ComfyUI Node

HSWQ Torch Compile

torch.compile for HSWQ models. Defaults: inductor + max-autotune-no-cudagraphs (safe with USDU / Distorch). Avoid cudagraphs with multi-tile USDU.

By ussoewwin·Created 8 months ago·Updated about 15 hours ago· 16
HSWQ Torch Compile
  • model
  • MODEL
backendinductor
fullgraphfalse
modemax-autotune-no-cudagraphs
dynamicfalse
dynamo_cache_size_limit64
force_parameter_static_shapestrue
patch_distorch_weight_casttrue
debug_compile_keysfalse
disable_dynamic_vramfalse
CategoryHSWQ/torchcompile

Inputs (10)

NameTypeDefaultDescription
modelMODEL
backendCOMBOinductorUse inductor with USDU. cudagraphs often fails on multi-tile / cudaMallocAsync pools.
fullgraphBOOLEANfalseFull graph mode. Keep False for HSWQ ConvRot / Distorch paths.
modeCOMBOmax-autotune-no-cudagraphsHSWQ default avoids CUDA graphs (USDU / dynamic shapes).
dynamicBOOLEANfalseDynamic shape tracing. Prefer off unless tile shapes vary every step.
dynamo_cache_size_limitINT640–1024torch._dynamo.config.cache_size_limit
force_parameter_static_shapesBOOLEANtruetorch._dynamo.config.force_parameter_static_shapes — helps avoid symbolic weight Size errors.
patch_distorch_weight_castBOOLEANtrueMark comfy.ops cast_bias_weight etc. as eager (Distorch / dynamic VRAM).
debug_compile_keysBOOLEANfalseLog compile keys
disable_dynamic_vramoptBOOLEANfalseClone model with disable_dynamic=True when ComfyUI supports it.

Outputs (1)

NameTypeDescription
MODELMODEL