Enum TensorStoragePolicy 

pub enum TensorStoragePolicy {
    Verbatim,
    Q8PerOutputChannel,
    Q8PerGroup64,
}

The storage recipe for one source tensor in a portable .fttsq artifact.

The conversion plan must state this policy for every tensor in its source manifest. That makes protected high-precision tensors an explicit, auditable choice rather than an accidental fallback, and it prevents a new checkpoint tensor from being silently omitted.

Variants

Verbatim

Preserve the source BF16 or F32 bytes exactly.

Q8PerOutputChannel

Quantize a rank-two-or-greater weight matrix with canonical per-output-channel Q8 scales.

Q8PerGroup64

Quantize with one canonical Q8 scale per Q8_GROUP_WIDTH-element group of each row.

For rows whose energy is uneven across the row (the cold text embedding’s common-token rows), a single row scale quantizes the quiet stretches at the loud stretch’s step size; per-group scales recover ~11 dB on the worst measured rows for ~3% payload overhead in scales. The quantization primitive is the same quantize_output_channel_q8, applied per group, so grouped bytes remain bit-consistent with the canonical recipe.

Trait Implementations

impl Clone for TensorStoragePolicy

fn clone(&self) -> TensorStoragePolicy

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for TensorStoragePolicy

impl Debug for TensorStoragePolicy

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for TensorStoragePolicy

impl PartialEq for TensorStoragePolicy

fn eq(&self, other: &TensorStoragePolicy) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for TensorStoragePolicy