apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 00:33:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[OPs] Universal optimization and Fix early_stop cuda 700 (#3375)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
Details

Browse Source 
* delete nonzero

* delete setup_ops_base.py

* check if

* check gcp infer_seed.cpu()

* fix repetition_early_stopper_kernel cuda 700
This commit is contained in:
chen
2025-08-14 22:40:44 +08:00
committed by GitHub
parent 09c979f3dd
commit f0f00a6025
15 changed files with 102 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
custom_ops/gpu_ops/cpp_extensions.cc
View File

@@ -776,6 +776,22 @@ void MergePrefillDecodeOutput(
const int head_dim,
const int max_token);
std::vector<paddle::Tensor> TopPSamplingReject(const paddle::Tensor &probs,
const paddle::Tensor &top_p,
const paddle::optional<paddle::Tensor> &top_k,
int64_t seed);
std::vector<paddle::Tensor> TopKRenorm(const paddle::Tensor &probs,
const paddle::Tensor &top_k);
std::vector<paddle::Tensor> MinPSamplingFromProbs(const paddle::Tensor &probs,
const paddle::Tensor &min_p);
void SaveOutMmsgStatic(const paddle::Tensor& x,
const paddle::Tensor& not_need_stop,
int64_t rank_id,
bool save_each_rank);
PYBIND11_MODULE(fastdeploy_ops, m) {
m.def("get_expert_token_num", &GetExpertTokenNum, py::arg("topk_ids"),
@@ -1128,4 +1144,12 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
m.def("speculate_step_paddle",&SpeculateStepPaddle, "speculate_step_paddle function");
m.def("merge_prefill_decode_output", &MergePrefillDecodeOutput, "merge_prefill_decode_output function");
m.def("rejection_top_p_sampling", &TopPSamplingReject, "rejection_top_p_sampling function");
m.def("top_k_renorm_probs", &TopKRenorm, "top_k_renorm_probs function");
m.def("min_p_sampling", &MinPSamplingFromProbs, "min_p_sampling function");
m.def("save_output", &SaveOutMmsgStatic, "save_output function");
}

8
custom_ops/gpu_ops/get_output_ep.cc
View File

@@ -109,11 +109,11 @@ void GetOutputEp(const paddle::Tensor& x,
return;
}
void GetOutputStatic(const paddle::Tensor& x, int64_t rank_id, bool wait_flag) {
void GetOutputEPStatic(const paddle::Tensor& x, int64_t rank_id, bool wait_flag) {
GetOutputEp(x, rank_id, wait_flag, 1);
}
void GetOutputDynamic(const paddle::Tensor& x,
void GetOutputEPDynamic(const paddle::Tensor& x,
int64_t rank_id,
bool wait_flag,
int msg_queue_id) {
@@ -125,11 +125,11 @@ PD_BUILD_STATIC_OP(get_output_ep)
.Attrs({"rank_id: int64_t", "wait_flag: bool"})
.Outputs({"x_out"})
.SetInplaceMap({{"x", "x_out"}})
.SetKernelFn(PD_KERNEL(GetOutputStatic));
.SetKernelFn(PD_KERNEL(GetOutputEPStatic));
PD_BUILD_STATIC_OP(get_output_ep_dynamic)
.Inputs({"x"})
.Attrs({"rank_id: int64_t", "wait_flag: bool", "msg_queue_id: int"})
.Outputs({"x_out"})
.SetInplaceMap({{"x", "x_out"}})
.SetKernelFn(PD_KERNEL(GetOutputDynamic));
.SetKernelFn(PD_KERNEL(GetOutputEPDynamic));

4
custom_ops/gpu_ops/sample_kernels/rejection_top_p_sampling.cu
View File

@@ -19,7 +19,7 @@
std::vector<paddle::Tensor> TopPSamplingReject(const paddle::Tensor &probs,
const paddle::Tensor &top_p,
const paddle::optional<paddle::Tensor> &top_k,
int seed) {
int64_t seed) {
std::vector<int64_t> probs_shape = probs.shape();
unsigned int batch_size = probs_shape[0];
unsigned int vocab_size = probs_shape[1];
@@ -82,7 +82,7 @@ TopPSamplingRejectInferDtype(const paddle::DataType &probs_dtype,
PD_BUILD_STATIC_OP(rejection_top_p_sampling)
.Inputs({"probs", "top_p", paddle::Optional("top_k")})
.Outputs({"samples"})
.Attrs({"seed: int"})
.Attrs({"seed: int64_t"})
.SetKernelFn(PD_KERNEL(TopPSamplingReject))
.SetInferShapeFn(PD_INFER_SHAPE(TopPSamplingRejectInferShape))
.SetInferDtypeFn(PD_INFER_DTYPE(TopPSamplingRejectInferDtype));