key
 
-    def _prepare_mapping_info(self, model: Optional[str]) -> InferenceProviderMapping:
+    def _prepare_mapping_info(self, model: str | None) -> InferenceProviderMapping:
         if model is None:
             raise ValueError("Please provide an OpenAI model ID, e.g. `gpt-4o` or `o1`.")
         return InferenceProviderMapping(

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

@@ -1,4 +1,4 @@
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._common import RequestParameters, _as_dict, _as_url
@@ -26,7 +26,7 @@ class ReplicateTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         mapped_model = provider_mapping_info.provider_id
         payload: dict[str, Any] = {"input": {"prompt": inputs, **filter_none(parameters)}}
         if ":" in mapped_model:
@@ -34,7 +34,7 @@ class ReplicateTask(TaskProviderHelper):
             payload["version"] = version
         return payload
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         response_dict = _as_dict(response)
         if response_dict.get("output") is None:
             raise TimeoutError(
@@ -53,8 +53,8 @@ class ReplicateTextToImageTask(ReplicateTask):
 
     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]
+    ) -> dict | None:
+        payload: dict = super()._prepare_payload_as_dict(inputs, parameters, provider_mapping_info)  # type: ignore
         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
@@ -66,8 +66,8 @@ class ReplicateTextToSpeechTask(ReplicateTask):
 
     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]
+    ) -> dict | None:
+        payload: dict = super()._prepare_payload_as_dict(inputs, parameters, provider_mapping_info)  # type: ignore
         payload["input"]["text"] = payload["input"].pop("prompt")  # rename "prompt" to "text" for TTS
         return payload
 
@@ -81,7 +81,7 @@ class ReplicateAutomaticSpeechRecognitionTask(ReplicateTask):
         inputs: Any,
         parameters: dict,
         provider_mapping_info: InferenceProviderMapping,
-    ) -> Optional[dict]:
+    ) -> dict | None:
         mapped_model = provider_mapping_info.provider_id
         audio_url = _as_url(inputs, default_mime_type="audio/wav")
 
@@ -97,7 +97,7 @@ class ReplicateAutomaticSpeechRecognitionTask(ReplicateTask):
 
         return payload
 
-    def get_response(self, response: Union[bytes, dict], request_params: Optional[RequestParameters] = None) -> Any:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         response_dict = _as_dict(response)
         output = response_dict.get("output")
 
@@ -111,7 +111,7 @@ class ReplicateAutomaticSpeechRecognitionTask(ReplicateTask):
             if isinstance(first_item, dict):
                 output = first_item
 
-        text: Optional[str] = None
+        text: str | None = None
         if isinstance(output, dict):
             transcription = output.get("transcription")
             if isinstance(transcription, str):
@@ -139,10 +139,19 @@ class ReplicateImageToImageTask(ReplicateTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         image_url = _as_url(inputs, default_mime_type="image/jpeg")
 
-        payload: dict[str, Any] = {"input": {"input_image": image_url, **filter_none(parameters)}}
+        # Different Replicate models expect the image in different keys
+        payload: dict[str, Any] = {
+            "input": {
+                "image": image_url,
+                "images": [image_url],
+                "input_image": image_url,
+                "input_images": [image_url],
+                **filter_none(parameters),
+            }
+        }
 
         mapped_model = provider_mapping_info.provider_id
         if ":" in mapped_model:

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

@@ -1,4 +1,4 @@
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._common import RequestParameters, _as_dict
@@ -11,7 +11,7 @@ class SambanovaConversationalTask(BaseConversationalTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         response_format_config = parameters.get("response_format")
         if isinstance(response_format_config, dict):
             if response_format_config.get("type") == "json_schema":
@@ -33,10 +33,10 @@ class SambanovaFeatureExtractionTask(TaskProviderHelper):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         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:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = 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

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Optional, Union
+from typing import Any
 
 from huggingface_hub.inference._common import RequestParameters, _as_dict
 
@@ -18,11 +18,11 @@ class ScalewayFeatureExtractionTask(TaskProviderHelper):
         return "/v1/embeddings"
 
     def _prepare_payload_as_dict(
-        self, inputs: Any, parameters: Dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[Dict]:
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> dict | None:
         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:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = 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

