Module sampling 

Sampling algorithms kernel.

Matches sampling-algorithms-v1.yaml. Greedy, top-k, top-p, and temperature sampling for autoregressive generation.

Each function provides one of three backends:

• fn {name}_scalar(...) – Pure Rust scalar reference (ground truth)
• unsafe fn {name}_avx2(...) – AVX2 SIMD implementation
• fn sampling_ptx() -> &'static str – PTX assembly source string

Functions

greedy_avx2^⚠

AVX2 greedy sampling – delegates to scalar.

greedy_scalar

Greedy sampling: return the index of the maximum logit.

sample_scalar

Full sampling pipeline: apply temperature, softmax, then greedy (scalar reference).

sampling_ptx

PTX assembly for greedy sampling (argmax reduction).

temperature_avx2^⚠

AVX2 temperature scaling – delegates to scalar.

temperature_scalar

Apply temperature scaling to logits in-place: logits[i] /= temperature.

top_k_scalar

Top-K filtering: zero out all probabilities except the K highest.

top_p_scalar

Top-P (nucleus) filtering: retain the minimal set of tokens whose cumulative probability exceeds threshold.