From 5b96a911562d8c04d934b123cbb4a09bd39c18ac Mon Sep 17 00:00:00 2001
From: root <root@localhost>
Date: Mon, 17 Aug 2026 14:17:03 +0000
Subject: [PATCH] =?UTF-8?q?[fix]=20baseline4=20=E4=B8=8D=E6=98=AF=20CUDA?=
 =?UTF-8?q?=20=E4=B8=8D=E5=8F=AF=E7=94=A8=EF=BC=8C=E6=98=AF=20.so=20?=
 =?UTF-8?q?=E5=9C=A8=E9=9D=9E=E4=B8=BB=E8=BF=9B=E7=A8=8B=E9=87=8C=E8=A2=AB?=
 =?UTF-8?q?=20dlopen=20=E6=97=B6=E6=AE=B5=E9=94=99=E8=AF=AF=E3=80=82?=
 =?UTF-8?q?=E9=9C=80=E8=A6=81=E7=94=A8=E8=BF=9B=E7=A8=8B=E7=BA=A7=E5=88=AB?=
 =?UTF-8?q?=E7=9A=84=E5=88=A4=E6=96=AD?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 qwen3_6_scripts/patch_ops.sh | 18 ++++++++----------
 qwen3_6_scripts/qwen3_5.py   |  6 ++++++
 2 files changed, 14 insertions(+), 10 deletions(-)

diff --git a/qwen3_6_scripts/patch_ops.sh b/qwen3_6_scripts/patch_ops.sh
index bd944641..3735d411 100755
--- a/qwen3_6_scripts/patch_ops.sh
+++ b/qwen3_6_scripts/patch_ops.sh
@@ -246,18 +246,15 @@ INIT_EOF
 """Apply ix_ops patches at vllm startup."""
 import logging
 _logger = logging.getLogger("ix_startup_patch")
+_applied = False
 def apply():
-    import sys, os
-    # Skip in subprocess inspection (no GPU context → dlopen .so crashes)
-    try:
-        import torch
-        if not torch.cuda.is_available():
-            return 0
-    except Exception:
+    global _applied
+    if _applied:
         return 0
+    _applied = True
+    import sys, os
     # Ensure ex_engine is importable
-    for p in ["/workspace/qwen3_6_scripts",
-              "/workspace"]:
+    for p in ["/workspace/qwen3_6_scripts", "/workspace"]:
         rp = os.path.realpath(p)
         if os.path.isdir(rp) and rp not in sys.path:
             sys.path.insert(0, rp)
@@ -279,7 +276,8 @@ def apply():
     except Exception as e:
         _logger.warning("ix_startup_patch: hot-path patches failed: %s", e)
     return n
-_n_patches = apply()
+# DO NOT call apply() at import time — registry subprocess would crash.
+# apply() is called from qwen3_5.py model init instead.
 STARTUP_EOF
     echo "[patch_ops] deployed ix_startup_patch.py"
 
diff --git a/qwen3_6_scripts/qwen3_5.py b/qwen3_6_scripts/qwen3_5.py
index 343e904a..408d1492 100644
--- a/qwen3_6_scripts/qwen3_5.py
+++ b/qwen3_6_scripts/qwen3_5.py
@@ -2201,6 +2201,12 @@ class Qwen3_5ForCausalLM(nn.Module, HasInnerState, SupportsLoRA,
         multimodal_config: Optional[MultiModalConfig] = None,
         prefix: str = "",
     ) -> None:
+        # Apply ix_bridge operator patches on first model init (safe: GPU is ready)
+        try:
+            from vllm import ix_startup_patch
+            ix_startup_patch.apply()
+        except Exception:
+            pass
         _bi100_model_trace("Qwen3_5ForCausalLM initialization begin")
         super().__init__()
         self.config = config