ComfyUI Node
SparkDiffusion Runtime
Configure the isolated SparkDiffusion runtime (backend, interpreter, model root, FP8, kernels).
SparkDiffusion Runtime
- runtime
◄backendauto►
◄quantizationfp8►
◄fused_kernelstrue►
◄torch_compilefalse►
◄sparkdiffusion_repo►
◄python►
◄model_root►
◄wsl_distro►
◄fp8_compatauto►
◄sla_src►
◄timeout_seconds3600►
CategorySparkDiffusion
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| backend | COMBO | auto | auto = windows-native on Windows, linux elsewhere. wsl2 runs the runtime inside a WSL2 distro (recommended for FP8 on Windows). |
| quantization | COMBO | fp8 | fp8 = upstream W8A8 FP8 (FFN + attention projections). bf16 = no quantization. |
| fused_kernels | BOOLEAN | true | Upstream Triton inference kernels for RoLA. Disabling requires an external thu-ml/SLA checkout (sla_src). |
| torch_compile | BOOLEAN | false | torch.compile the DiT (upstream default). Faster steps, but each ComfyUI run starts a fresh process and pays the compile time again. |
| sparkdiffusion_repoopt | STRING | SparkDiffusion checkout. Empty = SPARKDIFFUSION_REPO or sparkdiffusion_config.json. | |
| pythonopt | STRING | Runtime interpreter (its own venv). For wsl2 this is a Linux path inside the distro. | |
| model_rootopt | STRING | Directory holding the Wan assets and SparkWan checkpoints (see README). | |
| wsl_distroopt | STRING | WSL distro name (wsl2 backend only). | |
| fp8_compatopt | COMBO | auto | When PyTorch lacks row-wise FP8 GEMMs (Windows builds), auto emulates them with a tensor-wise GEMM + epilogue (experimental). off = fail instead. |
| sla_srcopt | STRING | thu-ml/SLA checkout (only for fused_kernels=off). | |
| timeout_secondsopt | INT | 36000–86400 | 0 = no timeout |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| runtime | SPARKDIFFUSION_RUNTIME | — |