[EC Connector] Add EC Transfer Params (#42433)

Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Co-authored-by: Or Ozeri <oro@il.ibm.com>

rust/src/chat/src/event.rs

@@ -202,5 +202,8 @@ pub enum ChatEvent {
         finish_reason: FinishReason,
         /// Connector-specific KV transfer parameters for disaggregated serving.
         kv_transfer_params: Option<serde_json::Value>,
+        /// Connector-specific encoder cache transfer parameters for
+        /// disaggregated serving.
+        ec_transfer_params: Option<serde_json::Value>,
     },
 }

rust/src/chat/src/output/default/unified.rs

@@ -257,6 +257,7 @@ pub(crate) async fn unified_event_stream(
                         usage: finished.usage,
                         finish_reason: finished.finish_reason,
                         kv_transfer_params: finished.kv_transfer_params,
+                        ec_transfer_params: finished.ec_transfer_params,
                     })
                     .await;
                 }
@@ -387,6 +388,7 @@ mod tests {
                 usage: vllm_llm::TokenUsage::default(),
                 finish_reason: crate::FinishReason::Stop(None),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         }
     }
@@ -628,6 +630,7 @@ mod tests {
                     usage: vllm_llm::TokenUsage::default(),
                     finish_reason: crate::FinishReason::Stop(None),
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 },
             ]
         );
@@ -671,6 +674,7 @@ mod tests {
                     usage: vllm_llm::TokenUsage::default(),
                     finish_reason: crate::FinishReason::Stop(None),
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 },
             ]
         );

rust/src/chat/src/output/harmony/mod.rs

@@ -370,6 +370,7 @@ async fn harmony_assistant_event_stream(
                         usage: finished.usage,
                         finish_reason: finished.finish_reason,
                         kv_transfer_params: finished.kv_transfer_params,
+                        ec_transfer_params: finished.ec_transfer_params,
                     })
                     .await;
                 }

rust/src/chat/src/output/harmony/tests.rs

@@ -52,6 +52,7 @@ fn finished() -> Finished {
         },
         finish_reason: FinishReason::stop_eos(),
         kv_transfer_params: None,
+        ec_transfer_params: None,
     }
 }
 
@@ -115,6 +116,7 @@ fn interrupted_final_message_is_preserved() {
             },
             finish_reason: FinishReason::stop_eos(),
             kv_transfer_params: None,
+            ec_transfer_params: None,
         })
     );
 }
@@ -175,6 +177,7 @@ fn interrupted_analysis_message_is_preserved() {
             },
             finish_reason: FinishReason::stop_eos(),
             kv_transfer_params: None,
+            ec_transfer_params: None,
         })
     );
 }

rust/src/chat/src/output/mod.rs

@@ -48,6 +48,9 @@ pub(crate) enum AssistantEvent {
         finish_reason: FinishReason,
         /// Connector-specific KV transfer parameters for disaggregated serving.
         kv_transfer_params: Option<serde_json::Value>,
+        /// Connector-specific encoder cache transfer parameters for
+        /// disaggregated serving.
+        ec_transfer_params: Option<serde_json::Value>,
     },
 }
 

rust/src/chat/src/output/structured.rs

@@ -146,6 +146,7 @@ impl StructuredEventState {
         usage: vllm_llm::TokenUsage,
         finish_reason: FinishReason,
         kv_transfer_params: Option<serde_json::Value>,
+        ec_transfer_params: Option<serde_json::Value>,
     ) -> Result<Vec<ChatEvent>> {
         let mut events = Vec::new();
         self.close_open_text_block(&mut events);
@@ -155,6 +156,7 @@ impl StructuredEventState {
             usage,
             finish_reason,
             kv_transfer_params,
+            ec_transfer_params,
         });
         Ok(events)
     }
@@ -296,8 +298,11 @@ pub(crate) async fn structured_chat_event_stream(
                 usage,
                 finish_reason,
                 kv_transfer_params,
+                ec_transfer_params,
             } => {
-                for next in state.finish(usage, finish_reason, kv_transfer_params)? {
+                for next in
+                    state.finish(usage, finish_reason, kv_transfer_params, ec_transfer_params)?
+                {
                     y.yield_ok(next).await;
                 }
             }
@@ -334,6 +339,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 
@@ -388,6 +394,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 
@@ -439,6 +446,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 
@@ -490,6 +498,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 
@@ -557,6 +566,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 

