refactor: route hosted_vllm Anthropic passthrough through ProviderConfigManager

- Move _should_skip_anthropic_translation to hosted_vllm/messages/transformation.py
- Register HostedVLLMAnthropicMessagesConfig in ProviderConfigManager instead of
  hard-coding a provider branch in the Anthropic messages handler
- Pass litellm_params to get_provider_anthropic_messages_config so the flag check
  can happen inside the abstraction used by all other providers
- Pin ProviderConfigManager mock in test_flag_disabled_uses_translation so the test
  is resilient to future hosted_vllm config registrations

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

litellm/llms/anthropic/experimental_pass_through/messages/handler.py

@@ -7,7 +7,6 @@
 
 import asyncio
 import contextvars
-import os
 from functools import partial
 from typing import Any, AsyncIterator, Coroutine, Dict, List, Optional, Union, cast
 
@@ -22,7 +21,7 @@ from litellm.llms.base_llm.anthropic_messages.transformation import (
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
 from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler
 from litellm.llms.hosted_vllm.messages.transformation import (
-    HostedVLLMAnthropicMessagesConfig,
+    _should_skip_anthropic_translation,  # re-exported for backwards compat
 )
 from litellm.types.llms.anthropic_messages.anthropic_request import AnthropicMetadata
 from litellm.types.llms.anthropic_messages.anthropic_response import (
@@ -54,23 +53,6 @@ def _should_route_to_responses_api(custom_llm_provider: Optional[str]) -> bool:
     return custom_llm_provider in _RESPONSES_API_PROVIDERS
 
 
-def _should_skip_anthropic_translation(litellm_params: GenericLiteLLMParams) -> bool:
-    """Return True when Anthropic→OpenAI translation should be bypassed for hosted_vllm.
-
-    Checked in priority order:
-    1. Per-deployment ``disable_anthropic_translation`` in litellm_params
-    2. Global env var ``DISABLE_HOSTED_VLLM_ANTHROPIC_TRANSLATION``
-    """
-    param_flag = litellm_params.get("disable_anthropic_translation")
-    if param_flag is not None:
-        return bool(param_flag)
-    return os.environ.get("DISABLE_HOSTED_VLLM_ANTHROPIC_TRANSLATION", "").lower() in (
-        "true",
-        "1",
-        "yes",
-    )
-
-
 ####### ENVIRONMENT VARIABLES ###################
 # Initialize any necessary instances or variables here
 base_llm_http_handler = BaseLLMHTTPHandler()
@@ -431,17 +413,14 @@ def anthropic_messages_handler(
 
     anthropic_messages_provider_config: Optional[BaseAnthropicMessagesConfig] = None
 
-    if custom_llm_provider == "hosted_vllm" and _should_skip_anthropic_translation(
-        litellm_params
-    ):
-        anthropic_messages_provider_config = HostedVLLMAnthropicMessagesConfig()
-    elif custom_llm_provider is not None and custom_llm_provider in [
+    if custom_llm_provider is not None and custom_llm_provider in [
         provider.value for provider in LlmProviders
     ]:
         anthropic_messages_provider_config = (
             ProviderConfigManager.get_provider_anthropic_messages_config(
                 model=model,
                 provider=litellm.LlmProviders(custom_llm_provider),
+                litellm_params=litellm_params,
             )
         )
     if anthropic_messages_provider_config is None:

litellm/llms/hosted_vllm/messages/transformation.py

@@ -7,6 +7,7 @@ Anthropic→OpenAI chat/completions translation and POSTs the original Anthropic
 directly to `{api_base}/v1/messages`.
 """
 
+import os
 from typing import Any, AsyncIterator, Dict, List, Optional, Tuple
 
 import httpx
@@ -21,6 +22,23 @@ from litellm.types.llms.anthropic_messages.anthropic_response import (
 from litellm.types.router import GenericLiteLLMParams
 
