tpt_gpu_model_optimizer/
lib.rs1pub mod activation_capture;
20pub mod benchmark;
21pub mod calibration;
22pub mod domain_mapper;
23pub mod export;
24pub mod import;
25pub mod kv_calculator;
26pub mod profiler;
27pub mod pruner;
28pub mod quant_allocator;
29pub mod sensitivity;
30pub mod streaming;
31pub mod tptf_format;
32
33pub use activation_capture::{
34 load_domain_maps_from_dir, ActivationCallback, ActivationCapture, ActivationMap,
35 LayerActivations,
36};
37pub use benchmark::QualityBenchmark;
38pub use calibration::CalibrationGenerator;
39pub use domain_mapper::DomainMapper;
40pub use export::detect::ModelFormat;
41pub use import::gguf::{GgufImporter, GgufModel};
42pub use kv_calculator::KvCacheCalculator;
43pub use profiler::HardwareProfiler;
44pub use pruner::SurgicalPruner;
45pub use quant_allocator::MixedPrecisionAllocator;
46pub use sensitivity::{LayerSensitivityMap, SensitivityConfig};
47pub use streaming::StreamingLoader;
48pub use tptf_format::{TensorBlock, TptfHeader, TptfWriter};
49
50use std::path::PathBuf;
51
52#[derive(Debug, Clone)]
54pub struct OptimizerConfig {
55 pub model_path: PathBuf,
57 pub output_path: PathBuf,
59 pub max_loss_fraction: f32,
61 pub domains_to_drop: Vec<String>,
63 pub force_streaming: bool,
65}
66
67impl OptimizerConfig {
68 pub fn new(model_path: impl Into<PathBuf>, output_path: impl Into<PathBuf>) -> Self {
69 OptimizerConfig {
70 model_path: model_path.into(),
71 output_path: output_path.into(),
72 max_loss_fraction: 0.05,
73 domains_to_drop: Vec::new(),
74 force_streaming: false,
75 }
76 }
77}
78
79#[derive(Debug, Clone)]
81pub struct OptimizationResult {
82 pub output_path: PathBuf,
83 pub per_layer_bits: Vec<u8>,
84 pub pruned_domains: Vec<String>,
85 pub size_before_gb: f64,
86 pub size_after_gb: f64,
87 pub compression_ratio: f32,
88 pub ppl_delta_pct: f32,
89}