Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Make the one scored completion GRPO test follow the TRL version (#14775)

Daniel Hancd7850ce23be
diff --git a/tests/test_grpo_sample_weight.py b/tests/test_grpo_sample_weight.py
index 4bed8c9eef0..c6b46159b72 100644
--- a/tests/test_grpo_sample_weight.py
+++ b/tests/test_grpo_sample_weight.py
@@ -280,14 +280,23 @@ def test_exclusion_follows_normalize_then_sum(monkeypatch):
assert _kept_groups(trainer, flat, monkeypatch) == [False]
-def test_group_with_one_scored_completion_is_kept(monkeypatch):
[email protected]("trl_version", [None, "1.5.1", "1.6.0"])
+def test_group_with_one_scored_completion_is_kept(monkeypatch, trl_version):
+ trl = pytest.importorskip("trl")
+ from packaging.version import Version
+
+ if trl_version is not None:
+ monkeypatch.setattr(trl, "__version__", trl_version)
+ masks_unscorable = Version(trl.__version__) >= Version("1.6.0")
rpf = torch.full((2 * G, 2), torch.nan)
rpf[0] = torch.tensor([1.0, 0.0])
rpf[G:] = torch.tensor([[0.0, 1.0], [1.0, 1.0], [2.0, 1.0], [0.0, 0.0]])
- for aggregation in ("sum_then_normalize", "normalize_then_sum"):
- trainer = _trainer("dapo", None)
- trainer.multi_objective_aggregation = aggregation
- assert _kept_groups(trainer, rpf, monkeypatch) == [True, True]
+ trainer = _trainer("dapo", None)
+ # TRL < 1.6 scores unscored rows 0, so normalize_then_sum zeroes the group (nan z-scores) and excludes it.
+ trainer.multi_objective_aggregation = "sum_then_normalize"
+ assert _kept_groups(trainer, rpf, monkeypatch) == [True, True]
+ trainer.multi_objective_aggregation = "normalize_then_sum"
+ assert _kept_groups(trainer, rpf, monkeypatch) == [masks_unscorable, True]
def test_unscored_completion_counts_as_zero_on_trl_before_1_6(monkeypatch):