Struct DequantizePerTokenPlan 

pub struct DequantizePerTokenPlan<TIn: Element, TOut: IntElement> { /* private fields */ }

dequantize_per_token plan.

x[n, d] = scale[n] * (q[n, d] - zero_point[n]). Inverse of QuantizePerTokenPlan.

When to use: FP recovery from W8A8-style per-row-quantized activations. Pair with DequantizePerTokenBackwardPlan.

Dtypes: input int {s8, u8}; output FP {f32, f64, f16, bf16}.

Shape limits: rank-2 [N, D]; per-row scale and zp of length N.

Workspace: none.

Precision guarantee: deterministic, bit-stable.

Implementations

impl<TIn: Element, TOut: IntElement> DequantizePerTokenPlan<TIn, TOut>

pub fn select( _stream: &Stream, desc: &DequantizePerTokenDescriptor, _pref: PlanPreference, ) -> Result<Self>

Pick a kernel for desc.

pub fn can_implement( &self, args: &DequantizePerTokenArgs<'_, TIn, TOut>, ) -> Result<()>

Validate args.

pub fn workspace_size(&self) -> usize

Workspace bytes — none.

pub fn sku(&self) -> KernelSku

Identity.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees.

pub fn run( &self, stream: &Stream, _workspace: Workspace<'_>, args: DequantizePerTokenArgs<'_, TIn, TOut>, ) -> Result<()>

Launch.