Patch trellis2/representations/mesh/base.py: guard cumesh and flex_gemm imports

cumesh and flex_gemm C++ extensions are compiled against torch 2.6.0 and fail
to load under torch 2.8.0 (ABI symbol mismatch). Both imports now have
try/except guards:
- cumesh: fill_holes/remove_faces/simplify become no-ops when unavailable.
o_voxel handles its own decimation; mesh may have minor holes but exports fine.
- flex_gemm grid_sample_3d: query_attrs raises RuntimeError if called, but
it is not invoked in the generation → GLB export pipeline path.

TRELLIS.2/trellis2/representations/mesh/base.py

@@ -1,8 +1,24 @@
 from typing import *
 import torch
 from ..voxel import Voxel
-import cumesh
-from flex_gemm.ops.grid_sample import grid_sample_3d
+
+# cumesh and flex_gemm are compiled against torch 2.6.0; they fail to load
+# under torch 2.8.0+ due to C++ ABI changes.  Guard both and provide minimal
+# stubs so the rest of the module (and the generation pipeline) still works.
+try:
+    import cumesh as _cumesh
+    _HAS_CUMESH = True
+except Exception:
+    _cumesh = None
+    _HAS_CUMESH = False
+
+try:
+    from flex_gemm.ops.grid_sample import grid_sample_3d as _grid_sample_3d
+    _HAS_FLEX_GRID = True
+except Exception:
+    _HAS_FLEX_GRID = False
+    def _grid_sample_3d(*_a, **_kw):
+        raise RuntimeError("flex_gemm grid_sample_3d unavailable (ABI mismatch)")
 
 
 class Mesh:
@@ -33,10 +49,12 @@ class Mesh:
         return self.to('cpu')
     
     def fill_holes(self, max_hole_perimeter=3e-2):
+        if not _HAS_CUMESH:
+            return  # skip; mesh may have small holes but GLB export still works
         vertices = self.vertices.cuda()
         faces = self.faces.cuda()
-        
-        mesh = cumesh.CuMesh()
+
+        mesh = _cumesh.CuMesh()
         mesh.init(vertices, faces)
         mesh.get_edges()
         mesh.get_boundary_info()
@@ -52,31 +70,37 @@ class Mesh:
             return
         mesh.fill_holes(max_hole_perimeter=max_hole_perimeter)
         new_vertices, new_faces = mesh.read()
-        
+
         self.vertices = new_vertices.to(self.device)
         self.faces = new_faces.to(self.device)
         
     def remove_faces(self, face_mask: torch.Tensor):
+        if not _HAS_CUMESH:
+            keep = ~face_mask.bool().to(self.faces.device)
+            self.faces = self.faces[keep]
+            return
         vertices = self.vertices.cuda()
         faces = self.faces.cuda()
-        
-        mesh = cumesh.CuMesh()
+
+        mesh = _cumesh.CuMesh()
         mesh.init(vertices, faces)
         mesh.remove_faces(face_mask)
         new_vertices, new_faces = mesh.read()
-        
+
         self.vertices = new_vertices.to(self.device)
         self.faces = new_faces.to(self.device)
         
     def simplify(self, target=1000000, verbose: bool=False, options: dict={}):
+        if not _HAS_CUMESH:
+            return  # skip; o_voxel.postprocess handles its own decimation
         vertices = self.vertices.cuda()
         faces = self.faces.cuda()
-        
-        mesh = cumesh.CuMesh()
+
+        mesh = _cumesh.CuMesh()
         mesh.init(vertices, faces)
         mesh.simplify(target, verbose=verbose, options=options)
         new_vertices, new_faces = mesh.read()
-        
+
         self.vertices = new_vertices.to(self.device)
         self.faces = new_faces.to(self.device)
 
@@ -221,7 +245,7 @@ class MeshWithVoxel(Mesh, Voxel):
         
     def query_attrs(self, xyz):
         grid = ((xyz - self.origin) / self.voxel_size).reshape(1, -1, 3)
-        vertex_attrs = grid_sample_3d(
+        vertex_attrs = _grid_sample_3d(
             self.attrs,
             torch.cat([torch.zeros_like(self.coords[..., :1]), self.coords], dim=-1),
             self.voxel_shape,