apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-25 17:40:35 +08:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

[Code Simplification] remove cum_offsets (#3410)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
Details
Publish Job / publish_pre_check (push) Has been cancelled
Details
Publish Job / print_publish_pre_check_outputs (push) Has been cancelled
Details
Publish Job / FD-Clone-Linux (push) Has been cancelled
Details
Publish Job / Show Code Archive Output (push) Has been cancelled
Details
Publish Job / BUILD_SM8090 (push) Has been cancelled
Details
Publish Job / BUILD_SM8689 (push) Has been cancelled
Details
Publish Job / PADDLE_PYPI_UPLOAD_8090 (push) Has been cancelled
Details
Publish Job / PADDLE_PYPI_UPLOAD_8689 (push) Has been cancelled
Details
Publish Job / Run FastDeploy Unit Tests and Coverage (push) Has been cancelled
Details
Publish Job / Run FastDeploy LogProb Tests (push) Has been cancelled
Details
Publish Job / Extracted partial CE model tasks to run in CI. (push) Has been cancelled
Details
Publish Job / Run Base Tests (push) Has been cancelled
Details
Publish Job / Run Accuracy Tests (push) Has been cancelled
Details

Browse Source
This commit is contained in:
lizexu123
2025-08-18 20:21:25 +08:00
committed by GitHub
parent 2cf96ddd68
commit 32b39620bc
9 changed files with 73 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
custom_ops/cpu_ops/rebuild_padding.cc
View File

