quantize - Module

`enum QuantMethod`

Source: quantize.joule:29

`struct QuantConfig`

Source: quantize.joule:68

`fn default() -> Self`

Source: quantize.joule:92

`fn int8() -> Self`

INT8 quantization preset

Source: quantize.joule:113

`fn int4() -> Self`

INT4 quantization preset

Source: quantize.joule:122

`fn gptq() -> Self`

GPTQ preset

Source: quantize.joule:132

`fn awq() -> Self`

AWQ preset

Source: quantize.joule:142

`fn nf4() -> Self`

NF4 preset (QLoRA-style)

Source: quantize.joule:152

`fn fp8() -> Self`

FP8 preset

Source: quantize.joule:162

`fn int3() -> Self`

INT3 preset

Source: quantize.joule:171

`fn int2() -> Self`

INT2 preset

Source: quantize.joule:181

`fn ternary() -> Self`

Ternary preset (-1, 0, +1) - effectively 1.58 bits

Source: quantize.joule:191

`fn binary() -> Self`

Binary preset (-1, +1) - 1 bit

Source: quantize.joule:202

`fn bitnet() -> Self`

BitNet b1.58 preset (ternary optimized for LLMs)

Source: quantize.joule:213

`fn one_bit() -> Self`

1-bit with learned scaling

Source: quantize.joule:225

`fn sub_one_bit() -> Self`

Sub-1-bit quantization (experimental)

Source: quantize.joule:235

`struct CalibrationConfig`

Source: quantize.joule:247

`fn default() -> Self`

Source: quantize.joule:263

`struct QuantStats`

Source: quantize.joule:281

`fn new() -> Self`

Create empty stats

Source: quantize.joule:300

`fn update(&mut self, tensor: &Tensor)`

Update stats with new tensor

Source: quantize.joule:313

`fn compute_scale_zp(&self, bits: u8, symmetric: bool) -> (f32, i32)`

Get scale and zero point for quantization

Source: quantize.joule:337

`struct LayerQuantParams`

Source: quantize.joule:355

`enum QuantError`

Source: quantize.joule:374

`fn from(e: FormatError) -> Self`

Source: quantize.joule:384

`fn from(e: std::io::Error) -> Self`

Source: quantize.joule:390

`struct Quantizer`

Main quantizer struct

Source: quantize.joule:396

`fn new(config: QuantConfig) -> Self`

Create new quantizer

Source: quantize.joule:405

`fn calibrate<D: Dataset>(`

Source: quantize.joule:417

`fn quantize(`

Source: quantize.joule:469

`fn quantize_gptq_with_hessian<D: Dataset>(`

Source: quantize.joule:530

`fn quantize_dynamic(`

Source: quantize.joule:573

`fn quantize_static(`

Source: quantize.joule:599

`fn quantize_gptq(`

Source: quantize.joule:627

`fn quantize_awq(`

Source: quantize.joule:680

`fn quantize_llm_int8(`

Source: quantize.joule:722

`fn quantize_nf4(`

Source: quantize.joule:766

`fn quantize_fp8(`

Source: quantize.joule:839

`fn quantize_ternary(`

Source: quantize.joule:878

`fn quantize_binary(`

Source: quantize.joule:936

`fn quantize_bitnet(`

Source: quantize.joule:998

`fn quantize_sub_one_bit(`

Source: quantize.joule:1054

`fn pack_ternary(&self, values: &[i8]) -> Vec<u8>`

Source: quantize.joule:1131

`fn kmeans_init(`

Source: quantize.joule:1151

`fn should_skip_layer(&self, name: &str) -> bool`

Source: quantize.joule:1208

`fn compute_scale_zp(&self, min_val: f32, max_val: f32) -> (f32, i32)`

Source: quantize.joule:1222

`fn quantize_tensor(&self, tensor: &Tensor, scale: f32, zp: i32) -> Tensor`

Source: quantize.joule:1238

`fn quantize_per_channel(`

Source: quantize.joule:1251

`fn quantize_tensor_with_mask(`

Source: quantize.joule:1285

`fn quantize_scales(&self, scales: &Tensor) -> Result<Tensor, QuantError>`

Source: quantize.joule:1317

`fn compute_hessian<D: Dataset>(`

Source: quantize.joule:1325

`fn gptq_layer(&self, weight: &Tensor, hessian: &Tensor) -> Result<QuantizedTensor, QuantError>`

