[fix bug] Fix the error that RotaryEmbedding is converted to a non-fp32 format during training, and add a compatible method for the llama model. (#239)

Co-authored-by: YWMditto <862779238@qq.com>

internlm/model/embedding.py

@@ -137,15 +137,13 @@ class RotaryEmbedding(torch.nn.Module):
         """ """
         super().__init__()
         # Generate and save the inverse frequency buffer (non trainable)
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim))
-        self.register_buffer("inv_freq", inv_freq)
+        self.inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim))
         self.scale_base = scale_base
-        scale = (
+        self.scale = (
             (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
             if scale_base > 0
             else None
         )
-        self.register_buffer("scale", scale)
 
         self._seq_len_cached = 0
         self._cos_cached = None
@@ -220,3 +218,15 @@ class RotaryEmbedding(torch.nn.Module):
                 self._cos_k_cached[seqlen_offset:],
                 self._sin_k_cached[seqlen_offset:],
             )
+
+    def _single_forward(self, x, indexes=0):
+        assert self.scale is None
+        self._update_cos_sin_cache(x, indexes)
+        x = x[None, ...]
+        ret = legacy_apply_rotary_embed(x, self._cos_cached[indexes], self._sin_cached[indexes]).squeeze(0)
+        return ret
+
+    def _single_eval_forward(self, x, seqlen_offset=0):
+        assert self.scale is None
+        self._update_cos_sin_cache(x, seqlen_offset + x.shape[1])
+        return legacy_apply_rotary_embed(x, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:])