Skip to content

Avoid unused quantized SDPA scratch allocation #22417

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

Draft
JacobSzwejbka wants to merge 1 commit into
base: perf/cpu-ring-attention-mask-skipping
Choose a base branch
from
Draft

Avoid unused quantized SDPA scratch allocation #22417

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
37 changes: 20 additions & 17 deletions extension/llm/custom_ops/op_sdpa_impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -846,22 +846,24 @@ void cpu_flash_attention(
buf_reduced = scratch_reduced.get();
}
}
int64_t size_per_thread_qdq_vec = kvSplitSize * headSize;
// Lets align size_per_thread_qdq_vec to 64 bytes, for coalesced cache reads,
// by padding with right number of per thread elements
int64_t size_per_thread_qdq_bytes = size_per_thread_qdq_vec * sizeof(accum_t);
int64_t size_qdq_bytes = size_per_thread_qdq_bytes * num_thread;
int64_t size_per_thread_qdq_vec = 0;
std::unique_ptr<char[]> allocated_buf_for_qdq;
accum_t* scratch_for_quant_dequant;
Result<void*> scratch_for_quant_dequant_res =
ctx.allocate_temp(size_qdq_bytes, 64);
if (!scratch_for_quant_dequant_res.ok()) {
allocated_buf_for_qdq = std::make_unique<char[]>(size_qdq_bytes);
scratch_for_quant_dequant =
reinterpret_cast<accum_t*>(allocated_buf_for_qdq.get());
} else {
scratch_for_quant_dequant =
reinterpret_cast<accum_t*>(scratch_for_quant_dequant_res.get());
accum_t* scratch_for_quant_dequant = nullptr;
if (is_quantized_sdpa) {
size_per_thread_qdq_vec = kvSplitSize * headSize;
const int64_t size_per_thread_qdq_bytes =
size_per_thread_qdq_vec * sizeof(accum_t);
const int64_t size_qdq_bytes = size_per_thread_qdq_bytes * num_thread;
Result<void*> scratch_for_quant_dequant_res =
ctx.allocate_temp(size_qdq_bytes, 64);
if (!scratch_for_quant_dequant_res.ok()) {
allocated_buf_for_qdq = std::make_unique<char[]>(size_qdq_bytes);
scratch_for_quant_dequant =
reinterpret_cast<accum_t*>(allocated_buf_for_qdq.get());
} else {
scratch_for_quant_dequant =
reinterpret_cast<accum_t*>(scratch_for_quant_dequant_res.get());
}
}

// Data ptrs
Expand Down Expand Up @@ -1067,8 +1069,9 @@ void cpu_flash_attention(
scalar_t* qk_reduced_data = is_reduced_type
? buf_reduced_data + ompIdx * qSplitSize * kvSplitSize
: nullptr;
accum_t* buf_qdq_ptr =
scratch_for_quant_dequant + ompIdx * size_per_thread_qdq_vec;
accum_t* buf_qdq_ptr = is_quantized_sdpa
? scratch_for_quant_dequant + ompIdx * size_per_thread_qdq_vec
: nullptr;

for (int64_t z = begin; z < end; z++) {
int64_t m = k * qSplitSize;
Expand Down
Loading