Fix typo in modeling_grok

#3
Files changed (1) hide show
  1. modeling_grok.py +1 -1
modeling_grok.py CHANGED
@@ -273,7 +273,7 @@ class GrokAttention(nn.Module):
273
  value_states = repeat_kv(value_states, self.num_key_value_groups)
274
 
275
  attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.attn_output_multiplier
276
- attn_logits = 30 * torch.tanh(attn_weights / 30)
277
 
278
  if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
279
  raise ValueError(
 
273
  value_states = repeat_kv(value_states, self.num_key_value_groups)
274
 
275
  attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.attn_output_multiplier
276
+ attn_weights = 30 * torch.tanh(attn_weights / 30)
277
 
278
  if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
279
  raise ValueError(