BalancedTernaryEngine.Pipeline (balanced_ternary_engine v0.1.0-dev)

End-to-end conversion: safetensors model → .tbin balanced ternary model.

{:ok, summary} =
  BalancedTernaryEngine.Pipeline.run("model.safetensors", "model.tbin",
    arch: :llama,
    mode: :absmean
  )

Steps:

Read all tensors from the source .safetensors file (BalancedTernaryEngine.Safetensors)
Classify every tensor via the architecture policy (BalancedTernaryEngine.Architectures) → ternary | keep FP16 | skip
Quantize ternary targets with the selected mode and pack them; sensitive tensors are stored as raw FP16 records
Export a version-2 .tbin file with provenance metadata

Options

:arch — :llama, :gpt2, :bert, :vit, :generic (default: :generic)
:mode — quantization mode passed to BalancedTernaryEngine.Quantizer.quantize/2: :delta, :twn, :absmean (default: :twn)
:delta — explicit threshold when mode is :delta
:compress — gzip the output (default: false)

Validate that a conversion input path exists.

@spec run(String.t(), String.t(), keyword()) :: {:ok, map()} | {:error, term()}

@spec validate_input(String.t()) :: :ok | {:error, term()}

Validate that a conversion input path exists.