Before you start
Windows plus ROCm is the newest and least settled TRELLIS.2 route. RDNA 4 has been confirmed running in ComfyUI and testers are being recruited, but AMD still ships only PyTorch on Windows rather than the full ROCm stack, and the upstream custom nodes depend on CUDA-compiled wheels. This guide separates the part that works from the part that is still being built.
Prerequisites
- Windows 11 with a current Adrenalin driver
- An RDNA 4 card for the best-reported path, or RDNA 3 to help test
- Python 3.12 — AMD's Windows PyTorch package is limited to it
- Willingness to file and read issue reports rather than expect a finished product
Step by step
Understand what Windows actually ships
AMD's Radeon documentation states that only PyTorch is currently available on Windows and that the rest of the ROCm stack is supported on Linux. Windows also carries documented limits: no machine-learning training, Python 3.12 only, and older transformers releases may be incompatible. Plan around inference, not training.
Know which version pair you are testing
The community test effort currently covers PyTorch 2.9.1 with ROCm 7.2.1, and PyTorch 2.12 with ROCm 7.14. These are pairs, not a menu. Mixing a PyTorch build with a mismatched ROCm runtime is the most common way this route fails before it starts.
Expect the CUDA wheel wall
The upstream ComfyUI custom nodes ship prebuilt wheels compiled against CUDA. On a Radeon card, cumesh and o_voxel look for cudart64_12.dll, nvdiffrast pulls c10_cuda.dll, and flex_gemm needs Triton CUDA kernels. Installing with no-deps or adding NVIDIA runtime packages does not fix this; the extensions have to be rebuilt with HIP. The upstream issue tracking this is still open.
Choose between testing and shipping
If you want to contribute, follow the community test call and report results with your card, gfx target and exact version pair. If you need working output today, the native C++ route publishes prebuilt Windows binaries with ROCm and Vulkan backends and skips the Python dependency chain entirely.
Verify output, not just startup
The maintainer of the Windows effort warns that silent bugs may still affect final output. A run that completes is not evidence of a correct run. Compare geometry and material channels against a known-good reference before trusting the pipeline.
Known trade-offs
- This is a work-in-progress route; the maintainer explicitly warns about silent bugs that can affect final output
- Only PyTorch is currently available on Windows — the wider ROCm stack remains Linux only
- The stock ComfyUI-Trellis2 wheels are CUDA-only and the ROCm rebuild issue is unresolved upstream
- AMD's Windows guidance includes disabling certain Windows security features; weigh that against your environment before proceeding
- WSL is not a free workaround: AMD documents missing tooling and lower inference performance versus native Linux