fix: is_offload_optimizer_states_in_train_step not passed to batch.meta_info in RLVR and agentic pipelines

[sanmuf]

Description

is_offload_optimizer_states_in_train_step controls whether optimizer states are offloaded to CPU after each train step (saving GPU memory at the cost of CPU-GPU transfer overhead). This field is defined in base_config.py with default False, and should be passed from each pipeline's config to batch.meta_info so that MegatronTrainStrategy.train_step can read it.

However, in the RLVR and agentic pipelines, this field was missing from batch.meta_info during training. As a result, MegatronTrainStrategy.train_step falls back to its hardcoded default True, rendering the user's yaml configuration ineffective.

Changes

• rlvr_pipeline.py: Added is_offload_optimizer_states_in_train_step to training batch's meta_info, reading from self.pipeline_config
• agentic_pipeline.py: Same fix

Impact

Users can now control is_offload_optimizer_states_in_train_step via yaml configuration in RLVR and agentic training workflows. Previously, the setting was silently ignored.

[CLAassistant]

All committers have signed the CLA.