121 lines
4.7 KiB
C++
121 lines
4.7 KiB
C++
#include "../llama-build-context.h"
|
|
#include "../llama-model.h"
|
|
#include "../llama-context.h"
|
|
|
|
ggml_cgraph * llm_build_context::build_gemma2() {
|
|
ggml_cgraph * gf = new_graph_custom();
|
|
|
|
const int64_t n_embd_head_k = hparams.n_embd_head_k_full;
|
|
|
|
struct ggml_tensor * cur;
|
|
struct ggml_tensor * inpL;
|
|
|
|
inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
|
|
|
|
inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
|
|
cb(inpL, "inp_scaled", -1);
|
|
|
|
// inp_pos - contains the positions
|
|
struct ggml_tensor * inp_pos = build_inp_pos();
|
|
|
|
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
|
|
// gemma 2 requires different mask for layers using sliding window (SWA)
|
|
struct ggml_tensor * KQ_mask = build_inp_KQ_mask(true);
|
|
struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(true);
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
// (il % 2) layers use SWA
|
|
struct ggml_tensor * KQ_mask_l = (il % 2 == 0) ? KQ_mask_swa : KQ_mask;
|
|
|
|
// norm
|
|
cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il);
|
|
cb(cur, "attn_norm", il);
|
|
|
|
// self-attention
|
|
{
|
|
auto [Qcur, Kcur, Vcur] = llm_build_mul_mat_qkv(gf, cur, model.layers[il].wq, nullptr,
|
|
model.layers[il].wk, nullptr,
|
|
model.layers[il].wv, nullptr, 0, il);
|
|
Qcur = ggml_rope_ext(
|
|
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr,
|
|
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
ext_factor, attn_factor, beta_fast, beta_slow);
|
|
cb(Qcur, "Qcur", il);
|
|
|
|
// ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
|
|
switch (model.type) {
|
|
case e_model::MODEL_2B:
|
|
case e_model::MODEL_9B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); break;
|
|
case e_model::MODEL_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break;
|
|
default: GGML_ABORT("fatal error");
|
|
};
|
|
cb(Qcur, "Qcur_scaled", il);
|
|
|
|
Kcur = ggml_rope_ext(
|
|
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
|
|
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
ext_factor, attn_factor, beta_fast, beta_slow);
|
|
cb(Kcur, "Kcur", il);
|
|
|
|
cur = llm_build_kv(ctx0, lctx, kv_self, gf,
|
|
model.layers[il].wo, NULL,
|
|
Kcur, Vcur, Qcur, KQ_mask_l, n_tokens, kv_head, n_kv, 1.0f, cb, il, nullptr,
|
|
KQ_mask_l == KQ_mask_swa ? hparams.n_swa : 0);
|
|
}
|
|
|
|
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, cb, il);
|
|
cb(cur, "attn_post_norm", il);
|
|
|
|
if (il == n_layer - 1) {
|
|
// skip computing output for unused tokens
|
|
struct ggml_tensor * inp_out_ids = build_inp_out_ids();
|
|
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
|
inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
|
|
}
|
|
|
|
struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
|
|
cb(sa_out, "sa_out", il);
|
|
|
|
// feed-forward network
|
|
{
|
|
cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, sa_out,
|
|
model.layers[il].ffn_up, NULL, NULL,
|
|
model.layers[il].ffn_gate, NULL, NULL,
|
|
model.layers[il].ffn_down, NULL, NULL,
|
|
NULL,
|
|
LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
|
|
cb(cur, "ffn_out", il);
|
|
}
|
|
|
|
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, cb, -1);
|
|
cb(cur, "ffn_post_norm", -1);
|
|
|
|
cur = ggml_add(ctx0, cur, sa_out);
|
|
cur = lctx.cvec.apply_to(ctx0, cur, il);
|
|
cb(cur, "l_out", il);
|
|
|
|
// input for next layer
|
|
inpL = cur;
|
|
}
|
|
|
|
cur = inpL;
|
|
|
|
cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM_RMS, cb, -1);
|
|
cb(cur, "result_norm", -1);
|
|
|
|
// lm_head
|
|
cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
|
|
|
|
// final logit soft-capping
|
|
cur = ggml_softcap(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping, hparams.f_final_logit_softcapping);
|
|
//cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
|
|
//cur = ggml_tanh(ctx0, cur);
|
|
//cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
|
|
|
|
cb(cur, "result_output", -1);
|
|
|
|
ggml_build_forward_expand(gf, cur);
|
|
|
|
return gf;
|
|
}
|