Skip to content

Fix bench_qwen_gqa_prefill comparing every mode against itself #87

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

Closed
lvyufeng wants to merge 1 commit into from
Closed

Fix bench_qwen_gqa_prefill comparing every mode against itself #87

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
8 changes: 4 additions & 4 deletions cpp_engine/tests/bench_qwen_gqa_prefill.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,11 +55,11 @@ double time_case(int rows, int position_offset, int max_context, int iters,
// baseline explicit so this benchmark cannot compare hpg6 against itself.
setenv("DSV4_QWEN_GQA_LONG_TILE", mode == Mode::Base ? "0" : "1", 1);
unsetenv("DSV4_QWEN_GQA_FLASH_TILE");
unsetenv("DSV4_QWEN_GQA_MMA_TILE");
if (mode == Mode::Flash) setenv("DSV4_QWEN_GQA_FLASH_TILE", "1", 1);
// The MMA dispatch is checked before flash and long, so this one variable
// selects it regardless of the others.
if (mode == Mode::Mma) setenv("DSV4_QWEN_GQA_MMA_TILE", "1", 1);
// MMA is the production default and is dispatched before flash and long, so
// unsetting its variable would silently run MMA for every mode and make all
// four rows identical. It has to be switched off explicitly instead.
setenv("DSV4_QWEN_GQA_MMA_TILE", mode == Mode::Mma ? "1" : "0", 1);
std::vector<uint16_t> host_q(q_elements);
std::vector<uint16_t> host_kv(kv_elements, 0);
for (uint16_t& value : host_q) {
Expand Down