fix: resume optimizer LR, step display, and consumed_samples alignment

2026-07-15 08:59:52 +08:00
parent 13998da15a
commit bb175fda91
3 changed files with 8 additions and 2 deletions
@@ -99,7 +99,12 @@ class TrainContextBuilder:
                    context.model_config = checkpoint.config
                context.epoch = checkpoint.epoch or cfg.start_epoch
                if checkpoint.consumed_samples > 0:
-                    context.consumed_samples = checkpoint.consumed_samples
+                    per_step = (
+                        cfg.batch_per_device * context.world_size * cfg.grad_accum_steps
+                    )
+                    context.consumed_samples = (
+                        checkpoint.consumed_samples // per_step
+                    ) * per_step
                else:
                    context.consumed_samples = cfg.start_samples * context.world_size
                context.checkpoint = checkpoint