Skip to content

[FEAT][executors]: implement stateless zero-KV scoring executor #94

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
Flink-ddd merged 3 commits into from
Jun 22, 2026
Merged

[FEAT][executors]: implement stateless zero-KV scoring executor #94

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
232ebef
feat(executors): add stateless reference reward scoring
inaniloquentee Jun 10, 2026
22378a8
Address stateless executor review feedback
inaniloquentee Jun 15, 2026
d911d55
Format stateless executor test
inaniloquentee Jun 15, 2026
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
481 changes: 481 additions & 0 deletions benchmarks/benchmark_stateless_executor.py
View file
Open in desktop

Large diffs are not rendered by default.

14 changes: 13 additions & 1 deletion rl_engine/executors/deepspeed_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,7 @@
RolloutStageResult,
TorchRLTrainingConfig,
TrainingStageResult,
objective_reference_logps,
)
from rl_engine.testing import (
compute_policy_ratio,
Expand Down Expand Up @@ -113,7 +114,7 @@ def train(self, rollout: RolloutStageResult) -> TrainingStageResult:
output_dtype=torch.float32,
)
old_logps = current_logps.detach() - 0.01
ref_logps = current_logps.detach() - 0.02
ref_logps = objective_reference_logps(current_logps, batch)
ratio = compute_policy_ratio(current_logps, old_logps, batch.completion_mask)
unclipped = ratio * batch.advantages.float()
clipped = torch.clamp(ratio, 0.8, 1.2) * batch.advantages.float()
Expand All @@ -133,6 +134,7 @@ def train(self, rollout: RolloutStageResult) -> TrainingStageResult:

finished_at = time.perf_counter()
published = self._next_published_weight_version(rollout.weight_version)
active_advantages = batch.advantages.float()[batch.completion_mask]
return TrainingStageResult(
iteration=rollout.iteration,
consumed_weight_version=rollout.weight_version,
Expand All @@ -145,6 +147,16 @@ def train(self, rollout: RolloutStageResult) -> TrainingStageResult:
"training_device": str(self.device),
"deepspeed_engine": type(self.engine).__name__,
"deepspeed_zero_stage": self.config.zero_stage,
"active_advantage_mean_global": (
float(active_advantages.mean().detach().cpu().item())
if active_advantages.numel()
else 0.0
),
"active_advantage_std_global": (
float(active_advantages.std(unbiased=False).detach().cpu().item())
if active_advantages.numel()
else 0.0
),
**payload_metrics,
},
started_at=started_at,
Expand Down
Loading
Loading