TRELLIS.2 outputs point cloud / voxel samples instead of a surface mesh (Blackwell GPU, CUDA 12.8)

I am running TRELLIS.2 successfully on a recent NVIDIA Blackwell GPU (RTX 5060 Ti–class), with CUDA 12.8 and PyTorch cu128.
The full pipeline executes without runtime errors and produces a 3D output, but the result appears to be a point cloud / voxel surface samples, not a triangulated mesh.

This looks like the pipeline is stopping at an intermediate representation (e.g. shape samples / occupancy / splats) and not performing or exporting the final meshing step.

Environment

• GPU: NVIDIA Blackwell (RTX 5060 Ti–class)
• Driver: recent NVIDIA driver (supports CUDA 12.8)
• CUDA: 12.8
• OS: Linux (Docker-based setup)
• Python: 3.10
• PyTorch: 2.7.0 + cu128
• Triton: dev version (from triton-lang repo)
• flash-attn: enabled
• Sparse conv backend: spconv (forced to avoid FlexGEMM/Triton MLIR crash)
• spconv-cu124 2.3.8

Note: The pipeline runs end-to-end without crashing after switching from flex_gemm to spconv.

Logs:

/opt/venv/lib/python3.10/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
/opt/venv/lib/python3.10/site-packages/timm/models/registry.py:4: FutureWarning: Importing from timm.models.registry is deprecated, please import via timm.models
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.models", FutureWarning)
* Running on local URL:  http://0.0.0.0:7860
* To create a public link, set `share=True` in `launch()`.
/opt/venv/lib/python3.10/site-packages/utils3d/torch/nerf.py:41: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(enabled=False):
/opt/venv/lib/python3.10/site-packages/torch/nn/functional.py:5015: UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
  warnings.warn(
/opt/venv/lib/python3.10/site-packages/cumesh/remeshing.py:220: UserWarning: Using torch.cross without specifying the dim arg is deprecated.
Please either pass the dim explicitly or simply use torch.linalg.cross.
The default value of dim will change to agree with that of linalg.cross in a future release. (Triggered internally at /pytorch/aten/src/ATen/native/Cross.cpp:62.)
  normals0 = torch.cross(mesh_vertices[atempt_triangles_0[:, 1]] - mesh_vertices[atempt_triangles_0[:, 0]], mesh_vertices[atempt_triangles_0[:, 2]] - mesh_vertices[atempt_triangles_0[:, 0]])
/opt/venv/lib/python3.10/site-packages/utils3d/torch/nerf.py:41: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(enabled=False):
/opt/venv/lib/python3.10/site-packages/torch/nn/functional.py:5015: UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
  warnings.warn(
/opt/venv/lib/python3.10/site-packages/utils3d/torch/nerf.py:41: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(enabled=False):
/opt/venv/lib/python3.10/site-packages/torch/nn/functional.py:5015: UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
  warnings.warn(

No repository files or tests are named. Start by tracing the pipeline's final output and export path in the Docker-based setup, then reproduce the reported Blackwell, CUDA 12.8 environment to determine where point-cloud or voxel samples replace the expected output. Done means the pipeline produces and exports a triangulated surface mesh.