Skip to content

Optimize kernel launch #832

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
Vinkle-hzt merged 4 commits into from
Apr 23, 2026
Merged

Optimize kernel launch #832

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
5707312
refactor: implement fused copy operations and enhance tensor handling…
Vinkle-hzt Apr 16, 2026
64e797c
test: add fused copy kernel tests
Vinkle-hzt Apr 16, 2026
edfaea1
chore: use pin mem for cp input
Vinkle-hzt Apr 17, 2026
f5e96ba
fix(fuse_copy): raise cap to 64 for micro-batch accumulation
Vinkle-hzt Apr 23, 2026
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
1 change: 1 addition & 0 deletions rtp_llm/cpp/core/BUILD
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -184,6 +184,7 @@ cc_library(
":event",
"//rtp_llm/cpp/config:config_modules",
"//rtp_llm/cpp/models:stats",
"//rtp_llm/models_py/bindings/common/kernels:fuse_copy_util",
] + torch_deps() + select({
"@//:using_rocm": ["@local_config_rocm//rocm:rocm_headers"],
"//conditions:default": [],
Expand Down
4 changes: 4 additions & 0 deletions rtp_llm/cpp/core/ExecOps.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
#include "rtp_llm/cpp/core/Event.h"
#include "rtp_llm/cpp/config/ConfigModules.h"
#include "rtp_llm/cpp/models/eplb/stats/ExpertStats.h"
#include "rtp_llm/models_py/bindings/common/kernels/fuse_copy_util.h"

#include <memory>
#include <atomic>
Expand Down Expand Up @@ -92,6 +93,9 @@ void execNoBlockCopy(const CopyParams& params);
void execBatchCopy(const BatchCopyParams& params);
void execMultiMergeCopy(const MultiMergeCopyParams& params);

void fusedCopy(const FusedD2DCopyParams& params);
void fusedStridedCopy(const FusedStridedCopyParams& params);

// ===================================================================
// Sample ops
// ===================================================================
Expand Down
6 changes: 5 additions & 1 deletion rtp_llm/cpp/cuda/ops/CudaFlashInfer.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,7 +76,10 @@ FlashInferAttnParams::allocateManyBuffer(const std::vector<std::vector<int64_t>>
auto buf_options = torch::TensorOptions(torch::kInt32);
if (atype == AllocationType::DEVICE) {
buf_options = buf_options.device(torch::kCUDA);
} else {
buf_options = buf_options.device(torch::kCPU).pinned_memory(true);
}

auto buf = torch::empty({(int64_t)total_size}, buf_options);

size_t offset = 0;
Expand Down Expand Up @@ -104,7 +107,8 @@ FlashInferAttnParams* FlashInferAttnParams::create(int batch_size, int input_tok
params->float_workspace_ =
torch::empty({128 * 1024 * 1024}, torch::TensorOptions(torch::kInt8).device(torch::kCUDA));
params->int_workspace_ = torch::empty({8 * 1024 * 1024}, torch::TensorOptions(torch::kInt8).device(torch::kCUDA));
params->int_host_workspace_ = torch::empty({8 * 1024 * 1024}, torch::kInt8);
params->int_host_workspace_ =
torch::empty({8 * 1024 * 1024}, torch::TensorOptions(torch::kInt8).device(torch::kCPU).pinned_memory(true));