@@ -1,4 +1,4 @@
// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
@@ -19,10 +19,11 @@
#define PD_BUILD_STATIC_OP(name) PD_BUILD_OP(static_op_##name)
#endif
template <typename T>
void RebuildPaddingCPUImpl(T *output_data,
const T *input_data,
const int *cum_offsets_data,
const int *cu_seqlens_q_data,
const int *seq_len_this_time_data,
const int *seq_lens_decoder_data,
const int *seq_lens_encoder_data,
@@ -40,11 +41,12 @@ void RebuildPaddingCPUImpl(T *output_data,
if (seq_lens_decoder_data[bi] == 0 && seq_lens_encoder_data[bi] == 0) {
continue;
}
if (seq_lens_encoder_data[bi] > 0) {
seq_id = seq_lens_encoder_data[bi] - 1;
}
const int ori_token_idx =
bi * max_input_length - cum_offsets_data[bi] + seq_id;
const int ori_token_idx = cu_seqlens_q_data[bi] + seq_id;
const int src_offset = ori_token_idx * dim_embed + bias_idx;
output_data[i] = input_data[src_offset];
@@ -54,7 +56,7 @@ void RebuildPaddingCPUImpl(T *output_data,
template <typename T>
void RebuildAppendPaddingCPUImpl(T *output_data,
const T *input_data,
const int *cum_offsets_data,
const int *cu_seqlens_q_data,
const int *seq_len_this_time_data,
const int *seq_lens_decoder_data,
const int *seq_lens_encoder_data,
@@ -69,30 +71,32 @@ void RebuildAppendPaddingCPUImpl(T *output_data,
int bi = ori_token_id / max_input_length;
if (seq_len_this_time_data[bi] == 0 ||
(seq_lens_decoder_data[bi] == 0 &&
seq_lens_encoder_data[bi] == 0)) {
continue;
}
seq_lens_encoder_data[bi] == 0)) {
continue;
}
int seq_id = 0;
if (seq_lens_encoder_data[bi] > 0) {
seq_id = seq_lens_encoder_data[bi] - 1;
}
int input_token_id = ori_token_id - cum_offsets_data[bi] + seq_id;
int input_token_id = cu_seqlens_q_data[bi] + seq_id;
int bias_idx = i % dim_embed;
int src_offset = input_token_id * dim_embed + bias_idx;
output_data[i] = input_data[src_offset];
}
}
std::vector<paddle::Tensor> RebuildPaddingCPU(
const paddle::Tensor &tmp_out,
const paddle::Tensor &cum_offsets,
const paddle::Tensor &cu_seqlens_q,
const paddle::Tensor &seq_len_this_time,
const paddle::Tensor &seq_lens_decoder,
const paddle::Tensor &seq_lens_encoder,
const paddle::optional<paddle::Tensor> &output_padding_offset,
int max_input_length) {
auto tmp_out_cpu = tmp_out.copy_to(paddle::CPUPlace(), true);
auto cum_offsets_cpu = cum_offsets.copy_to(paddle::CPUPlace(), true);
auto cu_seqlens_q_cpu = cu_seqlens_q.copy_to(paddle::CPUPlace(), true);
auto seq_len_this_time_cpu =
seq_len_this_time.copy_to(paddle::CPUPlace(), true);
auto seq_lens_decoder_cpu =
@@ -107,7 +111,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
int token_num = tmp_out_cpu.shape()[0];
int dim_embed = tmp_out_cpu.shape()[1];
int bsz = cum_offsets_cpu.shape()[0];
int bsz = cu_seqlens_q_cpu.shape()[0] - 1;
paddle::Tensor out;
if (output_padding_offset_cpu) {
@@ -128,7 +132,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
{bsz, dim_embed}, 0, tmp_out_cpu.dtype(), paddle::CPUPlace());
}
const int *cum_offsets_data = cum_offsets_cpu.data<int>();
const int *cu_seqlens_q_data = cu_seqlens_q_cpu.data<int>();
const int *seq_len_this_time_data = seq_len_this_time_cpu.data<int>();
const int *seq_lens_decoder_data = seq_lens_decoder_cpu.data<int>();
const int *seq_lens_encoder_data = seq_lens_encoder_cpu.data<int>();
@@ -141,7 +145,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
case paddle::DataType::FLOAT32:
RebuildAppendPaddingCPUImpl<float>(out.data<float>(),
tmp_out_cpu.data<float>(),
cum_offsets_data,
cu_seqlens_q_data,
seq_len_this_time_data,
seq_lens_decoder_data,
seq_lens_encoder_data,
@@ -154,7 +158,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
RebuildAppendPaddingCPUImpl<paddle::float16>(
out.data<paddle::float16>(),
tmp_out_cpu.data<paddle::float16>(),
cum_offsets_data,
cu_seqlens_q_data,
seq_len_this_time_data,
seq_lens_decoder_data,
seq_lens_encoder_data,
@@ -167,7 +171,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
RebuildAppendPaddingCPUImpl<paddle::bfloat16>(
out.data<paddle::bfloat16>(),
tmp_out_cpu.data<paddle::bfloat16>(),
cum_offsets_data,
cu_seqlens_q_data,
seq_len_this_time_data,
seq_lens_decoder_data,
seq_lens_encoder_data,
@@ -186,7 +190,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
case paddle::DataType::FLOAT32:
RebuildPaddingCPUImpl<float>(out.data<float>(),
tmp_out_cpu.data<float>(),
cum_offsets_data,
cu_seqlens_q_data,
seq_len_this_time_data,
seq_lens_decoder_data,
seq_lens_encoder_data,
@@ -198,7 +202,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
RebuildPaddingCPUImpl<paddle::float16>(
out.data<paddle::float16>(),
tmp_out_cpu.data<paddle::float16>(),
cum_offsets_data,
cu_seqlens_q_data,
seq_len_this_time_data,
seq_lens_decoder_data,
seq_lens_encoder_data,
@@ -207,11 +211,10 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
elem_nums);
break;
case paddle::DataType::BFLOAT16:
RebuildPaddingCPUImpl<paddle::bfloat16>(
out.data<paddle::bfloat16>(),
tmp_out_cpu.data<paddle::bfloat16>(),
cum_offsets_data,
cu_seqlens_q_data,
seq_len_this_time_data,
seq_lens_decoder_data,
seq_lens_encoder_data,
@@ -230,7 +233,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
std::vector<std::vector<int64_t>> RebuildPaddingInferShape(
const std::vector<int64_t> &tmp_out_shape,
const std::vector<int64_t> &cum_offsets_shape,
const std::vector<int64_t> &cu_seqlens_q_shape,
const std::vector<int64_t> &seq_len_this_time_shape,
const std::vector<int64_t> &seq_lens_decoder_shape,
const std::vector<int64_t> &seq_lens_encoder_shape,
@@ -239,14 +242,14 @@ std::vector<std::vector<int64_t>> RebuildPaddingInferShape(
if (output_padding_offset_shape) {
return {{-1, dim_embed}};
} else {
int64_t bsz = cum_offsets_shape[0];
int64_t bsz = cu_seqlens_q_shape[0] - 1;
return {{bsz, dim_embed}};
}
}
std::vector<paddle::DataType> RebuildPaddingInferDtype(
const paddle::DataType &tmp_out_dtype,
const paddle::DataType &cum_offsets_dtype,
const paddle::DataType &cu_seqlens_q_dtype,
const paddle::DataType &seq_len_this_time_dtype,
const paddle::DataType &seq_lens_decoder_dtype,
const paddle::DataType &seq_lens_encoder_dtype,
@@ -256,7 +259,7 @@ std::vector<paddle::DataType> RebuildPaddingInferDtype(
PD_BUILD_STATIC_OP(rebuild_padding_cpu)
.Inputs({"tmp_out",
"cum_offsets",
"cu_seqlens_q",
"seq_len_this_time",
"seq_lens_decoder",
"seq_lens_encoder",