rust/src/chat/src/stream.rs

@@ -22,6 +22,9 @@ pub struct CollectedAssistantMessage {
     pub finish_reason: FinishReason,
     /// Connector-specific KV transfer parameters for disaggregated serving.
     pub kv_transfer_params: Option<serde_json::Value>,
+    /// Connector-specific encoder cache transfer parameters for disaggregated
+    /// serving.
+    pub ec_transfer_params: Option<serde_json::Value>,
 }
 
 /// Per-request stream of chat events.
@@ -77,6 +80,7 @@ impl ChatEventStream {
                     usage,
                     finish_reason,
                     kv_transfer_params,
+                    ec_transfer_params,
                 } => {
                     return Ok(CollectedAssistantMessage {
                         message: done,
@@ -89,6 +93,7 @@ impl ChatEventStream {
                         usage,
                         finish_reason,
                         kv_transfer_params,
+                        ec_transfer_params,
                     });
                 }
                 ChatEvent::ToolCallEnd { call, .. } => {
@@ -194,6 +199,7 @@ mod tests {
                     },
                     finish_reason: FinishReason::stop_eos(),
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             ]),
         );
@@ -234,6 +240,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }
         );
     }

rust/src/chat/tests/chat.rs

@@ -50,6 +50,7 @@ fn request_output(
         stop_reason,
         events: None,
         kv_transfer_params: None,
+        ec_transfer_params: None,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,
@@ -75,6 +76,7 @@ fn request_output_with_logprobs(
         stop_reason,
         events: None,
         kv_transfer_params: None,
+        ec_transfer_params: None,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,

rust/src/chat/tests/roundtrip.rs

@@ -676,6 +676,7 @@ fn decoded_completion_stream(
                     usage: Default::default(),
                     finish_reason: FinishReason::stop_eos(),
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             }
         });
@@ -686,6 +687,7 @@ fn decoded_completion_stream(
                 usage: Default::default(),
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             });
             events.push(DecodedTextEvent::TextDelta {
                 delta: chunk.delta,

rust/src/engine-core-client/src/protocol/output.rs

@@ -97,6 +97,8 @@ pub struct EngineCoreOutput {
     #[serde(default)]
     pub kv_transfer_params: Option<serde_json::Value>,
     #[serde(default)]
+    pub ec_transfer_params: Option<serde_json::Value>,
+    #[serde(default)]
     pub trace_headers: Option<OpaqueValue>,
     /// Breakdown of the scheduled prefill computation, set on the first output
     /// of a newly scheduled prefill and elided for subsequent decode outputs.
@@ -374,6 +376,7 @@ mod tests {
                 stop_reason: Some(StopReason::Text("stop".to_string())),
                 events: None,
                 kv_transfer_params: None,
+                ec_transfer_params: None,
                 trace_headers: None,
                 prefill_stats: None,
                 routed_experts: None,
@@ -426,6 +429,7 @@ mod tests {
                             stop_reason: None,
                             events: None,
                             kv_transfer_params: None,
+                            ec_transfer_params: None,
                             trace_headers: None,
                             prefill_stats: None,
                             routed_experts: None,

rust/src/engine-core-client/src/tests/client.rs

@@ -226,6 +226,7 @@ fn request_output(
         stop_reason: None,
         events: None,
         kv_transfer_params: None,
+        ec_transfer_params: None,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,
@@ -2517,6 +2518,7 @@ fn python_msgpack_fixtures_match_rust_encoding() {
                         stop_reason: None,
                         events: None,
                         kv_transfer_params: None,
+                        ec_transfer_params: None,
                         trace_headers: None,
                         prefill_stats: None,
                         routed_experts: None,

