diff --git a/linalg/x86_64/fma/fma_mmm_i32_scalars.tmpliq b/linalg/x86_64/fma/fma_mmm_i32_scalars.tmpliq
index 427dbaa774..4de1a5d49c 100644
--- a/linalg/x86_64/fma/fma_mmm_i32_scalars.tmpliq
+++ b/linalg/x86_64/fma/fma_mmm_i32_scalars.tmpliq
@@ -8,18 +8,16 @@
 {% include "fma_mmm_ymm_scalar.tmpliq" label:"scalar_sub_flipped", op:"vpsubd", from:from, to:to, flipped: true%}
 
 {{L}}leaky_relu:
-    // can only use zmm12 to zmm15
+    // can only use ymm12 to ymm15
     // ymm15 <- alpha
-    vbroadcastss    zmm15, dword ptr [rdi + 8]
+    vbroadcastss    ymm15, dword ptr [rdi + 8]
     // ymm14 <- all zero
-    vpxorq          zmm14, zmm14, zmm14
+    vpxor          ymm14, ymm14, ymm14
 
     {% for reg in (from..to) %}
-        // ymm12 <- alpha * x
-        vpmulld     zmm12, zmm{{reg}}, zmm15
-        vpcmpd      k1, zmm14, zmm{{reg}}, 1 // 1 means LT
-        vblendmps   zmm{{reg}} {k1}, zmm12, zmm{{reg}}
+        vpmulld     ymm12, ymm{{reg}}, ymm15
+        vpcmpgtd    ymm13, ymm14, ymm{{reg}}
+        vblendvps   ymm{{reg}}, ymm{{reg}}, ymm12, ymm13
     {% endfor %}
-    // select muled of orginal
 
     jmp    {{L}}non_linear_loop