Skip to content

model : more uniform output id handling #14275

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 3 commits into from
Jun 20, 2025
Merged

model : more uniform output id handling #14275

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b8a2d10
model : more uniform output id handling
ggerganov Jun 19, 2025
6f02448
cont : revert n_outputs < n_tokens optimization
ggerganov Jun 19, 2025
2b940c0
cont : fix out_ids initialization
ggerganov Jun 19, 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
54 changes: 27 additions & 27 deletions src/llama-graph.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,36 +92,28 @@ void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) {
}

void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) {
if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
//GGML_ASSERT(out_ids && "every model that can must skip unused outputs");
GGML_ASSERT(out_ids);

if (!out_ids) {
LLAMA_LOG_WARN("%s: 'out_ids' is not created\n", __func__);
} else {
const int64_t n_tokens = ubatch->n_tokens;
const int64_t n_tokens = ubatch->n_tokens;

GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer));
int32_t * data = (int32_t *) out_ids->data;
GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer));
int32_t * data = (int32_t *) out_ids->data;

if (n_outputs == n_tokens) {
for (int i = 0; i < n_tokens; ++i) {
data[i] = i;
}
} else if (ubatch->output) {
int32_t n_outputs = 0;
for (int i = 0; i < n_tokens; ++i) {
if (ubatch->output[i]) {
data[n_outputs++] = i;
}
}
// the graph needs to have been passed the correct number of outputs
GGML_ASSERT(n_outputs == n_outputs);
} else if (n_outputs == 1) {
// only keep last output
data[0] = n_tokens - 1;
} else {
GGML_ASSERT(n_outputs == 0);
}
if (n_outputs == n_tokens) {
for (int i = 0; i < n_tokens; ++i) {
data[i] = i;
}

return;
}

GGML_ASSERT(ubatch->output);

int n_outputs = 0;

for (int i = 0; i < n_tokens; ++i) {
if (ubatch->output[i]) {
data[n_outputs++] = i;
}
}
}
Expand Down Expand Up @@ -874,6 +866,14 @@ ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
}

ggml_tensor * llm_graph_context::build_inp_out_ids() const {
// note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls,
// but this would make the graph topology depend on the number of output tokens, which can interere with
// features that require constant topology such as pipline parallelism
// ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471
//if (n_outputs < n_tokens) {
// return nullptr;
//}

auto inp = std::make_unique<llm_graph_input_out_ids>(hparams, cparams, n_outputs);

auto & cur = inp->out_ids;
Expand Down
Loading
Loading