Update modeling_deepseek.py

modeling_deepseek.py

@@ -58,9 +58,11 @@ from .configuration_deepseek import DeepseekV2Config
 import torch.distributed as dist
 import numpy as np
 
-if is_flash_attn_2_available():
+try:
     from flash_attn import flash_attn_func, flash_attn_varlen_func
     from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+except:
+    pass
 
 
 # This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.