fastchat/model/monkey_patch_non_inplace.py [78:92]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
        raise ValueError(
            f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
            f" {attn_weights.size()}"
        )

    if attention_mask is not None:
        if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
            raise ValueError(
                f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
            )
        attn_weights = attn_weights + attention_mask
        attn_weights = torch.max(
            attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
        )
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



fastchat/train/llama_xformers_attn_monkey_patch.py [97:111]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights + attention_mask
            attn_weights = torch.max(
                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
            )
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -