[ckpt] fix: edit esi doc (verl-project#2354)

plutoZZZZ · web-flow · commit a53fb3089ea3 · 2025-07-04T13:34:12.000+08:00
This PR addresses the "ESI" comprehension issue left by the previous PR (verl-project#2192). This PR refines `ppo_trainer.yaml` by expanding the esi_redundant_time comment to define ESI (Elastic Server Instance) and draw a parallel to a training plan. In `ray_trainer.py`, it clarifies ESI-related checkpoint-saving conditions. These edits boost code readability and maintainability.
diff --git a/verl/trainer/config/ppo_trainer.yaml b/verl/trainer/config/ppo_trainer.yaml
@@ -998,7 +998,11 @@ trainer:
   # Save frequency (by iteration) for model checkpoints
   save_freq: -1
 
-  # ESI redundant time (in seconds) for model checkpointsAdd commentMore actions
+  # ESI refers to the elastic server instance used during training, similar to the training plan. For example,
+  # if you purchase 10 hours of computing power, the ESI will automatically shut down after 10 hours of training.
+  # To ensure a checkpoint is saved before ESI shuts down, the system will start saving a checkpoint in advance.
+  # The advance time is calculated as: Advance Time = Longest historical step duration + Checkpoint save duration + esi_redundant_time.
+  # Here, esi_redundant_time is a user-defined value that further extends the advance time for added safety.
   esi_redundant_time: 0
 
   # Resume mode: "auto", "disable", or "resume_path"
diff --git a/verl/trainer/ppo/ray_trainer.py b/verl/trainer/ppo/ray_trainer.py
@@ -1318,10 +1318,17 @@ def fit(self):
                                 last_val_metrics = val_metrics
                         metrics.update(val_metrics)
 
+                    # Check if the ESI (Elastic Server Instance)/training plan is close to expiration.
                     esi_close_to_expiration = should_save_ckpt_esi(
                         max_steps_duration=self.max_steps_duration,
                         redundant_time=self.config.trainer.esi_redundant_time,
                     )
+                    # Check if the conditions for saving a checkpoint are met.
+                    # The conditions include a mandatory condition (1) and one of the following optional conditions (2, 3, or 4):
+                    # 1. The save frequency is set to a positive value.
+                    # 2. It's the last training step.
+                    # 3. The current step number is a multiple of the save frequency.
+                    # 4. The ESI(Elastic Server Instance)/training plan is close to expiration.
                     if self.config.trainer.save_freq > 0 and (
                         is_last_step
                         or self.global_steps % self.config.trainer.save_freq == 0
