BalancedTernaryEngine.RawLayer (balanced_ternary_engine v0.1.0-dev)

A dense weight matrix kept at half precision inside a converted model.

Quantization-sensitive tensors — token embeddings, LM head, layer norms — are stored as RawLayer instead of being ternarized (see BalancedTernaryEngine.Architectures). Weights are held as raw float16 bytes; activations are computed in f32.

Summary

Types

t()

Functions

forward_nx(layer, input)

Forward pass: dequantize to f32 and apply as a linear projection.

from_binary(binary)

Deserialize from binary. Returns {layer, rest_binary}.

from_tensor(tensor, name)

Build a RawLayer from a tensor of any float type (stored as f16).

packed_bytes(layer)

Byte size of the stored f16 payload (for stats).

to_binary(layer)

Serialize to binary: name, shape, f16 payload.

to_tensor(layer)

Materialize the weights as an f32 tensor.

Types

t()

@type t() :: %BalancedTernaryEngine.RawLayer{
  data: binary(),
  name: String.t(),
  shape: tuple()
}

Functions

forward_nx(layer, input)

@spec forward_nx(t(), Nx.Tensor.t()) :: Nx.Tensor.t()

Forward pass: dequantize to f32 and apply as a linear projection.

Handles both (out, in) (y = x·Wᵀ) and (in, out) (y = x·W) layouts — GPT-2-style checkpoints store Conv1D weights transposed, and the converter preserves whichever layout it was given. Intended for lm_head-style tensors; embeddings and norm vectors are consumed by runtimes via to_tensor/1.

from_binary(binary)

@spec from_binary(binary()) :: {t(), binary()}

Deserialize from binary. Returns {layer, rest_binary}.

from_tensor(tensor, name)

@spec from_tensor(Nx.Tensor.t(), String.t()) :: t()

Build a RawLayer from a tensor of any float type (stored as f16).

packed_bytes(layer)

@spec packed_bytes(t()) :: non_neg_integer()

Byte size of the stored f16 payload (for stats).

to_binary(layer)

@spec to_binary(t()) :: binary()

Serialize to binary: name, shape, f16 payload.

to_tensor(layer)

@spec to_tensor(t()) :: Nx.Tensor.t()

Materialize the weights as an f32 tensor.