feat: Decode -> Prefill cached kv transfer (#340)

examples/llm/components/prefill_worker.py

@@ -64,6 +64,12 @@ class PrefillWorker:
             print("Prefill must be done eagerly, setting to True")
             self.engine_args.enforce_eager = True
 
+        if self.engine_args.enable_prefix_caching is not False:
+            print(
+                "Prefix caching is not supported yet in prefill worker, setting to False"
+            )
+            self.engine_args.enable_prefix_caching = False
+
     @async_on_start
     async def async_init(self):
         self._engine_context = build_async_engine_client_from_engine_args(
@@ -115,6 +121,7 @@ class PrefillWorker:
             is_remote_decode=True,
             decode_block_ids=request.block_ids,
             decode_engine_id=request.engine_id,
+            decode_computed_block_ids=request.computed_block_ids,
         )
 
         # TODO check if metadata has changed

examples/llm/configs/disagg_router.yaml

@@ -30,22 +30,25 @@ Router:
 
 VllmWorker:
   model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
-  enforce-eager: true
   kv-transfer-config: '{"kv_connector":"DynamoNixlConnector"}'
   block-size: 64
   max-model-len: 16384
   max-num-batched-tokens: 16384
+  remote-prefill: true
   conditional-disagg: true
   max-local-prefill-length: 10
   max-prefill-queue-size: 2
   tensor-parallel-size: 1
   router: kv
   enable-prefix-caching: true
+  ServiceArgs:
+    workers: 1
+    resources:
+      gpu: 1
 
 # TODO - set all of these but model as default
 PrefillWorker:
   model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
-  enforce-eager: true
   kv-transfer-config: '{"kv_connector":"DynamoNixlConnector"}'
   block-size: 64
   max-model-len: 16384