Llama‘s freqs_cos data loss as for convert dtype

[WeiMa01]

• The freqs_cos of the Llama model structure of ExecuTorch is FP32 at the time of its creation
  calculation of freqs_cos by precompute_freqs_cis：
  freqs = 1.0 / ( theta ** (torch.arange(0, dim, 2, device="cpu")[: (dim // 2)].float() / dim) )

• The model is loaded by get_eager_model and is converted from FP32 to BF16 (Llama original model data type). This can lead to data loss and affect inferences
  ` def get_eager_model(self) -> torch.nn.Module:

    if self.dtype:
        return self.model_.to(self.dtype)`
  

• I think freqs_cos should always keep FP32

[kimishpatel]

Makes sense. Does this happen when you override dtype during model export via -d bf16?

[kimishpatel]

cc @jackzhxng

[jackzhxng]

@WeiMa01 Seems like the issue I was trying to fix in this PR: #8441 a while back. I close that PR in favor of this one - #8488, which I am trying to land today. This should address your problem

[WeiMa01]

@jackzhxng ，thank you， it seems that you remove dtype converting of get_eager_model. it can solve my issue

[jackzhxng]

@WeiMa01 fix merged in #9430