Drop flash-attn install (was broken stub); attn_implementation=sdpa; GPU duration 119->60 (free tier max)

app.py

@@ -1,7 +1,5 @@
 import os
 import gc
-import subprocess
-subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
 # Workaround for gradio_client crashes on bool schemas (e.g. additionalProperties: True/False).
 # Must run BEFORE `import gradio as gr` so the patched functions are used.
@@ -156,7 +154,7 @@ def clear_old_cache():
         gc.collect()
         torch.cuda.empty_cache()
 
-@spaces.GPU(duration=119)
+@spaces.GPU(duration=60)
 def generate_text_gpu(model_id_str, message, history, system, temp, top_p, top_k, max_tokens, rep_penalty):
     """Text generation with branch support"""
     global models_cache, stop_event, current_thread
@@ -196,7 +194,7 @@ def generate_text_gpu(model_id_str, message, history, system, temp, top_p, top_k
                 "quantization_config": bnb_config,
                 "device_map": "auto",
                 "trust_remote_code": True,
-                "attn_implementation": "flash_attention_2" if torch.cuda.is_available() else None,
+                "attn_implementation": "sdpa" if torch.cuda.is_available() else None,
                 "low_cpu_mem_usage": True
             }
             if branch: