Remove RuntimeErrors for NaN-checking in 20B (huggingface#17563)

src/transformers/models/gpt_neox/modeling_gpt_neox.py

@@ -193,8 +193,6 @@ def _attn(self, query, key, value, attention_mask=None, head_mask=None):
         query = query.view(batch_size * num_attention_heads, query_length, attn_head_size)
         key = key.view(batch_size * num_attention_heads, key_length, attn_head_size)
         attn_scores = torch.einsum("bik,bjk->bij", query, key) / self.norm_factor
-        if torch.isnan(attn_scores).any():
-            raise RuntimeError()
         attn_scores = attn_scores.view(batch_size, num_attention_heads, query_length, key_length)
 
         attn_scores = torch.where(causal_mask, attn_scores, self.masked_bias.to(attn_scores.dtype))
@@ -204,17 +202,13 @@ def _attn(self, query, key, value, attention_mask=None, head_mask=None):
             attn_scores = attn_scores + attention_mask
 
         attn_weights = nn.functional.softmax(attn_scores, dim=-1)
-        if torch.isnan(attn_weights).any():
-            raise RuntimeError()
         attn_weights = attn_weights.to(value.dtype)
 
         # Mask heads if we want to
         if head_mask is not None:
             attn_weights = attn_weights * head_mask
 
         attn_output = torch.matmul(attn_weights, value)
-        if torch.isnan(attn_output).any():
-            raise RuntimeError()
         return attn_output, attn_weights