params->float_workspace_d = params->float_workspace_;
params->int_workspace_d = params->int_workspace_;
Expand Down
3 changes: 3 additions & 0 deletions rtp_llm/cpp/cuda_graph/BUILD
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,9 +66,12 @@ cc_library(
deps = torch_deps() + [
":cuda_graph_base",
":cuda_graph_hdrs_lib",
"//rtp_llm/cpp/core:exec_ops_hdr",
"//rtp_llm/cpp/utils:core_utils",
"//rtp_llm/cpp/utils:profiling_scope",
"//rtp_llm/models_py/bindings:op_defs",
"//rtp_llm/models_py/bindings/common/kernels:fuse_copy_util",
"//rtp_llm/models_py/bindings/common:fuse_copy_op",
] + select({
"//:using_cuda": [
"//rtp_llm/cpp/cuda:cuda",
Expand Down
29 changes: 17 additions & 12 deletions rtp_llm/cpp/cuda_graph/cuda_graph_prefill.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,9 +23,10 @@ void CudaGraphRunner::capturePrefill() {
inputs.attention_inputs.prefix_lengths.fill_(0);
// Must set cu_seqlens/cu_kv_seqlens/input_lengths to match actual seq_len,
// otherwise FlashInfer plans for max_seq_len tokens but q/k/v only have seq_len tokens
inputs.attention_inputs.cu_seqlens.data_ptr<int>()[0] = 0;
inputs.attention_inputs.cu_seqlens.data_ptr<int>()[1] = seq_len;
inputs.attention_inputs.input_lengths.data_ptr<int>()[0] = seq_len;
inputs.attention_inputs.cu_seqlens_host[0] = 0;
inputs.attention_inputs.cu_seqlens_host[1] = seq_len;
inputs.attention_inputs.cu_seqlens.copy_(inputs.attention_inputs.cu_seqlens_host, false);
inputs.attention_inputs.input_lengths[0] = seq_len;
} else {
// Draft model prefill: distribute seq_len tokens across batches (max num_tokens_per_bs_ each).
// All max_bs_ batches get prefix to ensure buffer allocation covers worst-case replay.
Expand All @@ -36,22 +37,26 @@ void CudaGraphRunner::capturePrefill() {
// Active batches get real input tokens, inactive batches get 0 input tokens.
inputs.attention_inputs.input_lengths.fill_(0);
inputs.attention_inputs.prefix_lengths.fill_(prefix_len);
auto* input_lengths_ptr = inputs.attention_inputs.input_lengths.data_ptr<int>();
auto& input_lengths = inputs.attention_inputs.input_lengths;
for (int b = 0; b < active_bs; b++) {
int tokens = (b < active_bs - 1) ? num_tokens_per_bs_ : (seq_len - b * num_tokens_per_bs_);
input_lengths_ptr[b] = tokens;
input_lengths[b] = tokens;
}

// Build cu_seqlens and cu_kv_seqlens as cumulative sums
auto* prefix_lengths_ptr = inputs.attention_inputs.prefix_lengths.data_ptr<int>();
auto* cu_seqlens_ptr = inputs.attention_inputs.cu_seqlens.data_ptr<int>();
auto* cu_kv_seqlens_ptr = inputs.attention_inputs.cu_kv_seqlens.data_ptr<int>();
cu_seqlens_ptr[0] = 0;
cu_kv_seqlens_ptr[0] = 0;
auto cu_seqlens_host = inputs.attention_inputs.cu_seqlens_host;
auto cu_kv_seqlens_host = inputs.attention_inputs.cu_kv_seqlens.cpu();
auto prefix_lengths = inputs.attention_inputs.prefix_lengths;

cu_seqlens_host[0] = 0;
cu_kv_seqlens_host[0] = 0;
for (int b = 0; b < max_bs_; b++) {
cu_seqlens_ptr[b + 1] = cu_seqlens_ptr[b] + input_lengths_ptr[b];
cu_kv_seqlens_ptr[b + 1] = cu_kv_seqlens_ptr[b] + input_lengths_ptr[b] + prefix_lengths_ptr[b];
cu_seqlens_host[b + 1] = cu_seqlens_host[b].item<int>() + input_lengths[b].item<int>();
cu_kv_seqlens_host[b + 1] = cu_kv_seqlens_host[b].item<int>() + input_lengths[b].item<int>() + prefix_lengths[b].item<int>();
}

inputs.attention_inputs.cu_seqlens.copy_(cu_seqlens_host);
inputs.attention_inputs.cu_kv_seqlens.copy_(cu_kv_seqlens_host);
}

inputs.attention_inputs.context_total_kv_length = seq_len;
Expand Down
Loading
Loading