feat(hosted_vllm): translate Anthropic prefill (prefix: true) to continue_final_message

Closes the documented prefill-API coverage gap (#4881) for the hosted_vllm
provider so /v1/messages requests routed to vLLM honor the Anthropic
prefill semantic ("when the last message is role:assistant, the model
continues from where that message leaves off").

Two minimal changes:

* `litellm/llms/hosted_vllm/chat/transformation.py`: extend transform_request
  and async_transform_request to detect the unified prefill marker
  (`prefix: True` on the trailing assistant message) and inject vLLM's
  native continuation flags into extra_body — `continue_final_message:
  true` and `add_generation_prompt: false`. Idempotent (setdefault), no-op
  when the trigger is absent, never overrides caller-set values.

* `litellm/llms/anthropic/experimental_pass_through/adapters/handler.py`:
  in `_prepare_completion_kwargs` — which is reached only when /v1/messages
  is routed to a non-Anthropic provider — auto-tag a trailing
  role:"assistant" message with `prefix: True` so the provider-side
  translator above fires. This makes the Anthropic /v1/messages prefill
  contract portable across providers.

Tests added in test_hosted_vllm_chat_transformation.py:
  - prefill marker → continue_final_message injected
  - no marker → no-op
  - user-ending → no-op
  - existing extra_body preserved and merged
  - explicit caller-set continue_final_message:False respected

Verified locally: 13/13 hosted_vllm transformation tests pass (8 pre-existing
+ 5 new).

Refs:
- #4881 — closed; defines the unified prefill API but only Anthropic provider implemented it
- BerriAI/litellm#28580 — the bug report with empirical trace

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

@@ -184,6 +184,18 @@ class LiteLLMMessagesToCompletionTransformationHandler:
             Logging as LiteLLMLoggingObject,
         )
 
+        # Anthropic /v1/messages prefill contract: when the last message is
+        # role:"assistant" the model continues from there. The Anthropic
+        # provider config implements this via `prefix:true` (see #4881 +
+        # `litellm/llms/anthropic/chat/transformation.py::get_prefix_prompt`).
+        # When this adapter routes the call to a non-Anthropic provider (the
+        # whole reason this adapter exists), tag the trailing assistant
+        # message with prefix:true so provider configs that understand the
+        # unified marker — currently hosted_vllm — can translate it into
+        # their native continuation flag. Idempotent: setdefault only.
+        if messages and isinstance(messages[-1], dict) and messages[-1].get("role") == "assistant":
+            messages[-1].setdefault("prefix", True)
+
         request_data = {
             "model": model,
             "messages": messages,

litellm/llms/hosted_vllm/chat/transformation.py

@@ -92,6 +92,76 @@ class HostedVLLMChatConfig(OpenAIGPTConfig):
         params.extend(["reasoning_effort", "thinking"])
         return params
 
+    @staticmethod
+    def _apply_prefill_translation(
+        messages: List[AllMessageValues], optional_params: dict
+    ) -> None:
+        """Translate LiteLLM's unified prefill marker to vLLM's native flags.
+
+        When the last message is ``role:"assistant"`` and carries ``prefix: True``
+        (the unified prefill marker introduced in #4881), vLLM's chat completions
+        endpoint expects ``continue_final_message: true`` and
+        ``add_generation_prompt: false`` to actually continue the assistant turn
+        rather than start a fresh one. Neither flag is part of the standard
+        OpenAI surface, so we inject them via ``extra_body`` — the openai SDK
+        forwards extra_body verbatim into the HTTP request body, and vLLM
+        honors them.
+
+        Does not override caller-set values. No-op if the trigger pattern is
+        absent. Refs: #4881 (defines prefill API), upstream issue tracking the
+        hosted_vllm coverage gap.
+        """
+        if not messages:
+            return
+        last_message = messages[-1]
+        if not isinstance(last_message, dict):
+            return
+        if last_message.get("role") != "assistant":
+            return
+        if not last_message.get("prefix"):
+            return
+
+        extra_body = optional_params.get("extra_body")
+        if not isinstance(extra_body, dict):
+            extra_body = {}
+        extra_body.setdefault("continue_final_message", True)
+        extra_body.setdefault("add_generation_prompt", False)
+        optional_params["extra_body"] = extra_body
+
+    def transform_request(
+        self,
+        model: str,
+        messages: List[AllMessageValues],
+        optional_params: dict,
+        litellm_params: dict,
+        headers: dict,
+    ) -> dict:
+        self._apply_prefill_translation(messages, optional_params)
+        return super().transform_request(
+            model=model,
+            messages=messages,
+            optional_params=optional_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
+    async def async_transform_request(
+        self,
+        model: str,
+        messages: List[AllMessageValues],
+        optional_params: dict,
+        litellm_params: dict,
+        headers: dict,
+    ) -> dict:
+        self._apply_prefill_translation(messages, optional_params)
+        return await super().async_transform_request(
+            model=model,
+            messages=messages,
+            optional_params=optional_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
     def map_openai_params(
         self,
         non_default_params: dict,

