Update README.md

Also next morning I realized that Mamba module doesn't come with built-in normalization and residual as I thought,
which explains why BF16 attempts failed. Oops.

README.md

@@ -32,3 +32,7 @@ rather than last as if it was with [1, 1, 1, 0, 1, 1, 0, 0]. Intuition with that
 * I tried to use BF16 originally, but model went into nan (with default big LR) or gradients were so small weights didn't change(smaller LR). I switched back to F32, however some layers still initialize weight with factor x0.001 as I hoped it
 would stop model from going to nan.
 
+--------
+
+Also next morning I realized that Mamba module doesn't come with built-in normalization and residual as I thought, 
+which explains why BF16 attempts failed. Oops.