Add D-OPSD as a distillation handeling option for MiniMax H3 ref2va

jobs/process/BaseSDTrainProcess.py

@@ -2247,6 +2247,22 @@ class BaseSDTrainProcess(BaseTrainProcess):
 
         ### HOOk ###
         self.before_dataset_load()
+        if getattr(self.sd, 'require_pixel_tensor_cache', False):
+            # model needs pixel tensors at train time even with cached latents;
+            # storing them changes the latent cache key (first run re-caches)
+            for ds_list in [self.datasets, self.datasets_reg]:
+                if ds_list is None:
+                    continue
+                for ds in ds_list:
+                    if not (ds.cache_latents or ds.cache_latents_to_disk):
+                        # live-loading datasets already have pixels on the batch
+                        continue
+                    if not ds.cache_tensors_to_disk:
+                        print_acc(
+                            f"Model requires cached pixel tensors: forcing "
+                            f"cache_tensors_to_disk on dataset {ds.folder_path}"
+                        )
+                        ds.cache_tensors_to_disk = True
         # load datasets if passed in the root process
         if self.datasets is not None:
             self.data_loader = get_dataloader_from_datasets(self.datasets, self.train_config.batch_size, self.sd)