[CI] Add P2pNccl integration test + rename nixl_integration to pd_integration

.buildkite/test_areas/distributed.yaml

@@ -181,21 +181,33 @@ steps:
   num_devices: 4
   source_file_dependencies:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
+    - tests/v1/kv_connector/pd_integration/
   commands:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+    - bash v1/kv_connector/pd_integration/config_sweep_accuracy_test.sh
 
 - label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs)
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
+    - tests/v1/kv_connector/pd_integration/
   commands:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+    - DP_EP=1 bash v1/kv_connector/pd_integration/config_sweep_accuracy_test.sh
+
+- label: P2pNccl Connector PD accuracy (2 GPUs)
+  timeout_in_minutes: 25
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  optional: true
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/p2p/
+    - tests/v1/kv_connector/pd_integration/
+    - benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py
+  commands:
+    - GPU_MEMORY_UTILIZATION=0.4 bash v1/kv_connector/pd_integration/run_p2p_nccl_accuracy_test.sh
 
 - label: Pipeline + Context Parallelism (4 GPUs))
   timeout_in_minutes: 60

tests/v1/kv_connector/pd_integration/run_p2p_nccl_accuracy_test.sh

@@ -0,0 +1,71 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# P2pNccl Connector accuracy test script
+# Uses disagg_prefill_proxy_server.py which encodes KV addresses in request IDs
+
+set -xe
+
+# Required for NCCL on some systems (consistent with other distributed tests)
+export NCCL_CUMEM_HOST_ENABLE=0
+
+MODEL_NAME=${MODEL_NAME:-"Qwen/Qwen3-0.6B"}
+GPU_MEMORY_UTILIZATION=${GPU_MEMORY_UTILIZATION:-0.3}
+GIT_ROOT=$(git rev-parse --show-toplevel)
+
+# Get the host IP that vLLM will use for KV transfer
+# This must match what P2pNcclEngine binds to
+KV_HOST=$(python3 -c "from vllm.utils.network_utils import get_ip; print(get_ip())")
+
+# Trap the SIGINT signal (triggered by Ctrl+C)
+trap 'kill $(jobs -pr) 2>/dev/null' SIGINT SIGTERM EXIT
+
+wait_for_server() {
+  local port=$1
+  timeout 600 bash -c "until curl -s localhost:${port}/v1/completions > /dev/null; do sleep 1; done"
+}
+
+# Cleanup any existing instances
+pkill -f "vllm serve" || true
+sleep 2
+
+# Prefill instance (kv_producer) - port 8100, kv_port 14579
+CUDA_VISIBLE_DEVICES=0 vllm serve $MODEL_NAME \
+  --port 8100 \
+  --max-model-len 4096 \
+  --enforce-eager \
+  --gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
+  --kv-transfer-config '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_port":14579}' &
+
+# Decode instance (kv_consumer) - port 8200, kv_port 14580 (different!)
+CUDA_VISIBLE_DEVICES=1 vllm serve $MODEL_NAME \
+  --port 8200 \
+  --max-model-len 4096 \
+  --enforce-eager \
+  --gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
+  --kv-transfer-config '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_port":14580}' &
+
+echo "Waiting for prefill instance on port 8100 to start..."
+wait_for_server 8100
+echo "Waiting for decode instance on port 8200 to start..."
+wait_for_server 8200
+
+# Start proxy server with P2pNccl request ID encoding
+# This proxy encodes prefill and decode KV addresses in the request ID
+# Note: --kv-host must match the IP that vLLM's P2pNcclEngine binds to
+python3 ${GIT_ROOT}/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py \
+  --port 8192 \
+  --prefill-url http://localhost:8100 \
+  --decode-url http://localhost:8200 \
+  --kv-host ${KV_HOST} \
+  --prefill-kv-port 14579 \
+  --decode-kv-port 14580 &
+
+sleep 5
+
+# Run accuracy test
+TEST_MODEL=$MODEL_NAME python3 -m pytest -s -x ${GIT_ROOT}/tests/v1/kv_connector/pd_integration/test_accuracy.py
+
+pkill -f "vllm serve" || true
+echo "P2pNccl accuracy test completed!"