feat: remove --connector flag for vLLM backend (LLM-90) (#6450)

Signed-off-by: alec-flowers <aflowers@nvidia.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

examples/multimodal/launch/lora/lora_agg.sh

@@ -140,13 +140,11 @@ echo "Starting vLLM worker..."
 
 # --enable-lora: Enable LoRA adapter support in vLLM engine
 # --max-lora-rank: Maximum LoRA rank (increase if your adapters have higher rank)
-# --connector none: No KV transfer needed for aggregated serving
 CUDA_VISIBLE_DEVICES="$GPU_DEVICE" \
 DYN_SYSTEM_PORT="$SYSTEM_PORT" \
     python -m dynamo.vllm \
         --enable-multimodal \
         --model "$MODEL_NAME" \
-        --connector none \
         --enable-lora \
         --max-lora-rank "$MAX_LORA_RANK" \
         "${MODEL_SPECIFIC_ARGS[@]}" \

tests/frontend/test_prompt_embeds.py

@@ -73,8 +73,6 @@ class VllmPromptEmbedsWorkerProcess(ManagedProcess):
             "dynamo.vllm",
             "--model",
             TEST_MODEL,
-            "--connector",
-            "none",
             "--max-model-len",
             "4096",
             "--discovery-backend",

tests/frontend/test_vllm.py

@@ -93,8 +93,6 @@ class VllmWorkerProcess(ManagedProcess):
             "harmony",
             "--dyn-reasoning-parser",
             "gpt_oss",
-            "--connector",
-            "none",
         ]
 
         env = os.environ.copy()

tests/kvbm_integration/test_consolidator_router_e2e.py

@@ -354,8 +354,8 @@ def llm_worker(frontend_server, test_directory, runtime_services, engine_type):
             "dynamo.vllm",
             "--model",
             model_id,
-            "--connector",
-            "kvbm",
+            "--kv-transfer-config",
+            '{"kv_connector":"DynamoConnector","kv_connector_module_path":"kvbm.vllm_integration.connector","kv_role":"kv_both"}',
             "--enforce-eager",  # For faster startup in tests
         ]
     else:  # trtllm
@@ -777,8 +777,8 @@ class TestConsolidatorRouterE2E:
                     "dynamo.vllm",
                     "--model",
                     model_id,
-                    "--connector",
-                    "kvbm",
+                    "--kv-transfer-config",
+                    '{"kv_connector":"DynamoConnector","kv_connector_module_path":"kvbm.vllm_integration.connector","kv_role":"kv_both"}',
                     "--enforce-eager",
                     "--enable-prefix-caching",
                     "--num-gpu-blocks-override",

tests/kvbm_integration/test_determinism_disagg.py

@@ -143,8 +143,8 @@ class LLMServerManager:
             "16",
             "--max-model-len",
             "8000",  # required to fit on L4 GPU when using 8b model
-            "--connector",
-            "nixl",
+            "--kv-transfer-config",
+            '{"kv_connector":"NixlConnector","kv_role":"kv_both"}',
         ]
 
         # Construct prefiller command
@@ -160,9 +160,8 @@ class LLMServerManager:
             "16",
             "--max-model-len",
             "8000",  # required to fit on L4 GPU when using 8b model
-            "--connector",
-            "kvbm",
-            "nixl",
+            "--kv-transfer-config",
+            '{"kv_connector":"PdConnector","kv_role":"kv_both","kv_connector_extra_config":{"connectors":[{"kv_connector":"DynamoConnector","kv_connector_module_path":"kvbm.vllm_integration.connector","kv_role":"kv_both"},{"kv_connector":"NixlConnector","kv_role":"kv_both"}]},"kv_connector_module_path":"kvbm.vllm_integration.connector"}',
         ]
 
         # GPU blocks override

tests/lmcache/deploy-lmcache_enabled-dynamo.sh

@@ -27,4 +27,4 @@ echo "🔧 Starting dynamo worker with LMCache enabled..."
 
 python -m dynamo.frontend &
 
-python3 -m dynamo.vllm --model $MODEL_URL --connector lmcache
+python3 -m dynamo.vllm --model $MODEL_URL --kv-transfer-config '{"kv_connector":"LMCacheConnectorV1","kv_role":"kv_both"}'

tests/mm_router/test_vllm_mm_router_e2e.py

@@ -109,8 +109,6 @@ class VLLMWorkerProcess(ManagedProcess):
                 "0.85",
                 "--max-model-len",
                 "8192",
-                "--connector",
-                "none",
                 "--served-model-name",
                 f"{VLLM_MM_MODEL}__internal",
             ],

tests/router/test_router_e2e_with_vllm.py

@@ -148,7 +148,7 @@ class VLLMProcess:
         # - When data_parallel_size is set, launch one process per DP rank
         # - Each process gets --data-parallel-rank and --data-parallel-size
         # - Each process runs on its own GPU via CUDA_VISIBLE_DEVICES
-        # - --connector nixl enables KV cache transfer between ranks
+        # - --kv-transfer-config enables KV cache transfer between ranks
 
         for worker_idx in range(num_workers):
             # Calculate GPU device for this process
@@ -207,7 +207,7 @@ class VLLMProcess:
                         str(data_parallel_size),
                         # "--data-parallel-address", "127.0.0.1",  # Required for DP coordination
                         # "--data-parallel-rpc-port", "13345",  # RPC port for DP coordination
-                        # "--connector", "nixl",  # Required for KV transfer between DP ranks
+                        # "--kv-transfer-config", '{"kv_connector":"NixlConnector","kv_role":"kv_both"}',  # Required for KV transfer between DP ranks
                     ]
                 )