[P/D] Heterogeneous TP

tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh

@@ -8,7 +8,9 @@ MODELS=(
 
 # Number of prefill and decode instances to create
 NUM_PREFILL_INSTANCES=${NUM_PREFILL_INSTANCES:-1} # Default to 1
-NUM_DECODE_INSTANCES=${NUM_DECODE_INSTANCES:-2}   # Default to 2
+NUM_DECODE_INSTANCES=${NUM_DECODE_INSTANCES:-1}   # Default to 1
+PREFILLER_TP_SIZE=${PREFILLER_TP_SIZE:-1}
+DECODER_TP_SIZE=${DECODER_TP_SIZE:-1}
 
 # Find the git repository root directory
 GIT_ROOT=$(git rev-parse --show-toplevel)
@@ -74,9 +76,10 @@ run_tests_for_model() {
   for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs
     GPU_ID=$((i % $(get_num_gpus)))
+
     # Calculate port number (base port + instance number)
     PORT=$((8100 + i))
-    # Calculate side channel port
+    # Calculate side channel port. Avoid clash with with TP workers. 
     SIDE_CHANNEL_PORT=$((5559 + i))
 
     echo "Starting prefill instance $i on GPU $GPU_ID, port $PORT"
@@ -87,6 +90,7 @@ run_tests_for_model() {
     --enforce-eager \
     --disable-log-requests \
     --gpu-memory-utilization 0.2 \
+    --tensor-parallel-size $PREFILLER_TP_SIZE \
     --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\"}'"
 
     if [ -n "$model_args" ]; then
@@ -109,7 +113,7 @@ run_tests_for_model() {
     # Calculate port number (base port + instance number)
     PORT=$((8200 + i))
     # Calculate side channel port
-    SIDE_CHANNEL_PORT=$((5659 + i))
+    SIDE_CHANNEL_PORT=$((5659 + i * $DECODER_TP_SIZE))
 
     echo "Starting decode instance $i on GPU $GPU_ID, port $PORT"
 
@@ -119,6 +123,7 @@ run_tests_for_model() {
     --enforce-eager \
     --disable-log-requests \
     --gpu-memory-utilization 0.2 \
+    --tensor-parallel-size $DECODER_TP_SIZE \
     --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\"}'"
 
     if [ -n "$model_args" ]; then

tests/v1/kv_connector/nixl_integration/test_accuracy.py

@@ -14,6 +14,7 @@
 # Model-specific expected values
 EXPECTED_VALUES = {
     "Qwen/Qwen3-0.6B": 0.41,
+    "deepseek-ai/deepseek-vl2-small": 0.59
 }
 
 SIMPLE_PROMPT = "The best part about working on vLLM is that I got to meet so many people across various different organizations like UCB, Google, and Meta which means",  # noqa: E501

vllm/distributed/kv_transfer/kv_connector/utils.py

@@ -3,11 +3,12 @@
 """
 KV cache helper for store.
 """
+
 import torch
 
 import vllm.envs as envs
 from vllm import _custom_ops as ops
-from vllm.config import VllmConfig
+from vllm.config import VllmConfig, get_current_vllm_config
 from vllm.logger import init_logger
 
 logger = init_logger(__name__)
@@ -90,3 +91,18 @@ def put_kv_to_cache(self, model_executable: torch.nn.Module, keys, values,
                 layer.self_attn.attn._k_scale,
                 layer.self_attn.attn._v_scale,
             )
+
+
+def get_kv_connector_cache_layout():
+    vllm_config = get_current_vllm_config()
+    kv_config = vllm_config.kv_transfer_config
+    if vllm_config.model_config is None:
+        logger.warning("Unable to detect current VLLM config. " \
+        "Defaulting to NHD kv cache layout.")
+    else:
+        use_mla = vllm_config.model_config.use_mla
+        if not use_mla and kv_config.kv_connector == "NixlConnector":
+            logger.info("NixlConnector detected. Setting KV cache " \
+            "layout to HND for better xfer performance.")
+            return "HND"
+    return "NHD"