Update app.py

app.py

@@ -6,6 +6,12 @@ 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")
@@ -78,19 +84,14 @@ from ltx_pipelines.utils.media_io import decode_audio_from_file, encode_video
 from ltx_core.loader.primitives import LoraPathStrengthAndSDOps
 from ltx_core.loader.sd_ops import LTXV_LORA_COMFY_RENAMING_MAP
 
-from ltx_core.model.transformer import attention as _attn_mod
+from ltx_pipelines.utils.types import OffloadMode
 
+# xformers attention patch (same as the outpaint app).
+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
-    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
+    _attn_mod.memory_efficient_attention = _mea
     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}")
@@ -249,18 +250,10 @@ class LTX23DistilledA2VPipeline:
                 video_state=video_state,
                 audio_state=audio_state,
                 stepper=stepper,
-                denoise_fn=multi_modal_guider_denoising_func(
-                    video_guider=MultiModalGuider(
-                        params=video_guider_params,
-                        negative_context=v_context_n,
-                    ),
-                    audio_guider=MultiModalGuider(
-                        params=audio_guider_params,
-                        negative_context=a_context_n,
-                    ),
-                    v_context=v_context_p,
-                    a_context=a_context_p,
-                    transformer=transformer,
+                denoise_fn=simple_denoising_func(
+                    video_context=v_context_p,
+                    audio_context=a_context_p,
+                    transformer=transformer,  # noqa: F821
                 ),
             )