fix: Fix port collision and pass prefill param in kvbm connector (#4411)

Signed-off-by: krishung5 <krish@nvidia.com>

components/src/dynamo/vllm/args.py

@@ -325,10 +325,16 @@ def parse_args() -> Config:
 
 def create_kv_events_config(config: Config) -> Optional[KVEventsConfig]:
     """Create KVEventsConfig for prefix caching if needed."""
+    if config.is_decode_worker:
+        logger.info(
+            f"Decode worker detected (is_decode_worker={config.is_decode_worker}): "
+            f"kv_events_config disabled (decode workers don't publish KV events)"
+        )
+        return None
 
     # If prefix caching is not enabled, no events config needed
-    if not config.engine_args.enable_prefix_caching or config.is_decode_worker:
-        logger.info("No kv_events_config required")
+    if not config.engine_args.enable_prefix_caching:
+        logger.info("No kv_events_config required: prefix caching is disabled")
         return None
 
     # There is a bug with KV events publishing when LORA is enabled.

examples/backends/vllm/launch/disagg_kvbm_router.sh

@@ -10,19 +10,11 @@ export PYTHONHASHSEED=0
 # Common configuration
 MODEL="Qwen/Qwen3-0.6B"
 
-# run decode router with kv-overlap-score-weight 0 for pure load balancing
 python -m dynamo.frontend \
     --router-mode kv \
     --http-port 8000 \
-    --kv-overlap-score-weight 0 \
     --router-reset-states &
 
-# run standalone router service for prefill workers
-python -m dynamo.router \
-    --endpoint dynamo.prefill.generate \
-    --router-reset-states \
-    --no-track-active-blocks &
-
 # two decode workers (without KVBM)
 # --enforce-eager is added for quick deployment. for production use, need to remove this flag
 CUDA_VISIBLE_DEVICES=0 python3 -m dynamo.vllm \
@@ -30,6 +22,7 @@ CUDA_VISIBLE_DEVICES=0 python3 -m dynamo.vllm \
     --enforce-eager \
     --is-decode-worker &
 
+VLLM_NIXL_SIDE_CHANNEL_PORT=20096 \
 CUDA_VISIBLE_DEVICES=1 python3 -m dynamo.vllm \
     --model $MODEL \
     --enforce-eager \
@@ -37,6 +30,7 @@ CUDA_VISIBLE_DEVICES=1 python3 -m dynamo.vllm \
 
 # two prefill workers with KVBM enabled
 # Each worker needs unique ZMQ ports to avoid KVBM coordination conflicts
+VLLM_NIXL_SIDE_CHANNEL_PORT=20097 \
 DYN_KVBM_LEADER_ZMQ_PUB_PORT=56001 \
 DYN_KVBM_LEADER_ZMQ_ACK_PORT=56002 \
 CUDA_VISIBLE_DEVICES=2 DYN_KVBM_CPU_CACHE_GB=20 \
@@ -44,10 +38,10 @@ CUDA_VISIBLE_DEVICES=2 DYN_KVBM_CPU_CACHE_GB=20 \
     --model $MODEL \
     --enforce-eager \
     --is-prefill-worker \
-    --connector kvbm &
+    --connector kvbm nixl \
+    --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20081"}' &
 
-DYN_VLLM_KV_EVENT_PORT=20081 \
-VLLM_NIXL_SIDE_CHANNEL_PORT=20097 \
+VLLM_NIXL_SIDE_CHANNEL_PORT=20098 \
 DYN_KVBM_LEADER_ZMQ_PUB_PORT=56003 \
 DYN_KVBM_LEADER_ZMQ_ACK_PORT=56004 \
 CUDA_VISIBLE_DEVICES=3 DYN_KVBM_CPU_CACHE_GB=20 \
@@ -55,4 +49,5 @@ CUDA_VISIBLE_DEVICES=3 DYN_KVBM_CPU_CACHE_GB=20 \
     --model $MODEL \
     --enforce-eager \
     --is-prefill-worker \
-    --connector kvbm
+    --connector kvbm nixl \
+    --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20082"}'