ComfyUI Node
Torch Compile Speed Settings
torch.compile settings for maximum speed optimization. Speed Mode Features: - Uses inductor backend with max-autotune-no-cudagraphs - Enables dynamic compilation for better cache reuse - Disables CUDA graphs for flexibility - Enables all Triton autotune optimizations - First run: comprehensive autotune (slower) - Second run: cached execution (extremely fast) Author: eddy
Torch Compile Speed Settings
- torch_compile_args
◄backendinductor►
◄fullgraphfalse►
◄modespeed►
◄dynamicfalse►
◄dynamo_cache_size_limit64►
◄compile_transformer_blocks_onlytrue►
◄reuse_if_similartrue►
◄experimental_ptxfalse►
◄ptx_fast_mathtrue►
◄warmup_runs1►
◄ptx_cache_dir►
◄dynamo_recompile_limit128►
Categoryoptimization
Inputs (12)
| Name | Type | Default | Description |
|---|---|---|---|
| backend | COMBO | inductor | 2 options: inductor, cudagraphs |
| fullgraph | BOOLEAN | false | Enable full graph mode |
| mode | COMBO | speed | 5 options: default, max-autotune, max-autotune-no-cudagraphs, reduce-overhead, speed |
| dynamic | BOOLEAN | false | Enable dynamic mode |
| dynamo_cache_size_limit | INT | 640–1024 | torch._dynamo.config.cache_size_limit |
| compile_transformer_blocks_only | BOOLEAN | true | Compile only transformer blocks |
| reuse_if_similar | BOOLEAN | true | Reuse compiled artifacts when similar |
| experimental_ptx | BOOLEAN | false | Enable experimental PTX acceleration |
| ptx_fast_math | BOOLEAN | true | Enable fast math if available |
| warmup_runs | INT | 10–5 | — |
| ptx_cache_diropt | STRING | — | |
| dynamo_recompile_limitopt | INT | 1280–1024 | torch._dynamo.config.recompile_limit |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| torch_compile_args | WANCOMPILEARGS | — |