[Core][KVConnector] Support HMA+NixlConnector (#35758)

Signed-off-by: NickLucche <nlucches@redhat.com>

tests/v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh

@@ -12,6 +12,7 @@ tp_configs=(
   "GPU_MEMORY_UTILIZATION=0.8 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # MLA case
   "GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny"
   "GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=1 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny"
+  "GPU_MEMORY_UTILIZATION=0.8 MODEL_NAMES=google/gemma-3-4b-it VLLM_SERVE_EXTRA_ARGS=--max-model-len,8192" # SW model
 )
 dp_ep_configs=(
 "DP_EP=1 GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # MLA+P-TP1, D-DPEP=2 (TP=1)
@@ -26,6 +27,14 @@ else
   configs=("${tp_configs[@]}")
 fi
 
+if [[ -n "${ENABLE_HMA_FLAG:-}" ]]; then
+  # Append ENABLE_HMA_FLAG=1 to each config in the selected array
+  echo "ENABLE_HMA_FLAG is set, appending ENABLE_HMA_FLAG=1 to each config"
+  for i in "${!configs[@]}"; do
+    configs[$i]="ENABLE_HMA_FLAG=1 ${configs[$i]}"
+  done
+fi
+
 run_tests() {
   local label=$1
   local extra_args=$2

tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh

@@ -5,6 +5,12 @@ set -xe
 KV_BUFFER_DEVICE="cuda"  # Default to cuda
 ATTENTION_BACKEND=""  # Default to empty (use vllm default)
 CROSS_LAYERS_BLOCKS="False"
+ENABLE_HMA_VAR=""  # Default to empty (HMA disabled by default for kv connector)
+# Check for ENABLE_HMA_FLAG environment variable
+if [[ -n "${ENABLE_HMA_FLAG:-}" ]]; then
+  ENABLE_HMA_VAR="--no-disable-hybrid-kv-cache-manager"
+fi
+
 while [[ $# -gt 0 ]]; do
   case $1 in
     --kv_buffer_device)
@@ -31,6 +37,12 @@ echo "Running accuracy tests with kv_buffer_device=$KV_BUFFER_DEVICE"
 if [[ -n "$ATTENTION_BACKEND" ]]; then
   echo "Using attention backend: $ATTENTION_BACKEND"
 fi
+if [[ -n "$ENABLE_HMA_VAR" ]]; then
+  echo "HMA (Hybrid KV Cache Manager) enabled"
+fi
+if [[ -n "$VLLM_SERVE_EXTRA_ARGS" ]]; then
+  echo "vLLM serve extra args: $VLLM_SERVE_EXTRA_ARGS"
+fi
 
 DECODER_KV_LAYOUT=${DECODER_KV_LAYOUT:-"HND"} # Default to HND, optional NHD
 if [[ "$DECODER_KV_LAYOUT" == "NHD" ]]; then
@@ -70,6 +82,8 @@ DECODER_TP_SIZE=${DECODER_TP_SIZE:-1}
 GPU_MEMORY_UTILIZATION=${GPU_MEMORY_UTILIZATION:-0.2}
 PREFILL_BLOCK_SIZE=${PREFILL_BLOCK_SIZE:-128}
 DECODE_BLOCK_SIZE=${DECODE_BLOCK_SIZE:-128}
+# Comma-separated extra args for vllm serve (e.g. --max-model-len,2048)
+VLLM_SERVE_EXTRA_ARGS=${VLLM_SERVE_EXTRA_ARGS:-}
 
 # Find the git repository root directory
 GIT_ROOT=$(git rev-parse --show-toplevel)
@@ -151,14 +165,24 @@ run_tests_for_model() {
     --gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
     --tensor-parallel-size $PREFILLER_TP_SIZE \
     --kv-transfer-config '$KV_CONFIG'"
+    if [[ -n "$VLLM_SERVE_EXTRA_ARGS" ]]; then
+      IFS=',' read -r -a extra_args <<< "$VLLM_SERVE_EXTRA_ARGS"
+      for arg in "${extra_args[@]}"; do
+        BASE_CMD="${BASE_CMD} $arg"
+      done
+    fi
 
     # Add attention backend config if specified
     if [[ -n "$ATTENTION_BACKEND" ]]; then
       BASE_CMD="${BASE_CMD} --attention-backend=$ATTENTION_BACKEND"
     fi
 
+    # Add HMA flag if specified
+    if [[ -n "$ENABLE_HMA_VAR" ]]; then
+      BASE_CMD="${BASE_CMD} $ENABLE_HMA_VAR"
+    fi
+    
     FULL_CMD="$BASE_CMD"
-
     eval "$FULL_CMD &"
 
     # Store host and port for proxy configuration
@@ -193,12 +217,23 @@ run_tests_for_model() {
     --block-size ${DECODE_BLOCK_SIZE} \
     --gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
     --kv-transfer-config '$KV_CONFIG'"
+    if [[ -n "$VLLM_SERVE_EXTRA_ARGS" ]]; then
+      IFS=',' read -r -a extra_args <<< "$VLLM_SERVE_EXTRA_ARGS"
+      for arg in "${extra_args[@]}"; do
+        BASE_CMD="${BASE_CMD} $arg"
+      done
+    fi
 
     # Add attention backend config if specified
     if [[ -n "$ATTENTION_BACKEND" ]]; then
       BASE_CMD="${BASE_CMD} --attention-backend=$ATTENTION_BACKEND"
     fi
 
+    # Add HMA flag if specified
+    if [[ -n "$ENABLE_HMA_VAR" ]]; then
+      BASE_CMD="${BASE_CMD} $ENABLE_HMA_VAR"
+    fi
+
   # DP-EP attention mode
   if [[ -z "$DP_EP" ]]; then
     BASE_CMD="${BASE_CMD} --tensor-parallel-size $DECODER_TP_SIZE"

tests/v1/kv_connector/nixl_integration/test_accuracy.py

@@ -17,6 +17,7 @@ EXPECTED_VALUES = {
     "deepseek-ai/deepseek-vl2-small": 0.59,
     "deepseek-ai/deepseek-vl2-tiny": 0.19,
     "deepseek-ai/DeepSeek-V2-Lite-Chat": 0.65,
+    "google/gemma-3-4b-it": 0.74,
 }
 
 SIMPLE_PROMPT = (