Skip to content

mtmd: use RAII for setting and resetting non-causal attention #25723

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
ngxson merged 3 commits into from
Jul 22, 2026
Merged

mtmd: use RAII for setting and resetting non-causal attention #25723

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ae3a430
mtmd: use RAII for setting and resetting non-causal attention
nikwen Jul 15, 2026
0a363e5
mtmd: drop dependency on <optional>
nikwen Jul 16, 2026
2ee8ead
mtmd: shorten class and variable names
nikwen Jul 21, 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
37 changes: 24 additions & 13 deletions tools/mtmd/mtmd-helper.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -238,6 +238,29 @@ struct decode_embd_batch {
}
};

// Helper class to set non-causal attention via RAII
class scope_non_causal {
public:
scope_non_causal(llama_context * context, bool enabled) : context_(context), enabled_(enabled) {
if (enabled_) {
// TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
llama_set_causal_attn(context_, false);
}
}
~scope_non_causal() {
if (enabled_) {
llama_set_causal_attn(context_, true);
}
}

scope_non_causal(const scope_non_causal &) = delete;
scope_non_causal & operator=(const scope_non_causal &) = delete;

private:
llama_context * context_;
bool enabled_;
};

// Helper function for decoding an image whose embeddings have already been calculated
int32_t mtmd_helper_decode_image_chunk(
mtmd_context * ctx,
Expand Down Expand Up @@ -288,10 +311,7 @@ int32_t mtmd_helper_decode_image_chunk(
}

const bool use_non_causal = mtmd_decode_use_non_causal(ctx, chunk);
if (use_non_causal) {
llama_set_causal_attn(lctx, false);
// TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
}
const scope_non_causal non_causal(lctx, use_non_causal);

while (i_batch < n_img_batches) { // split into batches
int pos_offset = i_batch*n_batch;
Expand All @@ -304,19 +324,13 @@ int32_t mtmd_helper_decode_image_chunk(
int32_t ret = llama_decode(lctx, batch_embd_view);
if (ret != 0) {
LOG_ERR("failed to decode %s\n", name);
if (use_non_causal) {
llama_set_causal_attn(lctx, true);
}
return ret;
}

if (callback != nullptr) {
ret = callback(batch_embd_view, user_data);
if (ret != 0) {
LOG_ERR("post-decode callback failed\n");
if (use_non_causal) {
llama_set_causal_attn(lctx, true);
}
return ret;
}
}
Expand All @@ -329,9 +343,6 @@ int32_t mtmd_helper_decode_image_chunk(
n_past += mtmd_input_chunk_get_n_pos(chunk);
*new_n_past = n_past;

if (use_non_causal) {
llama_set_causal_attn(lctx, true);
}
return 0;
}

Expand Down
Loading