Fused Gumbel selection after native nucleus probability reductions.
Keep ATen's top-k threshold, sort, softmax and cumulative sum, including boundary ties. Random numbers remain indexed by original token id. Only the final mask, Gumbel transform and argmax are fused; no vocabulary scatter is needed. This avoids changing which tokens survive a nucleus boundary through different reduction orders.
sample_top_k_top_p_gumbel(
logits: Tensor,
uniforms: Tensor,
*,
top_k: int,
top_p: float,
) -> Tensor
Sample CUDA rows without scattering sorted logits back to the vocabulary.
Uniforms are caller-owned FP32 values with the same shape as logits; their generation and generator advancement are deliberately outside this kernel. Softmax/scan use the reference ATen operations. The Gumbel logarithms still use libdevice, so this is not a universal bitwise-equivalence claim.