Upload folder using huggingface_hub

app.py

@@ -5,7 +5,8 @@ using Microsoft's Mage-Flow model (Turbo variants for fast inference).
 """
 import os
 
-os.environ.setdefault("VF_HF_ATTN_IMPL", "sdpa")
+# Use flash_attention_2 for the HF text encoder (flash_attn is installed via wheel)
+os.environ.setdefault("VF_HF_ATTN_IMPL", "flash_attention_2")
 
 import spaces  # MUST be first (after env setup)
 import torch

requirements.txt

@@ -11,4 +11,7 @@ pydantic>=2.0
 pillow>=10.0
 numpy>=1.26
 loguru>=0.7.0
-typing_extensions>=4.0
+typing_extensions>=4.0
+
+# Flash Attention 2 (prebuilt Blackwell sm_120 wheel)
+https://huggingface.co/datasets/multimodalart/zerogpu-blackwell-wheels/resolve/main/wheels/pt211-cu130-cp312/flash_attn-2.8.3-cp312-cp312-linux_x86_64.whl