From 104c9904a52ee0c481394e957317423a743b37d0 Mon Sep 17 00:00:00 2001
From: Mikei386 <44135113+Mikei386@users.noreply.github.com>
Date: Wed, 26 Aug 2026 07:06:03 +0200
Subject: [PATCH] Stream llama responses without proxy buffering

---
 router/ai_profile_router.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/router/ai_profile_router.py b/router/ai_profile_router.py
index 0f37d80..ae3b849 100755
--- a/router/ai_profile_router.py
+++ b/router/ai_profile_router.py
@@ -2091,7 +2091,11 @@ class Handler(BaseHTTPRequestHandler):
         self.end_headers()
         try:
             while True:
-                chunk = resp.read(16384)
+                # ``read(n)`` may wait until the complete buffer is filled.
+                # That defeats SSE: a client sees no token for a long time and
+                # may time out while llama.cpp is already generating. read1()
+                # returns the next currently available wire chunk instead.
+                chunk = resp.read1(16384)
                 if not chunk:
                     break
                 self.wfile.write(chunk)