Major WanGP Launcher Update -- Modernization and Wider and Detailed System Support
The WanGP launcher has made a major update today.
The old version of this launcher had some stale parts, which didn't stay up to date with all the new features of WanGP. This is why many people recommended using @morpheus fork, until today.
All of Morpheus's improvements from Morpheus's fork have been merged into this featured launcher, and now every NVIDIA GPU user can use this launcher going forward.
This launcher is now fully up to date with all the new WanGP requirements.
Before
The old version of this Pinokio WanGP launcher was effectively NVIDIA-only on Windows and Linux, with one generic dependency stack per operating system. AMD GPUs and macOS were explicitly blocked.
After
The new version replaces that approach with hardware-specific dependency stacks for:
- more NVIDIA generations
- mac support (Apple Silicon path)
And migrates the environment to Python 3.11, and moves modern NVIDIA cards to PyTorch 2.10 with CUDA 13 (Previously used CUDA 12.8).
Soon even the AMD fork will be merged into this launcher so this one launcher can handle ALL GPUs, but for now, if you're an AMD GPU user, use Morpheus's fork here: https://pinokio.co/apps/github-com-6morpheus6-wan2gp-amd
System Support Comparison
Here's what the new launcher supports:
| System | Old launcher | Updated launcher | What changed |
|---|---|---|---|
| Windows — NVIDIA GTX 10 / GTX 16 | Generic Windows NVIDIA stack | Dedicated legacy compatibility stack | Stays on PyTorch 2.7.1 with CUDA 12.8. Exempt from the new NVIDIA R580 driver requirement. The new launcher does not explicitly install the newer SageAttention/FlashAttention bundle for these cards. |
| Windows — NVIDIA RTX 20 | Generic PyTorch 2.7.1/CUDA 12.8 stack | Dedicated modern stack | PyTorch 2.10 with CUDA 13, Triton, SageAttention 1.0.6, SpargeAttention, llama.cpp GGUF CUDA kernels, and bitsandbytes. Requires NVIDIA driver R580 or newer. |
| Windows — NVIDIA RTX 30 / RTX 40 | Generic PyTorch 2.7.1/CUDA 12.8 stack | Dedicated optimized stack | PyTorch 2.10 with CUDA 13, Triton, SageAttention 2.2, FlashAttention 2.8.3, SpargeAttention, GGUF CUDA kernels, and bitsandbytes. Requires R580 or newer. |
| Windows — NVIDIA RTX 50 | The same generic stack used by every NVIDIA GPU | Dedicated full acceleration stack | Adds explicit RTX 50 support with PyTorch 2.10/CUDA 13, SageAttention, FlashAttention, SpargeAttention, LightX2V kernels, Nunchaku, GGUF CUDA kernels, and bitsandbytes. Requires R580 or newer. |
| Windows — AMD | Explicitly blocked | New experimental support | Uses custom ROCm/TheRock gfx110x PyTorch wheels, plus bitsandbytes and a NumPy 1.26.4 compatibility pin. This should be described as support for compatible gfx110x hardware, not every AMD GPU. |
| Linux — NVIDIA GTX 10 / GTX 16 | Generic Linux NVIDIA stack | Dedicated legacy compatibility stack | PyTorch 2.7 with CUDA 12.8. Exempt from R580. Unlike the old generic Linux stack, the new branch does not explicitly install SageAttention and FlashAttention wheels. |
| Linux — NVIDIA RTX 20 | Generic PyTorch 2.7/CUDA 12.8 stack | Dedicated modern stack | PyTorch 2.10/CUDA 13 with SageAttention 2.2, SpargeAttention, bitsandbytes, and NumPy 2.1.2. Requires R580 or newer. |
| Linux — NVIDIA RTX 30 / RTX 40 | Generic PyTorch 2.7/CUDA 12.8 stack | Dedicated branch intended, but routing needs correction | The intended stack includes PyTorch 2.10/CUDA 13, SageAttention, FlashAttention, SpargeAttention, and bitsandbytes. As currently written, these cards are captured by the preceding generic Linux condition. |
| Linux — NVIDIA RTX 50 | Generic CUDA 12.8 stack | Dedicated full acceleration stack | PyTorch 2.10/CUDA 13 with SageAttention, FlashAttention, SpargeAttention, LightX2V, Nunchaku, and bitsandbytes. Requires R580 or newer. |
| Linux — AMD | Explicitly blocked | ROCm 7.1 branch added, but currently shadowed | The intended branch installs PyTorch 2.10 from the ROCm 7.1 index. Earlier generic Linux conditions currently capture AMD systems before this branch is reached. Do not announce Linux AMD as working until corrected and tested. |
| macOS — Apple Silicon | Explicitly blocked | New basic Apple Silicon path | The launcher now has an arm64 macOS branch that installs standard PyTorch, torchvision, and torchaudio wheels. It does not include the CUDA acceleration packages available on NVIDIA. |
| macOS — Intel | Blocked | Still unsupported | There is no darwin/x64 installation branch. |
| CPU-only Windows/Linux | No proper CPU-specific installation path | Still no dedicated CPU-only path | The new matrix remains GPU-oriented. Apple Silicon is the only non-CUDA path outside AMD ROCm. |
Biggest Practical Differences
1. AMD and macOS are no longer universally rejected
The old installer stopped immediately whenever it detected an AMD GPU or any macOS machine.
