fix(videos): re-stamp created video id with resolved model_id so status/content resolve per-model key

litellm/proxy/video_endpoints/endpoints.py

@@ -1,6 +1,6 @@
 #### Video Endpoints #####
 
-from typing import Any, Dict, Optional
+from typing import Any, Dict, Optional, cast
 
 import orjson
 from fastapi import APIRouter, Depends, File, Form, Request, Response, UploadFile
@@ -17,7 +17,9 @@ from litellm.proxy.common_utils.openai_endpoint_utils import (
 )
 from litellm.proxy.image_endpoints.endpoints import batch_to_bytesio
 from litellm.proxy.video_endpoints.utils import (
+    coerce_optional_str,
     encode_character_id_in_response,
+    encode_video_id_in_response,
     extract_model_from_target_model_names,
     get_custom_provider_from_data,
 )
@@ -88,7 +90,7 @@ async def video_generation(
     # Process request using ProxyBaseLLMRequestProcessing
     processor = ProxyBaseLLMRequestProcessing(data=data)
     try:
-        return await processor.base_process_llm_request(
+        response = await processor.base_process_llm_request(
             request=request,
             fastapi_response=fastapi_response,
             user_api_key_dict=user_api_key_dict,
@@ -106,6 +108,12 @@ async def video_generation(
             user_api_base=user_api_base,
             version=version,
         )
+        request_data = cast("dict[str, object]", data)
+        return encode_video_id_in_response(
+            response=cast(object, response),
+            fallback_provider=coerce_optional_str(request_data.get("custom_llm_provider")),
+            fallback_model_id=coerce_optional_str(request_data.get("model")),
+        )
     except Exception as e:
         raise await processor._handle_llm_api_exception(
             e=e,

litellm/proxy/video_endpoints/utils.py

@@ -2,7 +2,13 @@ from typing import Any, Dict, Optional
 
 import orjson
 
-from litellm.types.videos.utils import encode_character_id_with_provider
+from litellm.router_utils.add_retry_fallback_headers import get_hidden_params_dict
+from litellm.types.videos.main import VideoObject
+from litellm.types.videos.utils import (
+    decode_video_id_with_provider,
+    encode_character_id_with_provider,
+    encode_video_id_with_provider,
+)
 
 
 def extract_model_from_target_model_names(target_model_names: Any) -> Optional[str]:
@@ -52,3 +58,36 @@ def encode_character_id_in_response(response: Any, custom_llm_provider: str, mod
             model_id=model_id,
         )
     return response
+
+
+def coerce_optional_str(value: object) -> Optional[str]:
+    return value if isinstance(value, str) and value else None
+
+
+def encode_video_id_in_response(
+    response: object,
+    fallback_provider: Optional[str],
+    fallback_model_id: Optional[str],
+) -> object:
+    if not isinstance(response, VideoObject) or not response.id:
+        return response
+
+    hidden_params = get_hidden_params_dict(response)
+    model_id = coerce_optional_str(hidden_params.get("model_id")) or fallback_model_id
+
+    decoded = decode_video_id_with_provider(response.id)
+    provider = (
+        decoded.get("custom_llm_provider")
+        or coerce_optional_str(hidden_params.get("custom_llm_provider"))
+        or fallback_provider
+    )
+    if not provider:
+        return response
+
+    original_video_id = decoded.get("video_id") or response.id
+    response.id = encode_video_id_with_provider(
+        video_id=original_video_id,
+        provider=provider,
+        model_id=model_id,
+    )
+    return response

tests/test_litellm/proxy/video_endpoints/test_endpoints.py

@@ -44,7 +44,9 @@ from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing
 from litellm.proxy.utils import ProxyLogging
 from litellm.router import Router
+from litellm.types.videos.main import VideoObject
 from litellm.types.videos.utils import (
+    decode_video_id_with_provider,
     encode_character_id_with_provider,
     encode_video_id_with_provider,
 )
@@ -246,6 +248,60 @@ async def test_generation__input_reference_attached(harness):
     }
 
 
+def _video_response(video_id: str, hidden_params: Dict[str, Any]) -> VideoObject:
+    """A VideoObject as returned by the create transform: the id is already
+    provider-encoded but with an empty model_id (the proxy flow never has the
+    model at transform time), and the router-set model_id lives in _hidden_params."""
+    obj = VideoObject(id=video_id, object="video", status="queued", created_at=0)
+    obj._hidden_params = hidden_params
+    return obj
+
+
+@pytest.mark.asyncio
+async def test_generation__reencodes_id_with_hidden_params_model_id(harness):
+    """Regression for #33423: create encodes the video id with an empty model_id,
+    so GET status/content later cannot resolve the deployment (and thus the
+    per-model api_key) and fall back to a nonexistent provider key. The endpoint
+    must re-stamp the returned id with the model_id the router recorded in
+    _hidden_params so the id round-trips to a resolvable deployment."""
+    create_time_id = encode_video_id_with_provider("video_orig123", "azure", "")
+    # sanity: the id the transform produced really has an empty model_id
+    assert decode_video_id_with_provider(create_time_id)["model_id"] == ""
+
+    harness.base_process.return_value = _video_response(
+        create_time_id,
+        hidden_params={"model_id": VIDEO_MODEL_ID, "custom_llm_provider": "azure"},
+    )
+
+    resp = await call_generation(harness, body={"model": "sora-2", "prompt": "x"})
+
+    # the returned id now carries the real model_id + provider ...
+    decoded = decode_video_id_with_provider(resp.id)
+    assert decoded["model_id"] == VIDEO_MODEL_ID
+    assert decoded["custom_llm_provider"] == "azure"
+    assert decoded["video_id"] == "video_orig123"
+    # ... and is byte-for-byte what a subsequent status/content call decodes and
+    # resolves to a model name (the router resolver maps VIDEO_MODEL_ID).
+    assert resp.id == AZURE_VIDEO_ID
+    assert harness.resolve_model.side_effect(decoded["model_id"]) == "azure-sora"
+
+
+@pytest.mark.asyncio
+async def test_generation__reencodes_id_falls_back_to_request_model(harness):
+    """When the router did not surface a model_id in _hidden_params, the endpoint
+    falls back to the model from the request body so the id is still stamped."""
+    create_time_id = encode_video_id_with_provider("video_orig123", "openai", "")
+
+    harness.base_process.return_value = _video_response(create_time_id, hidden_params={})
+
+    resp = await call_generation(harness, body={"model": "sora-2", "prompt": "x"})
+
+    decoded = decode_video_id_with_provider(resp.id)
+    assert decoded["model_id"] == "sora-2"
+    assert decoded["custom_llm_provider"] == "openai"
+    assert decoded["video_id"] == "video_orig123"
+
+
 @pytest.mark.asyncio
 async def test_generation__exception_routed_through_handler(harness):
     harness.base_process.side_effect = ValueError("provider boom")