* llama: fix kv shifting and add tests * llama: add graph build_rope (#31693) * llama: add graph build_rope * improve * migrate new models * update review skill * fix ci * add --kv-shift to CI
212 lines
9.0 KiB
C++
212 lines
9.0 KiB
C++
#include "models.h"
|
|
#include "../llama-memory-hybrid-iswa.h"
|
|
|
|
void llama_model_lfm3::load_arch_hparams(llama_model_loader & ml) {
|
|
ml.get_key(LLM_KV_SHORTCONV_L_CACHE, hparams.n_shortconv_l_cache);
|
|
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
|
|
|
|
for (uint32_t il = 0; il < hparams.n_layer(); ++il) {
|
|
hparams.is_recr_impl[il] = hparams.n_head_kv(il) == 0;
|
|
}
|
|
|
|
// LFM3 always mixes full and sliding window attention layers
|
|
ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);
|
|
if (!hparams.is_swa_any()) {
|
|
throw std::runtime_error("LFM3 requires sliding window attention layers");
|
|
}
|
|
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
|
|
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
|
|
|
|
ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa);
|
|
|
|
// d1: full attention layers read the prompt both ways, sliding window layers stay causal
|
|
if (!hparams.causal_attn) {
|
|
hparams.non_causal_type = LLAMA_NON_CAUSAL_TYPE_DENSE_ONLY;
|
|
}
|
|
|
|
switch (hparams.n_ff()) {
|
|
case 7680: type = LLM_TYPE_1_7B; break;
|
|
default: type = LLM_TYPE_UNKNOWN;
|
|
}
|
|
}
|
|
|
|
void llama_model_lfm3::load_arch_tensors(llama_model_loader &) {
|
|
LLAMA_LOAD_LOCALS;
|
|
|
|
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
|
|
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
|
|
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
|
|
if (output == NULL) {
|
|
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
|
|
}
|
|
|
|
int rope_freqs_flag = 0;
|
|
|
|
for (int i = 0; i < n_layer; ++i) {
|
|
auto & layer = layers[i];
|
|
|
|
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
|
|
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
|
|
// absent when use_sandwich_norm is false
|
|
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
|
|
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
|
|
if ((layer.attn_post_norm == nullptr) != (layer.ffn_post_norm == nullptr)) {
|
|
throw std::runtime_error("LFM3 requires both post norms or neither");
|
|
}
|
|
|
|
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
|
|
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
|
|
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
|
|
|
|
if (!hparams.is_recr(i)) {
|
|
// note: n_embd_head_k * n_head may differ from n_embd
|
|
const int64_t n_embd_q = n_embd_head_k * n_head;
|
|
const int64_t n_embd_k = hparams.n_embd_k_gqa(i);
|
|
const int64_t n_embd_v = hparams.n_embd_v_gqa(i);
|
|
if (n_embd_k != n_embd_v) {
|
|
throw std::runtime_error("LFM3 requires equal key and value sizes");
|
|
}
|
|
|
|
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
|
|
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
|
|
|
|
create_tensor_qkv(layer, i, n_embd, n_embd_q, n_embd_k, n_embd_v, 0);
|
|
|
|
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_q, n_embd}, 0);
|
|
|
|
layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head}, 0);
|
|
|
|
if (!hparams.is_swa(i)) {
|
|
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_embd_head_k / 2}, rope_freqs_flag);
|
|
rope_freqs_flag = TENSOR_DUPLICATED;
|
|
}
|
|
} else {
|
|
layer.shortconv.conv = create_tensor(tn(LLM_TENSOR_SHORTCONV_CONV, "weight", i), {hparams.n_shortconv_l_cache, n_embd}, 0);
|
|
layer.shortconv.in_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_INPROJ, "weight", i), {n_embd, 3 * n_embd}, 0);
|
|
layer.shortconv.out_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_OUTPROJ, "weight", i), {n_embd, n_embd}, 0);
|
|
}
|
|
}
|
|
}
|
|
|
|
std::unique_ptr<llm_graph_context> llama_model_lfm3::build_arch_graph(const llm_graph_params & params) const {
|
|
if (!hparams.causal_attn) {
|
|
return std::make_unique<graph<true>>(params);
|
|
}
|
|
return std::make_unique<graph<false>>(params);
|
|
}
|
|
|
|
template <bool decision>
|
|