Source: quantize.joule:1355

`fn energy_scope(&self, op: &str) -> Option<EnergyScope>`

Source: quantize.joule:1410

`struct QuantizedTensor`

Source: quantize.joule:1421

`fn dequantize(&self) -> Tensor`

Source: quantize.joule:1437

`fn size_bytes(&self) -> usize`

Size in bytes

Source: quantize.joule:1457

`fn compression_ratio(&self) -> f32`

Compression ratio vs FP16

Source: quantize.joule:1469

`struct QuantizedModel`

Source: quantize.joule:1477

`fn size_bytes(&self) -> usize`

Total size in bytes

Source: quantize.joule:1488

`fn compression_ratio(&self) -> f32`

Average compression ratio

Source: quantize.joule:1493

`fn get(&self, name: &str) -> Option<&QuantizedTensor>`

Get quantized weight

Source: quantize.joule:1506

`fn dequantize_all(&self) -> HashMap<String, Tensor>`

Source: quantize.joule:1512

`fn save_gguf(&self, path: &Path) -> Result<(), QuantError>`

Source: quantize.joule:1520

`fn save_safetensors(&self, path: &Path) -> Result<(), QuantError>`

Source: quantize.joule:1533

`fn quantize_int8(model: HashMap<String, Tensor>) -> Result<QuantizedModel, QuantError>`

Source: quantize.joule:1551

`fn quantize_int4(model: HashMap<String, Tensor>) -> Result<QuantizedModel, QuantError>`

Source: quantize.joule:1558

`fn quantize_gptq<D: Dataset>(`

Source: quantize.joule:1565

`fn quantize_awq<D: Dataset>(`

Source: quantize.joule:1576

`fn quantize_int3(model: HashMap<String, Tensor>) -> Result<QuantizedModel, QuantError>`

Source: quantize.joule:1587

`fn quantize_int2(model: HashMap<String, Tensor>) -> Result<QuantizedModel, QuantError>`

Source: quantize.joule:1594

`fn quantize_ternary(model: HashMap<String, Tensor>) -> Result<QuantizedModel, QuantError>`

Source: quantize.joule:1601

`fn quantize_binary(model: HashMap<String, Tensor>) -> Result<QuantizedModel, QuantError>`

Source: quantize.joule:1608

`fn quantize_bitnet(model: HashMap<String, Tensor>) -> Result<QuantizedModel, QuantError>`

Source: quantize.joule:1615

`fn quantize_one_bit(model: HashMap<String, Tensor>) -> Result<QuantizedModel, QuantError>`

Source: quantize.joule:1622

`fn quantize_sub_one_bit(model: HashMap<String, Tensor>) -> Result<QuantizedModel, QuantError>`

Source: quantize.joule:1629

`fn estimate_energy_savings(`

Estimate energy savings from quantization

Source: quantize.joule:1635

`fn bits_per_weight(method: QuantMethod) -> f32`

Calculate bits per weight for different quantization methods

Source: quantize.joule:1651

`fn estimate_quantized_size(`

Estimate model size after quantization

Source: quantize.joule:1666

`fn recommend_quantization(`

Recommend best quantization for memory constraint

Source: quantize.joule:1676

`enum QuantComplexity`

Source: quantize.joule:1711

`struct QuantTimeEstimate`

Source: quantize.joule:1726

`fn quantization_complexity(method: QuantMethod) -> QuantComplexity`

Get complexity class for a quantization method

Source: quantize.joule:1740

`fn estimate_quantization_time(`

Estimate time to quantize a model Returns estimated duration based on: - Model size (parameters) - Quantization method complexity - Hardware (CPU vs GPU) - Calibration requirements

Source: quantize.joule:1778

`fn secs_to_duration(secs: f64) -> std::time::Duration`

Source: quantize.joule:1896

`fn format_duration(d: std::time::Duration) -> String`

Format duration for display

Source: quantize.joule:1901

`struct QuantTradeoff`

Source: quantize.joule:1914

`fn analyze_quantization_options(`

Analyze all quantization options for a model

Source: quantize.joule:1924

`fn print_quantization_analysis(`

Print quantization options as a table

Source: quantize.joule:1960

`fn recommend_quantization_balanced(`

Recommend quantization balancing time, quality, and size

Source: quantize.joule:1988