fix: prevent Qwen reasoning and tool loops

2026-08-24 20:06:26 +02:00
parent dec4709ae9
commit 335c9a8501
5 changed files with 35 additions and 4 deletions
@@ -176,6 +176,11 @@ services:
      - --jinja
      - --reasoning
      - auto
+      # Bound each individual thinking phase. Long agent jobs can still use
+      # many phases around tool calls, but one degenerate reasoning loop can
+      # no longer consume the complete response budget indefinitely.
+      - --reasoning-budget
+      - "8192"
      - --reasoning-preserve
      - --host
      - 0.0.0.0
@@ -189,9 +194,11 @@ services:
      - --no-mmap
      - --no-ui
      - --temperature
-      - "0.2"
+      # Qwen3.8's official thinking-mode sampler. The former 0.2 setting was
+      # overly deterministic and could lock reasoning into verbatim loops.
+      - "1.0"
      - --top-p
-      - "0.8"
+      - "0.95"
      - --top-k
      - "20"
      - --device