Update app.py

app.py

@@ -6,9 +6,6 @@ import sys
 os.environ["TORCH_COMPILE_DISABLE"] = "1"
 os.environ["TORCHDYNAMO_DISABLE"] = "1"
 
-# Install xformers for memory-efficient attention
-subprocess.run([sys.executable, "-m", "pip", "install", "xformers==0.0.32.post2", "--no-build-isolation"], check=False)
-
 # Clone LTX-2 repo and install packages
 LTX_REPO_URL = "https://github.com/Lightricks/LTX-2.git"
 LTX_REPO_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "LTX-2")
@@ -90,10 +87,18 @@ from ltx_pipelines.utils.media_io import decode_audio_from_file, encode_video
 
 # Force-patch xformers attention into the LTX attention module.
 from ltx_core.model.transformer import attention as _attn_mod
+
 print(f"[ATTN] Before patch: memory_efficient_attention={_attn_mod.memory_efficient_attention}")
 try:
     from xformers.ops import memory_efficient_attention as _mea
-    _attn_mod.memory_efficient_attention = _mea
+    from xformers.ops.fmha import cutlass
+
+    def _cutlass_memory_efficient_attention(*args, **kwargs):
+        # Force CUTLASS and avoid FlashAttention paths that are crashing.
+        kwargs["op"] = (cutlass.FwOp, cutlass.BwOp)
+        return _mea(*args, **kwargs)
+
+    _attn_mod.memory_efficient_attention = _cutlass_memory_efficient_attention
     print(f"[ATTN] After patch: memory_efficient_attention={_attn_mod.memory_efficient_attention}")
 except Exception as e:
     print(f"[ATTN] xformers patch FAILED: {type(e).__name__}: {e}")