ros/LlamaFactory
1
0
Fork 0
mirror of https://github.com/hiyouga/LlamaFactory.git synced 2026-02-02 20:43:38 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix generation bug #532

Browse Source 
Former-commit-id: c071121e67374e5f09798db57cfc8668617a36ae
This commit is contained in:
hiyouga
2023-08-17 22:21:34 +08:00
parent e993e717a5
commit fa1893b59c
5 changed files with 15 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/llmtuner/tuner/sft/trainer.py
View File

@@ -50,9 +50,10 @@ class Seq2SeqPeftTrainer(PeftTrainer):
loss, generated_tokens, labels = super().prediction_step(
model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys
)
generated_tokens = (
generated_tokens[:, max(prompt_len, label_len):] if generated_tokens is not None else None
)
if generated_tokens is not None:
generated_tokens[:, :max(prompt_len, label_len)] = (
self.tokenizer.pad_token_id * torch.ones_like(generated_tokens[:, :max(prompt_len, label_len)])
)
return (loss, generated_tokens, labels)
@@ -72,10 +73,7 @@ class Seq2SeqPeftTrainer(PeftTrainer):
assert self.tokenizer.padding_side == "left", "This method only accepts left-padded tensor."
pad_token_id = self.tokenizer.pad_token_id
else:
if self.model.config.pad_token_id is not None:
pad_token_id = self.model.config.pad_token_id
else:
raise ValueError("Pad_token_id must be set in the configuration of the model.")
raise ValueError("PAD token is required.")
padded_tensor = pad_token_id * torch.ones_like(tgt_tensor)
padded_tensor[:, -src_tensor.shape[-1]:] = src_tensor # adopt left-padding