remove flash_attn imports and usage (#1)

- remove flash_attn imports and usage (edc7c55e9ce3fe0ccd6c1a2af7a500f898e7faf4)
- Update modeling_minicpm.py (9fc8a74191cf8fd01d98d9cd988f0cdfc994ee18)

modeling_minicpm.py

@@ -51,11 +51,11 @@ from transformers.utils.import_utils import is_torch_fx_available
 from .configuration_minicpm import MiniCPMConfig
 import re
 
-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
+#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.
@@ -755,7 +755,7 @@ class MiniCPMSdpaAttention(MiniCPMAttention):
 
 MINICPM_ATTENTION_CLASSES = {
     "eager": MiniCPMAttention,
-    "flash_attention_2": MiniCPMFlashAttention2,
+    #"flash_attention_2": MiniCPMFlashAttention2,
     "sdpa": MiniCPMSdpaAttention,
 }