Skip to content

Change invalid indices mask to -1 to save some cmp insts #109

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Change invalid indices mask to -1 to save some cmp insts #109

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6dd9a6f
Update test_flash_mla_prefill.py make OOB indices = -1
mingyangHao Oct 13, 2025
e77c306
Update fwd.cu
mingyangHao Oct 13, 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
2 changes: 1 addition & 1 deletion csrc/sm90/prefill/sparse/fwd.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -562,7 +562,7 @@ sparse_attn_fwd_kernel(__grid_constant__ const SparsePrefillParams params, __gri
int offs = (block_idx+buf_idx)*B_TOPK + local_row*NUM_GROUPS + group_idx;
int t = __ldg(gIndices + offs);
token_indices[buf_idx][local_row] = t*(int64_t)params.stride_kv_s_kv; // We mult it with params.stride_kv_s_kv here since it's faster
is_token_valid[buf_idx][local_row] = t >= 0 && t < params.s_kv;
is_token_valid[buf_idx][local_row] = t >= 0;
}
}
};
Expand Down
2 changes: 1 addition & 1 deletion tests/test_flash_mla_prefill.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,7 +50,7 @@ def generate_testcase(t: TestParam) -> Testcase:
cur_indices = torch.randperm(t.s_kv)[:t.topk]
cur_indices[near_mask] = torch.randint(max(0, t.s_kv - 20000), t.s_kv - 1, (near_mask.sum().item(),))
if len(cur_indices) < t.topk:
cur_indices = torch.cat([cur_indices, torch.full((t.topk - len(cur_indices),), 2147480000)])
cur_indices = torch.cat([cur_indices, torch.full((t.topk - len(cur_indices),), -1)])
cur_indices = cur_indices[torch.randperm(t.topk)]
indices[b, s, h] = cur_indices
indices = indices.to(q.device)
Expand Down