@@ -1,20 +1,38 @@
 import base64
+import time
 from abc import ABC
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
-from huggingface_hub.inference._common import RequestParameters, _as_dict
+from huggingface_hub.inference._common import (
+    RequestParameters,
+    _as_dict,
+    _as_url,
+)
 from huggingface_hub.inference._providers._common import (
     BaseConversationalTask,
     BaseTextGenerationTask,
     TaskProviderHelper,
     filter_none,
 )
+from huggingface_hub.utils import get_session, hf_raise_for_status, logging
 
 
+logger = logging.get_logger(__name__)
+
 _PROVIDER = "together"
 _BASE_URL = "https://api.together.xyz"
 
+# Polling interval for async video generation (in seconds).
+_VIDEO_POLLING_INTERVAL = 2.0
+
+# Upper bound on status polls (initial response may already be terminal; each further poll is one attempt).
+_VIDEO_MAX_POLL_ATTEMPTS = 150  # ~5 minutes at _VIDEO_POLLING_INTERVAL
+
+# Job statuses that mean "keep polling". Together returns "queued" before transitioning to
+# "in_progress", so we must treat both as pending.
+_VIDEO_PENDING_STATUSES = {"queued", "in_progress"}
+
 
 class TogetherTask(TaskProviderHelper, ABC):
     """Base class for Together API tasks."""
@@ -23,12 +41,16 @@ class TogetherTask(TaskProviderHelper, ABC):
         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"
+        match self.task:
+            case "text-to-image" | "image-to-image":
+                return "/v1/images/generations"
+            case "text-to-speech":
+                return "/v1/audio/speech"
+            case "feature-extraction":
+                return "/v1/embeddings"
+            case "text-to-video" | "image-to-video":
+                # Video creation lives under /v2 (see https://docs.together.ai/reference/create-videos).
+                return "/v2/videos"
         raise ValueError(f"Unsupported task '{self.task}' for Together API.")
 
 
@@ -36,7 +58,7 @@ 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:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         output = _as_dict(response)["choices"][0]
         return {
             "generated_text": output["text"],
@@ -53,17 +75,23 @@ class TogetherConversationalTask(BaseConversationalTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         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"],
-                }
-
+        if payload is None:
+            return None
+        # Together accepts response_format `{type: "json_schema", schema: <schema>}` (flattened),
+        # so unwrap the OpenAI-style `{type: "json_schema", json_schema: {schema}}` envelope.
+        response_format = payload.get("response_format")
+        if (
+            isinstance(response_format, dict)
+            and response_format.get("type") == "json_schema"
+            and isinstance(response_format.get("json_schema"), dict)
+            and "schema" in response_format["json_schema"]
+        ):
+            payload["response_format"] = {
+                "type": "json_schema",
+                "schema": response_format["json_schema"]["schema"],
+            }
         return payload
 
