vollo_trees_compiler
Bindings for the Vollo tree compiler.
This module provides functions to compile models to vollo-trees programs, and to evaluate these programs.
- class vollo_trees_compiler.Config
Hardware configuration for compiling a forest to a program
- max_input_features
Maximum number of input features
- max_tree_depth
Maximum depth of a tree
- num_tree_units
Number of tree units
- unit_ram_depth
Maximum number of nodes in each tree unit
- class vollo_trees_compiler.Forest
An internal representation of a forest of decision trees
- static from_onnx(path: str) Forest
Convert an ONNX model with a TreeEnsembleRegressor (https://onnx.ai/onnx/operators/onnx_aionnxml_TreeEnsembleRegressor.html) node to a vollo-trees
Forestrepresentation
- max_depth() int
Get the maximum depth of the trees in the forest
- num_input_features() int
Get the number of input features expected by the forest
- num_trees() int
Get the number of trees in the forest
- to_program_bf16(config: Config) ProgramBf16
Compile the forest to a program where nodes have bf16 comparators, and the inputs and output are bf16
- to_program_f32(config: Config) ProgramF32
Compile the forest to a program where nodes have f32 comparators, and the inputs and output are f32
- class vollo_trees_compiler.ProgramF32
A compiled program where comparators and leaves are float32 values
- eval(input: List[float], model_ix: int = 0) float
Evaluate the program on a given input
- eval_with_cycle_estimate(&self, input: Vec<f32>, model_ix: int = 0) -> (f32, int)
Evaluate a program on a given input, and return the result along with the estimated number of cycles For a worst-case cycle count estimate see
pessimistic_cycle_estimate().
- static load(input_path: str) ProgramF32
Load the program from the specified file path
- pessimistic_cycle_estimate(model_ix: int = 0) int
Return a pessimistic estimate of the number of cycles required to calculate the result of an inference. This is calculated based on the maximum length branch in each tree. For an input-based cycle count see
eval_with_cycle_estimate().
- save(output_path: str)
Save the program to the specified file path
- class vollo_trees_compiler.ProgramBf16
A compiled program where comparators and leaves are bf16 valuess
- eval(input: List[float], model_ix: int = 0) float
Evaluate the program on a given input.
The input float32 values are converted to bf16 before evaluating the program, and the result is converted from bf16 to float32
- eval_with_cycle_estimate(&self, input: Vec<f32>, model_ix: int = 0) -> (f32, int)
Evaluate a program on a given input, and return the result along with the estimated number of cycles. For a worst-case cycle count estimate see
pessimistic_cycle_estimate().
- static load(input_path: str) ProgramBf16
Load the program from the specified file path
- pessimistic_cycle_estimate(model_ix: int = 0) int
Return a pessimistic estimate of the number of cycles required to calculate the result of an inference. This is calculated based on the maximum length branch in each tree. For an input-based cycle count see
eval_with_cycle_estimate().
- save(output_path: str)
Save the program to the specified file path