feat(keyvalue): Filesystem backed KeyValueStore

Running any of the components with `--store-kv file` will use the file
system instead of etcd for discovery. If there is a shared folder, or
for the single-node case, this allows running frontend + backend without
 etcd.

Author: grahamking

Cargo.lock

@@ -225,6 +225,12 @@ def parse_args():
         ),
         help=f"Interval in seconds for polling custom backend metrics. Set to > 0 to enable polling (default: 0=disabled, suggested: 9.2s which is less than typical Prometheus scrape interval). Can be set via {CUSTOM_BACKEND_METRICS_POLLING_INTERVAL_ENV_VAR} env var.",
     )
+    parser.add_argument(
+        "--store-kv",
+        type=str,
+        default=os.environ.get("DYN_STORE_KV", "etcd"),
+        help="Which key-value backend to use: etcd, mem, file. Etcd uses the ETCD_* env vars (e.g. ETCD_ENPOINTS) for connection details. File uses root dir from env var DYN_FILE_KV or defaults to $TMP/dynamo_store_kv.",
+    )
 
     flags = parser.parse_args()
 
@@ -252,8 +258,7 @@ async def async_main():
             os.environ["DYN_METRICS_PREFIX"] = flags.metrics_prefix
 
     loop = asyncio.get_running_loop()
-
-    runtime = DistributedRuntime(loop, is_static)
+    runtime = DistributedRuntime(loop, flags.store_kv, is_static)
 
     def signal_handler():
         asyncio.create_task(graceful_shutdown(runtime))

components/src/dynamo/frontend/main.py

@@ -204,6 +204,12 @@ def parse_args():
         default=False,
         help="Mark this as a decode worker which does not publish KV events and skips prefill cost estimation (default: False)",
     )
+    parser.add_argument(
+        "--store-kv",
+        type=str,
+        default=os.environ.get("DYN_STORE_KV", "etcd"),
+        help="Which key-value backend to use: etcd, mem, file. Etcd uses the ETCD_* env vars (e.g. ETCD_ENPOINTS) for connection details. File uses root dir from env var DYN_FILE_KV or defaults to $TMP/dynamo_store_kv.",
+    )
 
     args = parser.parse_args()
     validate_worker_type_args(args)

components/src/dynamo/mocker/args.py

@@ -72,7 +72,7 @@ async def launch_workers(args, extra_engine_args_path):
         logger.info(f"Creating mocker worker {worker_id + 1}/{args.num_workers}")
 
         # Create a separate DistributedRuntime for this worker (on same event loop)
-        runtime = DistributedRuntime(loop, False)
+        runtime = DistributedRuntime(loop, args.store_kv, False)
         runtimes.append(runtime)
 
         # Create EntrypointArgs for this worker

components/src/dynamo/mocker/main.py

@@ -93,6 +93,12 @@
         "default": None,
         "help": "Dump debug config to the specified file path. If not specified, the config will be dumped to stdout at INFO level.",
     },
+    "store-kv": {
+        "flags": ["--store-kv"],
+        "type": str,
+        "default": os.environ.get("DYN_STORE_KV", "etcd"),
+        "help": "Which key-value backend to use: etcd, mem, file. Etcd uses the ETCD_* env vars (e.g. ETCD_ENPOINTS) for connection details. File uses root dir from env var DYN_FILE_KV or defaults to $TMP/dynamo_store_kv.",
+    },
 }
 
 
@@ -102,6 +108,7 @@ class DynamoArgs:
     component: str
     endpoint: str
     migration_limit: int
+    store_kv: str
 
     # tool and reasoning parser options
     tool_call_parser: Optional[str] = None
@@ -329,6 +336,7 @@ async def parse_args(args: list[str]) -> Config:
         component=parsed_component_name,
         endpoint=parsed_endpoint_name,
         migration_limit=parsed_args.migration_limit,
+        store_kv=parsed_args.store_kv,
         tool_call_parser=tool_call_parser,
         reasoning_parser=reasoning_parser,
         custom_jinja_template=expanded_template_path,

components/src/dynamo/sglang/args.py

@@ -11,7 +11,7 @@
 
 from dynamo.common.config_dump import dump_config
 from dynamo.llm import ModelInput, ModelType
-from dynamo.runtime import DistributedRuntime, dynamo_worker
+from dynamo.runtime import DistributedRuntime
 from dynamo.runtime.logging import configure_dynamo_logging
 from dynamo.sglang.args import Config, DisaggregationMode, parse_args
 from dynamo.sglang.health_check import (
@@ -33,9 +33,12 @@
 configure_dynamo_logging()
 
 
-@dynamo_worker(static=False)
-async def worker(runtime: DistributedRuntime):
+async def worker():
+    config = await parse_args(sys.argv[1:])
+    dump_config(config.dynamo_args.dump_config_to, config)
+
     loop = asyncio.get_running_loop()
+    runtime = DistributedRuntime(loop, config.dynamo_args.store_kv, False)
 
     def signal_handler():
         asyncio.create_task(graceful_shutdown(runtime))
@@ -45,9 +48,6 @@ def signal_handler():
 
     logging.info("Signal handlers will trigger a graceful shutdown of the runtime")
 
