Skip to content

expand teacher prompt tensors by num_generations in _prepare_teacher_logprob_inputs #122

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
zhuxiaoxuhit wants to merge 1 commit into
base: main
Choose a base branch
from
Open

expand teacher prompt tensors by num_generations in _prepare_teacher_logprob_inputs #122

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 10 additions & 6 deletions west/trainer/kd_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -266,21 +266,25 @@ def _prepare_teacher_logprob_inputs(
)
prompt_length = prompt_processed["input_ids"].shape[1]

# Get prompt tensors
prompt_ids = prompt_processed["input_ids"]
prompt_mask = prompt_processed["attention_mask"]
# Get prompt tensors and expand for num_generations so batch dim matches
# generated_ids which has shape (B * num_generations, completion_len)
device = self.accelerator.device
prompt_ids = prompt_processed["input_ids"].repeat_interleave(self.num_generations, dim=0)
prompt_mask = prompt_processed["attention_mask"].repeat_interleave(self.num_generations, dim=0)
input_features = prompt_processed["input_features"].repeat_interleave(self.num_generations, dim=0)
feature_attention_mask = prompt_processed["feature_attention_mask"].repeat_interleave(
self.num_generations, dim=0)

# Concatenate: [prompt_ids, generated_ids]
# Note: generated_ids are from student, but we assume shared vocabulary
device = self.accelerator.device
full_input_ids = torch.cat([prompt_ids.to(device), generated_ids], dim=1)
full_attention_mask = torch.cat([prompt_mask.to(device), generated_mask], dim=1)

return {
"input_ids": full_input_ids,
"attention_mask": full_attention_mask,
"input_features": prompt_processed["input_features"].to(device),
"feature_attention_mask": prompt_processed["feature_attention_mask"].to(device),
"input_features": input_features.to(device),
"feature_attention_mask": feature_attention_mask.to(device),
}, prompt_length

def _compute_reverse_kl(
Expand Down
Loading