Upload BD3LM

modeling_bd3lm.py

@@ -603,8 +603,8 @@ class BD3LM(transformers.PreTrainedModel):
     for block in self.backbone.blocks:
       block.kv_cache = torch.zeros(
         eval_batch_size,
-        self.n,
-        self.config.model.hidden_size * 3,
+        self.config.model_length,
+        self.config.hidden_dim * 3,
         device='cuda',
         dtype=torch.bfloat16)
       block.cache_idx = 0