Skip to content

[Optimization] Skip memcpy(DtoH) capture in get_block_shape_and_split_kv_block #4988

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Nov 13, 2025
Merged

[Optimization] Skip memcpy(DtoH) capture in get_block_shape_and_split_kv_block #4988

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8285bf5
skip DtoH capture
Sunny-bot1 Nov 12, 2025
08945bf
Merge branch 'develop' into dtoh_skip
EmmonsCurse Nov 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 13 additions & 6 deletions custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@
#ifndef PADDLE_WITH_CUSTOM_DEVICE_METAX_GPU
#include "paddle/phi/core/memory/memcpy.h"
#endif
#include "paddle/phi/backends/gpu/cuda/cuda_graph_with_memory_pool.h"
#include "utils.cuh"

template <int THREADBLOCK_SIZE>
Expand Down Expand Up @@ -287,9 +288,12 @@ void GetBlockShapeAndSplitKVBlock(
seq_lens_encoder.data<int>(),
max_len_tensor_gpu.data<int>(),
bsz);

max_len_tensor_cpu.copy_(
max_len_tensor_gpu, max_len_tensor_cpu.place(), false);
// Note (sunxin): Skip capturing the DtoH copy (it's time-consuming); CPU data
// is only for branching in attention.
if (!phi::backends::gpu::IsCUDAGraphCapturing()) {
max_len_tensor_cpu.copy_(
max_len_tensor_gpu, max_len_tensor_cpu.place(), false);
}
Comment on lines +291 to +296
Copy link
Collaborator

@gongshaotian gongshaotian Nov 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这个工作会生效吗

Copy link
Collaborator Author

@Sunny-bot1 Sunny-bot1 Nov 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这个工作会生效吗

before:

image

after:

image


auto max_len_cpu_ptr = max_len_tensor_cpu.data<int>();
int max_len_this_time = max_len_cpu_ptr[0];
Expand Down Expand Up @@ -398,9 +402,12 @@ void GetBlockShapeAndSplitKVBlock(
bsz,
decoder_block_shape_q,
group_size);

decoder_num_blocks_cpu.copy_(
decoder_num_blocks_device, decoder_num_blocks_cpu.place(), false);
// Note (sunxin): Skip capturing the DtoH copy (it's time-consuming); CPU
// data is only for branching in attention.
if (!phi::backends::gpu::IsCUDAGraphCapturing()) {
decoder_num_blocks_cpu.copy_(
decoder_num_blocks_device, decoder_num_blocks_cpu.place(), false);
}
PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(
decoder_chunk_size_device.data<int>(), 64, sizeof(int32_t), stream));
}
Expand Down
Loading