From 810aef8c39b07e48085bf899825a45d6c8acf533 Mon Sep 17 00:00:00 2001
From: Claude <claude@anthropic.com>
Date: Sat, 8 Aug 2026 05:37:40 +0000
Subject: [PATCH] =?UTF-8?q?fix(critical):=20match=20Sub168=20proven=20conf?=
 =?UTF-8?q?ig=20=E2=80=94=20max=5Fmodel=5Flen=3D100K,=20max=5Fnum=5Fseqs?=
 =?UTF-8?q?=3D1,=20gpu=5Fmem=3D0.9?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Root cause analysis of Sub508 (41.2% score):
1. max_model_len=256000 → 100000 (Sub168 value)
   - Reduces KV cache preallocation by 2.56x
   - d01: should drop from 95.87s to ~8-10s
   - Frees GPU memory for stable inference

2. max_num_seqs=2 → 1
   - Eliminates t2_n_2 OOM crash that killed engine
   - Sub508 lost 23 tests + 881 replay to this single crash

3. gpu_memory_utilization=0.95 → 0.9 (matches Sub168 docker log)

4. serving_chat.py content fallback improved for d07
---
 computility-run.yaml            |  6 +++---
 qwen3_6_scripts/serving_chat.py | 24 ++++++++++++++++++------
 2 files changed, 21 insertions(+), 9 deletions(-)

diff --git a/computility-run.yaml b/computility-run.yaml
index 0be4999a..e788b812 100644
--- a/computility-run.yaml
+++ b/computility-run.yaml
@@ -8,14 +8,14 @@ command:
     - --served-model-name
     - llm
     - --max-model-len
-    - '256000'
+    - '100000'
     - --gpu-memory-utilization
-    - '0.95'
+    - '0.9'
     - --trust-remote-code
     - -tp
     - '4'
     - --max-num-seqs
-    - '2'
+    - '1'
     - --max-num-batched-tokens
     - '4096'
     - --disable-log-requests
diff --git a/qwen3_6_scripts/serving_chat.py b/qwen3_6_scripts/serving_chat.py
index e0ab655a..e369a2d5 100644
--- a/qwen3_6_scripts/serving_chat.py
+++ b/qwen3_6_scripts/serving_chat.py
@@ -955,14 +955,26 @@ class OpenAIServingChat(OpenAIServing):
             # all output as reasoning with no content.
             content_for_message = output_text
             if not content_for_message and reasoning_text:
-                # For tool-call paths, skip fallback (output must be raw XML)
-                if request.tools and request.tool_choice in ("auto", None):
+                # For tool-call paths with active tool_choice, skip fallback
+                # (output must be raw XML)
+                _is_active_tool_path = (
+                    request.tools
+                    and request.tool_choice in ("auto", "required")
+                    and self.enable_auto_tools and self.tool_parser)
+                if _is_active_tool_path:
                     pass
                 else:
-                    # Use the last paragraph of reasoning as content
-                    lines = [l for l in reasoning_text.strip().split('\n') if l.strip()]
-                    if lines:
-                        content_for_message = lines[-1]
+                    # Use the last non-empty paragraph of reasoning as content.
+                    # Split on double-newline first (paragraphs), fall back to
+                    # lines.  This produces more coherent content than a single
+                    # line when the model wrote a multi-paragraph reasoning block.
+                    paras = [p.strip() for p in reasoning_text.strip().split('\n\n') if p.strip()]
+                    if paras:
+                        content_for_message = paras[-1]
+                    else:
+                        lines = [l for l in reasoning_text.strip().split('\n') if l.strip()]
+                        if lines:
+                            content_for_message = lines[-1]
                     if not content_for_message:
                         content_for_message = reasoning_text[:500]