fix: add find_unused_parameters=True to DDP for mixed text-image batches

When a batch contains only text samples (pixel_values is None),
vision_proj parameters do not participate in the forward pass,
causing DDP to raise:
  'Expected to have finished reduction in the prior iteration'

Adding find_unused_parameters=True resolves this.

trainer/train_sft_vlm.py

@@ -150,7 +150,7 @@ if __name__ == "__main__":
         Logger('torch.compile enabled')
     if dist.is_initialized():
         model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}
-        model = DistributedDataParallel(model, device_ids=[local_rank])
+        model = DistributedDataParallel(model, device_ids=[local_rank], find_unused_parameters=True)
     
     # ========== 8. 开始训练 ==========
     for epoch in range(start_epoch, args.epochs):