[DeepSeek v3.2] Make top-k work for any logit values. (#27568)

Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Daniel Cámpora
2025-12-08 06:55:58 -08:00
committed by GitHub
co-authored by Michael Goin Cyrus Leung
parent eb1051fb95
commit 184076c3fe
5 changed files with 643 additions and 224 deletions
+8 -5
View File
@@ -102,13 +102,16 @@ void apply_repetition_penalties_(torch::Tensor& logits,
const torch::Tensor& output_mask,
const torch::Tensor& repetition_penalties);
void top_k_per_row(const torch::Tensor& logits, const torch::Tensor& rowStarts,
const torch::Tensor& rowEnds, torch::Tensor& indices,
int64_t numRows, int64_t stride0, int64_t stride1);
void top_k_per_row_prefill(const torch::Tensor& logits,
const torch::Tensor& rowStarts,
const torch::Tensor& rowEnds, torch::Tensor& indices,
int64_t numRows, int64_t stride0, int64_t stride1,
int64_t topK);
void top_k_per_row_decode(const torch::Tensor& logits, int64_t next_n,
const torch::Tensor& seq_lens, torch::Tensor& indices,
int64_t numRows, int64_t stride0, int64_t stride1);
const torch::Tensor& seqLens, torch::Tensor& indices,
int64_t numRows, int64_t stride0, int64_t stride1,
int64_t topK);
void rms_norm_static_fp8_quant(torch::Tensor& out, torch::Tensor& input,
torch::Tensor& weight, torch::Tensor& scale,