Skip to content

fix: adjust temperature scaling logic based on engine version #660

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

Merged
terrykong merged 2 commits into from
Jul 16, 2025
Merged

fix: adjust temperature scaling logic based on engine version #660

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c4a0414
fix: adjust temperature scaling logic based on engine version
jubick1337 Jul 12, 2025
08c4d24
Merge branch 'main' into mnovikov/vllm_temp_fix
jubick1337 Jul 16, 2025
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
7 changes: 6 additions & 1 deletion nemo_rl/models/policy/dtensor_policy_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -578,7 +578,12 @@ def train(
"generation" in self.cfg
and self.cfg["generation"] is not None
):
logits.div_(self.cfg["generation"]["temperature"])
# The V1 engine returns raw logits before temperature scaling.
# The V0 engine (when VLLM_USE_V1 is not '1') returns scaled logits.
# Therefore, we only divide if we are NOT using the V1 engine.
use_v1_engine = os.environ.get("VLLM_USE_V1") == "1"
if not use_v1_engine:
logits.div_(self.cfg["generation"]["temperature"])

if self.cp_size > 1:
seq_index_dtensor = (
Expand Down
Loading