fix(http_handler): remove __del__ to prevent in-flight streaming failures on TTL eviction

Fixes #24929

When the TTL cache evicts an httpx client, the __del__ handler calls
close() which propagates into the TCP connection pool, killing in-flight
streaming responses. Clients observe periodic httpx.ReadTimeout spikes
aligned with _DEFAULT_TTL_FOR_HTTPX_CLIENTS (3600s).

The __del__ handlers are unnecessary — Python's GC already cleans up
the underlying httpx.AsyncClient/httpx.Client resources via their own
finalizers. The explicit close() in __del__ is harmful because:

1. TTL eviction drops the last reference, triggering __del__
2. __del__ calls close() which closes the TCP connection pool
3. In-flight streaming responses lose their connection mid-transfer
4. Clients experience ReadTimeout bursts every TTL interval

Removing __del__ from both AsyncHTTPHandler and HTTPHandler lets the
httpx clients be reclaimed by GC naturally without disrupting active
connections.

Added 2 tests verifying that deleting a handler does NOT close the
underlying httpx client.

litellm/llms/custom_httpx/http_handler.py

@@ -725,12 +725,6 @@ class AsyncHTTPHandler:
         response.raise_for_status()
         return response
 
-    def __del__(self) -> None:
-        try:
-            asyncio.get_running_loop().create_task(self.close())
-        except Exception:
-            pass
-
     @staticmethod
     def _create_async_transport(
         ssl_context: Optional[ssl.SSLContext] = None,
@@ -1182,12 +1176,6 @@ class HTTPHandler:
         except Exception as e:
             raise e
 
-    def __del__(self) -> None:
-        try:
-            self.close()
-        except Exception:
-            pass
-
     def _create_sync_transport(self) -> Optional[HTTPTransport]:
         """
         Create an HTTP transport with IPv4 only if litellm.force_ipv4 is True.

tests/test_litellm/llms/custom_httpx/test_http_handler.py

@@ -658,3 +658,49 @@ async def test_httpx_handler_uses_env_user_agent(monkeypatch):
         assert req.headers.get("User-Agent") == "Claude Code"
     finally:
         await handler.close()
+
+
+class TestDelHandlerRemoved:
+    """
+    Verify that AsyncHTTPHandler and HTTPHandler do NOT close the underlying
+    httpx client when the wrapper is garbage-collected.
+
+    Prior to this fix, __del__ called self.close(), which propagated into
+    the TCP connection pool and killed in-flight streaming responses whenever
+    the TTL cache evicted a client.  See #24929.
+    """
+
+    @pytest.mark.asyncio
+    async def test_async_handler_no_close_on_del(self):
+        """AsyncHTTPHandler should NOT close the httpx client on __del__."""
+        handler = AsyncHTTPHandler(timeout=httpx.Timeout(timeout=10.0))
+
+        # Grab reference to the underlying client
+        client = handler.client
+
+        # Simulate GC / TTL eviction
+        del handler
+
+        # The client should still be open (not closed by __del__)
+        assert not client.is_closed
+
+        # Cleanup
+        await client.aclose()
+
+    def test_sync_handler_no_close_on_del(self):
+        """HTTPHandler should NOT close the httpx client on __del__."""
+        from litellm.llms.custom_httpx.http_handler import HTTPHandler
+
+        handler = HTTPHandler(timeout=httpx.Timeout(timeout=10.0))
+
+        # Grab reference to the underlying client
+        client = handler.client
+
+        # Simulate GC / TTL eviction
+        del handler
+
+        # The client should still be open (not closed by __del__)
+        assert not client.is_closed
+
+        # Cleanup
+        client.close()