diff --git a/sgl-kernel/python/sgl_kernel/gemm.py b/sgl-kernel/python/sgl_kernel/gemm.py
index 5cfdd412b..3af4751fc 100644
--- a/sgl-kernel/python/sgl_kernel/gemm.py
+++ b/sgl-kernel/python/sgl_kernel/gemm.py
@@ -465,7 +465,7 @@ def scaled_fp4_experts_quant(
     # larger models.
     import os
 
-    MAX_TOKENS_PER_EXPERT = os.environ.get("MODELOPT_MAX_TOKENS_PER_EXPERT", 65536)
+    MAX_TOKENS_PER_EXPERT = int(os.environ.get("MODELOPT_MAX_TOKENS_PER_EXPERT", 65536))
     assert m_numtopk <= MAX_TOKENS_PER_EXPERT * topk, (
         f"m_numtopk must be less than MAX_TOKENS_PER_EXPERT("
         f"{MAX_TOKENS_PER_EXPERT})"