[Feat][DSV4] Fuse q pad into deepseek v4 fused kernel (#43162)

This commit is contained in:
Yongye Zhu
2026-05-26 05:12:54 -10:00
committed by GitHub
parent 445ded18c1
commit 6ab6ffb428
11 changed files with 338 additions and 150 deletions
+4 -3
View File
@@ -70,10 +70,11 @@ void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
void fused_add_rms_norm(torch::Tensor& input, torch::Tensor& residual,
torch::Tensor& weight, double epsilon);
void fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
torch::Tensor& q, torch::Tensor const& kv, torch::Tensor& k_cache,
torch::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
torch::Tensor const& q_in, torch::Tensor const& kv, torch::Tensor& k_cache,
torch::Tensor const& slot_mapping, torch::Tensor const& position_ids,
torch::Tensor const& cos_sin_cache, double eps, int64_t cache_block_size);
torch::Tensor const& cos_sin_cache, int64_t q_head_padded, double eps,
int64_t cache_block_size);
void apply_repetition_penalties_(torch::Tensor& logits,
const torch::Tensor& prompt_mask,