llama_model_lfm3::graph<decision>::graph(const llm_graph_params & params) :
|
|
llm_build_lfm_base(params) {
|
|
auto build_attn_block = [this](ggml_tensor * cur_in,
|
|
ggml_tensor * inp_pos,
|
|
auto * inp_attn,
|
|
int il) -> ggml_tensor * {
|
|
const int64_t n_embd_head = hparams.n_embd_head_k(il);
|
|
const int64_t n_head_il = hparams.n_head(il);
|
|
const int64_t n_head_kv = hparams.n_head_kv(il);
|
|
|
|
auto [q, k, v] = build_qkv(model.layers[il], cur_in, n_embd_head, n_head_il, n_head_kv, il);
|
|
|
|
q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
|
|
cb(q, "model.layers.{}.self_attn.q_layernorm", il);
|
|
k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
|
|
cb(k, "model.layers.{}.self_attn.k_layernorm", il);
|
|
|
|
q = build_rope(q, inp_pos, il);
|
|
k = build_rope(k, inp_pos, il);
|
|
|
|
ggml_tensor * attn_out = build_attn(inp_attn,
|
|
NULL, NULL, NULL,
|
|
q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
|
|
cb(attn_out, "attn_out", il);
|
|
|
|
ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, cur_in, model.layers[il].wqkv_gate_s);
|
|
gate = ggml_sigmoid(ctx0, gate);
|
|
cb(gate, "model.layers.{}.self_attn.gate", il);
|
|
|
|
ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, attn_out, n_embd_head, n_head_il, n_tokens);
|
|
ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate, 1, n_head_il, n_tokens);
|
|
attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);
|
|
attn_out = ggml_reshape_2d(ctx0, attn_3d, n_embd_head * n_head_il, n_tokens);
|
|
|
|
ggml_tensor * out = build_lora_mm(model.layers[il].wo, attn_out, model.layers[il].wo_s);
|
|
cb(out, "model.layers.{}.self_attn.out_proj", il);
|
|
|
|
return out;
|
|
};
|
|
|
|
ggml_tensor * cur = build_inp_embd(model.tok_embd);
|
|
cb(cur, "model.embed_tokens", -1);
|
|
|
|
ggml_build_forward_expand(gf, cur);
|
|
|
|
// decision: no memory, the whole prompt is read in one batch
|
|
auto * inp_hybrid = decision ? nullptr : build_inp_mem_hybrid_iswa();
|
|
auto * inp_attn = decision ? build_attn_inp_no_cache() : nullptr;
|
|
|
|
ggml_tensor * inp_pos = build_inp_pos();
|
|
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
|
|
|
for (int il = 0; il < n_layer; ++il) {
|
|
res->t_layer_inp[il] = cur;
|
|
|
|
auto * prev_cur = cur;
|
|
|
|
cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
|
|
cb(cur, "model.layers.{}.operator_norm", il);
|
|
|
|
if (hparams.is_recr(il)) {
|
|
if constexpr (decision) {
|
|
cur = build_lfm_shortconv(cur, il);
|
|
} else {
|
|
const auto * mctx_recr = static_cast<const llama_memory_hybrid_iswa_context *>(mctx)->get_recr();
|
|
cur = build_lfm_shortconv(mctx_recr, inp_hybrid->get_recr(), cur, il);
|
|
}
|
|
} else if constexpr (decision) {
|
|
cur = build_attn_block(cur, inp_pos, inp_attn, il);
|
|
} else {
|
|
cur = build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);
|
|
}
|
|
|
|
if (model.layers[il].attn_post_norm) {
|
|
cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
|
|
cb(cur, "model.layers.{}.post_operator_norm", il);
|
|
}
|
|
|
|
if (il == n_layer - 1 && inp_out_ids) {
|
|
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
|
prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);
|
|
}
|
|
|
|
cur = ggml_add(ctx0, prev_cur, cur);
|
|
|
|
auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
|
|
cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);
|
|
|
|
ggml_tensor * ffn_out = build_lfm_dense_ffn(ffn_norm_out, il);
|
|
|
|
if (model.layers[il].ffn_post_norm) {
|
|
ffn_out = build_norm(ffn_out, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
|
|
cb(ffn_out, "model.layers.{}.post_ffn_norm", il);
|
|
}
|
|
|
|
cur = ggml_add(ctx0, cur, ffn_out);
|
|
|
|
cur = build_cvec(cur, il);
|
|
cb(cur, "l_out", il);
|
|
}
|
|
|
|
cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
|
|
cb(cur, "result_norm", -1);
|
|
res->t_embd = cur;
|
|
|
|
cur = build_lora_mm(model.output, cur, model.output_s);
|
|
cb(cur, "result_output", -1);
|
|
|
|
res->t_logits = cur;
|
|
}
|
|
|
|
template struct llama_model_lfm3::graph<false>;
|
|
template struct llama_model_lfm3::graph<true>;
|