rust/src/engine-core-client/src/tests/python_compat.py

@@ -91,6 +91,7 @@ class EngineCoreOutput(
     stop_reason: int | str | None = None
     events: object | None = None
     kv_transfer_params: object | None = None
+    ec_transfer_params: object | None = None
     trace_headers: object | None = None
     prefill_stats: object | None = None
     routed_experts: object | None = None

rust/src/llm/src/output.rs

@@ -38,6 +38,9 @@ pub struct CollectedGenerateOutput {
     pub usage: TokenUsage,
     /// Connector-specific KV transfer parameters for disaggregated serving.
     pub kv_transfer_params: Option<serde_json::Value>,
+    /// Connector-specific encoder cache transfer parameters for disaggregated
+    /// serving.
+    pub ec_transfer_params: Option<serde_json::Value>,
 }
 
 /// Prompt-scoped metadata emitted only once on the first [`GenerateOutput`] for
@@ -146,6 +149,9 @@ pub struct GenerateOutput {
     pub cached_token_count: usize,
     /// Connector-specific KV transfer parameters for disaggregated serving.
     pub kv_transfer_params: Option<serde_json::Value>,
+    /// Connector-specific encoder cache transfer parameters for disaggregated
+    /// serving.
+    pub ec_transfer_params: Option<serde_json::Value>,
 }
 
 impl GenerateOutput {
@@ -192,6 +198,7 @@ impl GenerateOutput {
             finish_reason,
             cached_token_count: 0,
             kv_transfer_params: None,
+            ec_transfer_params: None,
         }
     }
 }
@@ -280,6 +287,7 @@ impl Stream for GenerateOutputStream {
             finish_reason,
             cached_token_count,
             kv_transfer_params: raw.kv_transfer_params,
+            ec_transfer_params: raw.ec_transfer_params,
         };
 
         Poll::Ready(Some(Ok(output)))
@@ -362,6 +370,7 @@ impl<T: Stream<Item = Result<GenerateOutput>> + Send> T {
                             cached_token_count,
                         },
                         kv_transfer_params: None,
+                        ec_transfer_params: None,
                     });
                 }
 
@@ -374,6 +383,7 @@ impl<T: Stream<Item = Result<GenerateOutput>> + Send> T {
                         cached_token_count,
                     };
                     collected.kv_transfer_params = output.kv_transfer_params;
+                    collected.ec_transfer_params = output.ec_transfer_params;
                     return Ok(collected);
                 }
             }

rust/src/llm/tests/generate.rs

@@ -51,6 +51,7 @@ fn request_output_with_events(
         stop_reason: None,
         events,
         kv_transfer_params: None,
+        ec_transfer_params: None,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,
@@ -75,6 +76,7 @@ fn request_output_with_logprobs(
         stop_reason: None,
         events: None,
         kv_transfer_params: None,
+        ec_transfer_params: None,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,
@@ -89,6 +91,7 @@ fn request_output_with_logprobs_and_kv(
     new_logprobs: Option<Logprobs>,
     prompt_logprobs: Option<Logprobs>,
     kv_transfer_params: Option<serde_json::Value>,
+    ec_transfer_params: Option<serde_json::Value>,
 ) -> EngineCoreOutput {
     EngineCoreOutput {
         request_id: request_id.to_string(),
@@ -100,6 +103,7 @@ fn request_output_with_logprobs_and_kv(
         stop_reason: None,
         events: None,
         kv_transfer_params,
+        ec_transfer_params,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,
@@ -356,6 +360,7 @@ async fn collect_output_aggregates_raw_tokens_logprobs_and_terminal_metadata() {
                                 Some(logprobs_for_position(44, -0.3, 1, 88, -0.4)),
                                 None,
                                 Some(serde_json::json!({"connector": "x"})),
+                                None,
                             ),
                         ],
                         ..Default::default()

