disable flash_attention

app_dialogue.py

@@ -2,11 +2,11 @@ import os
 import subprocess
 
 # Install flash attention
-subprocess.run(
-    "pip install flash-attn --no-build-isolation",
-    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
-    shell=True,
-)
+# subprocess.run(
+#     "pip install flash-attn --no-build-isolation",
+#     env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
+#     shell=True,
+# )
 
 
 import copy
@@ -31,7 +31,7 @@ MODELS = {
     "xgen-mm-phi3-mini-instruct-interleave-r-v1.5": AutoModelForVision2Seq.from_pretrained(
         "Salesforce/xgen-mm-phi3-mini-instruct-interleave-r-v1.5",
         torch_dtype=torch.bfloat16,
-        _attn_implementation="flash_attention_2",
+        # _attn_implementation="flash_attention_2",
         trust_remote_code=True
     ).to(DEVICE),
 }