Skip to content

Load all MoE experts during warmup and make warmup 1 token #198

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
Feb 10, 2025
Merged

Load all MoE experts during warmup and make warmup 1 token #198

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ca4e8e5
Load all MoE experts during warmup
saood06 Feb 9, 2025
3702743
Unify warmup to one token
saood06 Feb 9, 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
6 changes: 4 additions & 2 deletions common/common.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2169,8 +2169,10 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
if (bos != -1) {
tmp.push_back(bos);
}
tmp.push_back(eos);

else
{
tmp.push_back(eos);
}
if (llama_model_has_encoder(model)) {
llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0));
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
Expand Down
2 changes: 1 addition & 1 deletion examples/llama-bench/llama-bench.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1586,7 +1586,7 @@ int main(int argc, char ** argv) {
if (params.warmup) {
if (t.n_prompt > 0) {
//test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
test_prompt(ctx, 1, 0, t.n_batch, t.n_threads);
}
if (t.n_gen > 0) {
test_gen(ctx, 1, 0, t.n_threads);
Expand Down
19 changes: 12 additions & 7 deletions src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3784,7 +3784,7 @@ static size_t llama_model_max_nodes(const llama_model & /*model*/) {
// return 32768;
//}

return 8192;
return 65536;
}

struct llama_model_loader {
Expand Down Expand Up @@ -8879,7 +8879,8 @@ struct llm_build_context {
llama_context & lctx,
const llama_batch & batch,
const llm_build_cb & cb,
bool worst_case) :
bool worst_case,
bool warmup) :
model (lctx.model),
lctx (lctx),
hparams (model.hparams),
Expand All @@ -8897,7 +8898,7 @@ struct llm_build_context {
n_embd_head_v (hparams.n_embd_head_v),
n_embd_v_gqa (hparams.n_embd_v_gqa()),
n_expert (hparams.n_expert),
n_expert_used (hparams.n_expert_used),
n_expert_used (warmup ? hparams.n_expert : hparams.n_expert_used),
freq_base (cparams.rope_freq_base),
freq_scale (cparams.rope_freq_scale),
ext_factor (cparams.yarn_ext_factor),
Expand Down Expand Up @@ -14433,7 +14434,7 @@ static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const

llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };

struct llm_build_context llm(lctx, dummy, cb, false);
struct llm_build_context llm(lctx, dummy, cb, false, false);

llm.init();

Expand All @@ -14450,7 +14451,7 @@ static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) {

llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };

struct llm_build_context llm(lctx, dummy, cb, false);
struct llm_build_context llm(lctx, dummy, cb, false, false);

llm.init();

Expand All @@ -14467,7 +14468,7 @@ static struct ggml_cgraph * llama_build_graph_s_copy(llama_context & lctx) {

llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };

struct llm_build_context llm(lctx, dummy, cb, false);
struct llm_build_context llm(lctx, dummy, cb, false, false);

llm.init();

Expand Down Expand Up @@ -14517,7 +14518,11 @@ static struct ggml_cgraph * llama_build_graph(

struct ggml_cgraph * result = NULL;

struct llm_build_context llm(lctx, batch, cb, worst_case);
const llama_vocab * vocab = llama_get_vocab(&lctx);
llama_token bos = llama_token_bos_impl(*vocab);
llama_token eos = llama_token_eos_impl(*vocab);
bool is_warming_up = (batch.n_tokens == 1 && batch.token[0] == bos);
struct llm_build_context llm(lctx, batch, cb, worst_case, is_warming_up);

llm.init();

Expand Down