[P/D] AnthropicMessages add kv_transfer_params for PD disaggregation (#37535)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>

vllm/entrypoints/anthropic/protocol.py

@@ -5,7 +5,7 @@
 import time
 from typing import Any, Literal
 
-from pydantic import BaseModel, field_validator, model_validator
+from pydantic import BaseModel, Field, field_validator, model_validator
 
 
 class AnthropicError(BaseModel):
@@ -112,6 +112,12 @@ class AnthropicMessagesRequest(BaseModel):
     top_k: int | None = None
     top_p: float | None = None
 
+    # vLLM-specific fields that are not in Anthropic spec
+    kv_transfer_params: dict[str, Any] | None = Field(
+        default=None,
+        description="KVTransfer parameters used for disaggregated serving.",
+    )
+
     @field_validator("model")
     @classmethod
     def validate_model(cls, v):
@@ -181,6 +187,11 @@ class AnthropicMessagesResponse(BaseModel):
     stop_sequence: str | None = None
     usage: AnthropicUsage | None = None
 
+    # vLLM-specific fields that are not in Anthropic spec
+    kv_transfer_params: dict[str, Any] | None = Field(
+        default=None, description="KVTransfer parameters."
+    )
+
     def model_post_init(self, __context):
         if not self.id:
             self.id = f"msg_{int(time.time() * 1000)}"

vllm/entrypoints/anthropic/serving.py

@@ -331,6 +331,7 @@ class AnthropicServingMessages(OpenAIServingChat):
             temperature=anthropic_request.temperature,
             top_p=anthropic_request.top_p,
             top_k=anthropic_request.top_k,
+            kv_transfer_params=anthropic_request.kv_transfer_params,
         )
 
     @classmethod
@@ -441,6 +442,7 @@ class AnthropicServingMessages(OpenAIServingChat):
                 input_tokens=generator.usage.prompt_tokens,
                 output_tokens=generator.usage.completion_tokens,
             ),
+            kv_transfer_params=generator.kv_transfer_params,
         )
         choice = generator.choices[0]
         if choice.finish_reason == "stop":