fix(sglang): disable piecewise CUDA graph in launch scripts (#8609)

examples/backends/sglang/launch/agg.sh

@@ -85,6 +85,7 @@ python3 -m "$WORKER_MODULE" \
   --trust-remote-code \
   --skip-tokenizer-init \
   --enable-metrics \
+  --disable-piecewise-cuda-graph \
   $GPU_MEM_ARGS \
   "${TRACE_ARGS[@]}" \
   "${EXTRA_ARGS[@]}" &

examples/backends/sglang/launch/agg_router.sh

@@ -86,6 +86,7 @@ python3 -m dynamo.sglang \
   --trust-remote-code \
   "${KV_EVENTS_ARGS_1[@]}" \
   --enable-metrics \
+  --disable-piecewise-cuda-graph \
   $GPU_MEM_ARGS \
   "${TRACE_ARGS[@]}" &
 
@@ -98,6 +99,7 @@ CUDA_VISIBLE_DEVICES=1 python3 -m dynamo.sglang \
   --trust-remote-code \
   "${KV_EVENTS_ARGS_2[@]}" \
   --enable-metrics \
+  --disable-piecewise-cuda-graph \
   $GPU_MEM_ARGS \
   "${TRACE_ARGS[@]}" &
 

examples/backends/sglang/launch/disagg.sh

@@ -81,6 +81,7 @@ python3 -m dynamo.sglang \
   --port 40000 \
   --disaggregation-transfer-backend nixl \
   --enable-metrics \
+  --disable-piecewise-cuda-graph \
   $GPU_MEM_ARGS \
   "${TRACE_ARGS[@]}" &
 
@@ -97,6 +98,7 @@ CUDA_VISIBLE_DEVICES=1 python3 -m dynamo.sglang \
   --host 0.0.0.0 \
   --disaggregation-transfer-backend nixl \
   --enable-metrics \
+  --disable-piecewise-cuda-graph \
   $GPU_MEM_ARGS \
   "${TRACE_ARGS[@]}" &
 

examples/backends/sglang/launch/disagg_router.sh

@@ -76,6 +76,7 @@ python3 -m dynamo.sglang \
   --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}' \
   --disaggregation-transfer-backend nixl \
   --enable-metrics \
+  --disable-piecewise-cuda-graph \
   "${TRACE_ARGS[@]}" &
 
 # run prefill worker
@@ -91,6 +92,7 @@ CUDA_VISIBLE_DEVICES=1 python3 -m dynamo.sglang \
   --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5558"}' \
   --disaggregation-transfer-backend nixl \
   --enable-metrics \
+  --disable-piecewise-cuda-graph \
   "${TRACE_ARGS[@]}" &
 
 # run decode worker
@@ -106,6 +108,7 @@ CUDA_VISIBLE_DEVICES=3 python3 -m dynamo.sglang \
   --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5560"}' \
   --disaggregation-transfer-backend nixl \
   --enable-metrics \
+  --disable-piecewise-cuda-graph \
   "${TRACE_ARGS[@]}" &
 
 # run decode worker
@@ -121,6 +124,7 @@ CUDA_VISIBLE_DEVICES=2 python3 -m dynamo.sglang \
   --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5559"}' \
   --disaggregation-transfer-backend nixl \
   --enable-metrics \
+  --disable-piecewise-cuda-graph \
   "${TRACE_ARGS[@]}" &
 
 # Wait for any worker to exit (keeps script running)