fix(batches): read batch input files for rate limiting from the batch deployment

litellm/proxy/hooks/batch_rate_limiter.py

@@ -542,11 +542,13 @@ class _PROXY_BatchRateLimiter(CustomLogger):
         and the upstream provider returns 404.
         """
         from litellm.proxy.openai_files_endpoints.common_utils import (
+            batch_deployment_id,
             decode_model_from_file_id,
             explicit_request_credentials,
             get_credentials_for_model,
             get_original_file_id,
             get_team_provider_credentials,
+            hinted_batch_deployment_id,
         )
         from litellm.proxy.proxy_server import llm_router
 
@@ -560,7 +562,8 @@ class _PROXY_BatchRateLimiter(CustomLogger):
                 try:
                     credentials = get_credentials_for_model(
                         llm_router=llm_router,
-                        model_id=model_from_file_id,
+                        model_id=hinted_batch_deployment_id(llm_router, model_from_file_id, file_id)
+                        or model_from_file_id,
                         operation_context="batch input file read (rate limiting)",
                     )
                     fetch_kwargs.update(extract_file_access_credentials(credentials))
@@ -578,7 +581,7 @@ class _PROXY_BatchRateLimiter(CustomLogger):
                 try:
                     credentials = get_credentials_for_model(
                         llm_router=llm_router,
-                        model_id=request_model,
+                        model_id=batch_deployment_id(llm_router, request_model) or request_model,
                         operation_context="batch input file read (rate limiting)",
                     )
                     fetch_kwargs.update(extract_file_access_credentials(credentials))

litellm/router.py

@@ -12030,11 +12030,6 @@ class Router:
         return Deployment(**first_usable) if isinstance(first_usable, dict) else first_usable
 
     def get_batch_deployment_id(self, model_group_name: str, preferred_deployment_id: str | None = None) -> str | None:
-        """
-        The shared deployment of ``model_group_name`` that file and batch calls
-        should reach when the group mixes batch and non-batch deployments, or
-        None when the batch rule drops nothing and the group name resolves as before.
-        """
         group: Final = tuple(
             deployment
             for deployment in self.deployments_named(model_group_name) or ()

tests/unit/proxy/hooks/test_batch_file_validation.py

@@ -1353,6 +1353,84 @@ def test_resolve_fetch_params_model_embedded_fails_open_on_credential_error():
     assert fetch_kwargs == {"custom_llm_provider": "openai"}
 
 
+
+def _mixed_azure_chat_and_batch_router():
+    from litellm import Router
+
+    return Router(
+        model_list=[
+            {
+                "model_name": "gpt-mixed",
+                "litellm_params": {
+                    "model": "azure/gpt-chat",
+                    "api_base": "https://chat-resource.openai.azure.com",
+                    "api_key": "chat-key",
+                    "api_version": "2025-04-01-preview",
+                },
+                "model_info": {"id": "chat-id", "mode": "chat"},
+            },
+            {
+                "model_name": "gpt-mixed",
+                "litellm_params": {
+                    "model": "azure/gpt-batch",
+                    "api_base": "https://batch-resource.openai.azure.com",
+                    "api_key": "batch-key",
+                    "api_version": "2025-04-01-preview",
+                },
+                "model_info": {"id": "batch-id", "mode": "batch"},
+            },
+        ]
+    )
+
+
+@pytest.mark.parametrize(
+    ("file_id", "data", "expected_api_base"),
+    [
+        pytest.param(
+            "file-hinted",
+            {},
+            "https://batch-resource.openai.azure.com",
+            id="hinted-file-id",
+        ),
+        pytest.param(
+            "file-legacy",
+            {},
+            "https://chat-resource.openai.azure.com",
+            id="group-only-file-id",
+        ),
+        pytest.param(
+            "file-raw-provider-id",
+            {"model": "gpt-mixed"},
+            "https://batch-resource.openai.azure.com",
+            id="request-model",
+        ),
+    ],
+)
+def test_resolve_fetch_params_reads_a_mixed_group_file_where_the_batch_runs(
+    file_id, data, expected_api_base
+):
+    from litellm.proxy.openai_files_endpoints.common_utils import (
+        encode_file_id_with_model,
+    )
+
+    proxy_file_id = {
+        "file-hinted": encode_file_id_with_model(
+            "file-orig", "gpt-mixed", deployment_id="batch-id"
+        ),
+        "file-legacy": encode_file_id_with_model("file-orig", "gpt-mixed"),
+    }.get(file_id, file_id)
+    rate_limiter = _make_rate_limiter()
+    with patch(
+        "litellm.proxy.proxy_server.llm_router", _mixed_azure_chat_and_batch_router()
+    ):
+        _, fetch_kwargs = rate_limiter._resolve_batch_input_file_fetch_params(
+            file_id=proxy_file_id,
+            custom_llm_provider="openai",
+            data=data,
+        )
+    assert fetch_kwargs["api_base"] == expected_api_base
+    assert fetch_kwargs["model"] == "gpt-mixed"
+
 @pytest.mark.asyncio
 async def test_check_and_increment_computes_descriptors_when_not_passed():
     from litellm.proxy.hooks.batch_rate_limiter import (