tests/test_litellm/llms/hosted_vllm/chat/test_hosted_vllm_chat_transformation.py

@@ -319,3 +319,94 @@ def test_hosted_vllm_custom_tools_use_top_level_input_schema():
     assert tools[0]["function"]["name"] == "search"
     assert tools[0]["function"]["description"] == "Search docs"
     assert tools[0]["function"]["parameters"] == input_schema
+
+
+def test_hosted_vllm_prefill_translation_injects_continue_final_message():
+    """Trailing assistant message with prefix:True translates into vLLM's
+    continue_final_message + add_generation_prompt flags inside extra_body."""
+    config = HostedVLLMChatConfig()
+    messages = [
+        {"role": "user", "content": "Count to three, comma-separated."},
+        {"role": "assistant", "content": "1, 2,", "prefix": True},
+    ]
+    transformed = config.transform_request(
+        model="hosted_vllm/qwen3.6-35b-a3b",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+    )
+    assert "extra_body" in transformed
+    assert transformed["extra_body"]["continue_final_message"] is True
+    assert transformed["extra_body"]["add_generation_prompt"] is False
+
+
+def test_hosted_vllm_prefill_translation_no_prefix_marker_is_noop():
+    """A trailing assistant message WITHOUT prefix:True must not inject
+    continuation flags — the caller did not opt in."""
+    config = HostedVLLMChatConfig()
+    messages = [
+        {"role": "user", "content": "Hi"},
+        {"role": "assistant", "content": "Hello!"},
+    ]
+    transformed = config.transform_request(
+        model="hosted_vllm/qwen3.6-35b-a3b",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+    )
+    assert "extra_body" not in transformed or transformed.get("extra_body", {}) == {}
+
+
+def test_hosted_vllm_prefill_translation_user_ending_is_noop():
+    """Requests ending in a user message must be unaffected."""
+    config = HostedVLLMChatConfig()
+    messages = [{"role": "user", "content": "Say hello in one word."}]
+    transformed = config.transform_request(
+        model="hosted_vllm/qwen3.6-35b-a3b",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+    )
+    assert "extra_body" not in transformed or transformed.get("extra_body", {}) == {}
+
+
+def test_hosted_vllm_prefill_translation_preserves_existing_extra_body():
+    """If the caller already set extra_body keys, prefill translation must
+    merge — not overwrite."""
+    config = HostedVLLMChatConfig()
+    messages = [
+        {"role": "user", "content": "Continue."},
+        {"role": "assistant", "content": "X", "prefix": True},
+    ]
+    transformed = config.transform_request(
+        model="hosted_vllm/qwen3.6-35b-a3b",
+        messages=messages,
+        optional_params={"extra_body": {"guided_choice": ["yes", "no"]}},
+        litellm_params={},
+        headers={},
+    )
+    eb = transformed["extra_body"]
+    assert eb["continue_final_message"] is True
+    assert eb["add_generation_prompt"] is False
+    assert eb["guided_choice"] == ["yes", "no"]
+
+
+def test_hosted_vllm_prefill_translation_respects_explicit_caller_value():
+    """If the caller already pinned continue_final_message=False, the helper
+    must not override their explicit choice (setdefault semantics)."""
+    config = HostedVLLMChatConfig()
+    messages = [
+        {"role": "user", "content": "Continue."},
+        {"role": "assistant", "content": "X", "prefix": True},
+    ]
+    transformed = config.transform_request(
+        model="hosted_vllm/qwen3.6-35b-a3b",
+        messages=messages,
+        optional_params={"extra_body": {"continue_final_message": False}},
+        litellm_params={},
+        headers={},
+    )
+    assert transformed["extra_body"]["continue_final_message"] is False