rust/src/server/src/grpc/convert.rs

@@ -69,6 +69,11 @@ pub fn to_text_request(
             let map = sampling_params.vllm_xargs.get_or_insert_with(Default::default);
             map.insert("kv_transfer_params".to_string(), kv_json);
         }
+        if let Some(ec_struct) = kv.ec_transfer_params.as_ref() {
+            let ec_json = proto_struct_to_json(ec_struct);
+            let map = sampling_params.vllm_xargs.get_or_insert_with(Default::default);
+            map.insert("ec_transfer_params".to_string(), ec_json);
+        }
         if kv.bypass_prefix_cache {
             sampling_params.skip_reading_prefix_cache = Some(true);
         }
@@ -343,6 +348,7 @@ fn to_finish_info(finished: &Finished, token_ids: &[u32]) -> pb::FinishInfo {
         finish_reason,
         stop_reason,
         kv_transfer_params: finished.kv_transfer_params.as_ref().and_then(json_to_proto_struct),
+        ec_transfer_params: finished.ec_transfer_params.as_ref().and_then(json_to_proto_struct),
     }
 }
 
@@ -586,6 +592,7 @@ mod tests {
             },
             finish_reason: reason,
             kv_transfer_params: None,
+            ec_transfer_params: None,
         }
     }
 

rust/src/server/src/grpc/mod.rs

@@ -71,6 +71,7 @@ impl pb::generate_server::Generate for GenerateServiceImpl {
             usage: collected.usage,
             finish_reason: collected.finish_reason,
             kv_transfer_params: collected.kv_transfer_params,
+            ec_transfer_params: collected.ec_transfer_params,
         };
 
         let outputs = convert::to_sequence_output(

rust/src/server/src/grpc/tests.rs

@@ -104,6 +104,7 @@ fn request_output(
         stop_reason: None,
         events: None,
         kv_transfer_params: None,
+        ec_transfer_params: None,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,

rust/src/server/src/routes/http_client_tests.rs

@@ -100,6 +100,7 @@ fn request_output(
         stop_reason: None,
         events: None,
         kv_transfer_params: None,
+        ec_transfer_params: None,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,

rust/src/server/src/routes/inference/generate.rs

@@ -266,6 +266,7 @@ fn collect_generate(
         }],
         prompt_logprobs,
         kv_transfer_params: collected.kv_transfer_params,
+        ec_transfer_params: collected.ec_transfer_params,
     })
 }
 
@@ -404,6 +405,7 @@ mod tests {
                 finish_reason: None,
                 cached_token_count: 0,
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
             Ok(GenerateOutput {
                 request_id: String::new(),
@@ -416,6 +418,7 @@ mod tests {
                 finish_reason: Some(FinishReason::stop_eos()),
                 cached_token_count: 2,
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 

rust/src/server/src/routes/inference/generate/convert.rs

@@ -4,7 +4,7 @@ use super::types::GenerateRequest;
 use super::validate;
 use crate::error::ApiError;
 use crate::lora::LoraModelResolution;
-use crate::utils::{ResolvedRequestContext, merge_kv_transfer_params};
+use crate::utils::{ResolvedRequestContext, merge_ec_transfer_params, merge_kv_transfer_params};
 
 /// Lowered generate request plus the response request ID.
 #[derive(Debug, Clone, PartialEq)]
@@ -56,6 +56,10 @@ pub(super) fn prepare_generate_request(
         sampling_params.vllm_xargs,
         request.kv_transfer_params.as_ref(),
     );
+    sampling_params.vllm_xargs = merge_ec_transfer_params(
+        sampling_params.vllm_xargs,
+        request.ec_transfer_params.as_ref(),
+    );
 
     let text_request = TextRequest {
         request_id: ctx.request_id.clone(),

rust/src/server/src/routes/inference/generate/types.rs

@@ -22,6 +22,7 @@ pub struct GenerateRequest {
     #[serde(default)]
     pub priority: i32,
     pub kv_transfer_params: Option<HashMap<String, Value>>,
+    pub ec_transfer_params: Option<HashMap<String, Value>>,
     #[serde(flatten)]
     pub other: Map<String, Value>,
 }
@@ -66,6 +67,7 @@ pub(super) struct GenerateResponse {
     pub choices: Vec<GenerateResponseChoice>,
     pub prompt_logprobs: Option<Vec<Option<HashMap<u32, GenerateLogprob>>>>,
     pub kv_transfer_params: Option<Value>,
+    pub ec_transfer_params: Option<Value>,
 }
 
