Skip to content

llama: fix ASAN error with M-RoPE #16848

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
merged 1 commit into from
Oct 29, 2025
Merged

llama: fix ASAN error with M-RoPE #16848

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
13 changes: 8 additions & 5 deletions src/llama-batch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -669,10 +669,8 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u

auto udata = std::make_shared<llama_ubatch::data_t>();

const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1;

const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0;
const int64_t n_pos_all = (int64_t) n_tokens*n_pos_cur;
const int64_t n_pos_all = (int64_t) n_tokens*n_pos_per_embd;

udata->token .resize(n_tokens);
udata->embd .resize(n_embd_all);
Expand All @@ -694,8 +692,13 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
memcpy(udata->embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
}

for (int j = 0; j < n_pos_cur; ++j) {
udata->pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
for (size_t j = 0; j < (size_t)n_pos_per_embd; ++j) {
// if we are using M-RoPE
// if the current batch is text, we need to broadcast the same position across all RoPE sections
Copy link
Collaborator Author

@ngxson ngxson Oct 29, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

please note that even though only 3 dims are used in QwenVL (while n_pos_per_embd = 4), the last dim is never actually processed by ggml, as its hparams.mrope_section will be 0. Therefore, it's safe to broadcast the all dims here.

// otherwise, the input batch is image embeddings, we copy the positions as-is
// if we are not using M-RoPE, there is only one position per token (this loop runs only once)
size_t src_off = batch.token ? 0 : j*batch.n_tokens;
udata->pos[j*n_tokens + i] = batch.pos[src_off + idxs[i]];
}

udata->n_seq_id[i] = batch.n_seq_id[idxs[i]];
Expand Down
Loading