Crate boostr 

boostr

ML framework built on numr — attention, quantization, model architectures.

boostr extends numr’s foundational numerical computing with ML-specific operations, quantized tensor support, and model building blocks. It uses numr’s runtime, tensors, and ops directly — no reimplementation, no wrappers.

Relationship to numr

┌─────────────────────────────────────────────────────────┐
│                    boostr ◄── YOU ARE HERE               │
│   (attention, RoPE, MoE, quantization, model loaders)   │
└──────────────────────────┬──────────────────────────────┘
│                      numr                                │
│     (tensors, ops, runtime, autograd, linalg, FFT)       │
└─────────────────────────────────────────────────────────┘

Design

• Extension traits: ML ops (AttentionOps, RoPEOps) implemented on numr’s clients
• QuantTensor: Separate type for block-quantized data (GGUF formats)
• impl_generic: Composite ops composed from numr primitives, same on all backends
• Custom kernels: Dequant, quantized matmul, fused attention (SIMD/PTX/WGSL)

Re-exports

pub use nn::Init;

pub use nn::VarBuilder;

pub use nn::VarMap;

pub use nn::Weight;

pub use ops::AttentionOps;

pub use ops::DeviceGrammarDfa;

pub use ops::FlashAttentionOps;

pub use ops::FusedFp8TrainingOps;

pub use ops::FusedOptimizerOps;

pub use ops::FusedQkvOps;

pub use ops::GrammarDfaOps;

pub use ops::KvCacheOps;

pub use ops::MlaOps;

pub use ops::PagedAttentionOps;

pub use ops::RoPEOps;

pub use ops::SamplingOps;

pub use ops::var_flash_attention;

pub use quant::DecomposedQuantLinear;

pub use quant::DecomposedQuantMethod;

pub use quant::DecomposedQuantTensor;

pub use quant::DequantOps;

pub use quant::FusedQuantOps;

pub use quant::QuantFormat;

pub use quant::QuantMatmulOps;

pub use quant::QuantTensor;

pub use model::ExpertWeights;

pub use format::GgufTokenizer;

pub use model::encoder::EmbeddingPipeline;

pub use model::encoder::Encoder;

pub use model::encoder::EncoderClient;

pub use model::encoder::EncoderConfig;

pub use model::encoder::Pooling;

Modules

autograd

Automatic differentiation (autograd)

data

distributed

error

boostr error types

format

inference

model

nn

ops

optimizer

quant

runtime

Runtime backends for tensor computation

tensor

Tensor types and operations

trainer

Structs

CpuClient

CPU client for operation dispatch

CpuDevice

CPU device (there’s only one: the host CPU)

CpuRuntime

CPU compute runtime

Tensor

N-dimensional array stored on a compute device

Enums

DType

Data types supported by numr tensors

NumrError

Errors that can occur in numr operations

Traits

ActivationOps

Activation operations

BinaryOps

Element-wise binary operations on tensors.

ConvOps

Convolution operations.

IndexingOps

Indexing operations

NormalizationOps

Normalization operations

Runtime

Core trait for compute backends

RuntimeClient

Trait for runtime clients that handle operation dispatch

ScalarOps

Scalar operations trait for tensor-scalar operations

TensorOps

Core tensor operations trait

TypeConversionOps

Type conversion operations

UnaryOps

Unary operations

Type Aliases

NumrResult

Result type alias using numr’s Error