diff --git a/unsloth/kernels/swiglu.py b/unsloth/kernels/swiglu.py
index 4e9b7ba2ae..3106656f32 100644
--- a/unsloth/kernels/swiglu.py
+++ b/unsloth/kernels/swiglu.py
@@ -24,11 +24,11 @@ def _fg_kernel(e, g, h, n_elements, BLOCK_SIZE : tl.constexpr,):
     offsets = block_idx*BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
     mask = offsets < n_elements
 
-    e_row = tl.load(e + offsets, mask = mask, other = 0).to(tl.float32)
+    e_row = tl.load(e + offsets, mask = mask, other = 0)#.to(tl.float32)
     g_row = tl.load(g + offsets, mask = mask, other = 0)#.to(tl.float32)
 
     # f = e * sigmoid(e)
-    f_row = e_row / (1 + tl.exp(-e_row))
+    f_row = e_row / (1 + tl.exp(-e_row.to(tl.float32)).to(g_row.dtype))
     f_row = f_row.to(g_row.dtype) # Exact copy from HF
     # h = f * g
     h_row = f_row * g_row