From 356454def6b5efed7a6afea69f5f059468227ea7 Mon Sep 17 00:00:00 2001
From: "devin-ai-integration[bot]"
 <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Fri, 16 Oct 2026 17:13:10 +0000
Subject: [PATCH] refactor(proxy): streaming loop hold budget is
 _STREAM_LOOP_HOLD_MS = 50

---
 litellm/proxy/proxy_server.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index f8902f6276a..97d2f8a3386 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -10012,7 +10012,7 @@ def _pop_complete_sse_frame(buffer: str) -> tuple[str | None, str]:
 _STREAM_KEEPALIVE: Final = object()
 
 _KEEPALIVE_MIN_SECONDS: Final = 1.0
-_STREAM_LOOP_HOLD_SECONDS: Final = 0.05
+_STREAM_LOOP_HOLD_MS: Final = 50
 _KEEPALIVE_MAX_SECONDS: Final = 300.0
 _EMPTY_MAPPING: Final[Mapping[str, object]] = MappingProxyType({})
 _EMPTY_HEADERS: Final[Mapping[str, str]] = MappingProxyType({})
@@ -10292,7 +10292,7 @@ async def async_data_generator(
         async for item in stream_source:
             # Unblock the event loop: a stream whose chunks are already buffered never awaits
             # between chunks, so without this yield one long stream holds the loop for its whole body
-            if time.monotonic() - loop_released_at >= _STREAM_LOOP_HOLD_SECONDS:
+            if (time.monotonic() - loop_released_at) * 1000 >= _STREAM_LOOP_HOLD_MS:
                 await asyncio.sleep(0)
                 loop_released_at = time.monotonic()
             if item is _STREAM_KEEPALIVE: