Update acwm_runner.py

acwm_runner.py

@@ -60,10 +60,35 @@ def launch_acwm_training_task(
     optimizer_class = get_optimizer_class(customized_optimizer)
     optimizer = optimizer_class(
         model.trainable_modules(), lr=learning_rate, weight_decay=weight_decay)
-    scheduler = torch.optim.lr_scheduler.ConstantLR(optimizer)
     dataloader = torch.utils.data.DataLoader(
         dataset, shuffle=True, collate_fn=lambda x: x[0], num_workers=num_workers)
 
+    # ---- lr scheduler (constant / cosine / linear, with optional warmup) ----
+    lr_scheduler_type = getattr(args, "lr_scheduler", "constant") if args is not None else "constant"
+    warmup_steps = getattr(args, "warmup_steps", 0) if args is not None else 0
+    total_steps = len(dataloader) * num_epochs
+    if lr_scheduler_type in ("cosine", "linear"):
+        import math
+        from torch.optim.lr_scheduler import LambdaLR
+
+        def _lr_lambda(step):
+            if warmup_steps > 0 and step < warmup_steps:
+                return step / max(1, warmup_steps)
+            progress = (step - warmup_steps) / max(1, total_steps - warmup_steps)
+            progress = min(max(progress, 0.0), 1.0)
+            if lr_scheduler_type == "cosine":
+                return 0.5 * (1.0 + math.cos(math.pi * progress))
+            return 1.0 - progress   # linear
+
+        scheduler = LambdaLR(optimizer, _lr_lambda)
+        if accelerator.is_main_process:
+            print(f"[lr] {lr_scheduler_type} decay, base_lr={learning_rate}, "
+                  f"warmup={warmup_steps}, total_steps={total_steps}")
+    else:
+        scheduler = torch.optim.lr_scheduler.ConstantLR(optimizer, factor=1.0, total_iters=0)
+        if accelerator.is_main_process:
+            print(f"[lr] constant, lr={learning_rate}")
+
     model.to(device=accelerator.device)
     model, optimizer, dataloader, scheduler = accelerator.prepare(
         model, optimizer, dataloader, scheduler)