Before you start
AMD support for TRELLIS.2 moved during 2026 from isolated experiments to several parallel routes: a ROCm source fork, a patched ComfyUI stack, an in-progress Windows effort, and a native C++ implementation with ROCm and Vulkan backends. None of them are official support. This page keeps them separated and labels each card and runtime combination by how much evidence stands behind it.
Prerequisites
- Your exact card model and its VRAM size, not just the family name
- Your GPU's gfx compile target, since kernels are built per architecture
- A decision on Linux or Windows before choosing a runtime
- Acceptance that every AMD route here is community-maintained
Compatibility matrix
Columns are runtimes, rows are cards. A cell states how much evidence exists for that combination — not that the run will finish on your machine. Card names link to a page with the full reasoning.
| GPU | VRAM | ISA | ROCm forkLinux | ComfyUI LinuxLinux | ComfyUI WindowsWindows 11 | trellis.cpp ROCmLinux · Windows | trellis.cpp VulkanLinux · Windows |
|---|---|---|---|---|---|---|---|
| RDNA 4 gfx1200 · gfx1201 | |||||||
| RX 9070 XT | 16 GB | gfx1201 | ●Tested | ○Expected | ◐Reported | ○Expected | ○Expected |
| RX 9070 | 16 GB | gfx1201 | ○Expected | ○Expected | ○Expected | ○Expected | ○Expected |
| RX 9060 XT | 16 GB | gfx1200 | ○Expected | –Unknown | –Unknown | ○Expected | ○Expected |
| RDNA 3 gfx1100 · gfx1101 · gfx1102 | |||||||
| RX 7900 XTX | 24 GB | gfx1100 | ○Expected | ●Tested | –Unknown | ○Expected | ○Expected |
| RX 7900 XT | 20 GB | gfx1100 | ○Expected | ○Expected | –Unknown | ○Expected | ○Expected |
| RX 7800 XT | 16 GB | gfx1101 | ○Expected | ○Expected | –Unknown | ○Expected | ○Expected |
| RX 7700 XT | 12 GB | gfx1101 | –Unknown | ○Expected | –Unknown | ○Expected | ○Expected |
| RX 7600 XT | 16 GB | gfx1102 | –Unknown | ○Expected | –Unknown | ○Expected | ○Expected |
| RDNA 3.5 gfx1150 · gfx1151 | |||||||
| Ryzen AI Max+ 395 | 96 GB | gfx1151 | –Unknown | ◐Reported | ◐Reported | ○Expected | ●Tested |
| RDNA 2 gfx1030 · gfx1031 · gfx1032 | |||||||
| RX 6900 XT | 16 GB | gfx1030 | –Unknown | –Unknown | ×Blocked | –Unknown | ○Expected |
- ●Tested
A named project or maintainer reports this exact card completing the pipeline end to end.
- ◐Reported
Community reports exist but come from work in progress, partial runs, or a single tester.
- ○Expected
Inferred from a shared instruction set with a tested card. No direct report yet.
- ×Blocked
A known dependency, kernel, or runtime gap prevents this path today.
- –Unknown
No usable report either way. Treat as untested rather than as a failure.
Every AMD route on this page is community or third-party work. Neither Microsoft nor AMD lists TRELLIS.2 as an officially supported workload, and AMD's own documentation still limits Windows to PyTorch rather than the full ROCm stack.
TRELLIS.2 ROCm source fork
A fork of the upstream repository that ships HIP builds of FlexGEMM, CuMesh, nvdiffrast and nvdiffrec, plus a setup script that detects CUDA or ROCm and installs the matching dependencies. Validated on an RX 9070 XT 16 GB.
- OS
- Linux
- Stack
- ROCm 7.2 · PyTorch rocm7.2 · Python 3.10+
- Python runtime
- Required
ComfyUI + ROCm on Linux
The community ComfyUI wrapper plus a patch set that fixes the hardcoded GPU architecture flag, package paths and checkpoint downloads. Reported working end to end on a 7900 XTX for both shape-only and textured runs.