 
@@ -73,16 +101,191 @@ class TogetherTextToImageTask(TogetherTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         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:
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
         response_dict = _as_dict(response)
         return base64.b64decode(response_dict["data"][0]["b64_json"])
+
+
+class TogetherImageToImageTask(TogetherTask):
+    def __init__(self):
+        super().__init__("image-to-image")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> dict | None:
+        mapped_model = provider_mapping_info.provider_id
+        image_url = _as_url(inputs, default_mime_type="image/jpeg")
+
+        # Filter `None` values first: the client always passes `"prompt": None` when the user
+        # omits the argument, so popping before filtering would yield `None` instead of the
+        # `""` default and send `"prompt": null` to Together (rejected by Flux Kontext).
+        parameters = filter_none(parameters)
+        prompt = parameters.pop("prompt", "")
+        if "num_inference_steps" in parameters:
+            parameters["steps"] = parameters.pop("num_inference_steps")
+
+        # Together exposes two mutually-exclusive image inputs (see
+        # https://docs.together.ai/docs/image-to-image): FLUX.1 Kontext only accepts
+        # `image_url`; FLUX.2 [dev] and Google models (Gemini 3 Pro Image, Flash Image
+        # 2.5) only accept `reference_images`. FLUX.2 [pro]/[flex] accept either but
+        # `reference_images` is the documented default. Use `image_url` only for
+        # FLUX.1 Kontext models and `reference_images` for everything else.
+        lowered = mapped_model.lower()
+        use_image_url = "kontext" in lowered and "flux.1" in lowered
+        image_field: dict[str, Any] = {"image_url": image_url} if use_image_url else {"reference_images": [image_url]}
+        return {
+            "prompt": prompt,
+            **image_field,
+            "response_format": "base64",
+            **parameters,
+            "model": mapped_model,
+        }
+
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
+        response_dict = _as_dict(response)
+        return base64.b64decode(response_dict["data"][0]["b64_json"])
+
+
+class TogetherFeatureExtractionTask(TogetherTask):
+    def __init__(self):
+        super().__init__("feature-extraction")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> dict | None:
+        return {
+            "input": inputs,
+            "model": provider_mapping_info.provider_id,
+            **filter_none(parameters),
+        }
+
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
+        return [item["embedding"] for item in _as_dict(response)["data"]]
+
+
+class TogetherTextToSpeechTask(TogetherTask):
+    def __init__(self):
+        super().__init__("text-to-speech")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> dict | None:
+        # `voice` is required by the Together API and is model-specific
+        # (see https://docs.together.ai/docs/text-to-speech#supported-voices),
+        # so we don't set a default and let the API surface a clear error if missing.
+        return {
+            "input": inputs,
+            "model": provider_mapping_info.provider_id,
+            **filter_none(parameters),
+        }
+
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
+        if isinstance(response, bytes):
+            return response
+        raise ValueError(f"Expected raw audio bytes for text-to-speech, got {type(response).__name__}.")
+
+
+def _normalize_video_parameters(parameters: dict) -> dict:
+    """Map HF inference-client conventions onto Together's video API parameter names."""
+    parameters = filter_none(parameters)
+    if "num_inference_steps" in parameters:
+        parameters["steps"] = parameters.pop("num_inference_steps")
+    if "target_size" in parameters:
+        target_size = parameters.pop("target_size")
+        if "width" in target_size:
+            parameters["width"] = target_size["width"]
+        if "height" in target_size:
+            parameters["height"] = target_size["height"]
+    return parameters
+
+
+class TogetherVideoTask(TogetherTask, ABC):
+    """Base class for Together's asynchronous video generation tasks."""
+
+    def get_response(self, response: bytes | dict, request_params: RequestParameters | None = None) -> Any:
+        if request_params is None:
+            raise ValueError("A `RequestParameters` object is required to poll Together video jobs.")
+
+        job = _as_dict(response)
+        job_id = job.get("id")
+        if not job_id:
+            raise ValueError("No job ID found in Together video generation response.")
+
+        # Status polling lives at the same /v2/videos URL with the job ID appended.
+        status_url = f"{request_params.url}/{job_id}"
+
+        logger.info("Generating video, polling for completion...")
+        # Together usually returns `status: "queued"` on the initial POST, but the field is
+        # optional per the spec — treat a missing status as "still pending" and poll, rather
+        # than falling through to the "unexpected status" error below.
+        status = job.get("status")
+        for _ in range(_VIDEO_MAX_POLL_ATTEMPTS):
+            if status is not None and status not in _VIDEO_PENDING_STATUSES:
+                break
+            time.sleep(_VIDEO_POLLING_INTERVAL)
+            status_response = get_session().get(status_url, headers=request_params.headers)
+            hf_raise_for_status(status_response)
+            job = status_response.json()
+            status = job.get("status")
+            if status is not None and status not in _VIDEO_PENDING_STATUSES:
+                break
+        else:
+            raise ValueError(
+                "Timed out while waiting for Together video generation "
+                f"— aborting after {_VIDEO_MAX_POLL_ATTEMPTS} status polls"
+            )
+
+        if status == "failed":
+            error = job.get("error") or {}
+            raise RuntimeError(f"Together video generation failed: {error.get('message') or 'Unknown error'}")
+        if status != "completed":
+            raise RuntimeError(f"Unexpected Together video job status: {status!r}")
+
+        video_url = (job.get("outputs") or {}).get("video_url")
+        if not video_url:
+            raise ValueError("No video URL found in completed Together video job.")
+
+        video_response = get_session().get(video_url)
+        hf_raise_for_status(video_response)
+        return video_response.content
+
+
+class TogetherTextToVideoTask(TogetherVideoTask):
+    def __init__(self):
+        super().__init__("text-to-video")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> dict | None:
+        return {
+            "prompt": inputs,
+            "model": provider_mapping_info.provider_id,
+            **_normalize_video_parameters(parameters),
+        }
+
+
+class TogetherImageToVideoTask(TogetherVideoTask):
+    def __init__(self):
+        super().__init__("image-to-video")
+
+    def _prepare_payload_as_dict(
+        self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
+    ) -> dict | None:
+        # Together expects each keyframe as `{input_image, frame: "first" | "last"}`
+        # for i2v models. See https://docs.together.ai/docs/inference/videos/reference-and-keyframes.
+        # Note: `input_image` accepts a data URL or an HTTP(S) URL but the field is capped
+        # at ~60KB — users with larger inputs should host the image and pass `frame_images`
+        # directly via `extra_body`.
+        return {
+            "model": provider_mapping_info.provider_id,
+            "frame_images": [{"input_image": _as_url(inputs, default_mime_type="image/png"), "frame": "first"}],
+            **_normalize_video_parameters(parameters),
+        }

