echodict/llama.cpp

1#include "models.h"2 3#include "llama-memory-recurrent.h"4 5llm_build_plamo2::llm_build_plamo2(const llama_model & model, const llm_graph_params & params) :6    llm_build_mamba_base(params) {7    ggml_tensor * cur;8    ggml_tensor * inpL;9 10    // {n_embd, n_tokens}11    inpL = build_inp_embd(model.tok_embd);12    cb(inpL, "embedding_output", -1);13 14    ggml_tensor * inp_pos = build_inp_pos();15 16    auto * inp_hybrid = build_inp_mem_hybrid();17 18    ggml_tensor * inp_out_ids = build_inp_out_ids();19 20    for (int il = 0; il < n_layer; ++il) {21        ggml_tensor * residual = inpL;22 23        // ggml_graph_add_node(gf, model.layers[il].attn_norm);24        // cb(model.layers[il].attn_norm, "attn_norm", il);25 26        // pre_mixer_norm27        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);28 29        // check if this layer is Mamba or Attention30        const bool is_mamba_layer = hparams.is_recurrent(il);31 32        if (is_mamba_layer) {33            // PLaMo-2 Mamba layer34            cur = build_plamo2_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il);35        } else {36            // PLaMo-2 Attention layer37            cur = build_plamo2_attn_layer(inp_hybrid->get_attn(), inp_pos, cur, model, il);38        }39 40        // post_mixer_norm41        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);42        cb(cur, "attn_post_norm", il);43 44        // residual connection45        cur = ggml_add(ctx0, cur, residual);46        cb(cur, "attn_residual", il);47        residual = cur;48 49        // pre-ffn norm50        cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);51        cb(cur, "ffn_pre_norm", il);52 53        // feed-forward network54        cur = build_ffn(cur,55                model.layers[il].ffn_up, NULL, NULL,56                NULL, NULL, NULL,57                model.layers[il].ffn_down, NULL, NULL,58                NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);59        cb(cur, "ffn_out", il);60 61        // post ffn norm62        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);63        cb(cur, "ffn_post_norm", il);64 65        if (il == n_layer - 1 && inp_out_ids) {66            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);67            residual = ggml_get_rows(ctx0, residual, inp_out_ids);68        }69 70        // residual connection71        cur = ggml_add(ctx0, cur, residual);72        cb(cur, "ffn_residual", il);73 74        // input for next layer75        inpL = cur;76    }77 78    cur = inpL;79 80    // final norm81    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);82    cb(cur, "result_norm", -1);83 84    res->t_embd = cur;85 86    // lm_head87    cur = build_lora_mm(model.output, cur);88    cb(cur, "result_output", -1);89 90    // Explicitly mark as output tensor to ensure proper backend assignment91    ggml_set_output(cur);92 93    res->t_logits = cur;94 95    ggml_build_forward_expand(gf, cur);96}97 98ggml_tensor * llm_build_plamo2::build_plamo2_attn_layer(llm_graph_input_attn_kv * inp,99                                                        ggml_tensor *             inp_pos,100                                                        ggml_tensor *             cur,101                                                        const llama_model &       model,102                                                        int                       il) {103    // self-attention104    {105        // PLaMo-2 uses combined QKV tensor106        ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);107        cb(qkv, "wqkv", il);108 109        // split QKV tensor into Q, K, V110        const int64_t n_embd_head_q = hparams.n_embd_head_k();111        const int64_t n_embd_head_k = hparams.n_embd_head_k();112        const int64_t n_embd_head_v = hparams.n_embd_head_v();113        int32_t       n_head        = hparams.n_head(il);114        int32_t       n_head_kv     = hparams.n_head_kv(il);115 116        const int64_t q_offset = 0;117        const int64_t k_offset = n_embd_head_q * n_head;118        const int64_t v_offset = k_offset + n_embd_head_k * n_head_kv;119 120        ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_q, n_head, n_tokens, n_embd_head_q * sizeof(float),121                                          qkv->nb[1], q_offset * ggml_element_size(qkv));122        ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv, n_tokens, n_embd_head_k * sizeof(float),123                                          qkv->nb[1], k_offset * ggml_element_size(qkv));124        ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv, n_tokens, n_embd_head_v * sizeof(float),125                                          qkv->nb[1], v_offset * ggml_element_size(qkv));126 127        cb(Qcur, "Qcur", il);128        cb(Kcur, "Kcur", il);129        cb(Vcur, "Vcur", il);130 131        Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);132        cb(Qcur, "Qcur_normed", il);133 134        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,135                             ext_factor, attn_factor, beta_fast, beta_slow);136 137        Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);138        