Struct Team 

pub struct Team { /* private fields */ }

The process-wide team. Armed by default at min(6, cores) partitions on native (FTTS_INT8_THREADS overrides; 1 disarms), and explicitly by the host on wasm.

Implementations

impl Team

pub fn linear_f32( &self, x: &[f32], weight: &[f32], bias: Option<&[f32]>, m: usize, k: usize, n: usize, out: &mut [f32], )

Runs one f32 dense linear across the team, bit-identically to the serial packed kernel.

Panics

Panics on shape mismatches, exactly as the serial kernel does.

The argument count mirrors the serial kernel’s signature exactly, which is the point: a caller swaps one call for the other with no reshaping, so any divergence would be a compile error rather than a silent behavioural difference.

pub fn linear_q8( &self, x_q: &[i8], x_scales: &[f32], weight: &QuantizedMatrix, bias: Option<&[f32]>, m: usize, out: &mut [f32], tier: Int8Tier, )

Runs one W8A8 linear across the team. Bit-identical to the serial path per element.

Panics

Panics on shape mismatches, exactly as the serial kernel does.

pub fn gqa_attention( &self, queries: &[f32], keys: &[f32], values: &[f32], mask: &[f32], query_positions: usize, key_positions: usize, q_heads: usize, kv_heads: usize, head_dim: usize, out: &mut [f32], )

Runs the default-arithmetic GQA attention across the team, partitioned over query heads. Bit-identical to the serial f32ref::gqa_attention (same extracted loop).

Panics

Panics on shape mismatches, exactly as the serial reference does.