Skip to content

llama : (mrope) allow using normal 1D position for text token #13138

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 2 commits into from
Apr 28, 2025
Merged

llama : (mrope) allow using normal 1D position for text token #13138

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bd310ff
llama : (mrope) use normal position for text token
ngxson Apr 27, 2025
9cd16a3
rm n_pos_per_embd from llm_graph_input_attn_temp
ngxson Apr 28, 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
8 changes: 0 additions & 8 deletions examples/llava/qwen2vl-cli.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,20 +92,12 @@ static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct lla

static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past, int * st_pos_id) {
int N = (int) tokens.size();
std::vector<llama_pos> pos;
for (int i = 0; i < N; i += n_batch) {
int n_eval = (int) tokens.size() - i;
if (n_eval > n_batch) {
n_eval = n_batch;
}
auto batch = llama_batch_get_one(&tokens[i], n_eval);
// TODO: add mrope pos ids somewhere else
pos.resize(batch.n_tokens * 4);
std::fill(pos.begin(), pos.end(), 0);
for (int j = 0; j < batch.n_tokens * 3; j ++) {
pos[j] = *st_pos_id + (j % batch.n_tokens);
}
batch.pos = pos.data();

if (llama_decode(ctx_llama, batch)) {
LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
Expand Down
26 changes: 19 additions & 7 deletions src/llama-graph.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,7 +55,18 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
if (ubatch->pos && pos) {
const int64_t n_tokens = ubatch->n_tokens;

ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_token*ggml_element_size(pos));
if (ubatch->token && n_pos_per_embd > 1) {
// in case we're using M-RoPE with text tokens, convert the 1D positions to 4D
// the other dimensions are all 0, they are unused for text tokens
std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd, 0);
// copy the first dimension
for (int i = 0; i < n_tokens; ++i) {
pos_data[i] = ubatch->pos[i];
}
ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
} else {
ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_embd*ggml_element_size(pos));
}
}
}

Expand All @@ -71,7 +82,7 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
) * f_attn_temp_scale + 1.0;
}

ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*n_pos_per_token*ggml_element_size(attn_scale));
ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*n_pos_per_embd*ggml_element_size(attn_scale));
}
}

Expand Down Expand Up @@ -592,7 +603,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
res (std::make_unique<llm_graph_result>()) {
}

int64_t llm_graph_context::n_pos_per_token() const {
int64_t llm_graph_context::n_pos_per_embd() const {
return arch == LLM_ARCH_QWEN2VL ? 4 : 1;
}

Expand Down Expand Up @@ -1018,11 +1029,11 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
}

ggml_tensor * llm_graph_context::build_inp_pos() const {
auto inp = std::make_unique<llm_graph_input_pos>(n_pos_per_token());
auto inp = std::make_unique<llm_graph_input_pos>(n_pos_per_embd());

auto & cur = inp->pos;

cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_token());
cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_embd());
ggml_set_input(cur);

res->add_input(std::move(inp));
Expand All @@ -1031,11 +1042,12 @@ ggml_tensor * llm_graph_context::build_inp_pos() const {
}

ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
auto inp = std::make_unique<llm_graph_input_attn_temp>(n_pos_per_token(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
auto inp = std::make_unique<llm_graph_input_attn_temp>(n_pos_per_embd(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
Copy link
Collaborator Author

@ngxson ngxson Apr 27, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

@ggerganov Because build_inp_attn_scale is currently used exclusively by llama 4, do you think we should get rid of n_pos_per_embd and replace it with a GGML_ASSERT(n_pos_per_embd() == 1) ?

The main motivation is to make this code looks less complicated, as there is ~0% chance Qwen model gonna use this

Copy link
Member

Choose a reason for hiding this comment

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

Yes, we can do that.

Copy link
Collaborator Author

Choose a reason for hiding this comment

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

On second thought, build_inp_attn_scale should work well even in the case of N pos per token.

That's because the scale is applied per embedding, and the number of embedding is independent from N pos per token.

In any cases, I removed the n_pos_per_embd in 9cd16a3 , merging this PR once the CI is green


auto & cur = inp->attn_scale;

cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens*n_pos_per_token());
// this need to be 1x1xN for broadcasting
cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens*n_pos_per_embd());
ggml_set_input(cur);

res->add_input(std::move(inp));
Expand Down
12 changes: 6 additions & 6 deletions src/llama-graph.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,28 +90,28 @@ class llm_graph_input_embd : public llm_graph_input_i {

class llm_graph_input_pos : public llm_graph_input_i {
public:
llm_graph_input_pos(int64_t n_pos_per_token) : n_pos_per_token(n_pos_per_token) {}
llm_graph_input_pos(int64_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
virtual ~llm_graph_input_pos() = default;

void set_input(const llama_ubatch * ubatch) override;

ggml_tensor * pos = nullptr; // I32 [n_batch]

const int64_t n_pos_per_token = 1;
const int64_t n_pos_per_embd = 1;
};

// temperature tuning, used by llama4
class llm_graph_input_attn_temp : public llm_graph_input_i {
public:
llm_graph_input_attn_temp(int64_t n_pos_per_token, uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
: n_pos_per_token(n_pos_per_token), n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
llm_graph_input_attn_temp(int64_t n_pos_per_embd, uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
: n_pos_per_embd(n_pos_per_embd), n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
virtual ~llm_graph_input_attn_temp() = default;

void set_input(const llama_ubatch * ubatch) override;

ggml_tensor * attn_scale = nullptr; // F32 [n_batch]

const int64_t n_pos_per_token = 1;
const int64_t n_pos_per_embd = 1;

const uint32_t n_attn_temp_floor_scale;
const float f_attn_temp_scale;
Expand Down Expand Up @@ -419,7 +419,7 @@ struct llm_graph_context {

llm_graph_context(const llm_graph_params & params);

int64_t n_pos_per_token() const;
int64_t n_pos_per_embd() const;

void cb(ggml_tensor * cur, const char * name, int il) const;

Expand Down
Loading