docs: Update dynamo serve disagg example (#212)

Co-authored-by: ptarasiewicz@nvidia.com <Piotr Tarasiewicz>

examples/llm/components/prefill_worker.py

@@ -25,7 +25,6 @@ from vllm.entrypoints.openai.api_server import (
     build_async_engine_client_from_engine_args,
 )
 from vllm.inputs.data import TokensPrompt
-from vllm.logger import logger as vllm_logger
 from vllm.remote_prefill import RemotePrefillParams, RemotePrefillRequest
 
 from dynamo.sdk import (
@@ -76,7 +75,6 @@ class PrefillWorker:
         if self.engine_args.enforce_eager is not True:
             print("Prefill must be done eagerly, setting to True")
             self.engine_args.enforce_eager = True
-        print("PrefillWorker initialized")
 
     @async_on_start
     async def async_init(self):
@@ -93,6 +91,7 @@ class PrefillWorker:
         await self._metadata_store.put(metadata.engine_id, metadata)
         task = asyncio.create_task(self.prefill_queue_handler())
         task.add_done_callback(lambda _: print("prefill queue handler created"))
+        print("PrefillWorker initialized")
 
     async def prefill_queue_handler(self):
         print("[DEBUG] prefill queue handler entered")
@@ -115,7 +114,7 @@ class PrefillWorker:
                 # need to test and check how much overhead it is
                 prefill_request = await prefill_queue.dequeue_prefill_request()
                 if prefill_request is not None:
-                    vllm_logger.info(f"Dequeued prefill request: {prefill_request}")
+                    print(f"Dequeued prefill request: {prefill_request.request_id}")
                     async for _ in self.generate(prefill_request):
                         pass
 

examples/llm/components/worker.py

@@ -175,12 +175,12 @@ class VllmWorker:
                 is_remote_prefill=True,
                 remote_prefill_request_callback=self.get_remote_prefill_request_callback(),
             )
-            vllm_logger.info(
+            print(
                 f"Prefilling remotely for request {request.request_id} with length {len(request.engine_prompt['prompt_token_ids'])}"
             )
         else:
             remote_prefill_params = None
-            vllm_logger.info(
+            print(
                 f"Prefilling locally for request {request.request_id} with length {len(request.engine_prompt['prompt_token_ids'])}"
             )
 

examples/llm/configs/disagg.yaml

@@ -21,28 +21,24 @@ Frontend:
 
 Processor:
   model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
-  block-size: 64
-  max-model-len: 16384
   router: round-robin
 
 VllmWorker:
+  # vllm enging args
   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
+
+  # dynamo args
+  remote-prefill: true
   conditional-disagg: true
-  tensor-parallel-size: 1
-  remote_prefill: true
-  max_local_prefill_length: 10
+  max-local-prefill-length: 10
 
-# TODO - set all of these but model as default
 PrefillWorker:
+  # vllm enging args
   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
+
+  # dynamo arg for local deployment
   cuda-visible-device-offset: 1