rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-06-05 21:59:24 +02:00
Code Issues Projects Releases Wiki Activity

Clamp input_ids in place

Browse Source 
Apparently transformers maintains an internal reference to input_ids
(to use for repetition penalty) so we have to clamp the internal
version, too, because otherwise transformers will throw an out-of-bounds
error upon attempting to access token IDs that are not in the
vocabulary.
This commit is contained in:
Gnome Ann
2021-10-28 16:52:39 -04:00
parent 24d5d63c9f
commit 40b4631f6c
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
aiserver.py
View File

@@ -518,13 +518,15 @@ if(not vars.model in ["InferKit", "Colab", "OAI", "ReadOnly"]):
input_ids = kwargs.get('input_ids').to(self.device) input_ids = kwargs.get('input_ids').to(self.device)
assert input_ids is not None assert input_ids is not None
kwargs['input_ids'] = None kwargs['input_ids'] = None
inputs_embeds = self.transformer.wte(input_ids.clamp(max=self.config.vocab_size-1)) if(vars.sp is not None):
input_ids = input_ids - self.config.vocab_size # Don't use the -= operator here, you'll get a cryptic error message shifted_input_ids = input_ids - self.config.vocab_size
input_ids.clamp_(max=self.config.vocab_size-1)
inputs_embeds = self.transformer.wte(input_ids)
if(vars.sp is not None): if(vars.sp is not None):
vars.sp = vars.sp.to(inputs_embeds.dtype).to(inputs_embeds.device) vars.sp = vars.sp.to(inputs_embeds.dtype).to(inputs_embeds.device)
inputs_embeds = torch.where( inputs_embeds = torch.where(
(input_ids >= 0)[:, :, None], (shifted_input_ids >= 0)[:, :, None],
vars.sp[input_ids.clamp(min=0)], vars.sp[shifted_input_ids.clamp(min=0)],
inputs_embeds, inputs_embeds,
) )
kwargs['inputs_embeds'] = inputs_embeds kwargs['inputs_embeds'] = inputs_embeds