mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-06-28 04:30:15 -05:00
327 lines
12 KiB
C++
327 lines
12 KiB
C++
#include "../llama-build-context.h"
|
|
#include "../llama-model.h"
|
|
#include "../llama-context.h"
|
|
|
|
ggml_cgraph * llm_build_context::build_qwen2() {
|
|
ggml_cgraph * gf = new_graph_custom();
|
|
|
|
const int64_t n_embd_head = hparams.n_embd_head_v(0);
|
|
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k(0));
|
|
GGML_ASSERT(n_embd_head == hparams.n_rot);
|
|
|
|
struct ggml_tensor * cur;
|
|
struct ggml_tensor * inpL;
|
|
|
|
inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
|
|
|
|
// 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)
|
|
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
struct ggml_tensor * inpSA = inpL;
|
|
|
|
// 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, model.layers[il].bq,
|
|
model.layers[il].wk, model.layers[il].bk,
|
|
model.layers[il].wv, model.layers[il].bv, 0.f, il);
|
|
|
|
Qcur = ggml_rope_ext(
|
|
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, 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);
|
|
|
|
Kcur = ggml_rope_ext(
|
|
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, 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, model.layers[il].bo,
|
|
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, 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);
|
|
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
|
}
|
|
|
|
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
|
cb(ffn_inp, "ffn_inp", il);
|
|
|
|
// feed-forward network
|
|
cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp,
|
|
model.layers[il].ffn_up, NULL, NULL,
|
|
model.layers[il].ffn_gate, NULL, NULL,
|
|
model.layers[il].ffn_down, NULL, NULL,
|
|
NULL,
|
|
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
|
|
cb(cur, "ffn_out", il);
|
|
|
|
cur = ggml_add(ctx0, cur, ffn_inp);
|
|
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);
|
|
cb(cur, "result_output", -1);
|
|
|
|
ggml_build_forward_expand(gf, cur);
|
|
|
|
return gf;
|
|
}
|
|
|
|
ggml_cgraph * llm_build_context::build_qwen2vl() {
|
|
ggml_cgraph * gf = new_graph_custom();
|
|
|
|
const int64_t n_embd_head = hparams.n_embd_head_v(0);
|
|
|
|
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k(0));
|
|
GGML_ASSERT(n_embd_head == hparams.n_rot);
|
|
|
|
ggml_tensor * cur;
|
|
ggml_tensor * inpL;
|
|
|
|
inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
|
|
|
|
// 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)
|
|
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
|
|
|
|
//auto * inp_attn = build_attn_inp_kv();
|
|
|
|
int sections[4];
|
|
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
|
|
|
|
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
ggml_tensor * inpSA = inpL;
|
|
|
|
// 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, model.layers[il].bq,
|
|
model.layers[il].wk, model.layers[il].bk,
|
|
model.layers[il].wv, model.layers[il].bv, 0.f, il);
|
|
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
|
|
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
|
|
|
|
Qcur = ggml_rope_multi(
|
|
ctx0, Qcur, inp_pos, nullptr,
|
|
n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
ext_factor, attn_factor, beta_fast, beta_slow
|
|
);
|
|
|
|
Kcur = ggml_rope_multi(
|
|
ctx0, Kcur, inp_pos, nullptr,
|
|
n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
|
|
ext_factor, attn_factor, beta_fast, beta_slow
|
|
);
|
|
|
|
cb(Qcur, "Qcur", il);
|
|
cb(Kcur, "Kcur", il);
|
|
|
|
cur = llm_build_kv(ctx0, lctx, kv_self, gf,
|
|
model.layers[il].wo, model.layers[il].bo,
|
|
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
|
|
}
|
|
|
|
if (il == n_layer - 1 && inp_out_ids) {
|
|
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
|
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
|
}
|
|
|
|
