update

inspiremusic/transformer/qwen_encoder.py

@@ -163,7 +163,7 @@ class QwenInputOnlyEncoder(nn.Module):
         else:
             self.dtype = torch.float32
         from transformers import Qwen2ForCausalLM
-        model = Qwen2ForCausalLM.from_pretrained(pretrain_path, device_map="auto", attn_implementation="flash_attention_2", torch_dtype=self.dtype)
+        model = Qwen2ForCausalLM.from_pretrained(pretrain_path, device_map="cpu")
         self.embed = model.model.embed_tokens
         for p in self.embed.parameters():
             p.requires_grad = False