apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 19 Packages Projects Releases Wiki Activity

[Optimization][DeepSeekV3.2]Reducing slot_mapping compute frequency from twice per layer to a single pre-processing step. (#7367)

Browse Source
This commit is contained in:
ShaneGZhu
2026-04-16 19:54:12 +08:00
committed by GitHub
parent d2d633b05c
commit 2d8338f9e4
10 changed files with 73 additions and 146 deletions
Download Patch File Download Diff File Expand all files Collapse all files
custom_ops/gpu_ops/cpp_extensions.cc
+4 -6
View File
@@ -540,12 +540,10 @@ std::vector<paddle::Tensor> count_tokens_per_expert_func(
const paddle::Tensor& topk_ids,
int64_t num_experts,
bool compute_padded_cumsum = false);
void GetPositionIdsAndMaskEncoderBatch(
const paddle::Tensor& seq_lens_encoder,
const paddle::Tensor& seq_lens_decoder,
const paddle::Tensor& seq_lens_this_time,
const paddle::Tensor& position_ids,
const paddle::Tensor& mask_encoder_batch);
void GetPositionIdsAndMaskEncoderBatch(const paddle::Tensor& seq_lens_encoder,
const paddle::Tensor& seq_lens_decoder,
const paddle::Tensor& seq_lens_this_time,
const paddle::Tensor& position_ids);
std::vector<paddle::Tensor> DecodeMLAWriteCacheKernel(
const paddle::Tensor& kv_nope,