[mxfp8 moe training] simplify e8m0 -> fp32 calc

stack-info: PR: #3201, branch: danielvegamyhre/stack/80

Author: danielvegamyhre

torchao/prototype/mx_formats/kernels.py

@@ -1371,10 +1371,9 @@ def _dequant_mxfp8_kernel(
 
     @triton.jit
     def _e8m0_to_fp32(scale_e8m0):
-        e8m0_exponent_bias = 127
         e8m0_nan_val = 255
-        s_offset = scale_e8m0.to(tl.int16) - e8m0_exponent_bias
-        s_fp = tl.exp2(s_offset.to(tl.float32))
+        fp32_mantissa_bits = 23
+        s_fp = scale_e8m0 << fp32_mantissa_bits
         s_fp = tl.where(scale_e8m0 != e8m0_nan_val, s_fp, float("nan"))
         return s_fp.to(tl.float32)