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
|
cb(ffn_inp, "ffn_inp", il);
|
|
|
|
// feed-forward network
|
|
cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp,
|
|
model.layers[il].ffn_up, NULL, NULL,
|
|
model.layers[il].ffn_gate, NULL, NULL,
|
|
model.layers[il].ffn_down, NULL, NULL,
|
|
NULL,
|
|
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
|
|
cb(cur, "ffn_out", il);
|
|
|
|
cur = ggml_add(ctx0, cur, ffn_inp);
|
|
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);
|
|
cb(cur, "result_output", -1);
|
|
|
|
ggml_build_forward_expand(gf, cur);
|
|
|
|
return gf;
|
|
|
|
}
|
|
|
|
ggml_cgraph * llm_build_context::build_qwen2moe() {
|
|
ggml_cgraph * gf = new_graph_custom();
|
|
|
|
// mutable variable, needed during the last layer of the computation to skip unused tokens
|
|
int32_t n_tokens = this->n_tokens;
|
|
|
|
const int64_t n_embd_head = hparams.n_embd_head_v(0);
|
|
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k(0));
|
|
GGML_ASSERT(n_embd_head == hparams.n_rot);
|
|
|
|
struct ggml_tensor * cur;
|
|
struct ggml_tensor * inpL;
|
|
|
|
inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
|
|
|
|
// 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)
|
|
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
struct ggml_tensor * inpSA = inpL;
|
|
|
|
// 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, model.layers[il].bq,
|
|
model.layers[il].wk, model.layers[il].bk,
|
|
model.layers[il].wv, model.layers[il].bv, 0.f, il);
|
|
|
|
Qcur = ggml_rope_ext(
|
|
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, 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);
|
|
|
|
Kcur = ggml_rope_ext(
|
|
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, 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, model.layers[il].bo,
|
|
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
|
|
}
|
|
|
|
if (il == n_layer - 1) {
|
|
// skip computing output for unused tokens
|
|
struct ggml_tensor * inp_out_ids = build_inp_out_ids();
|
|
n_tokens = n_outputs;
|
|
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
|
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
|
}
|
|
|
|
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
|
cb(ffn_inp, "ffn_inp", il);
|
|
|
|
// MoE branch
|
|
cur = llm_build_norm(ctx0, ffn_inp, hparams, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, cb, il);
|
|
cb(cur, "ffn_norm", il);
|
|
|
|
ggml_tensor * moe_out =
|
|
llm_build_moe_ffn(ctx0, lctx, cur,
|
|
model.layers[il].ffn_gate_inp,
|
|
model.layers[il].ffn_up_exps,
|
|
model.layers[il].ffn_gate_exps,
|
|
model.layers[il].ffn_down_exps,
|
|
nullptr,
|
|
n_expert, n_expert_used,
|
|
LLM_FFN_SILU, false,
|
|
false, 0.0,
|
|
LLM_EXPERT_GATING_FUNC_SOFTMAX,
|
|
cb, il, gf, false, model.layers[il].ffn_up_gate_exps);
|
|
cb(cur, "ffn_moe_out", il);
|
|
|
|
// FFN shared expert
|
|
{
|
|
ggml_tensor * cur_gate_inp = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_gate_inp_shexp, cur);
|
|
cb(cur_gate_inp, "ffn_shexp_gate_inp", il);
|
|
|
|
// sigmoid
|
|
ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp);
|
|
cb(cur_gate, "ffn_shexp_gate", il);
|
|
|
|
ggml_tensor * cur_ffn = llm_build_ffn(ctx0, lctx, nullptr, cur,
|
|
model.layers[il].ffn_up_shexp, NULL, NULL,
|
|
model.layers[il].ffn_gate_shexp, NULL, NULL,
|
|
model.layers[il].ffn_down_shexp, NULL, NULL,
|
|
NULL,
|
|
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
|
|
cb(cur_ffn, "ffn_shexp", il);
|
|
|
|
ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate);
|
|
cb(ffn_shexp_out, "ffn_shexp_out", il);
|
|
|
|
moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out);
|
|
cb(moe_out, "ffn_out", il);
|
|
|
|
cur = moe_out;
|
|
}
|
|
|
|
cur = ggml_add(ctx0, cur, ffn_inp);
|
|
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);
|
|
cb(cur, "result_output", -1);
|
|
|
|
ggml_build_forward_expand(gf, cur);
|
|
|
|
return gf;
|
|
}
|
|
|