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

@@ -1,7 +1,7 @@
 import base64
 import time
 from abc import ABC
-from typing import Any, Optional, Union
+from typing import Any
 from urllib.parse import urlparse
 
 from huggingface_hub.hf_api import InferenceProviderMapping
@@ -26,8 +26,8 @@ class WavespeedAITask(TaskProviderHelper, ABC):
 
     def get_response(
         self,
-        response: Union[bytes, dict],
-        request_params: Optional[RequestParameters] = None,
+        response: bytes | dict,
+        request_params: RequestParameters | None = None,
     ) -> Any:
         response_dict = _as_dict(response)
         data = response_dict.get("data", {})
@@ -91,7 +91,7 @@ class WavespeedAITextToImageTask(WavespeedAITask):
         inputs: Any,
         parameters: dict,
         provider_mapping_info: InferenceProviderMapping,
-    ) -> Optional[dict]:
+    ) -> dict | None:
         return {"prompt": inputs, **filter_none(parameters)}
 
 
@@ -109,7 +109,7 @@ class WavespeedAIImageToImageTask(WavespeedAITask):
         inputs: Any,
         parameters: dict,
         provider_mapping_info: InferenceProviderMapping,
-    ) -> Optional[dict]:
+    ) -> dict | None:
         # Convert inputs to image (URL or base64)
         if isinstance(inputs, str) and inputs.startswith(("http://", "https://")):
             image = inputs

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

@@ -1,6 +1,6 @@
 import time
 from abc import ABC
-from typing import Any, Optional, Union
+from typing import Any
 
 from huggingface_hub.hf_api import InferenceProviderMapping
 from huggingface_hub.inference._common import RequestParameters, _as_dict
@@ -50,7 +50,7 @@ class ZaiTextToImageTask(ZaiTask):
 
     def _prepare_payload_as_dict(
         self, inputs: Any, parameters: dict, provider_mapping_info: InferenceProviderMapping
-    ) -> Optional[dict]:
+    ) -> dict | None:
         width = parameters.pop("width", None)
         height = parameters.pop("height", None)
         size = None
@@ -69,8 +69,8 @@ class ZaiTextToImageTask(ZaiTask):
 
     def get_response(
         self,
-        response: Union[bytes, dict],
-        request_params: Optional[RequestParameters] = None,
+        response: bytes | dict,
+        request_params: RequestParameters | None = None,
     ) -> Any:
         """Handle async response by polling for results."""
         response_dict = _as_dict(response)