block flash_attn

.idea/.name

@@ -0,0 +1 @@
+inference.py

requirements.txt

@@ -1,7 +1,7 @@
 torch==2.1.2+cu121
 torchvision==0.17.0
 torchaudio==2.0.2
-transformers==4.34.1
+transformers==4.40.0
 decord==0.6.0
 peft==0.4.0
 huggingface_hub==0.23.4
@@ -10,4 +10,4 @@ einops
 scipy
 numpy==1.24.4
 tqdm
-flash-attn
+

robohusky/model/modeling_husky_embody2.py

@@ -39,15 +39,15 @@ from transformers.utils import (
     add_start_docstrings_to_model_forward,
     logging,
     replace_return_docstrings,
-    is_flash_attn_available
+#    is_flash_attn_available
 )
 from transformers import AutoModelForCausalLM, GenerationConfig
 
 from .configuration_husky import HuskyConfig, HuskyQFormerConfig, HuskyVisionConfig
 
-if is_flash_attn_available():
-    from flash_attn import flash_attn_func
-    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+#if is_flash_attn_available():
+#    from flash_attn import flash_attn_func
+#    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
 
 try:
     from apex.normalization import FusedLayerNorm as LayerNorm