内存泄漏修复

model_training_nnn_tpu/dataset_tf.py

@@ -417,9 +417,11 @@ class BrainToTextDatasetTF:
         )
 
         # Apply TPU-optimized transformations
-        if self.split == 'train':
-            # For training, add shuffling
-            dataset = dataset.shuffle(buffer_size=min(1000, self.n_batches))
+        # 🚨 GPU版本策略：不需要在Dataset级别shuffle!
+        # GPU版本在 _create_batch_index_train() 中已经做了随机采样（第107-118行）
+        # 这里再shuffle会导致内存爆炸（1000 batch × 256 trials = 256,000 trials同时在内存）
+        # if self.split == 'train':
+        #     dataset = dataset.shuffle(buffer_size=min(1000, self.n_batches))  # ← 注释掉内存杀手
 
         # Prefetch for better performance
         dataset = dataset.prefetch(self.prefetch_buffer)