diff --git a/deepinfer/neuron_ops/neuron_compile.py b/deepinfer/neuron_ops/neuron_compile.py
index a671df6..76ed23f 100644
--- a/deepinfer/neuron_ops/neuron_compile.py
+++ b/deepinfer/neuron_ops/neuron_compile.py
@@ -43,16 +43,19 @@ def _encode_scalar(value: float) -> int:
 
 
 def _decode(x_u8: torch.Tensor) -> torch.Tensor:
-    """Exact e4m3fn bits -> fp32, elementwise (no table, no bitcast): (8 + m) * 2^(e - 10), subnormals m * 2^-9."""
-    b = x_u8.to(torch.int32)
-    e = (b >> 3) & 15
-    m = (b & 7).to(torch.float32)
-    p = torch.full_like(m, 2.0**-10)
-    for bit, factor in ((1, 2.0), (2, 4.0), (4, 16.0), (8, 256.0)):
-        p = p * torch.where((e & bit) != 0, factor, 1.0)
-    v = torch.where(e == 0, m * 2.0**-9, (m + 8.0) * p)
-    v = torch.where(b >= 128, -v, v)
-    return torch.where((b & 127) == 127, torch.full_like(v, float("nan")), v)
+    """Exact e4m3fn bits -> fp32, elementwise (no table, no bitcast): (8 + m) * 2^(e - 10), subnormals m * 2^-9.
+
+    Float arithmetic only (Vector / Scalar engines): integer shifts and masks run on GpSimd, and this decode is in
+    every graph that reads an fp8 activation or cache. ``exp2`` is exact on integers on the device (see
+    nki_attention._pow2)."""
+    b = x_u8.to(torch.float32)
+    neg = b >= 128.0
+    mag = torch.where(neg, b - 128.0, b)
+    e = torch.floor(mag * 0.125)
+    m = mag - 8.0 * e
+    v = torch.where(e == 0.0, m * 2.0**-9, (m + 8.0) * torch.exp2(e - 10.0))
+    v = torch.where(neg, -v, v)
+    return torch.where(mag == 127.0, torch.full_like(v, float("nan")), v)
 
 
 def _decode_as(dtype: torch.dtype) -> Callable[[torch.Tensor], torch.Tensor]:
@@ -65,7 +68,7 @@ def _decode_as(dtype: torch.dtype) -> Callable[[torch.Tensor], torch.Tensor]:
 def _rne(m: torch.Tensor) -> torch.Tensor:
     f = torch.floor(m)
     d = m - f
-    odd = torch.remainder(f, 2.0) == 1.0
+    odd = (f - 2.0 * torch.floor(f * 0.5)) == 1.0  # not torch.remainder: it lowers to GpSimd
     return f + ((d > 0.5) | ((d == 0.5) & odd)).to(m.dtype)