Struct Q8SectionSink 

pub struct Q8SectionSink<'a, W> { /* private fields */ }

A bounded bridge from canonical Q8 matrix rows into one streaming .fttsq section.

.fttsq keeps the Q8 tensor contiguous, followed by the contiguous F32 scale tensor that its directory names. The sink therefore streams every Q8 row immediately, retaining only the scale tail (at most one MiB for the pinned inventory) until the values have completed. It also keeps one byte-per-row scratch buffer for the signed-to-wire byte conversion, so the total working set remains bounded by the row adapter’s 320 KiB plus at most 64 KiB of value bytes and one MiB of scales — never by the full matrix size.

Implementations

impl<'a, W: Write + Seek> Q8SectionSink<'a, W>

pub fn new( writer: &'a mut FttsqStreamingWriter<W>, section: impl Into<String>, expected_rows: usize, ) -> Result<Self, Q8SectionSinkError>

Starts writing one Q8 matrix section with a fixed number of output channels.

Errors

Returns Q8SectionSinkError::OutputChannelCountTooLarge before allocating when the planned F32 scale tail exceeds this adapter’s one-MiB memory ceiling.

pub fn finish(self) -> Result<(), Q8SectionSinkError>

Appends the scale tail after all Q8 values have been streamed.

Errors

Returns a named refusal when a prior row failed or a stream write cannot complete.

Trait Implementations

impl<W: Write + Seek> Q8RowSink for Q8SectionSink<'_, W>

type Error = Q8SectionSinkError

Failure produced while accepting a quantized row.

fn write_q8_row( &mut self, row: usize, scale: f32, values: &[i8], ) -> Result<(), Self::Error>

Accepts one output channel of a matrix.