guard against exploded loss (loss_cap/pred_cap) + adamw eps 1e-6

scripts/train_continue.py

@@ -69,6 +69,8 @@ def build_parser():
     ap.add_argument("--alpha", type=int, default=16)
     ap.add_argument("--seed", type=int, default=0)
     ap.add_argument("--max_grad_norm", type=float, default=1.0)
+    ap.add_argument("--loss_cap", type=float, default=100.0)
+    ap.add_argument("--pred_cap", type=float, default=1000.0)
     ap.add_argument("--keep_checkpoints", action="store_true")
     return ap
 
@@ -261,7 +263,7 @@ def main() -> None:
 )
     saved_cfg = build_adapter_config(lora_targets, args.rank, args.alpha, args.base_model)
 
-    optimizer = torch.optim.AdamW(trainable, lr=args.lr, weight_decay=1e-2)
+    optimizer = torch.optim.AdamW(trainable, lr=args.lr, weight_decay=1e-2, eps=1e-6)
     total_updates = args.max_train_steps
 
     def warmup_cosine(step: int) -> float:
@@ -359,11 +361,18 @@ def main() -> None:
                 ).sample
             pred = pred.float()
             loss = torch.nn.functional.mse_loss(pred, noise.float()) / args.grad_accum
-            if not torch.isfinite(loss):
+            loss_val = float(loss)
+            pred_max = float(pred.abs().max())
+            if (
+                not torch.isfinite(loss)
+                or not torch.isfinite(pred).all().item()
+                or loss_val > args.loss_cap
+                or pred_max > args.pred_cap
+            ):
                 skipped += 1
                 print(
-                    f"[warn] non-finite loss at step {global_step} micro {_micro} "
-                    f"row {pos - 1} - skipped",
+                    f"[warn] unstable loss at step {global_step} micro {_micro} "
+                    f"row {pos - 1} loss={loss_val:.3f} pred_max={pred_max:.3f} - skipped",
                     flush=True,
                 )
                 continue