API Reference

tensor

Types

pub type Tensor {
  Tensor(data: List(Float), shape: List(Int))
}

Creation

Function	Description
`from_list(data, shape)`	Create from list
`zeros(shape)`	Zero tensor
`ones(shape)`	Ones tensor
`random_uniform(shape)`	Values [0,1]

Operations

Function	Description
`add(a, b)`	Element-wise sum
`mul(a, b)`	Element-wise multiply
`scale(t, s)`	Scale by constant
`sum(t)`	Total sum
`mean(t)`	Mean

nf4

Types

pub type NF4Config {
  NF4Config(block_size: Int, double_quant: Bool)
}

pub type NF4Tensor {
  NF4Tensor(
    blocks: List(NF4Block),
    shape: List(Int),
    num_elements: Int,
    memory_bytes: Int,
    compression_ratio: Float,
  )
}

Functions

Function	Description
`default_config()`	Default config (block=64)
`quantize(t, config)`	Quantize tensor
`dequantize(nf4)`	Restore FP32
`nf4_levels()`	List 16 levels

awq

Types

pub type AWQConfig {
  AWQConfig(
    bits: Int,
    group_size: Int,
    alpha: Float,
    zero_point: Bool,
  )
}

Functions

Function	Description
`default_config()`	Default config
`quantize_awq(w, cal, cfg)`	Quantize with calibration
`dequantize_awq(awq)`	Restore FP32
`collect_activation_stats(data)`	Collect stats

flash_attention

Types

pub type FlashConfig {
  FlashConfig(block_size: Int, use_causal_mask: Bool)
}

Functions

Function	Description
`default_config()`	Default config
`flash_attention_forward(q, k, v, cfg)`	Forward pass

sparsity

Functions

Function	Description
`apply_2_4_sparsity(t)`	Apply 2:4 pattern
`compress_sparse(s)`	Remove zeros
`decompress_sparse(c)`	Restore

metrics

Functions

Function	Description
`compute_sqnr(orig, quant)`	SQNR in dB
`compute_mse(orig, quant)`	Mean squared error
`theoretical_sqnr(bits)`	Theoretical max SQNR