apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

c++ code format (#4527)

Browse Source
This commit is contained in:
zhupengyang
2025-10-22 17:59:50 +08:00
committed by GitHub
parent d7bcedf421
commit 3a6883ac1a
97 changed files with 8760 additions and 7382 deletions
Download Patch File Download Diff File Expand all files Collapse all files

110
custom_ops/cpu_ops/get_padding_offset.cc
View File

@@ -20,13 +20,13 @@ void remove_padding(int64_t *output_data,
const int *cum_offsets,
const int sequence_length,
const int bsz) {
for (int bi = 0; bi < bsz; ++bi) {
for (int i = 0; i < seq_lens[bi]; ++i) {
const int tgt_seq_id = bi * sequence_length - cum_offsets[bi] + i;
const int src_seq_id = bi * sequence_length + i;
output_data[tgt_seq_id] = input_data[src_seq_id];
}
for (int bi = 0; bi < bsz; ++bi) {
for (int i = 0; i < seq_lens[bi]; ++i) {
const int tgt_seq_id = bi * sequence_length - cum_offsets[bi] + i;
const int src_seq_id = bi * sequence_length + i;
output_data[tgt_seq_id] = input_data[src_seq_id];
}
}
}
void get_padding_offset_kernel(int *padding_offset,
@@ -37,56 +37,53 @@ void get_padding_offset_kernel(int *padding_offset,
const int *seq_lens,
const int max_seq_len,
const int bsz) {
for (int bi = 0; bi < bsz; ++bi) {
int cum_offset = bi == 0 ? 0 : cum_offsets[bi - 1];
auto seq_len_now = seq_lens[bi];
for (int i = 0; i < seq_len_now; ++i) {
padding_offset[bi * max_seq_len - cum_offset + i] = cum_offset;
}
cum_offsets_out[bi] = cum_offset;
int cum_seq_len = (bi + 1) * max_seq_len - cum_offsets[bi];
cu_seqlens_q[bi + 1] = cum_seq_len;
cu_seqlens_k[bi + 1] = cum_seq_len;
for (int bi = 0; bi < bsz; ++bi) {
int cum_offset = bi == 0 ? 0 : cum_offsets[bi - 1];
auto seq_len_now = seq_lens[bi];
for (int i = 0; i < seq_len_now; ++i) {
padding_offset[bi * max_seq_len - cum_offset + i] = cum_offset;
}
cum_offsets_out[bi] = cum_offset;
int cum_seq_len = (bi + 1) * max_seq_len - cum_offsets[bi];
cu_seqlens_q[bi + 1] = cum_seq_len;
cu_seqlens_k[bi + 1] = cum_seq_len;
}
}
std::vector<paddle::Tensor> GetPaddingOffset(const paddle::Tensor &input_ids,
const paddle::Tensor &cum_offsets,
const paddle::Tensor &token_num,
const paddle::Tensor &seq_len) {
std::vector<int64_t> input_ids_shape = input_ids.shape();
const int bsz = seq_len.shape()[0];
const int seq_length = input_ids_shape[1];
auto cum_offsets_out = cum_offsets.copy_to(paddle::CPUPlace(), false);
auto cpu_token_num = token_num.copy_to(paddle::CPUPlace(), false);
std::vector<int64_t> input_ids_shape = input_ids.shape();
const int bsz = seq_len.shape()[0];
const int seq_length = input_ids_shape[1];
auto cum_offsets_out = cum_offsets.copy_to(paddle::CPUPlace(), false);
auto cpu_token_num = token_num.copy_to(paddle::CPUPlace(), false);
const int token_num_data = cpu_token_num.data<int64_t>()[0];
auto x_remove_padding = paddle::empty(
{token_num_data}, paddle::DataType::INT64, input_ids.place());
auto padding_offset = paddle::empty(
{token_num_data}, paddle::DataType::INT32, input_ids.place());
auto cu_seqlens_q =
paddle::full({bsz + 1}, 0, paddle::DataType::INT32, input_ids.place());
auto cu_seqlens_k =
paddle::full({bsz + 1}, 0, paddle::DataType::INT32, input_ids.place());
get_padding_offset_kernel(padding_offset.data<int>(),
cum_offsets_out.data<int>(),
cu_seqlens_q.data<int>(),
cu_seqlens_k.data<int>(),
cum_offsets.data<int>(),
seq_len.data<int>(),
seq_length,
bsz);
remove_padding(x_remove_padding.data<int64_t>(),
input_ids.data<int64_t>(),
seq_len.data<int>(),
cum_offsets_out.data<int>(),
seq_length,
bsz);
return {x_remove_padding,
padding_offset,
cu_seqlens_q,
cu_seqlens_k};
const int token_num_data = cpu_token_num.data<int64_t>()[0];
auto x_remove_padding = paddle::empty(
{token_num_data}, paddle::DataType::INT64, input_ids.place());
auto padding_offset = paddle::empty(
{token_num_data}, paddle::DataType::INT32, input_ids.place());
auto cu_seqlens_q =
paddle::full({bsz + 1}, 0, paddle::DataType::INT32, input_ids.place());
auto cu_seqlens_k =
paddle::full({bsz + 1}, 0, paddle::DataType::INT32, input_ids.place());
get_padding_offset_kernel(padding_offset.data<int>(),
cum_offsets_out.data<int>(),
cu_seqlens_q.data<int>(),
cu_seqlens_k.data<int>(),
cum_offsets.data<int>(),
seq_len.data<int>(),
seq_length,
bsz);
remove_padding(x_remove_padding.data<int64_t>(),
input_ids.data<int64_t>(),
seq_len.data<int>(),
cum_offsets_out.data<int>(),
seq_length,
bsz);
return {x_remove_padding, padding_offset, cu_seqlens_q, cu_seqlens_k};
}
std::vector<std::vector<int64_t>> GetPaddingOffsetInferShape(
@@ -94,9 +91,9 @@ std::vector<std::vector<int64_t>> GetPaddingOffsetInferShape(
const std::vector<int64_t> &cum_offsets_shape,
const std::vector<int64_t> &token_num_shape,
const std::vector<int64_t> &seq_len_shape) {
int64_t bsz = seq_len_shape[0];
int64_t seq_len = input_ids_shape[1];
return {{-1}, {-1}, {bsz + 1}, {bsz + 1}};
int64_t bsz = seq_len_shape[0];
int64_t seq_len = input_ids_shape[1];
return {{-1}, {-1}, {bsz + 1}, {bsz + 1}};
}
std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
@@ -104,18 +101,13 @@ std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
const paddle::DataType &cum_offsets_dtype,
const paddle::DataType &token_num_dtype,
const paddle::DataType &seq_len_dtype) {
return {input_ids_dtype,
seq_len_dtype,
seq_len_dtype,
seq_len_dtype};
return {input_ids_dtype, seq_len_dtype, seq_len_dtype, seq_len_dtype};
}
PD_BUILD_STATIC_OP(get_padding_offset_cpu)
.Inputs({"input_ids", "cum_offsets", "token_num", "seq_len"})
.Outputs({"x_remove_padding",
"padding_offset",
"cu_seqlens_q",
"cu_seqlens_k"})
.Outputs(
{"x_remove_padding", "padding_offset", "cu_seqlens_q", "cu_seqlens_k"})
.SetKernelFn(PD_KERNEL(GetPaddingOffset))
.SetInferShapeFn(PD_INFER_SHAPE(GetPaddingOffsetInferShape))
.SetInferDtypeFn(PD_INFER_DTYPE(GetPaddingOffsetInferDtype));