Struct DynamicRangeQuantizePlan 

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

dynamic_range_quantize plan.

Composes per-row max-abs reduction + scale computation + per-row quantize into a single fused kernel launch (the dynamic-range recipe IS the fused composition — running the reduce and quantize as one kernel avoids a scale-buffer round-trip).

The trailblazer kernel is symmetric per-token only. Future fanout adds asymmetric mode (requires xmin + xmax reductions) and the other three scopes (tensor / channel / group) by orchestrating existing primitives.

When to use: post-training activation quantization at inference — compute scale from runtime range and quantize in one launch. No BW (inference-only).

Dtypes (trailblazer): TIn ∈ {f32, f64}, TOut = S8. f16 / bf16 activation, u8 output, and asymmetric mode gated as Unsupported until follow-up milestones wire the xmin/xmax reductions and offset-compute kernel.

Shape limits: rank-2 [N, D]; N ≤ 65535 (block-per-row grid cap, lifts when row tiling lands); q_max > 0 (symmetric divisor); q_max ≥ q_min.

Workspace: none — single-launch fused kernel.

Precision guarantee: deterministic, bit-stable. One block per row, no atomics; block-tree reduction is associative-stable on a single GPU.

Implementations

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

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

Pick a kernel for desc.

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

Validate args at run time.

pub fn workspace_size(&self) -> usize

Workspace bytes — none. The kernel is single-launch.

pub fn sku(&self) -> KernelSku

Identity of the selected kernel.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees.

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

Launch.