Fix NaN loss: cast LoRA params to fp32 for stable bf16 backward

scripts/training/train_flux_lora.py

@@ -186,6 +186,11 @@ def main():
                 set_peft_model_state_dict(transformer, state_dict)
                 print(f"  Loaded LoRA weights from checkpoint")
 
+    # Cast LoRA params to fp32 to prevent NaN in bf16 backward pass
+    for name, p in transformer.named_parameters():
+        if p.requires_grad:
+            p.data = p.data.float()
+
     transformer.to(train_device)
     transformer.print_trainable_parameters()
     transformer.train()
@@ -264,6 +269,7 @@ def main():
             noisy_packed = pack_latents(noisy_latents)
             target = pack_latents(noise - latents)
 
+            # Flux expects raw timestep 0-1
             timesteps = t
 
             b, seq_len, _ = noisy_packed.shape