1#include "models.h"
  2
  3#include "../llama-memory-hybrid.h"
  4
  5
  6llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :
  7    llm_graph_context(params),
  8    model(model) {
  9    ggml_tensor * cur = build_inp_embd(model.tok_embd);
 10    cb(cur, "model.embed_tokens", -1);
 11
 12    ggml_build_forward_expand(gf, cur);
 13
 14    ggml_tensor * inp_pos     = build_inp_pos();
 15    auto *        inp_hybrid  = build_inp_mem_hybrid();
 16    ggml_tensor * inp_out_ids = build_inp_out_ids();
 17
 18    for (int il = 0; il < n_layer; ++il) {
 19        const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);
 20
 21        auto * prev_cur = cur;
 22        cur             = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
 23        cb(cur, "model.layers.{}.operator_norm", il);
 24
 25        cur = hparams.is_recurrent(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) :
 26                                         build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);
 27
 28        if (il == n_layer - 1 && inp_out_ids) {
 29            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);
 30            prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);
 31        }
 32
 33        cur = ggml_add(ctx0, prev_cur, cur);
 34
 35        auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
 36        cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);
 37
 38        ggml_tensor * ffn_out =
 39            is_moe_layer ? build_moe_feed_forward(ffn_norm_out, il) : build_dense_feed_forward(ffn_norm_out, il);
 40        cb(ffn_norm_out, "model.layers.{}.ffn_out", il);
 41
 42        cur = ggml_add(ctx0, cur, ffn_out);
 43    }
 44
 45    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 46    cb(cur, "result_norm", -1);
 47    res->t_embd = cur;
 48
 49    cur = build_lora_mm(model.output, cur);
 50    cb(cur, "result_output", -1);
 51
 52    res->t_logits = cur;
 53
 54    ggml_build_forward_expand(gf, cur);
 55}
 56
 57ggml_tensor * llm_build_lfm2::build_moe_feed_forward(ggml_tensor * cur, int il) const {
 58    return build_moe_ffn(cur,
 59                        model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
 60                        model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
 61                        model.layers[il].ffn_exp_probs_b, n_expert, n_expert_used, LLM_FFN_SILU, true, false, 0.0,
 62                        static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func), il);
 63}
 64
 65ggml_tensor * llm_build_lfm2::build_dense_feed_forward(ggml_tensor * cur, int il) const {
 66    GGML_ASSERT(!model.layers[il].ffn_up_b);
 67    GGML_ASSERT(!model.layers[il].ffn_gate_b);
 68    GGML_ASSERT(!model.layers[il].ffn_down_b);
 69    return build_ffn(cur,
 70        model.layers[il].ffn_up, NULL, NULL,
 71        model.layers[il].ffn_gate, NULL, NULL,
 72        model.layers[il].ffn_down, NULL, NULL,
 73        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
 74}
 75
 76ggml_tensor * llm_build_lfm2::build_attn_block(ggml_tensor *             cur,
 77                                               ggml_tensor *             inp_pos,
 78                                               llm_graph_input_attn_kv * inp_attn,
 79                                               int                       il) const {
 80    GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
 81    const auto n_embd_head = hparams.n_embd_head_v;
 82    const auto n_head_kv   = hparams.n_head_kv(il);
 83
 84    auto * q = build_lora_mm(model.layers[il].wq, cur);
 85    cb(q, "model.layers.{}.self_attn.q_proj", il);
 86    auto * k = build_lora_mm(model.layers[il].wk, cur);
 87    cb(k, "model.layers.{}.self_attn.k_proj", il);
 88    auto * v = build_lora_mm(model.layers[il].wv, cur);
 89    cb(v, "model.layers.{}.self_attn.v_proj", il);
 90
 91    q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens);
 92    k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens);
 93    v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens);
 94
 95    // qk norm
 96    q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
 97    cb(q, "model.layers.{}.self_attn.q_layernorm", il);
 98    k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
 99    cb(k, "model.layers.{}.self_attn.k_layernorm", il);
100
101    // RoPE
102    q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
103                      attn_factor, beta_fast, beta_slow);
104    k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
105                      attn_factor, beta_fast, beta_slow);
106
107    cur = build_attn(inp_attn,
108            model.layers[il].wo, NULL,
109            q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
110
111    cb(cur, "model.layers.{}.self_attn.out_proj", il);
112
113    return cur;
114}
115
116ggml_tensor * llm_build_lfm2::build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il) {
117    const auto *   mctx_cur     = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
118    const uint32_t kv_head      = mctx_cur->get_head();
119    const int64_t  n_seq_tokens = ubatch.n_seq_tokens;
120    const int64_t  n_seqs       = ubatch.n_seqs;
121    GGML_ASSERT(n_seqs != 0);
122    GGML_ASSERT(ubatch.equal_seqs());
123    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
124
125    GGML_ASSERT(hparams.n_shortconv_l_cache > 1);
126    const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;
127
128    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
129    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
130
131    auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);
132    cb(bcx, "model.layers.{}.conv.in_proj", il);
133
134    constexpr auto n_chunks = 3;
135    GGML_ASSERT(bcx->ne[0] % n_chunks == 0);
136    const auto chunk_size = bcx->ne[0] / n_chunks;
137    auto *     b          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
138                                         0 * chunk_size * ggml_element_size(bcx));
139    auto *     c          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
140                                         1 * chunk_size * ggml_element_size(bcx));
141    auto *     x          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
142                                         2 * chunk_size * ggml_element_size(bcx));
143
144    auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));
145
146    // read conv state
147    auto * conv_state = mctx_cur->get_r_l(il);
148    auto * conv_rs    = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
149    auto * conv       = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);
150
151    bx = ggml_concat(ctx0, conv, bx, 0);
152    GGML_ASSERT(bx->ne[0] > conv->ne[0]);
153
154    // last d_conv columns is a new conv state
155    auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],
156                                   (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));
157    GGML_ASSERT(ggml_are_same_shape(conv, new_conv));
158
159    // write new conv conv state
160    ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,
161                                           ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),
162                                                        kv_head * d_conv * n_embd * ggml_element_size(new_conv))));
163
164    auto * conv_kernel = model.layers[il].shortconv.conv;
165    auto * conv_out    = ggml_ssm_conv(ctx0, bx, conv_kernel);
166    cb(conv_out, "model.layers.{}.conv.conv", il);
167
168    auto * y = ggml_mul(ctx0, c, conv_out);
169    y        = build_lora_mm(model.layers[il].shortconv.out_proj, y);
170    cb(y, "model.layers.{}.conv.out_proj", il);
171    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
172    y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);
173
174    return y;
175}