 /// Mirrors the Python vLLM `Logprob` class used in prompt-logprobs payloads.

rust/src/server/src/routes/openai/chat_completions.rs

@@ -144,6 +144,7 @@ async fn collect_chat_completion(
         usage,
         finish_reason,
         kv_transfer_params,
+        ec_transfer_params,
     } = collected;
     let stop_reason = finish_reason.as_stop_reason().map(stop_reason_to_json);
     let saw_tool_calls = message.tool_calls().next().is_some();
@@ -224,6 +225,7 @@ async fn collect_chat_completion(
         prompt_logprobs,
         prompt_token_ids: return_token_ids.then(|| prompt_token_ids.to_vec()),
         kv_transfer_params,
+        ec_transfer_params,
     })
 }
 
@@ -951,6 +953,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 
@@ -1031,6 +1034,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 
@@ -1086,6 +1090,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 
@@ -1167,6 +1172,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 
@@ -1300,6 +1306,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 
@@ -1381,6 +1388,7 @@ mod tests {
                 },
                 finish_reason: FinishReason::stop_eos(),
                 kv_transfer_params: None,
+                ec_transfer_params: None,
             }),
         ]);
 

rust/src/server/src/routes/openai/chat_completions/convert.rs

@@ -13,7 +13,9 @@ use crate::routes::openai::utils::structured_outputs::convert_from_response_form
 use crate::routes::openai::utils::types::{
     ChatMessage, ContentPart, MessageContent, Tool, ToolChoice, ToolChoiceValue,
 };
-use crate::utils::{ResolvedRequestContext, convert_logit_bias, merge_kv_transfer_params};
+use crate::utils::{
+    ResolvedRequestContext, convert_logit_bias, merge_ec_transfer_params, merge_kv_transfer_params,
+};
 
 /// Lowered chat request plus the public response metadata carried by every SSE
 /// chunk.
@@ -132,7 +134,7 @@ pub(super) fn prepare_chat_request(
             structured_outputs,
             skip_reading_prefix_cache: None,
             vllm_xargs: merge_kv_transfer_params(
-                request.vllm_xargs,
+                merge_ec_transfer_params(request.vllm_xargs, request.ec_transfer_params.as_ref()),
                 request.kv_transfer_params.as_ref(),
             ),
         },

rust/src/server/src/routes/openai/chat_completions/types.rs

@@ -232,6 +232,9 @@ pub struct ChatCompletionRequest {
     /// KV transfer parameters for disaggregated serving
     pub kv_transfer_params: Option<HashMap<String, Value>>,
 
+    /// Encoder cache transfer parameters for disaggregated serving
+    pub ec_transfer_params: Option<HashMap<String, Value>>,
+
     /// Additional request parameters with string or numeric values for custom
     /// extensions
     pub vllm_xargs: Option<HashMap<String, Value>>,
@@ -299,6 +302,7 @@ impl Default for ChatCompletionRequest {
             return_token_ids: None,
             cache_salt: None,
             kv_transfer_params: None,
+            ec_transfer_params: None,
             vllm_xargs: None,
             repetition_detection: None,
         }
@@ -346,6 +350,7 @@ pub(super) struct ChatCompletionResponse {
     pub prompt_logprobs: Option<Vec<Option<HashMap<String, f32>>>>,
     pub prompt_token_ids: Option<Vec<u32>>,
     pub kv_transfer_params: Option<Value>,
+    pub ec_transfer_params: Option<Value>,
 }
 
 /// Mirrors the Python vLLM `ChatCompletionResponseChoice` class.

rust/src/server/src/routes/openai/completions.rs

@@ -212,6 +212,7 @@ async fn collect_completion(
         usage: Some(usage),
         system_fingerprint: None,
         kv_transfer_params: collected.kv_transfer_params,
+        ec_transfer_params: collected.ec_transfer_params,
     })
 }
 
