From 01002acf36a8aefae1ce8533243142db27b7d9b2 Mon Sep 17 00:00:00 2001
From: "sherif.waly"
 <sherif.waly@slurm-login-sherif-waly-54ac345-0.slurm-login-all-users.tenant-slurm.svc.cluster.local>
Date: Mon, 13 Apr 2026 21:30:35 +0000
Subject: [PATCH] fix(aiohttp_transport): set total=None in per-request
 ClientTimeout to prevent silent 300s cap
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

When LiteLLMAiohttpTransport builds a per-request ClientTimeout, it sets
sock_connect, sock_read, and connect from the httpx timeout extensions
but leaves `total` unset (sentinel). aiohttp resolves the sentinel to
DEFAULT_TIMEOUT.total (300 seconds), which silently caps every request
at 300s — even when the caller configured a much longer timeout.

This is particularly problematic when combined with retries: a caller
setting timeout=3600s sees requests fail after 600s (300s × 2 attempts)
instead of the expected 3600s.

Setting total=None explicitly disables the overall duration limit,
letting the per-operation timeouts (sock_read, sock_connect, connect)
control the deadline individually. The caller's own asyncio.wait_for
provides the overall deadline.

The existing streaming test (test_handle_async_request_streaming_does_not_timeout_on_total_duration)
continues to pass, since total=None is the correct behavior for both
streaming and non-streaming requests.
---
 .../llms/custom_httpx/aiohttp_transport.py    |  1 +
 .../custom_httpx/test_aiohttp_transport.py    | 68 +++++++++++++++++++
 2 files changed, 69 insertions(+)

diff --git a/litellm/llms/custom_httpx/aiohttp_transport.py b/litellm/llms/custom_httpx/aiohttp_transport.py
index 132191c946c..111158b852b 100644
--- a/litellm/llms/custom_httpx/aiohttp_transport.py
+++ b/litellm/llms/custom_httpx/aiohttp_transport.py
@@ -272,6 +272,7 @@ class LiteLLMAiohttpTransport(AiohttpTransport):
             "allow_redirects": False,
             "auto_decompress": False,
             "timeout": ClientTimeout(
+                total=None,
                 sock_connect=timeout.get("connect"),
                 sock_read=timeout.get("read"),
                 connect=timeout.get("pool"),
diff --git a/tests/test_litellm/llms/custom_httpx/test_aiohttp_transport.py b/tests/test_litellm/llms/custom_httpx/test_aiohttp_transport.py
index 6e2e60ba0dd..27d836e9af3 100644
--- a/tests/test_litellm/llms/custom_httpx/test_aiohttp_transport.py
+++ b/tests/test_litellm/llms/custom_httpx/test_aiohttp_transport.py
@@ -544,3 +544,71 @@ async def test_handle_session_closed_during_request():
     assert counts["requests"] == 2  # First request failed, second succeeded
     assert counts["sessions"] == 2  # Created 2 sessions for retry
     assert response.status_code == 200
+
+
+@pytest.mark.asyncio
+async def test_per_request_timeout_sets_total_to_none():
+    """
+    Verify that the per-request ClientTimeout sets total=None explicitly.
+
+    When total is left as sentinel (unset), aiohttp resolves it to
+    DEFAULT_TIMEOUT.total (300s).  Since the per-request timeout replaces
+    the session-level timeout entirely, this silently caps every request
+    at 300 seconds — even when the caller configured a much longer timeout
+    (e.g. 3600s via sock_read).  Setting total=None disables the overall
+    duration limit so that per-operation timeouts (sock_read, etc.) and
+    the caller's own asyncio.wait_for deadline control the lifetime.
+    """
+    captured_timeouts: list = []
+
+    class CapturingSession:
+        def __init__(self):
+            self.closed = False
+            try:
+                self._loop = asyncio.get_running_loop()
+            except RuntimeError:
+                self._loop = None
+
+        def request(self, *args, **kwargs):
+            captured_timeouts.append(kwargs.get("timeout"))
+
+            class Resp:
+                status = 200
+                headers = {}
+
+                async def __aenter__(self):
+                    return self
+
+                async def __aexit__(self, *a):
+                    pass
+
+                @property
+                def content(self):
+                    class C:
+                        async def iter_chunked(self, size):
+                            yield b""
+
+                    return C()
+
+            return Resp()
+
+    transport = LiteLLMAiohttpTransport(client=lambda: CapturingSession())  # type: ignore
+
+    request = httpx.Request("GET", "http://example.com")
+    request.extensions["timeout"] = {
+        "connect": 3600.0,
+        "read": 3600.0,
+        "pool": 3600.0,
+    }
+
+    await transport.handle_async_request(request)
+
+    assert len(captured_timeouts) == 1
+    ct = captured_timeouts[0]
+    assert ct.total is None, (
+        f"Expected total=None but got {ct.total}; sentinel total resolves "
+        f"to aiohttp DEFAULT_TIMEOUT.total (300s), silently capping requests"
+    )
+    assert ct.sock_read == 3600.0
+    assert ct.sock_connect == 3600.0
+    assert ct.connect == 3600.0