Skip to content

fix: rename samples to seq_lens and remove dead code #1453

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
JackXu0 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix: rename samples to seq_lens and remove dead code #1453

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: 7 additions & 9 deletions slime/backends/megatron_utils/data.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -308,14 +308,14 @@ def _generate_data_iterator(rollout_data, micro_batch_size, micro_batch_indices=
data_iterator = _generate_data_iterator(rollout_data, args.micro_batch_size)
else:
assert args.max_tokens_per_gpu is not None
# calculate the number of mirobatches for each step
samples = rollout_data["total_lengths"]
assert len(samples) == num_local_samples
# calculate the number of microbatches for each step
seq_lens = rollout_data["total_lengths"]
assert len(seq_lens) == num_local_samples
num_microbatches = []
for i in range(num_steps_per_rollout):
start, end = i * num_local_gbs, (i + 1) * num_local_gbs
num_microbatches.append(
get_minimum_num_micro_batch_size(samples[start:end], args.max_tokens_per_gpu * cp_size)
get_minimum_num_micro_batch_size(seq_lens[start:end], args.max_tokens_per_gpu * cp_size)
)

num_microbatches = torch.tensor(num_microbatches, dtype=torch.int, device=torch.cuda.current_device())
Expand All @@ -330,14 +330,12 @@ def _generate_data_iterator(rollout_data, micro_batch_size, micro_batch_indices=

num_microbatches = num_microbatches.tolist()

# balance the each micro batch
samples = rollout_data["total_lengths"]
# balance the number of mirobatches across steps
# balance the number of microbatches across steps
micro_batch_indices = []
for i, num_mbs in enumerate(num_microbatches):
start, end = i * num_local_gbs, (i + 1) * num_local_gbs
samples = rollout_data["total_lengths"][start:end]
partitions = get_seqlen_balanced_partitions(samples, num_mbs, equal_size=False)
seq_lens = rollout_data["total_lengths"][start:end]
partitions = get_seqlen_balanced_partitions(seq_lens, num_mbs, equal_size=False)
for j in range(num_mbs):
for k in range(len(partitions[j])):
partitions[j][k] += start
Expand Down