[Serve][LLM] Tighten body-forwarding gate per /simplify

- router.py: drop the redundant else-branch (defaults already match).
- haproxy.py: `str(bool).lower()` instead of "true"/"false" ternary.
- haproxy_templates.py: flatten nested {% if %}{% if %} for tune.bufsize.
- lua.tmpl, test: drop WHAT-comment narration.

Signed-off-by: Seiji Eicher <seiji@anyscale.com>

Author: eicherseiji

python/ray/llm/_internal/serve/core/ingress/router.py

@@ -94,15 +94,11 @@ async def check_health(self):
 
     @router_app.post("/internal/route")
     async def route(self, request: Request):
-        # HAProxy only forwards a body when
-        # RAY_SERVE_INGRESS_REQUEST_ROUTER_FORWARD_BODY=1, so skip the read
-        # entirely on the default path. See haproxy.py / lua template.
+        body = None
+        body_truncated = False
         if RAY_SERVE_INGRESS_REQUEST_ROUTER_FORWARD_BODY:
             body = await request.body()
             body_truncated = _BODY_TRUNCATED_HEADER in request.headers
-        else:
-            body = None
-            body_truncated = False
         try:
             host, port, replica_id = self._pick_replica(
                 request_body=body, body_truncated=body_truncated

python/ray/serve/_private/haproxy.py

@@ -825,9 +825,7 @@ def _write_ingress_request_router_lua(
 
         content = _load_lua_template().substitute(
             TIMEOUT_S=RAY_SERVE_HAPROXY_INGRESS_REQUEST_ROUTER_TIMEOUT_S,
-            FORWARD_BODY="true"
-            if RAY_SERVE_INGRESS_REQUEST_ROUTER_FORWARD_BODY
-            else "false",
+            FORWARD_BODY=str(RAY_SERVE_INGRESS_REQUEST_ROUTER_FORWARD_BODY).lower(),
             ROUTERS=_format_routers_lua(routers),
             REPLICA_TARGETS=_format_replica_targets_lua(targets),
         )

python/ray/serve/_private/haproxy_templates.py

@@ -28,9 +28,9 @@
     nbthread {{ config.nbthread }}
     {%- if has_ingress_request_router %}
     lua-load-per-thread {{ ingress_request_router_lua_path }}
-    {%- if ingress_request_router_forward_body %}
-    tune.bufsize {{ ingress_request_router_bufsize }}
     {%- endif %}
+    {%- if has_ingress_request_router and ingress_request_router_forward_body %}
+    tune.bufsize {{ ingress_request_router_bufsize }}
     {%- endif %}
     {%- if config.enable_hap_optimization %}
     server-state-base {{ config.server_state_base }}

python/ray/serve/_private/ingress_request_router.lua.tmpl

@@ -2,13 +2,9 @@
 -- router. Templated at config-reload time; placeholders are filled in by
 -- _write_ingress_request_router_lua in haproxy.py.
 --
--- Body forwarding is gated by FORWARD_BODY (rendered from
--- RAY_SERVE_INGRESS_REQUEST_ROUTER_FORWARD_BODY). When false (default), we
--- POST an empty body to /internal/route and skip the
--- `wait-for-body`/`tune.bufsize` cost on every routed request. When true,
--- we forward the (possibly truncated) body so body-aware routing policies
--- can read it; bodies exceeding tune.bufsize are sent with X-Body-Truncated
--- since prefix-routing only needs the head.
+-- Bodies exceeding tune.bufsize are truncated; we forward what we have with
+-- X-Body-Truncated since prefix-routing only needs the head of the body.
+-- See RAY_SERVE_INGRESS_REQUEST_ROUTER_FORWARD_BODY in constants.py.
 
 local ROUTER_REQUEST_TIMEOUT_S = ${TIMEOUT_S}
 local FORWARD_BODY = ${FORWARD_BODY}

python/ray/serve/tests/test_haproxy_api.py

@@ -645,8 +645,6 @@ async def test_ingress_request_router_end_to_end(haproxy_api_cleanup, monkeypatc
     """Run actual HAProxy against a fake router + two replicas; verify a POST
     is pinned to the replica the router selects, while a GET (which doesn't
     trigger the router-routed path) is not."""
-    # Body forwarding to /internal/route is opt-in. This test asserts the body
-    # made it through the router, so flip it on.
     monkeypatch.setattr(
         "ray.serve._private.haproxy.RAY_SERVE_INGRESS_REQUEST_ROUTER_FORWARD_BODY",
         True,