base: 35 RPS; safety, 75 RPS

# What does this PR do?


## Test Plan
# What does this PR do?


## Test Plan

docs/source/distributions/k8s-benchmark/perf_test_config.yaml

@@ -0,0 +1,19 @@
+version: '2'
+image_name: perf-test-demo
+apis:
+- inference
+providers:
+  inference:
+  - provider_id: vllm-inference
+    provider_type: remote::vllm
+    config:
+      url: ${env.VLLM_URL:=http://localhost:8001/v1}
+      max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
+      api_token: ${env.VLLM_API_TOKEN:=fake}
+      tls_verify: ${env.VLLM_TLS_VERIFY:=false}
+models:
+- model_id: ${env.INFERENCE_MODEL}
+  provider_id: vllm-inference
+  model_type: llm
+server:
+  port: 8322

docs/source/distributions/k8s-benchmark/stack_run_config.yaml

@@ -3,6 +3,7 @@ image_name: kubernetes-benchmark-demo
 apis:
 - agents
 - inference
+- safety
 - telemetry
 - tool_runtime
 - vector_io
@@ -30,6 +31,11 @@ providers:
         db: ${env.POSTGRES_DB:=llamastack}
         user: ${env.POSTGRES_USER:=llamastack}
         password: ${env.POSTGRES_PASSWORD:=llamastack}
+  safety:
+  - provider_id: llama-guard
+    provider_type: inline::llama-guard
+    config:
+      excluded_categories: []
   agents:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
@@ -95,6 +101,8 @@ models:
 - model_id: ${env.INFERENCE_MODEL}
   provider_id: vllm-inference
   model_type: llm
+shields:
+- shield_id: ${env.SAFETY_MODEL:=meta-llama/Llama-Guard-3-1B}
 vector_dbs: []
 datasets: []
 scoring_fns: []

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(