riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-08-01 02:31:11 +06:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

make build_mpt_alibi_tensor a method of MptModel so that deepspeed co… (#25193)

Browse Source 
make build_mpt_alibi_tensor a method of MptModel so that deepspeed could override it to make autoTP work

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
This commit is contained in:
Wang, Yi 2023-08-01 13:35:49 +08:00 committed by GitHub
parent 0fd8d2aa2c
commit 4033ea7167
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/transformers/models/mpt/modeling_mpt.py
View File

@ -413,6 +413,9 @@ class MptModel(MptPreTrainedModel):
def get_input_embeddings(self):
return self.wte
def build_mpt_alibi_tensor(self, num_heads, sequence_length, alibi_bias_max=8, device=None):
return build_mpt_alibi_tensor(num_heads, sequence_length, alibi_bias_max, device)
def _prepare_attn_mask(
self, attention_mask: torch.Tensor, input_shape: Tuple[int, int], past_key_values_length: int
) -> torch.BoolTensor:
@ -507,7 +510,7 @@ class MptModel(MptPreTrainedModel):
else:
attention_mask = attention_mask.to(hidden_states.device)
alibi = build_mpt_alibi_tensor(self.num_heads, self.config.max_seq_len, device=hidden_states.device)
alibi = self.build_mpt_alibi_tensor(self.num_heads, self.config.max_seq_len, device=hidden_states.device)
causal_mask = self._prepare_attn_mask(
attention_mask,