@@ -682,6 +683,7 @@ mod tests {
                         "repetition_detected".to_string(),
                     ))),
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             }),
         ]);
@@ -786,6 +788,7 @@ mod tests {
                     },
                     finish_reason: FinishReason::Length,
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             }),
         ]);
@@ -837,6 +840,7 @@ mod tests {
                     },
                     finish_reason: FinishReason::Length,
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             }),
         ]);
@@ -891,6 +895,7 @@ mod tests {
                     },
                     finish_reason: FinishReason::Length,
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             }),
         ]);
@@ -962,6 +967,7 @@ mod tests {
                     },
                     finish_reason: FinishReason::Length,
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             }),
         ]);
@@ -1035,6 +1041,7 @@ mod tests {
                     },
                     finish_reason: FinishReason::Length,
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             }),
         ]);

rust/src/server/src/routes/openai/completions/convert.rs

@@ -7,7 +7,9 @@ use crate::error::ApiError;
 use crate::lora::LoraModelResolution;
 use crate::routes::openai::completions::validate;
 use crate::routes::openai::utils::structured_outputs::convert_from_response_format_value;
-use crate::utils::{ResolvedRequestContext, convert_logit_bias, merge_kv_transfer_params};
+use crate::utils::{
+    ResolvedRequestContext, convert_logit_bias, merge_ec_transfer_params, merge_kv_transfer_params,
+};
 
 /// Lowered completion request plus the public response metadata carried by
 /// every SSE chunk.
@@ -127,7 +129,7 @@ pub(super) fn prepare_completion_request(
             structured_outputs,
             skip_reading_prefix_cache: None,
             vllm_xargs: merge_kv_transfer_params(
-                request.vllm_xargs,
+                merge_ec_transfer_params(request.vllm_xargs, request.ec_transfer_params.as_ref()),
                 request.kv_transfer_params.as_ref(),
             ),
         },

rust/src/server/src/routes/openai/completions/types.rs

@@ -174,6 +174,9 @@ pub struct CompletionRequest {
     /// KV transfer parameters for disaggregated serving
     pub kv_transfer_params: Option<HashMap<String, Value>>,
 
+    /// Encoder cache transfer parameters for disaggregated serving
+    pub ec_transfer_params: Option<HashMap<String, Value>>,
+
     /// Additional request parameters with string or numeric values for custom
     /// extensions
     pub vllm_xargs: Option<HashMap<String, Value>>,
@@ -209,6 +212,7 @@ pub(super) struct CompletionResponse {
     pub usage: Option<Usage>,
     pub system_fingerprint: Option<String>,
     pub kv_transfer_params: Option<Value>,
+    pub ec_transfer_params: Option<Value>,
 }
 
 /// Mirrors the Python vLLM `CompletionResponseChoice` class.

rust/src/server/src/routes/tests.rs

@@ -76,6 +76,7 @@ fn request_output_with_stop_reason(
         stop_reason,
         events: None,
         kv_transfer_params: None,
+        ec_transfer_params: None,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,
@@ -101,6 +102,7 @@ fn request_output_with_logprobs(
         stop_reason,
         events: None,
         kv_transfer_params: None,
+        ec_transfer_params: None,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,
@@ -116,6 +118,7 @@ fn request_output_with_logprobs_and_kv(
     new_logprobs: Option<Logprobs>,
     new_prompt_logprobs_tensors: Option<Logprobs>,
     kv_transfer_params: Option<serde_json::Value>,
+    ec_transfer_params: Option<serde_json::Value>,
 ) -> EngineCoreOutput {
     EngineCoreOutput {
         request_id: request_id.to_string(),
@@ -127,6 +130,7 @@ fn request_output_with_logprobs_and_kv(
         stop_reason,
         events: None,
         kv_transfer_params,
+        ec_transfer_params,
         trace_headers: None,
         prefill_stats: None,
         routed_experts: None,
@@ -3634,6 +3638,7 @@ async fn non_stream_raw_generate_returns_token_output_envelope() {
                                 Some(sample_logprobs_for_token(44, 45)),
                                 None,
                                 Some(json!({"connector": "x"})),
+                                None,
                             ),
                         ],
                         ..Default::default()

