fix: keep async rollouts version-consistent
- serialize shared-model optimizer updates with generation - reject future or over-lagged rollout results after asynchronous scoring - close cache publication races - persist policy versions in online checkpoints
This commit is contained in:
@@ -94,8 +94,7 @@ class Trainer:
|
||||
|
||||
if executor.sync_gradients:
|
||||
self._call_callbacks("before_optimizer_step", context)
|
||||
context.optimizer.step()
|
||||
context.strategy.on_optimizer_step()
|
||||
context.strategy.optimizer_step(context.optimizer)
|
||||
context.optimizer.zero_grad()
|
||||
|
||||
if context.scheduler:
|
||||
|
||||
Reference in New Issue
Block a user