fix: rename loop variable to avoid shadowing outer data in CriticWorker.train_step

[dubin555]

Problem

In CriticWorker.train_step (roll/pipeline/base_worker.py) and ActorWorker.train_step (roll/pipeline/diffusion/reward_fl/actor_worker.py), the for-loop variable data shadows the outer data parameter:

data = data.to(current_platform.device_type)  # outer data = full batch on GPU
dataloader = data.make_iterator(...)

for batch_idx, data in tqdm(enumerate(dataloader)):  # shadows outer data
    self.strategy.train_step(batch=data, ...)

data.to("cpu")  # BUG: only moves last mini-batch, full batch stays on GPU

After the loop, data points to the last yielded mini-batch instead of the original full batch. The subsequent data.to("cpu") only moves the last mini-batch to CPU, leaving the original full batch allocated on GPU — causing a memory leak that accumulates over training steps.

Fix

Rename the loop variable from data to mini_batch so the outer data reference is preserved:

for batch_idx, mini_batch in tqdm(enumerate(dataloader)):
    self.strategy.train_step(batch=mini_batch, ...)

data.to("cpu")  # now correctly moves the full original batch

Files Changed

• roll/pipeline/base_worker.py — CriticWorker.train_step
• roll/pipeline/diffusion/reward_fl/actor_worker.py — ActorWorker.train_step
• tests/test_train_step_variable_shadow.py — AST-based regression test

[CLAassistant]

All committers have signed the CLA.