rust/src/server/src/utils.rs

@@ -45,6 +45,24 @@ pub fn merge_kv_transfer_params(
     xargs
 }
 
+/// Merge `ec_transfer_params` into the `vllm_xargs` map, mirroring the Python
+/// vLLM behavior where `ec_transfer_params` is injected into `extra_args` for
+/// engine-core consumption.
+pub fn merge_ec_transfer_params(
+    mut xargs: Option<HashMap<String, Value>>,
+    ec_transfer_params: Option<&HashMap<String, Value>>,
+) -> Option<HashMap<String, Value>> {
+    if let Some(ec_params) = ec_transfer_params {
+        let map = xargs.get_or_insert_with(HashMap::new);
+        map.insert(
+            "ec_transfer_params".to_string(),
+            // This is safe because we know that `ec_params` is already valid JSON.
+            serde_json::to_value(ec_params).unwrap(),
+        );
+    }
+    xargs
+}
+
 /// Convert OpenAI-style `logit_bias` with string token-ID keys into the
 /// internal `HashMap<u32, f32>` representation, validating that every key
 /// parses as a `u32`.

rust/src/text/src/output/decoded.rs

@@ -44,6 +44,9 @@ pub struct Finished {
     pub finish_reason: FinishReason,
     /// Connector-specific KV transfer parameters for disaggregated serving.
     pub kv_transfer_params: Option<serde_json::Value>,
+    /// Connector-specific encoder cache transfer parameters for disaggregated
+    /// serving.
+    pub ec_transfer_params: Option<serde_json::Value>,
 }
 
 /// Internal decoded-text event emitted before higher-level assistant
@@ -151,6 +154,7 @@ pub async fn decoded_text_event_stream(
         let decoder = decoder.as_mut().unwrap();
 
         let kv_transfer_params = output.kv_transfer_params;
+        let ec_transfer_params = output.ec_transfer_params;
         let mut finish_reason = output.finish_reason;
         let mut stop_str_matched = false;
         let suppress_terminal_stop_token = finish_reason.as_ref().is_some_and(|r| r.is_stop())
@@ -275,6 +279,7 @@ pub async fn decoded_text_event_stream(
                     },
                     finish_reason: reason,
                     kv_transfer_params,
+                    ec_transfer_params,
                 }),
             })
             .await;

rust/src/text/src/output/mod.rs

@@ -26,6 +26,9 @@ pub struct CollectedTextOutput {
     pub usage: vllm_llm::TokenUsage,
     /// Connector-specific KV transfer parameters for disaggregated serving.
     pub kv_transfer_params: Option<serde_json::Value>,
+    /// Connector-specific encoder cache transfer parameters for disaggregated
+    /// serving.
+    pub ec_transfer_params: Option<serde_json::Value>,
 }
 
 #[allow(clippy::manual_async_fn, reason = "specify `Send` bound")]
@@ -77,6 +80,7 @@ impl<T: TextOutputStream> T {
                                 finish_reason: FinishReason::Error,
                                 usage: vllm_llm::TokenUsage::default(),
                                 kv_transfer_params: None,
+                                ec_transfer_params: None,
                             })
                         };
 
@@ -85,6 +89,7 @@ impl<T: TextOutputStream> T {
                             collected.finish_reason = finished.finish_reason;
                             collected.usage = finished.usage;
                             collected.kv_transfer_params = finished.kv_transfer_params;
+                            collected.ec_transfer_params = finished.ec_transfer_params;
                             return Ok(collected);
                         }
                     }
@@ -156,6 +161,7 @@ mod tests {
                     },
                     finish_reason: FinishReason::stop_eos(),
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             }),
         ]);
@@ -273,6 +279,7 @@ mod tests {
                     },
                     finish_reason: FinishReason::stop_eos(),
                     kv_transfer_params: None,
+                    ec_transfer_params: None,
                 }),
             }),
         ]);