Forward Anthropic vllm_xargs to sampling params (#53308)

Signed-off-by: Maroon Ayoub <mayoub@redhat.com>

docs/features/custom_arguments.md

@@ -19,7 +19,7 @@ This allows arguments which are not already part of `SamplingParams` to be passe
 
 ## Online Custom Arguments
 
-The vLLM REST API allows custom arguments to be passed to the vLLM server via `vllm_xargs`. The example below integrates custom arguments into a vLLM REST API request:
+The OpenAI-compatible REST API and Anthropic-compatible `/v1/messages` endpoint allow custom arguments to be passed to the vLLM server via `vllm_xargs`. The example below integrates custom arguments into a vLLM REST API request:
 
 ``` bash
 curl http://localhost:8000/v1/completions \

tests/entrypoints/anthropic/test_anthropic_messages_conversion.py

@@ -182,6 +182,56 @@ class TestImageContentBlocks:
         }
 
 
+# ======================================================================
+# vllm_xargs pass-through
+# ======================================================================
+
+
+class TestVllmXargs:
+    def test_vllm_xargs_passed_through(self):
+        request = _make_request(
+            [{"role": "user", "content": "Hello"}],
+            vllm_xargs={
+                "kv_cache_report_mode": "full",
+                "existing_extension": 7,
+            },
+        )
+
+        result = _convert(request)
+
+        assert result.vllm_xargs == {
+            "kv_cache_report_mode": "full",
+            "existing_extension": 7,
+        }
+
+    def test_vllm_xargs_reaches_sampling_params_with_kv_transfer(self):
+        kv_transfer_params = {
+            "do_remote_decode": True,
+            "do_remote_prefill": False,
+        }
+        request = _make_request(
+            [{"role": "user", "content": "Hello"}],
+            vllm_xargs={
+                "kv_cache_report_mode": "full",
+                "existing_extension": "kept",
+            },
+            kv_transfer_params=kv_transfer_params,
+        )
+
+        converted = _convert(request)
+        sampling_params = converted.to_sampling_params(
+            max_tokens=converted.max_completion_tokens or 0,
+            default_sampling_params={},
+        )
+
+        assert converted.kv_transfer_params == kv_transfer_params
+        assert sampling_params.extra_args == {
+            "kv_cache_report_mode": "full",
+            "existing_extension": "kept",
+            "kv_transfer_params": kv_transfer_params,
+        }
+
+
 # ======================================================================
 # tool_result content handling
 # ======================================================================

vllm/entrypoints/anthropic/protocol.py

@@ -160,6 +160,13 @@ class AnthropicMessagesRequest(BaseModel):
             "ECTransfer parameters used for encoder-cache disaggregated serving."
         ),
     )
+    vllm_xargs: dict[str, str | int | float | list[str | int | float]] | None = Field(
+        default=None,
+        description=(
+            "Additional request parameters with (list of) string or "
+            "numeric values, used by custom extensions."
+        ),
+    )
     chat_template_kwargs: dict[str, Any] | None = Field(
         default=None,
         description=(

vllm/entrypoints/anthropic/serving.py

@@ -489,6 +489,7 @@ class AnthropicServingMessages(OpenAIServingChat):
             cache_salt=anthropic_request.cache_salt,
             kv_transfer_params=anthropic_request.kv_transfer_params,
             ec_transfer_params=anthropic_request.ec_transfer_params,
+            vllm_xargs=anthropic_request.vllm_xargs,
             chat_template_kwargs=anthropic_request.chat_template_kwargs,
         )