[PD] turn on KV cache preemption.swap mode (#1925)

must work with hpu extention commit
HabanaAI/vllm-hpu-extension@0bc6b42

Author: hlin99

pd_xpyd/1p_start_prefill.sh

@@ -45,7 +45,7 @@ CMD=(
 # Check if XPYD_LOG is set
 if [ -n "$XPYD_LOG" ]; then
     timestamp=$(date +"%Y%m%d_%H%M%S")
-    log_file="$XPYD_LOG/ProxyServer_${timestamp}.log"
+    log_file="$XPYD_LOG/Prefill_${timestamp}.log"
     echo "Logging to $log_file..."
 
     # Execute command and log stdout+stderr using tee

pd_xpyd/dp_d_env.sh

@@ -45,6 +45,8 @@ env | grep VLLM_DECODE_BS
 env | grep VLLM_DECODE_BLOCK
 # ***************************************  bucketing ends ************************************* #
 
+SWAP_SPACE=64 # GB, memory per rank for preemption.swap.
+
 # decode specific settings
 export VLLM_DP_SIZE=2
 export VLLM_USE_V1=0

pd_xpyd/dp_start_decode.sh

@@ -65,6 +65,8 @@ do
     --distributed_executor_backend mp
     --enable-reasoning
     --reasoning-parser deepseek_r1
+    --preemption-mode swap
+    --swap-space "$SWAP_SPACE"
     $kv_cache_dtype_arg
     --kv-transfer-config '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_consumer"}'
   )