fix: propagate vLLM --stream-interval to Dynamo frontend (#8101)

Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

components/src/dynamo/frontend/vllm_processor.py

@@ -466,12 +466,30 @@ class EngineFactory:
 
         input_processor = InputProcessor(vllm_config)
         tokenizer = input_processor.get_tokenizer()
+
+        # Resolve stream_interval: env var override > backend config > default (20)
+        stream_interval = self.stream_interval
+        if not os.getenv("DYN_VLLM_STREAM_INTERVAL"):
+            backend_interval = (
+                mdc.runtime_config().get("runtime_data", {}).get("stream_interval")
+            )
+            if backend_interval is not None:
+                try:
+                    stream_interval = max(1, int(backend_interval))
+                except (TypeError, ValueError):
+                    logger.warning(
+                        "Invalid stream_interval=%r from backend runtime_config, "
+                        "using default=%d",
+                        backend_interval,
+                        stream_interval,
+                    )
+
         output_processor = OutputProcessor(
             tokenizer,
             log_stats=False,
-            stream_interval=self.stream_interval,
+            stream_interval=stream_interval,
         )
-        logger.info("vLLM OutputProcessor stream_interval=%d", self.stream_interval)
+        logger.info("vLLM OutputProcessor stream_interval=%d", stream_interval)
 
         tool_parser_name = self.flags.tool_call_parser or mdc.runtime_config().get(
             "tool_call_parser"

components/src/dynamo/vllm/args.py

@@ -126,10 +126,10 @@ def cross_validate_config(
     """Validate dynamo and engine config together. This should not modify the configs."""
 
     if hasattr(engine_config, "stream_interval") and engine_config.stream_interval != 1:
-        logger.warning(
-            "--stream-interval is currently not respected in Dynamo. "
-            "Dynamo uses its own post-processing implementation on the frontend, "
-            "bypassing vLLM's OutputProcessor buffering."
+        logger.info(
+            "--stream-interval=%d will be propagated to the Dynamo frontend. "
+            "Set DYN_VLLM_STREAM_INTERVAL env var to override.",
+            engine_config.stream_interval,
         )
 
     # Validate --gms-shadow-mode requires --load-format gms

components/src/dynamo/vllm/main.py

@@ -656,6 +656,13 @@ async def register_vllm_model(
         config.exclude_tools_when_tool_choice_none
     )
 
+    # Propagate stream_interval so the frontend can respect --stream-interval.
+    # set_engine_specific requires a JSON-encoded string (the Rust binding
+    # parses it with serde_json::from_str); str(int) happens to be valid JSON.
+    stream_interval = getattr(config.engine_args, "stream_interval", None)
+    if stream_interval is not None:
+        runtime_config.set_engine_specific("stream_interval", str(stream_interval))
+
     # Get data_parallel_size from vllm_config (defaults to 1)
     dp_range = get_dp_range_for_worker(vllm_config)
     runtime_config.data_parallel_start_rank = dp_range[0]