vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : prompt processing optimizations in DeepSeek V2

Browse source
This commit is contained in:
Stanisław Szymczyk 2025-01-28 19:26:54 +01:00
parent 8ff0991eed
commit 8a887decd3
1 changed files with 26 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

38
src/llama.cpp
View file

@ -6403,6 +6403,10 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads) // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
// whether to use n_tokens as the matrix dimension during multiplication or n_head
// n_tokens is higher during prompt processing, this allows to optimize for this case
bool pp_opt = n_tokens > n_head;
for (int il = 0; il < n_layer; ++il) { for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL; struct ggml_tensor * inpSA = inpL;
@ -6535,14 +6539,18 @@ struct llm_build_context {
struct ggml_tensor * q_nope2 = ggml_mul_mat(ctx0, wk_b, q_nope_perm); struct ggml_tensor * q_nope2 = ggml_mul_mat(ctx0, wk_b, q_nope_perm);
cb(q_nope2, "q_nope2", il); cb(q_nope2, "q_nope2", il);
struct ggml_tensor * q_nope2_perm = ggml_permute(ctx0, q_nope2, 0, 2, 1, 3); if (!pp_opt) {
cb(q_nope2_perm, "q_nope2_perm", il); q_nope2 = ggml_permute(ctx0, q_nope2, 0, 2, 1, 3);
cb(q_nope2, "q_nope2_perm", il);
}
struct ggml_tensor * kq_nope = ggml_mul_mat(ctx0, kv_cache, q_nope2_perm); struct ggml_tensor * kq_nope = ggml_mul_mat(ctx0, kv_cache, q_nope2);
cb(kq_nope, "kq_nope", il); cb(kq_nope, "kq_nope", il);
struct ggml_tensor * q_pe_perm = ggml_permute(ctx0, q_pe, 0, 3, 2, 1); if (pp_opt) {
cb(q_pe_perm, "q_pe_perm", il); q_pe = ggml_permute(ctx0, q_pe, 0, 2, 1, 3);
cb(q_pe, "q_pe_perm", il);
}
struct ggml_tensor * kq_pe = ggml_mul_mat(ctx0, kr_cache, q_pe); struct ggml_tensor * kq_pe = ggml_mul_mat(ctx0, kr_cache, q_pe);
cb(kq_pe, "kq_pe", il); cb(kq_pe, "kq_pe", il);
@ -6550,20 +6558,26 @@ struct llm_build_context {
struct ggml_tensor * kq = ggml_add(ctx0, kq_nope, kq_pe); struct ggml_tensor * kq = ggml_add(ctx0, kq_nope, kq_pe);
cb(kq, "kq", il); cb(kq, "kq", il);
kq = ggml_cont(ctx0, ggml_permute(ctx0, kq, 0, 2, 1, 3)); if (!pp_opt) {
cb(kq, "kq_perm", il); kq = ggml_cont(ctx0, ggml_permute(ctx0, kq, 0, 2, 1, 3));
cb(kq, "kq_perm", il);
}
kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, kq_scale, hparams.f_max_alibi_bias); kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, kq_scale, hparams.f_max_alibi_bias);
cb(kq, "kq_soft_max_ext", il); cb(kq, "kq_soft_max_ext", il);
struct ggml_tensor * kq_perm = ggml_permute(ctx0, kq, 0, 2, 1, 3); if (!pp_opt) {
cb(kq_perm, "kq_soft_max_ext_perm", il); kq = ggml_permute(ctx0, kq, 0, 2, 1, 3);
cb(kq, "kq_soft_max_ext_perm", il);
}
struct ggml_tensor * kqv_compressed = ggml_mul_mat(ctx0, kv_cache_trans, kq_perm); struct ggml_tensor * kqv_compressed = ggml_mul_mat(ctx0, kv_cache_trans, kq);
cb(kqv_compressed, "kqv_compressed", il); cb(kqv_compressed, "kqv_compressed", il);
kqv_compressed = ggml_permute(ctx0, kqv_compressed, 0, 2, 1, 3); if (!pp_opt) {
cb(kqv_compressed, "kqv_compressed_perm", il); kqv_compressed = ggml_permute(ctx0, kqv_compressed, 0, 2, 1, 3);
cb(kqv_compressed, "kqv_compressed_perm", il);
}
struct ggml_tensor * wv_b = ggml_view_3d(ctx0, model.layers[il].wv_b, kv_lora_rank, n_embd_head_v, n_head, ggml_row_size(model.layers[il].wv_b->type, kv_lora_rank), ggml_row_size(model.layers[il].wv_b->type, kv_lora_rank * n_embd_head_v), 0); struct ggml_tensor * wv_b = ggml_view_3d(ctx0, model.layers[il].wv_b, kv_lora_rank, n_embd_head_v, n_head, ggml_row_size(model.layers[il].wv_b->type, kv_lora_rank), ggml_row_size(model.layers[il].wv_b->type, kv_lora_rank * n_embd_head_v), 0);
cb(wv_b, "wv_b", il); cb(wv_b, "wv_b", il);