base: 35 RPS; safety, 75 RPS

# What does this PR do?


## Test Plan
# What does this PR do?


## Test Plan

llama_stack/core/routers/inference.py

@@ -525,9 +525,8 @@ class InferenceRouter(Inference):
 
         response = await self._nonstream_openai_chat_completion(provider, params)
 
-        # Store the response with the ID that will be returned to the client
         if self.store:
-            await self.store.store_chat_completion(response, messages)
+            asyncio.create_task(self.store.store_chat_completion(response, messages))
 
         if self.telemetry:
             metrics = self._construct_metrics(
@@ -855,4 +854,4 @@ class InferenceRouter(Inference):
                     object="chat.completion",
                 )
                 logger.debug(f"InferenceRouter.completion_response: {final_response}")
-                await self.store.store_chat_completion(final_response, messages)
+                asyncio.create_task(self.store.store_chat_completion(final_response, messages))

llama_stack/core/server/server.py

@@ -73,6 +73,7 @@ from llama_stack.providers.inline.telemetry.meta_reference.telemetry import (
     TelemetryAdapter,
 )
 from llama_stack.providers.utils.telemetry.tracing import (
+    BACKGROUND_LOGGER,
     CURRENT_TRACE_CONTEXT,
     end_trace,
     setup_logger,
@@ -204,6 +205,10 @@ async def sse_generator(event_gen_coroutine):
 
 
 async def log_request_pre_validation(request: Request):
+    # Skip expensive body parsing if debug logging is disabled
+    if not logger.isEnabledFor(logging.DEBUG):
+        return
+        
     if request.method in ("POST", "PUT", "PATCH"):
         try:
             body_bytes = await request.body()
@@ -305,6 +310,10 @@ class TracingMiddleware:
             logger.debug(f"Bypassing custom routing for FastAPI built-in path: {path}")
             return await self.app(scope, receive, send)
 
+        # Quick exit if telemetry is disabled - skip expensive route matching and tracing
+        if BACKGROUND_LOGGER is None:
+            return await self.app(scope, receive, send)
+
         if not hasattr(self, "route_impls"):
             self.route_impls = initialize_route_impls(self.impls, self.external_apis)
 

llama_stack/providers/remote/inference/vllm/vllm.py

@@ -362,7 +362,9 @@ class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
         return AsyncOpenAI(
             base_url=self.config.url,
             api_key=self.config.api_token,
-            http_client=httpx.AsyncClient(verify=self.config.tls_verify),
+            http_client=httpx.AsyncClient(
+                verify=self.config.tls_verify, limits=httpx.Limits(max_connections=1000, max_keepalive_connections=1000)
+            ),
         )
 
     async def completion(