TorchCompileModelFluxAdvancedV2

The TorchCompileModelFluxAdvancedV2 node is deprecated – use TorchCompileModelAdvanced instead. In its technical implementation, it clones the input model and applies torch.compile to its internal diffusion model part.

TorchCompileModelFluxAdvancedV2 is a node used for optimizing and accelerating model inference. It compiles the given diffusion model using PyTorch's torch.compile functionality, aiming to improve the model's runtime efficiency during image generation. This node provides various compilation options, allowing users to adjust optimization strategies based on hardware and requirements.

Node Functionality

In its technical implementation, this node clones the input model and applies torch.compile to its internal diffusion model part. It supports multiple backends (e.g., inductor) and compilation modes (e.g., max-autotune), and allows control over whether to compile specific modules within the model (such as single/double attention blocks). Furthermore, it can configure underlying parameters like PyTorch Dynamo's cache size and shape forcing staticization for more granular performance tuning.

Node Parameter Description - TorchCompileModelFluxAdvancedV2

Connection Inputs (Inputs)

Parameter Name	Data Type	Required	Default Value	Value Range/Options	Description
`model`	MODEL	Yes	-	-	Receives the model input to be optimized. This parameter is the core input of the node, typically coming from an upstream model loading node.

Control Parameters (Parameters)

Parameter Name	Data Type	Required	Default Value	Value Range/Options	Description
`backend`	COMBO	Yes	inductor	"inductor", "cudagraphs"	Selects the backend engine used for PyTorch compilation. Options are `inductor` (default) or `cudagraphs`, specifying the underlying optimization implementation.
`mode`	COMBO	Yes	"default"	"default", "max-autotune", "max-autotune-no-cudagraphs", "reduce-overhead"	`max-autotune`, `max-autotune-no-cudagraphs`, and `reduce-overhead`, used to trade off between compilation time, memory usage, and runtime speed.
`fullgraph`	BOOLEAN	Yes	False	-	Enable full graph mode
`double_blocks`	BOOLEAN	Yes	True	-	Compile double blocks
`single_blocks`	BOOLEAN	Yes	True	-	Compile single blocks
`dynamic`	BOOLEAN	Yes	False	-	Enable dynamic mode
`dynamo_cache_size_limit`	INT	No	64	0 - 1024 (step: 1)	torch._dynamo.config.cache_size_limit
`force_parameter_static_shapes`	BOOLEAN	No	True	-	torch._dynamo.config.force_parameter_static_shapes

Output

Parameter Name	Data Type	Description
MODEL	MODEL	Outputs the model optimized by `torch.compile`, ready for use by subsequent sampling or processing nodes.

Usage Scenarios

In an image generation workflow, you can place this node after the model loading node and before the sampler node. For example, after loading a Stable Diffusion model, compile and optimize it using this node, then feed the optimized model into the sampler for image generation, aiming for faster generation speeds.

Notes

Please note that this node has been marked as "deprecated," and the developer recommends using its replacement node, TorchCompileModelAdvanced. It is still in an experimental stage, and its optimization effectiveness may vary significantly depending on the model, hardware, and specific parameters.