Remove unnecessary model attribute assignment on 'freqs_cis' (#1766)

Summary:
Fixes #1767, more details there.

Upstream PR link meta-llama/llama#349

Pull Request resolved: #1766

Reviewed By: msaroufim

Differential Revision: D47556245

Pulled By: xuzhao9

fbshipit-source-id: 5ad541ff3e38a36b6515b1b7135b3da92109c9f5

Author: BowenBao

torchbenchmark/models/llama/model.py

@@ -224,8 +224,9 @@ def forward(self, tokens: torch.Tensor, start_pos: int):
 
         h = self.tok_embeddings(tokens)
 
-        self.freqs_cis = self.freqs_cis.to(h.device)
-        freqs_cis = self.freqs_cis[start_pos : start_pos + seqlen]
+        # Reference: https://github.com/facebookresearch/llama/pull/349
+        freqs_cis = self.freqs_cis.to(h.device)
+        freqs_cis = freqs_cis[start_pos : start_pos + seqlen]
 
         mask = None