rixty
/
KoboldAI-Client
mirror of https://github.com/KoboldAI/KoboldAI-Client.git
1
1
Fork 0
Code Issues Projects Releases Wiki Activity

Put the XGLM embedding patch behind a version check

This commit is contained in:
Gnome Ann 2022-03-04 19:10:15 -05:00
parent 8e12b7df61
commit a00dede610
1 changed files with 21 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
aiserver.py
View File

@ -1015,27 +1015,28 @@ if(not vars.model in ["InferKit", "Colab", "OAI", "ReadOnly", "TPUMeshTransforme
import transformers.generation_utils import transformers.generation_utils
from transformers import __version__ as transformers_version from transformers import __version__ as transformers_version
# Temporary fix for XGLM positional embedding issues until # Some versions of transformers 4.17.0.dev0 are affected by
# https://github.com/huggingface/transformers/issues/15736 # https://github.com/huggingface/transformers/issues/15736
# is resolved # This is a workaround for those versions of transformers.
try: if(transformers_version == "4.17.0.dev0"):
from transformers.models.xglm.modeling_xglm import XGLMSinusoidalPositionalEmbedding try:
except ImportError: from transformers.models.xglm.modeling_xglm import XGLMSinusoidalPositionalEmbedding
pass except ImportError:
else: pass
@torch.no_grad() else:
def new_forward(self, input_ids: torch.Tensor = None, inputs_embeds: torch.Tensor = None, past_key_values_length: int = 0): @torch.no_grad()
bsz, seq_len = inputs_embeds.size()[:-1] def new_forward(self, input_ids: torch.Tensor = None, inputs_embeds: torch.Tensor = None, past_key_values_length: int = 0):
input_shape = inputs_embeds.size()[:-1] bsz, seq_len = inputs_embeds.size()[:-1]
sequence_length = input_shape[1] input_shape = inputs_embeds.size()[:-1]
position_ids = torch.arange( sequence_length = input_shape[1]
past_key_values_length + self.padding_idx + 1, past_key_values_length + sequence_length + self.padding_idx + 1, dtype=torch.long, device=inputs_embeds.device position_ids = torch.arange(
).unsqueeze(0).expand(input_shape).contiguous() past_key_values_length + self.padding_idx + 1, past_key_values_length + sequence_length + self.padding_idx + 1, dtype=torch.long, device=inputs_embeds.device
max_pos = self.padding_idx + 1 + seq_len + past_key_values_length ).unsqueeze(0).expand(input_shape).contiguous()
if max_pos > self.weights.size(0): max_pos = self.padding_idx + 1 + seq_len + past_key_values_length
self.make_weights(max_pos + self.offset, self.embedding_dim, self.padding_idx) if max_pos > self.weights.size(0):
return self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, -1).detach() self.make_weights(max_pos + self.offset, self.embedding_dim, self.padding_idx)
XGLMSinusoidalPositionalEmbedding.forward = new_forward return self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, -1).detach()
XGLMSinusoidalPositionalEmbedding.forward = new_forward
# Patch transformers to use our soft prompt # Patch transformers to use our soft prompt
def patch_causallm(cls): def patch_causallm(cls):