- OS
- Linux
- Stack
- ROCm 7.2.2 · PyTorch 2.11+rocm7.2 · Python 3.10–3.12
- Python runtime
- Required
ComfyUI + ROCm on Windows
Work in progress. RDNA 4 is confirmed running and testers are being recruited; RDNA 3 and RDNA 3.5 are next in line. The maintainer warns about silent bugs that can change the final output without raising an error.
- OS
- Windows 11
- Stack
- PyTorch 2.9.1 + ROCm 7.2.1, or PyTorch 2.12 + ROCm 7.14 · Python 3.12
- Python runtime
- Required
trellis.cpp — ROCm build
A native C++/GGML reimplementation of the whole TRELLIS.2 pipeline. Prebuilt ROCm archives are published for Linux and Windows, which removes the CUDA-only Python wheels from the problem entirely.
- OS
- Linux · Windows
- Stack
- Prebuilt ROCm/HIP binaries · no Python runtime
- Python runtime
- Not required
trellis.cpp — Vulkan build
The path with the fewest prerequisites: a Vulkan-capable driver replaces the entire ROCm stack. The project reports Vulkan as its fastest backend on some integrated GPUs.
- OS
- Linux · Windows
- Stack
- Vulkan driver only · no ROCm, no Python runtime
- Python runtime
- Not required
| Weight format | Weights on disk | Practical VRAM | Quality note | Where it comes from |
|---|---|---|---|---|
| f16 / bf16 (default) | ≈ 16.5 GB | 24 GB comfortable | Reference output. Everything else is measured against this. | trellis.cpp default; safetensors in the ComfyUI and source routes. |
| Q8 | ≈ 9.5 GB | 16 GB practical | Described by the project as visually near-lossless. | trellis.cpp v0.5.4 and later. |
| Q4 | ≈ 6 GB | 12 GB practical, 8 GB tight | Slight texture graininess, aimed explicitly at lower-VRAM systems. | trellis.cpp v0.5.4 and later. |
| GGUF K-quants (Q4_K_M – Q6_K) | ≈ 6 GB – 8 GB | 8 GB reported workable | Loader-dependent. Compatibility follows the custom node, not the file name. | ComfyUI GGUF loaders in the ROCm install guide. |
Step by step
Identify the compile target, not just the card
ROCm builds kernels per instruction set. Navi 31 is gfx1100, Navi 32 is gfx1101, Navi 33 is gfx1102, Navi 48 is gfx1201 and Navi 44 is gfx1200. Two cards in the same marketing tier can need different builds, and at least one community wrapper ships with a hardcoded architecture flag that has to be corrected by hand.
Pick the runtime before the model
Four routes exist. A ROCm source fork of the upstream repository targets Linux. A patched ComfyUI stack also targets Linux and has the most detailed public bug list. A Windows ComfyUI effort is under active testing. A native C++ reimplementation ships prebuilt ROCm and Vulkan binaries for both Linux and Windows with no Python runtime at all.
Match memory to a weight format
The native runtime documents roughly 16.5 GB for its f16 set, 9.5–10 GB for Q8, and 6 GB for Q4 before working memory. Smaller sets increase the chance that a workflow fits, but the result still depends on resolution, backend, texture settings, and live buffers; the smallest route also carries the largest stated quality risk.
Validate with the smallest run
Use a supplied sample image at the lowest practical resolution and confirm both the shape stage and the textured export before touching settings. On AMD this matters more than on NVIDIA, because several reported failures appear only in the texture stage.
Record the working combination
Write down card, gfx target, OS, ROCm version, PyTorch build, runtime commit and weight format together. AMD routes here break by version pairing far more often than by any single component being wrong.
Known trade-offs
- Neither AMD nor Microsoft lists TRELLIS.2 as a supported workload; every route on this page is third-party work
- AMD's own documentation states that only PyTorch is currently available on Windows and the rest of the ROCm stack remains Linux only
- The Windows ComfyUI effort is explicitly work in progress and its maintainer warns of silent bugs that can change output without raising an error
- A card marked expected rather than tested shares an instruction set with a tested card; that is an inference, not a report
- Weight file size is not peak VRAM — the texture stage and working buffers sit on top