-    config = await parse_args(sys.argv[1:])
-    dump_config(config.dynamo_args.dump_config_to, config)
-
     if config.dynamo_args.embedding_worker:
         await init_embedding(runtime, config)
     elif config.dynamo_args.multimodal_processor:

components/src/dynamo/sglang/main.py

@@ -39,7 +39,7 @@
 from dynamo.common.config_dump import dump_config
 from dynamo.common.utils.prometheus import register_engine_metrics_callback
 from dynamo.llm import ModelInput, ModelRuntimeConfig, ModelType, register_llm
-from dynamo.runtime import DistributedRuntime, dynamo_worker
+from dynamo.runtime import DistributedRuntime
 from dynamo.runtime.logging import configure_dynamo_logging
 from dynamo.trtllm.engine import TensorRTLLMEngine, get_llm_engine
 from dynamo.trtllm.health_check import TrtllmHealthCheckPayload
@@ -102,11 +102,13 @@ async def get_engine_runtime_config(
         return runtime_config
 
 
-@dynamo_worker(static=False)
-async def worker(runtime: DistributedRuntime):
-    # Set up signal handler for graceful shutdown
+async def worker():
+    config = cmd_line_args()
+
     loop = asyncio.get_running_loop()
+    runtime = DistributedRuntime(loop, config.store_kv, False)
 
+    # Set up signal handler for graceful shutdown
     def signal_handler():
         # Schedule the shutdown coroutine instead of calling it directly
         asyncio.create_task(graceful_shutdown(runtime))
@@ -116,7 +118,6 @@ def signal_handler():
 
     logging.info("Signal handlers set up for graceful shutdown")
 
-    config = cmd_line_args()
     await init(runtime, config)
 
 

components/src/dynamo/trtllm/main.py

@@ -58,6 +58,7 @@ def __init__(self) -> None:
         self.tool_call_parser: Optional[str] = None
         self.dump_config_to: Optional[str] = None
         self.custom_jinja_template: Optional[str] = None
+        self.store_kv: str = ""
 
     def __str__(self) -> str:
         return (
@@ -88,7 +89,8 @@ def __str__(self) -> str:
             f"reasoning_parser={self.reasoning_parser}, "
             f"tool_call_parser={self.tool_call_parser}, "
             f"dump_config_to={self.dump_config_to},"
-            f"custom_jinja_template={self.custom_jinja_template}"
+            f"custom_jinja_template={self.custom_jinja_template}",
+            f"store_kv={self.store_kv}",
         )
 
 
@@ -278,6 +280,12 @@ def cmd_line_args():
         default=None,
         help="Path to a custom Jinja template file to override the model's default chat template. This template will take precedence over any template found in the model repository.",
     )
+    parser.add_argument(
+        "--store-kv",
+        type=str,
+        default=os.environ.get("DYN_STORE_KV", "etcd"),
+        help="Which key-value backend to use: etcd, mem, file. Etcd uses the ETCD_* env vars (e.g. ETCD_ENPOINTS) for connection details. File uses root dir from env var DYN_FILE_KV or defaults to $TMP/dynamo_store_kv.",
+    )
 
     args = parser.parse_args()
 
@@ -337,6 +345,7 @@ def cmd_line_args():
     config.reasoning_parser = args.dyn_reasoning_parser
     config.tool_call_parser = args.dyn_tool_call_parser
     config.dump_config_to = args.dump_config_to
+    config.store_kv = args.store_kv
 
     # Handle custom jinja template path expansion (environment variables and home directory)
     if args.custom_jinja_template:

components/src/dynamo/trtllm/utils/trtllm_utils.py

@@ -50,6 +50,7 @@ class Config:
     kv_port: Optional[int] = None
     port_range: DynamoPortRange
     custom_jinja_template: Optional[str] = None
+    store_kv: str
 
     # mirror vLLM
     model: str
@@ -188,6 +189,12 @@ def parse_args() -> Config:
             "'USER: <image> please describe the image ASSISTANT:'."
         ),
     )
+    parser.add_argument(
+        "--store-kv",
+        type=str,
+        default=os.environ.get("DYN_STORE_KV", "etcd"),
+        help="Which key-value backend to use: etcd, mem, file. Etcd uses the ETCD_* env vars (e.g. ETCD_ENPOINTS) for connection details. File uses root dir from env var DYN_FILE_KV or defaults to $TMP/dynamo_store_kv.",
+    )
     add_config_dump_args(parser)
 
     parser = AsyncEngineArgs.add_cli_args(parser)

components/src/dynamo/vllm/args.py

@@ -26,7 +26,7 @@
     fetch_llm,
     register_llm,
 )
-from dynamo.runtime import DistributedRuntime, dynamo_worker
+from dynamo.runtime import DistributedRuntime
 from dynamo.runtime.logging import configure_dynamo_logging
 from dynamo.vllm.multimodal_handlers import (
     EncodeWorkerHandler,
@@ -71,16 +71,16 @@ async def graceful_shutdown(runtime):
     logging.info("DistributedRuntime shutdown complete")
 
 
-@dynamo_worker(static=False)
-async def worker(runtime: DistributedRuntime):
+async def worker():
     config = parse_args()
 
+    loop = asyncio.get_running_loop()
+    runtime = DistributedRuntime(loop, config.store_kv, False)
+
     await configure_ports(runtime, config)
     overwrite_args(config)
 
     # Set up signal handler for graceful shutdown
-    loop = asyncio.get_running_loop()
-
     def signal_handler():
         asyncio.create_task(graceful_shutdown(runtime))