From ff5d40aad59bfb7d0ee3efb588db2a4311ee7158 Mon Sep 17 00:00:00 2001 From: sora <2075279110@qq.com> Date: Fri, 11 Sep 2026 09:24:16 +0000 Subject: [PATCH] =?UTF-8?q?Read=20timeout=20300=E2=86=92600s=20base=20(GLM?= =?UTF-8?q?=20gateway=2030+s=20startup=20on=20100k+=20token=20inputs)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude --- evalharness/model/adapter.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/evalharness/model/adapter.py b/evalharness/model/adapter.py index 3b17c16..feebd20 100644 --- a/evalharness/model/adapter.py +++ b/evalharness/model/adapter.py @@ -481,7 +481,9 @@ class OpenAICompatible(ModelAdapter): # read timeout scales with the generation budget: a 32k-token # CoT legitimately takes 10+ minutes; a fixed 300s timeout would # kill and retry it forever (looks like a hang) - _rt = max(self.extra.get('timeout', 300), + # generous: GLM gateway takes 30+ seconds to start responding + # on 100k+ token inputs, even before any generation begins + _rt = max(self.extra.get('timeout', 600), int(payload.get('max_tokens') or 0) * 0.15) async with httpx.AsyncClient(timeout=httpx.Timeout( connect=self.extra.get('connect_timeout', 15),