Release: 0.18.1

setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = trl
-version = 0.18.0
+version = 0.18.1
 description = Train transformer language models with reinforcement learning.
 long_description = file: README.md
 long_description_content_type = text/markdown
@@ -89,7 +89,6 @@ dev =
     %(quantization)s
     %(scikit)s
     %(test)s
-    %(vllm)s
     %(vlm)s
 
 [options.entry_points]

trl/__init__.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "0.18.0"
+__version__ = "0.18.1"
 
 from typing import TYPE_CHECKING
 

trl/trainer/grpo_trainer.py

@@ -1229,7 +1229,7 @@ class GRPOTrainer(Trainer):
         # Identify sequences that terminated with EOS and log their lengths
         agg_terminated_with_eos = self.accelerator.gather(is_eos.any(dim=1))
         term_completion_lengths = agg_completion_lengths[agg_terminated_with_eos]
-        clipped_completions_ratio = 1 - len(term_completion_lengths) / len(completion_lengths)
+        clipped_completions_ratio = 1 - len(term_completion_lengths) / len(agg_completion_lengths)
         self._metrics[mode]["completions/clipped_ratio"].append(clipped_completions_ratio)
         if len(term_completion_lengths) == 0:  # edge case where no terminated sequences are found
             term_completion_lengths = torch.zeros(1, device=device)