 
+def _should_skip_anthropic_translation(litellm_params: GenericLiteLLMParams) -> bool:
+    """Return True when Anthropic→OpenAI translation should be bypassed for hosted_vllm.
+
+    Checked in priority order:
+    1. Per-deployment ``disable_anthropic_translation`` in litellm_params
+    2. Global env var ``DISABLE_HOSTED_VLLM_ANTHROPIC_TRANSLATION``
+    """
+    param_flag = litellm_params.get("disable_anthropic_translation")
+    if param_flag is not None:
+        return bool(param_flag)
+    return os.environ.get("DISABLE_HOSTED_VLLM_ANTHROPIC_TRANSLATION", "").lower() in (
+        "true",
+        "1",
+        "yes",
+    )
+
+
 class HostedVLLMAnthropicMessagesConfig(BaseAnthropicMessagesConfig):
     """
     Sends Anthropic-format /v1/messages requests directly to a vLLM instance.

litellm/utils.py

@@ -8499,7 +8499,19 @@ class ProviderConfigManager:
     def get_provider_anthropic_messages_config(
         model: str,
         provider: LlmProviders,
+        litellm_params: Optional[Any] = None,
     ) -> Optional[BaseAnthropicMessagesConfig]:
+        if provider == litellm.LlmProviders.HOSTED_VLLM:
+            from litellm.llms.hosted_vllm.messages.transformation import (
+                HostedVLLMAnthropicMessagesConfig,
+                _should_skip_anthropic_translation,
+            )
+
+            if litellm_params is not None and _should_skip_anthropic_translation(
+                litellm_params
+            ):
+                return HostedVLLMAnthropicMessagesConfig()
+            return None
         return ProviderConfigManager._get_provider_anthropic_messages_config_cached(
             model=model, provider=provider
         )

tests/litellm/llms/hosted_vllm/test_anthropic_passthrough.py

@@ -17,8 +17,6 @@ import pytest
 import litellm
 from litellm.llms.hosted_vllm.messages.transformation import (
     HostedVLLMAnthropicMessagesConfig,
-)
-from litellm.llms.anthropic.experimental_pass_through.messages.handler import (
     _should_skip_anthropic_translation,
 )
 from litellm.types.router import GenericLiteLLMParams
@@ -204,11 +202,13 @@ class TestHandlerRouting:
 
     @pytest.mark.asyncio
     async def test_flag_enabled_uses_passthrough_config(self):
-        """When disable_anthropic_translation=True, the handler must call
-        base_llm_http_handler.anthropic_messages_handler (native path) instead of
+        """When disable_anthropic_translation=True, ProviderConfigManager returns
+        HostedVLLMAnthropicMessagesConfig and the native path is used instead of
         LiteLLMMessagesToCompletionTransformationHandler."""
         from litellm.llms.anthropic.experimental_pass_through.messages import handler as h
+        from litellm.utils import ProviderConfigManager
 
+        passthrough_config = HostedVLLMAnthropicMessagesConfig()
         mock_response = MagicMock()
         mock_response.id = "msg_test"
 
@@ -219,6 +219,7 @@ class TestHandlerRouting:
                 "anthropic_messages_handler",
             ) as mock_translate,
             patch("litellm.get_llm_provider", return_value=("qwen36-27b-fp8", "hosted_vllm", "key", "http://vllm/v1")),
+            patch.object(ProviderConfigManager, "get_provider_anthropic_messages_config", return_value=passthrough_config),
         ):
             h.anthropic_messages_handler(
                 max_tokens=100,
@@ -232,7 +233,6 @@ class TestHandlerRouting:
         mock_native.assert_called_once()
         mock_translate.assert_not_called()
 
-        # Confirm HostedVLLMAnthropicMessagesConfig was passed in
         call_kwargs = mock_native.call_args.kwargs
         assert isinstance(
             call_kwargs.get("anthropic_messages_provider_config"),
@@ -244,6 +244,7 @@ class TestHandlerRouting:
         """When disable_anthropic_translation is absent (default), the handler must
         route to LiteLLMMessagesToCompletionTransformationHandler."""
         from litellm.llms.anthropic.experimental_pass_through.messages import handler as h
+        from litellm.utils import ProviderConfigManager
 
         mock_response = MagicMock()
 
@@ -254,6 +255,7 @@ class TestHandlerRouting:
                 return_value=mock_response,
             ) as mock_translate,
             patch("litellm.get_llm_provider", return_value=("qwen36-27b-fp8", "hosted_vllm", "key", "http://vllm/v1")),
+            patch.object(ProviderConfigManager, "get_provider_anthropic_messages_config", return_value=None),
         ):
             h.anthropic_messages_handler(
                 max_tokens=100,