From 8cc40578336d26b2ed4805d69480022ee17c17af Mon Sep 17 00:00:00 2001
From: Victor Oliveira <victormatheus@gmail.com>
Date: Mon, 14 Sep 2026 18:59:04 +0000
Subject: [PATCH] Fix Qwen3 ONNX export bug related to 0/1 specialization

---
 tensorrt_edgellm/models/default/modeling_default.py | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/tensorrt_edgellm/models/default/modeling_default.py b/tensorrt_edgellm/models/default/modeling_default.py
index 44fb3c0d7..f12b9db0d 100644
--- a/tensorrt_edgellm/models/default/modeling_default.py
+++ b/tensorrt_edgellm/models/default/modeling_default.py
@@ -68,9 +68,11 @@
 # ONNX export spec
 # ---------------------------------------------------------------------------
 
-_BATCH_SIZE = 1
-_SEQ_LEN = 1
-_PAST_LEN = 1
+# PyTorch applies 0/1 specialization which breaks ONNX dynamic batching
+# Using size = 2 avoids this issue
+_BATCH_SIZE = 2
+_SEQ_LEN = 2
+_PAST_LEN = 2
 _MAX_POS = 4096