Update app.py: add xformers stub to avoid missing dependency

app.py

@@ -39,6 +39,7 @@ import numpy as np
 import tempfile
 
 import zipfile
+import types
 
 # ---------------------------------------------------------------------------
 # NOTE
@@ -74,6 +75,50 @@ def _ensure_hi3dgen_available():
 # Make sure the hi3dgen package is available before importing it
 _ensure_hi3dgen_available()
 
+# ---------------------------------------------------------------------------
+# xformers stub
+#
+# Some modules in the Hi3DGen pipeline import `xformers.ops.memory_efficient_attention`
+# to compute multi-head attention.  The official `xformers` library is not
+# installed in this Space (and requires GPU-only build), so we provide a
+# minimal in-memory stub that exposes a compatible API backed by PyTorch's
+# built-in scaled dot-product attention.  This stub is lightweight and
+# CPU-friendly.  It registers both the `xformers` and `xformers.ops` modules
+# in sys.modules so that subsequent imports succeed.
+# ---------------------------------------------------------------------------
+
+def _ensure_xformers_stub():
+    import sys
+    # If xformers is already available, do nothing.
+    if 'xformers.ops' in sys.modules:
+        return
+    import torch.nn.functional as F
+    # Create a new module object for xformers and its ops submodule
+    xformers_mod = types.ModuleType('xformers')
+    ops_mod = types.ModuleType('xformers.ops')
+
+    def memory_efficient_attention(query, key, value, attn_bias=None):
+        """
+        Fallback implementation of memory_efficient_attention for CPU environments.
+        This wraps torch.nn.functional.scaled_dot_product_attention.
+        """
+        # PyTorch expects the attention mask (bias) to be additive with shape
+        # broadcastable to (batch, num_heads, seq_len_query, seq_len_key).  If
+        # attn_bias is provided and is non-zero, pass it through; otherwise
+        # supply None to avoid unnecessary allocations.
+        return F.scaled_dot_product_attention(query, key, value, attn_bias)
+
+    # Populate the ops module with our fallback function
+    ops_mod.memory_efficient_attention = memory_efficient_attention
+    # Expose ops as an attribute of xformers
+    xformers_mod.ops = ops_mod
+    # Register modules
+    sys.modules['xformers'] = xformers_mod
+    sys.modules['xformers.ops'] = ops_mod
+
+# Ensure the xformers stub is registered before importing Hi3DGen
+_ensure_xformers_stub()
+
 from hi3dgen.pipelines import Hi3DGenPipeline
 import trimesh
 MAX_SEED = np.iinfo(np.int32).max