Skip to content

CUDA: check for buffer overlap before fusing #21566

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
am17an merged 2 commits into from
Apr 7, 2026
+71 −67
Merged

CUDA: check for buffer overlap before fusing #21566

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fc889c1
CUDA: check for buffer overlap before fusing
am17an Apr 7, 2026
04e27cf
use ggml_cuda_check_fusion_memory_ranges
am17an Apr 7, 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
25 changes: 25 additions & 0 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2164,6 +2164,31 @@ static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
return false;
}

// Check that the fused output does not overlap with source tensors
auto overlaps = [](const ggml_tensor * a, const ggml_tensor * b) -> bool {
if (!a->data || !b->data) {
return false;
}
const uintptr_t a_start = (uintptr_t) a->data;
const uintptr_t a_end = a_start + ggml_nbytes(a);
const uintptr_t b_start = (uintptr_t) b->data;
const uintptr_t b_end = b_start + ggml_nbytes(b);
Copy link
Copy Markdown
Member

@ggerganov ggerganov Apr 7, 2026

Choose a reason for hiding this comment

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

For future-proof-ness it's better to use ggml_backend_buft_get_alloc_size() instead of ggml_nbytes(). Same in ggml_cuda_check_fusion_memory_ranges().

return a_start < b_end && b_start < a_end;
};

if (overlaps(glu, ffn_up->src[0]) || overlaps(glu, ffn_up->src[1]) || overlaps(glu, ffn_gate->src[0])) {
return false;
}

if (has_bias) {
const ggml_tensor * up_bias_data = ffn_up_bias->src[0] == ffn_up ? ffn_up_bias->src[1] : ffn_up_bias->src[0];
const ggml_tensor * gate_bias_data =
ffn_gate_bias->src[0] == ffn_gate ? ffn_gate_bias->src[1] : ffn_gate_bias->src[0];
if (overlaps(glu, up_bias_data) || overlaps(glu, gate_bias_data)) {
return false;
}
}

return true;
}

Expand Down
Loading