The new installer removes that blanket rejection:
- Windows AMD receives a specific TheRock/ROCm dependency stack.
- Linux AMD has a proposed ROCm 7.1 dependency stack.
- Apple Silicon receives its own PyTorch installation path.
- Intel Macs remain unsupported.
Important qualifications:
- Windows AMD: has a selectable branch, but the wheels specifically target
gfx110x, so this is not universal AMD support. - Linux AMD: the required branch exists, but the current conditional ordering needs to be corrected.
- Apple Silicon: is now installable, but does not receive the CUDA acceleration libraries available on NVIDIA.
- Intel Mac: remains unsupported.
2. Modern RTX cards now use generation-aware stacks
Previously, every Windows NVIDIA card received the same PyTorch 2.7.1/CUDA 12.8 environment, while every Linux NVIDIA card received the same PyTorch 2.7/CUDA 12.8 environment.
The updated launcher differentiates between:
- GTX 10 and GTX 16
- RTX 20
- RTX 30 and RTX 40
- RTX 50
This matters because different GPU generations support different acceleration libraries.
In practical terms:
- RTX 30/40: receive newer PyTorch, CUDA, FlashAttention, SageAttention, and SpargeAttention.
- RTX 50: receives the broadest stack, including LightX2V and Nunchaku.
- RTX 20: receives a more conservative modern stack suited to that generation.
- GTX 10/16: remains supported through a legacy CUDA 12.8 path rather than being forced onto CUDA 13.
3. Modern NVIDIA cards now require an R580-or-newer driver
RTX 20, RTX 30, RTX 40, and RTX 50 systems require an NVIDIA R580-or-newer driver because their updated stacks use CUDA 13.
GTX 10 and GTX 16 cards are explicitly exempt because they remain on CUDA 12.8.
The installer now detects an older NVIDIA driver and stops with a clear upgrade message instead of continuing with an incompatible dependency stack.
4. This is a full environment migration, not a normal update
The launcher now:
- Deletes the old
app/envenvironment. - Creates a new
app/venvenvironment. - Explicitly uses Python 3.11.
- Installs the dependency stack selected for the detected operating system and GPU.
- Reinstalls WanGP's current requirements.
That is why this update may take significantly longer than a normal WanGP update. It is rebuilding the runtime, not merely pulling new application code.
Important Linux Routing Issue
The current Linux branch ordering is effectively:
Linux GTX 10/16: specific condition
Linux RTX 20: specific condition
Linux RTX 50: platform === "linux"
Linux RTX 30/40: platform === "linux"
Linux AMD: platform === "linux" && gpu === "amd"
Because the supposed RTX 50 condition is only platform === "linux", it also matches RTX 30/40 and AMD Linux systems. The later RTX 30/40 and AMD branches therefore cannot be selected as intended.
The conditions should be closer to:
// Linux RTX 50
platform === "linux" &&
gpu === "nvidia" &&
kernel.gpu_model &&
/ 50.+/.test(kernel.gpu_model)
// Linux RTX 30 / RTX 40 fallback
platform === "linux" &&
gpu === "nvidia"
The later condition:
platform === "linux" && gpu === "amd"
can then correctly select the Linux AMD ROCm branch.
Summary
The Pinokio WanGP launcher has been completely rebuilt with GPU-specific dependency stacks for NVIDIA GTX 10/16 through RTX 50 on Windows and Linux, plus new experimental Windows AMD and Apple Silicon support. Modern RTX cards now receive the latest PyTorch 2.10/CUDA 13 acceleration stack instead of the old one-size-fits-all environment.