Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
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
2 changes: 1 addition & 1 deletion nemo_rl/algorithms/loss/loss_functions.py
Original file line number Diff line number Diff line change
Expand Up @@ -696,7 +696,7 @@ def __call__(
# See: docs/guides/grpo.md#sampling-importance-ratio
if self.sequence_level_importance_ratios:
sample_importance_ratio = masked_mean(
actor_importance_weights,
actor_importance_weights.squeeze(-1),
sample_mask,
global_normalization_factor=global_valid_seqs,
)
Expand Down
27 changes: 27 additions & 0 deletions tests/unit/algorithms/test_loss_functions.py
Original file line number Diff line number Diff line change
Expand Up @@ -2013,6 +2013,33 @@ def test_clipped_pg_loss_gspo_batch_size_2():
torch.testing.assert_close(actual_loss, expected_loss)


def test_clipped_pg_loss_sequence_importance_ratio_averages_per_sample():
"""Sequence-level importance-ratio metrics average one weight per sample."""
data, batch_size, seq_len, _ = _setup_clipped_pg_test_data(
batch_size=2, seq_len=3, device="cpu"
)
data["generation_logprobs"][:, 1] = -torch.log(torch.tensor([2.0, 4.0]))

loss_fn = ClippedPGLossFn(
ClippedPGLossConfig(
reference_policy_kl_penalty=0.0,
use_importance_sampling_correction=True,
sequence_level_importance_ratios=True,
token_level_loss=False,
)
)
_, metrics = loss_fn(
next_token_logprobs=torch.zeros((batch_size, seq_len - 1)),
data=data,
global_valid_seqs=data["sample_mask"].sum(),
global_valid_toks=(
data["token_mask"][:, 1:] * data["sample_mask"].unsqueeze(-1)
).sum(),
)

assert metrics["sampling_importance_ratio"] == pytest.approx(3.0)


def test_clipped_pg_loss_gspo_importance_sampling_correction():
"""Tests GSPO w/ importance sampling correction in ClippedPGLossFn."""
if not torch.cuda.is_available():
Expand Down
Loading