Avoid xformers Dynamo import in LTX app

app.py

@@ -8,9 +8,6 @@ import sys
 import tempfile
 from pathlib import Path
 
-os.environ["TORCH_COMPILE_DISABLE"] = "1"
-os.environ["TORCHDYNAMO_DISABLE"] = "1"
-
 import gradio as gr
 import numpy as np
 import spaces
@@ -43,8 +40,6 @@ def _run(cmd, cwd=None, check=True):
 
 
 def install_ltx_runtime():
-    _run([sys.executable, "-m", "pip", "install", "xformers==0.0.32.post2", "--no-build-isolation"], check=False)
-
     if not (LTX_REPO_DIR / ".git").exists():
         LTX_REPO_DIR.parent.mkdir(parents=True, exist_ok=True)
         if LTX_REPO_DIR.exists():
@@ -86,24 +81,6 @@ from ltx_pipelines.utils.args import ImageConditioningInput
 from ltx_pipelines.utils.media_io import encode_video
 
 
-try:
-    from ltx_core.model.transformer import attention as _attn_mod
-    from xformers.ops import memory_efficient_attention as _memory_efficient_attention
-
-    _attn_mod.memory_efficient_attention = _memory_efficient_attention
-    print("[ATTN] xformers memory_efficient_attention enabled")
-except Exception as exc:
-    print(f"[ATTN] xformers patch failed: {type(exc).__name__}: {exc}")
-
-try:
-    from xformers.ops.fmha import _set_use_fa3
-
-    _set_use_fa3(False)
-    print("[ATTN] xformers FA3 dispatch disabled")
-except Exception as exc:
-    print(f"[ATTN] FA3 disable failed: {type(exc).__name__}: {exc}")
-
-
 _SAFETENSORS_DTYPE_MAP = {
     "F64": torch.float64,
     "F32": torch.float32,
@@ -282,7 +259,6 @@ def update_resolution(image, high_res):
 
 
 @spaces.GPU
-@torch.inference_mode()
 def generate_video(
     input_image,
     prompt,
@@ -296,45 +272,46 @@ def generate_video(
 ):
     current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
     try:
-        if torch.cuda.is_available():
-            torch.cuda.reset_peak_memory_stats()
-        log_memory("start")
-
-        frame_rate = DEFAULT_FRAME_RATE
-        num_frames = int(float(duration) * frame_rate) + 1
-        num_frames = ((num_frames - 1 + 7) // 8) * 8 + 1
-        height = int(height)
-        width = int(width)
-
-        print(f"[generate] {width}x{height}, frames={num_frames}, seed={current_seed}")
-
-        images = []
-        if input_image is not None:
-            OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
-            input_path = OUTPUT_DIR / f"input_{current_seed}.png"
-            input_image.save(input_path)
-            images = [ImageConditioningInput(path=str(input_path), frame_idx=0, strength=1.0)]
-
-        tiling_config = TilingConfig.default()
-        chunks = get_video_chunks_number(num_frames, tiling_config)
-        log_memory("before pipeline")
-
-        video, audio = pipeline(
-            prompt=prompt,
-            seed=current_seed,
-            height=height,
-            width=width,
-            num_frames=num_frames,
-            frame_rate=frame_rate,
-            images=images,
-            tiling_config=tiling_config,
-            enhance_prompt=bool(enhance_prompt),
-        )
+        with torch.no_grad():
+            if torch.cuda.is_available():
+                torch.cuda.reset_peak_memory_stats()
+            log_memory("start")
+
+            frame_rate = DEFAULT_FRAME_RATE
+            num_frames = int(float(duration) * frame_rate) + 1
+            num_frames = ((num_frames - 1 + 7) // 8) * 8 + 1
+            height = int(height)
+            width = int(width)
+
+            print(f"[generate] {width}x{height}, frames={num_frames}, seed={current_seed}")
+
+            images = []
+            if input_image is not None:
+                OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
+                input_path = OUTPUT_DIR / f"input_{current_seed}.png"
+                input_image.save(input_path)
+                images = [ImageConditioningInput(path=str(input_path), frame_idx=0, strength=1.0)]
+
+            tiling_config = TilingConfig.default()
+            chunks = get_video_chunks_number(num_frames, tiling_config)
+            log_memory("before pipeline")
+
+            video, audio = pipeline(
+                prompt=prompt,
+                seed=current_seed,
+                height=height,
+                width=width,
+                num_frames=num_frames,
+                frame_rate=frame_rate,
+                images=images,
+                tiling_config=tiling_config,
+                enhance_prompt=bool(enhance_prompt),
+            )
 
-        output_path = tempfile.mktemp(suffix=".mp4")
-        encode_video(video=video, fps=frame_rate, audio=audio, output_path=output_path, video_chunks_number=chunks)
-        log_memory("after encode")
-        return output_path, current_seed
+            output_path = tempfile.mktemp(suffix=".mp4")
+            encode_video(video=video, fps=frame_rate, audio=audio, output_path=output_path, video_chunks_number=chunks)
+            log_memory("after encode")
+            return output_path, current_seed
     except Exception as exc:
         import traceback