Remove incompatible attention extensions

app.py

@@ -7,12 +7,16 @@ from diffusers import DiffusionPipeline
 import random
 import numpy as np
 import os
-import subprocess
+from unittest.mock import patch
+from transformers.dynamic_module_utils import get_imports
 from huggingface_hub import hf_hub_download
 from llm_inference import LLMInferenceNode
 
-# Install flash-attn
-subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+def fixed_get_imports(filename):
+    imports = get_imports(filename)
+    if str(filename).replace("\\", "/").endswith("/modeling_florence2.py"):
+        imports = [name for name in imports if name != "flash_attn"]
+    return imports
 
 # Initialize models
 device = "cuda" if torch.cuda.is_available() else "cpu"
@@ -23,9 +27,10 @@ huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
 # SD3.5 model
 pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-3.5-large", torch_dtype=dtype, use_safetensors=True, variant="fp16", token=huggingface_token).to(device)
 
-# Initialize Florence model
-florence_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()
-florence_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)
+# Initialize Florence model without requiring its optional FlashAttention path.
+with patch("transformers.dynamic_module_utils.get_imports", fixed_get_imports):
+    florence_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()
+    florence_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)
 
 # Prompt Enhancer
 enhancer_long = pipeline("summarization", model="gokaygokay/Lamini-Prompt-Enchance-Long", device=device)

requirements.txt

@@ -4,8 +4,6 @@ torch
 transformers==4.46.3
 sentencepiece
 spaces
-xformers
-sentencepiece
 timm
 einops
 openai