Use try-except for flash_attn

This PR is to avoid transformers hard check failure on import dependencies.
If user doesn't install flash_attn, they will not pass the import check for flash_attn. Also same for non-cuda users.

To solve this, we can see that try-except content will be filtered out: https://github.com/huggingface/transformers/blob/main/src/transformers/dynamic_module_utils.py#L155.

modeling_deepseek.py

@@ -48,7 +48,6 @@ from transformers.pytorch_utils import (
 from transformers.utils import (
  add_start_docstrings,
  add_start_docstrings_to_model_forward,
- is_flash_attn_2_available,
  is_flash_attn_greater_or_equal_2_10,
  logging,
  replace_return_docstrings,
@@ -58,10 +57,11 @@ 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 ImportError:
+ pass
 
 # This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
 # It means that the function will not be traced through and simply appear as a node in the graph.