update http.py coverage

Author: viraatc

src/inference_endpoint/endpoint_client/http.py

@@ -663,7 +663,7 @@ class HttpRequestTemplate:
         static_prefix: Pre-merged request line + host header bytes
     """
 
-    __slots__ = ("static_prefix", "_extra_headers_cache", "extra_cached_headers")
+    __slots__ = ("static_prefix", "cached_headers")
 
     # Pre-encoded general headers
     HEADERS_STREAMING = (
@@ -675,8 +675,7 @@ class HttpRequestTemplate:
 
     def __init__(self, static_prefix: bytes):
         self.static_prefix = static_prefix
-        self._extra_headers_cache: dict[frozenset, bytes] = {}
-        self.extra_cached_headers = b""
+        self.cached_headers = b""
 
     @classmethod
     def from_url(cls, host: str, port: int, path: str = "/") -> HttpRequestTemplate:
@@ -706,20 +705,14 @@ def from_url(cls, host: str, port: int, path: str = "/") -> HttpRequestTemplate:
 
     def cache_headers(self, headers: dict[str, str]) -> None:
         """
-        Pre-cache extra headers to avoid first-call encoding overhead.
-
-        Call this during setup to prime the cache for headers that will be
-        used repeatedly at runtime.
+        Pre-encode extra headers during setup to avoid encoding in the hotpath.
 
         Args:
             headers: Headers to pre-encode and cache
         """
-        cache_key = frozenset(headers.items())
-        if cache_key not in self._extra_headers_cache:
-            self._extra_headers_cache[cache_key] = "".join(
-                f"{k}: {v}\r\n" for k, v in headers.items()
-            ).encode("utf-8", "surrogateescape")
-            self.extra_cached_headers = b"".join(self._extra_headers_cache.values())
+        self.cached_headers += "".join(
+            f"{k}: {v}\r\n" for k, v in headers.items()
+        ).encode("utf-8", "surrogateescape")
 
     def build_request(
         self,
@@ -748,25 +741,22 @@ def build_request(
             return b"".join(
                 [
                     self.static_prefix,
-                    self.extra_cached_headers,
+                    self.cached_headers,
                     content_type_headers,
                     content_length,
                     body,
                 ]
             )
 
-        # Slow path: extra headers (~1us uncached, ~50ns per extra-header cached)
-        cache_key = frozenset(extra_headers.items())
-        if (extra := self._extra_headers_cache.get(cache_key)) is None:
-            extra = "".join(f"{k}: {v}\r\n" for k, v in extra_headers.items()).encode(
-                "utf-8", "surrogateescape"
-            )
-            self._extra_headers_cache[cache_key] = extra
-
+        # Slow path: extra headers are encoded per-call;
+        # use cache_headers() at setup time for headers that repeat every request.
+        extra = "".join(f"{k}: {v}\r\n" for k, v in extra_headers.items()).encode(
+            "utf-8", "surrogateescape"
+        )
         return b"".join(
             [
                 self.static_prefix,
-                self.extra_cached_headers,
+                self.cached_headers,
                 content_type_headers,
                 extra,
                 content_length,