From fc8a1202857c32f5672b5944d4227b4ed2705037 Mon Sep 17 00:00:00 2001 From: Shashank Rajput Date: Tue, 3 Dec 2024 22:25:26 -0800 Subject: [PATCH] .. --- llmfoundry/models/mpt/modeling_mpt.py | 1 - 1 file changed, 1 deletion(-) diff --git a/llmfoundry/models/mpt/modeling_mpt.py b/llmfoundry/models/mpt/modeling_mpt.py index 824a7fc929..969a8b56a2 100644 --- a/llmfoundry/models/mpt/modeling_mpt.py +++ b/llmfoundry/models/mpt/modeling_mpt.py @@ -196,7 +196,6 @@ def gen_sequence_id_info( attn_uses_sequence_id (bool): Whether the attention uses sequence id based masking. attn_impl (str): Attention implementation. This function is only creates attention_mask_in_length for flash attention. attention_mask (Union[torch.Tensor, None]): Attention mask tensor of shape (batch_size, seq_len) - return_pos_in_seq (bool): Whether to return the position in sequence tensor instead of attention mask in length. Returns: attention_mask_in_length: (batch, seqlen), int, a nonzero number (e.g., 1, 2, 3, etc.) means length of concatenated sequence in b-th batch, and 0 means none. For example, if batch = 3 and seqlen = 6, the attention_mask_in_length is: