ik_llama_opt/src/graphs/build_step35.cpp

219 lines
9.7 KiB
C++

#include "../llama-build-context.h"
#include "../llama-model.h"
#include "../llama-context.h"
ggml_cgraph * llm_build_context::build_step35() {
ggml_cgraph * gf = new_graph_custom();
ggml_tensor * cur;
auto inp_pos = build_inp_pos();
if (cparams.mtp_op_type != MTP_OP_NONE) {
GGML_ASSERT(model.mtp && hparams.nextn_predict_layers > 0);
GGML_ASSERT(batch.token && "Step35 MTP requires token batches");
const int n_layer_base = hparams.n_layer > hparams.nextn_predict_layers
? hparams.n_layer - hparams.nextn_predict_layers : hparams.n_layer;
const int n_heads_model = (int) hparams.nextn_predict_layers;
const int n_heads = lctx.mtp_n_heads > 0
? std::max(1, std::min((int) lctx.mtp_n_heads, n_heads_model)) : n_heads_model;
const int step = std::max(0, std::min((int) lctx.mtp_step_idx, n_heads - 1));
const int il = n_layer_base + step;
ggml_tensor * hidden_states = build_inp_mtp_states(n_embd);
const bool step_independent_warmup = model.arch == LLM_ARCH_STEP35 &&
(cparams.mtp_op_type == MTP_OP_WARMUP ||
cparams.mtp_op_type == MTP_OP_UPDATE_ACCEPTED) && n_heads > 1;
if (step_independent_warmup) {
for (int i = n_heads - 1; i >= 0; --i) {
const int head_il = n_layer_base + i;
const bool is_first = i == 0;
const bool emit_logits = is_first && cparams.mtp_op_type == MTP_OP_UPDATE_ACCEPTED;
cur = build_step35_mtp(model.layers[head_il], hidden_states, gf, inp_pos,
/*reduce_output=*/is_first, emit_logits);
ggml_build_forward_expand(gf, cur);
}
return gf;
}
const bool reduce_mtp_output = cparams.mtp_op_type != MTP_OP_NONE;
const bool emit_mtp_logits = cparams.mtp_op_type == MTP_OP_DRAFT_GEN ||
cparams.mtp_op_type == MTP_OP_UPDATE_ACCEPTED;
cur = build_step35_mtp(model.layers[il], hidden_states, gf, inp_pos,
reduce_mtp_output, emit_mtp_logits);
ggml_build_forward_expand(gf, cur);
return gf;
}
auto inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
auto inp_out_ids = build_inp_out_ids();
auto KQ_mask = build_inp_KQ_mask();
auto KQ_mask_swa = build_inp_KQ_mask_swa();
//const float kq_scale = 1.0f / sqrtf(float(n_rot));
const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k));
const int n_layer_base = hparams.n_layer > hparams.nextn_predict_layers
? hparams.n_layer - hparams.nextn_predict_layers : hparams.n_layer;
for (int il = 0; il < n_layer_base; ++il) {
bool is_swa = hparams.swa_layers[il];
auto & layer = const_cast<llama_layer&>(model.layers[il]);
ggml_tensor * rope_factors = nullptr;
const uint32_t apply_mask = hparams.rope_scaling_apply_mask;
if ((is_swa && (apply_mask & 0x2)) || (!is_swa && (apply_mask & 0x1))) {
rope_factors = build_rope_factors(il);
}
auto rope_freqs = layer.rope_freqs;
layer.rope_freqs = nullptr;
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL,
inp_pos, il == n_layer_base - 1 && n_tokens > 1 && !cparams.mtp ? inp_out_ids : nullptr,
rope_factors, is_swa ? KQ_mask_swa : KQ_mask, nullptr, nullptr, kq_scale, 0.0f, is_swa ? hparams.n_swa : 0,
il, true, false, true);
layer.rope_freqs = rope_freqs;
if (model.layers[il].ffn_gate_inp == nullptr) {
// dense FFN
cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL,
nullptr,
LLM_FFN_SILU, LLM_FFN_PAR, cb, il, gf, true);
cb(cur, "ffn_out", il);
} else {
const bool norm_w = hparams.expert_weights_norm;
const float w_scale = hparams.expert_weights_scale;
const bool scale_w = w_scale != 0.0f;
cur = llm_build_std_moe_ffn(ctx0, lctx, model.layers[il].ffn_norm, cur,
model.layers[il].ffn_gate_inp, model.layers[il].ffn_gate_inp_b,
model.layers[il].ffn_up_exps, model.layers[il].ffn_up_exps_b,
model.layers[il].ffn_gate_exps, model.layers[il].ffn_gate_exps_b,
model.layers[il].ffn_down_exps, model.layers[il].ffn_down_exps_b,
model.layers[il].ffn_exp_probs_b,
model.layers[il].ffn_up_shexp, nullptr, // we don't have shared expert biases?
model.layers[il].ffn_gate_shexp, nullptr,
model.layers[il].ffn_down_shexp, nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, norm_w, scale_w, w_scale,
LLM_EXPERT_GATING_FUNC_SIGMOID,
//(llm_expert_gating_func_type) hparams.expert_gating_func,
LLM_FFN_SILU, cb, il, gf, true, model.layers[il].ffn_up_gate_exps);
}
cur = lctx.cvec.apply_to(ctx0, cur, il);
cb(cur, "l_out", il);
inpL = cur;
}
if (cparams.mtp) {
ggml_tensor * mtp_embd = inpL->type == GGML_TYPE_F32 ? inpL : ggml_cast(ctx0, inpL, GGML_TYPE_F32);
cb(mtp_embd, "result_mtp_embd", -1);
ggml_set_output(mtp_embd);
ggml_build_forward_expand(gf, mtp_embd);
if (inp_out_ids) {
inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
}
}
cur = build_output(lctx, ctx0, inpL, model.output, model.output_norm, cb);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
ggml_tensor * llm_build_context::build_step35_mtp(
const llama_layer & mtp_layer,
ggml_tensor * hidden_states_from_main_model,
ggml_cgraph * gf,
ggml_tensor * inp_pos,
bool reduce_output,
bool emit_logits,
ggml_tensor ** hidden_out) {
const int il = (int) (&mtp_layer - model.layers.data());
GGML_ASSERT(mtp_layer.nextn.eh_proj && mtp_layer.nextn.enorm && mtp_layer.nextn.hnorm);
GGML_ASSERT(mtp_layer.wq && mtp_layer.wk && mtp_layer.wv && mtp_layer.wo);
ggml_tensor * inp_out_ids = (n_tokens > 1 && n_outputs < n_tokens) ? build_inp_out_ids() : nullptr;
ggml_tensor * tok_embd_w = mtp_layer.nextn.embed_tokens ? mtp_layer.nextn.embed_tokens : model.tok_embd;
ggml_tensor * tok_embd = build_inp_embd_mtp(tok_embd_w);
ggml_tensor * cur = build_mtp_input(mtp_layer, hidden_states_from_main_model,
tok_embd, il, "mtp_eh_proj");
const bool is_swa = hparams.swa_layers[il];
ggml_tensor * rope_factors = nullptr;
const uint32_t apply_mask = hparams.rope_scaling_apply_mask;
if ((is_swa && (apply_mask & 0x2)) || (!is_swa && (apply_mask & 0x1))) {
rope_factors = build_rope_factors(il);
}
auto KQ_mask = is_swa ? build_inp_KQ_mask_swa() : build_inp_KQ_mask();
const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k(il)));
cur = build_std_attention(gf, mtp_layer.attn_norm, cur, inp_pos, nullptr,
rope_factors, KQ_mask, nullptr, nullptr, kq_scale, 0.0f,
is_swa ? hparams.n_swa : 0, il, true, false, true, false, false, nullptr, il);
if (mtp_layer.ffn_gate_inp == nullptr) {
cur = llm_build_ffn(ctx0, lctx, mtp_layer.ffn_norm, cur,
mtp_layer.ffn_up, nullptr, nullptr,
mtp_layer.ffn_gate, nullptr, nullptr,
mtp_layer.ffn_down, nullptr, nullptr,
nullptr, LLM_FFN_SILU, LLM_FFN_PAR, cb, il, gf, true);
} else {
cur = llm_build_std_moe_ffn(ctx0, lctx, mtp_layer.ffn_norm, cur,
mtp_layer.ffn_gate_inp, nullptr,
mtp_layer.ffn_up_exps, nullptr,
mtp_layer.ffn_gate_exps, nullptr,
mtp_layer.ffn_down_exps, nullptr,
mtp_layer.ffn_exp_probs_b,
mtp_layer.ffn_up_shexp, nullptr,
mtp_layer.ffn_gate_shexp, nullptr,
mtp_layer.ffn_down_shexp, nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm, hparams.expert_weights_scale != 0.0f,
hparams.expert_weights_scale,
(llm_expert_gating_func_type) hparams.expert_gating_func,
LLM_FFN_SILU, cb, il, gf, true, mtp_layer.ffn_up_gate_exps);
}
cur = lctx.cvec.apply_to(ctx0, cur, il);
cb(cur, "mtp_post_ffn", il);
if (hidden_out) {
*hidden_out = cur;
}
ggml_tensor * output_hidden = cur;
if (reduce_output) {
if (cparams.mtp_op_type != MTP_OP_NONE && n_tokens > 1) {
output_hidden = ggml_view_2d(ctx0, cur, n_embd, 1,
cur->nb[1], (size_t) (n_tokens - 1) * cur->nb[1]);
} else if (inp_out_ids) {
output_hidden = ggml_get_rows(ctx0, cur, inp_out_ids);
}
}
if (reduce_output) {
ggml_tensor * mtp_embd = output_hidden->type == GGML_TYPE_F32 ? output_hidden : ggml_cast(ctx0, output_hidden, GGML_TYPE_F32);
cb(mtp_embd, "result_mtp_embd", -1);
ggml_set_output(mtp_embd);
ggml_build_forward_expand(gf, mtp_embd);
}
if (!emit_logits) {
return output_hidden;
}
ggml_tensor * head_norm = mtp_layer.nextn.shared_head_norm
? mtp_layer.nextn.shared_head_norm : model.output_norm;
ggml_tensor * head = mtp_layer.nextn.shared_head_head
? mtp_layer.nextn.shared_head_head : model.output;
GGML_ASSERT(head_norm && head);
cur = llm_build_context::build_output(lctx, ctx0, output_hidden, head, head_norm, cb);
cb(cur, "result_output", -1);
return cur;
}