Fix NaN: autocast forward, NaN skip, show cur_loss for debug

scripts/training/train_flux_lora.py

@@ -283,18 +283,28 @@ def main():
 
             txt_ids = torch.zeros(encoder_hidden_states.shape[1], 3, device=train_device, dtype=torch.bfloat16)
 
-            # Forward
-            model_pred = transformer(
-                hidden_states=noisy_packed,
-                timestep=timesteps,
-                encoder_hidden_states=encoder_hidden_states,
-                pooled_projections=pooled_prompt_embeds,
-                img_ids=img_ids,
-                txt_ids=txt_ids,
-                return_dict=False,
-            )[0]
-
+            # Forward with autocast for numerical stability
+            with torch.amp.autocast("cuda", dtype=torch.bfloat16):
+                model_pred = transformer(
+                    hidden_states=noisy_packed,
+                    timestep=timesteps,
+                    encoder_hidden_states=encoder_hidden_states,
+                    pooled_projections=pooled_prompt_embeds,
+                    img_ids=img_ids,
+                    txt_ids=txt_ids,
+                    return_dict=False,
+                )[0]
+
+            # Compute loss in fp32
             loss = F.mse_loss(model_pred.float(), target.float())
+
+            # Check for NaN and skip batch if needed
+            if torch.isnan(loss):
+                print(f"  WARNING: NaN loss at accum_count={accum_count}, skipping batch", flush=True)
+                optimizer.zero_grad()
+                accum_count += 1
+                continue
+
             scaled_loss = loss / args.gradient_accumulation
             scaled_loss.backward()
 
@@ -312,10 +322,12 @@ def main():
                     elapsed = time.time() - t0
                     steps_done = global_step - resume_step
                     steps_per_sec = steps_done / elapsed if elapsed > 0 else 0
-                    avg_loss = accum_loss / (50 * args.gradient_accumulation)
+                    avg_loss = accum_loss / (50 * args.gradient_accumulation) if accum_loss != 0 else float('nan')
+                    cur_loss = loss.item()
                     print(
                         f"  Step {global_step} | "
-                        f"Loss: {avg_loss:.4f} | "
+                        f"Loss(avg): {avg_loss:.4f} | "
+                        f"Loss(cur): {cur_loss:.4f} | "
                         f"LR: {lr_scheduler.get_last_lr()[0]:.2e} | "
                         f"Speed: {steps_per_sec:.2f} steps/s | "
                         f"Elapsed: {elapsed/3600:.1f}h",