Skip to content

early stop when all sequence reach EOS #57

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
wants to merge 4 commits into
base: main
Choose a base branch
from
Open

early stop when all sequence reach EOS #57

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
55a1c73
fix: early stop when all sequence reach EOS
je1lee Apr 9, 2024
488e5f2
style: tab in line
je1lee Apr 9, 2024
815a0c9
fix: (xla) early stop when all sequence reach EOS
je1lee Apr 9, 2024
e1d6092
Merge branch 'google:main' into fix/earlystop
je1lee Apr 16, 2024
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
12 changes: 12 additions & 0 deletions gemma/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -469,6 +469,7 @@ def generate(

batch_size = len(prompts)
prompt_tokens = [self.tokenizer.encode(prompt) for prompt in prompts]
prompt_length = [len(p) for p in prompt_tokens]
min_prompt_len = min(len(p) for p in prompt_tokens)
max_prompt_len = max(len(p) for p in prompt_tokens)
max_seq_len = max_prompt_len + output_len
Expand Down Expand Up @@ -511,6 +512,7 @@ def generate(
top_ks_tensor = torch.LongTensor([top_k] * batch_size).to(device)
output_index = torch.tensor(min_prompt_len, dtype=torch.int64).to(
device)
eos_flags_tensor = torch.tensor([False] * batch_size).to(device)

# Prefill up to min_prompt_len tokens, then treat other prefill as
# decode and ignore output.
Expand Down Expand Up @@ -543,6 +545,16 @@ def generate(
device)
output_index = output_index + 1

# Check if all sequences have reached EOS.
batch_eos_idx = (next_token_ids == self.tokenizer.eos_id).nonzero(
as_tuple=True)[0]
for eos_idx in batch_eos_idx:
if output_index >= prompt_length[eos_idx]:
eos_flags_tensor[eos_idx] = True

if eos_flags_tensor.all():
break

# Detokenization.
token_ids = token_ids_tensor.tolist()
results = []
Expand Down
14 changes: 13 additions & 1 deletion scripts/run_xla.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,6 +134,7 @@ def generate(
input_token_ids_tensor = torch.full((batch_size, min_prompt_len),
tokenizer.pad_id,
dtype=torch.int64)
prompt_length = [len(p) for p in prompt_tokens]
for i, p in enumerate(prompt_tokens):
token_ids_tensor[i, :len(p)] = torch.tensor(p)
input_token_ids_tensor[i, :min_prompt_len] = torch.tensor(
Expand All @@ -152,9 +153,10 @@ def generate(
top_ps_tensor = torch.FloatTensor(top_ps).to(device)
top_ks_tensor = torch.LongTensor(top_ks).to(device)
output_index = torch.tensor(min_prompt_len, dtype=torch.int64).to(device)
eos_flags_tensor = torch.tensor([False] * batch_size).to(device)

if not USE_CUDA:
xm.mark_step()

# Prefill up to min_prompt_len tokens, then treat other prefill as decode and ignore output.
for i in range(max_seq_len - min_prompt_len):
next_token_ids = model(
Expand Down Expand Up @@ -184,6 +186,16 @@ def generate(
if not USE_CUDA:
xm.mark_step()

# Check if all sequences have reached EOS.
batch_eos_idx = (next_token_ids == tokenizer.eos_id).nonzero(
as_tuple=True)[0]
for eos_idx in batch_eos_idx:
if output_index >= prompt_length[eos_idx]:
eos_flags_tensor[eos_idx] = True

if eos_flags_tensor.all():
break

# Detokenization.
token_ids = token_ids_tensor.tolist()
results = []
Expand Down