cb(Kcur, "Kcur_normed", il);139 140        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,141                             ext_factor, attn_factor, beta_fast, beta_slow);142 143        cur = build_attn(inp,144            model.layers[il].wo, NULL, model.layers[il].wo_s,145            Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f / sqrtf(float(n_embd_head_v)), il);146    }147 148    cb(cur, "attn_out", il);149 150    return cur;151}152 153ggml_tensor * llm_build_plamo2::build_plamo2_mamba_layer(llm_graph_input_rs * inp,154                                                         ggml_tensor *        cur,155                                                         const llama_model &  model,156                                                         const llama_ubatch & ubatch,157                                                         int                  il) {158    const auto * mctx_cur = inp->mctx;159 160    const auto kv_head = mctx_cur->get_head();161 162    const int64_t d_conv   = hparams.ssm_d_conv;163    const int64_t d_inner  = hparams.ssm_d_inner;164    const int64_t d_state  = hparams.ssm_d_state;165    const int64_t n_heads  = hparams.ssm_dt_rank;166    const int64_t head_dim = d_inner / n_heads;167    const int64_t n_group  = hparams.ssm_n_group;168    const int64_t n_seqs   = ubatch.n_seqs;169 170    const int64_t n_seq_tokens = ubatch.n_seq_tokens;171 172    GGML_ASSERT(n_seqs != 0);173    GGML_ASSERT(ubatch.equal_seqs());174    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);175    GGML_ASSERT(d_inner % n_head == 0);176    GGML_ASSERT(n_group == 0);177 178    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);179    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);180 181    ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);182    conv               = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs);183 184    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}185    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);186 187    // in_proj: {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}188    ggml_tensor * zx = build_lora_mm(model.layers[il].ssm_in, cur);189    cb(zx, "mamba_in_proj", il);190    // {8192, 5, 1, 1} -> {8192, 1, 5, 1}191    zx = ggml_permute(ctx0, zx, 0, 2, 1, 3);192    zx = ggml_cont_4d(ctx0, zx, head_dim * 2, n_heads, n_seq_tokens, n_seqs);193    cb(zx, "mamba_in_proj_out", il);194 195    // split into z and x196    // => {head_dim * n_heads, n_seq_tokens, n_seqs}197    ggml_tensor * x = ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3],198                                   head_dim * ggml_element_size(zx));199    x               = ggml_cont_3d(ctx0, x, head_dim * n_heads, n_seq_tokens, n_seqs);200    // x = ggml_permute(ctx0, x, 0, 2, 1, 3);201    cb(x, "mamba_x_split", il);202 203    ggml_tensor * z =204        ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], 0);205    cb(z, "mamba_z_split", il);206 207    // conv1d208    {209        // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}210        ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0);211        cb(conv_x, "mamba_conv1d_input", il);212 213        // copy last (d_conv - 1) columns back into the state cache214        ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2],215                                               n_seq_tokens * (conv_x->nb[0]));216 217        ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv,218                                               ggml_view_1d(ctx0, conv_states_all,219                                                            (d_conv - 1) * (d_inner + 2 * n_group * d_state) * (n_seqs),220                                                            kv_head * (d_conv - 1) * (d_inner + 2 * n_group * d_state) *221                                                                ggml_element_size(conv_states_all))));222        cb(conv_states_all, "mamba_conv1d_state", il);223 224        // 1D convolution225        x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);226        cb(x, "mamba_conv1d", il);227 228        x = ggml_silu(ctx0, x);229        cb(x, "mamba_conv1d_silu", il);230    }231 232    // SSM233    {234        // bcdt_proj: {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs}235        ggml_tensor * x_bcdt = build_lora_mm(model.layers[il].ssm_x, x);236        cb(x_bcdt, "mamba_bcdt_proj", il);237 238        // split into dt, B, C239        const int64_t dt_dim = std::max(64, int(hparams.n_embd / 16));240        ggml_tensor * B  = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], 0);241        ggml_tensor * C  = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2],242                                        ggml_element_size(x_bcdt) * d_state);243        ggml_tensor * dt = ggml_view_3d(ctx0, x_bcdt, dt_dim, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2],244                                        ggml_element_size(x_bcdt) * (2 * d_state));245        cb(B, "mamba_B_raw", il);246        cb(C, "mamba_C_raw", il);247        cb(dt, "mamba_dt_raw", il);248 249        // Apply RMS norm to dt, B, C (PLaMo-2 specific)250        B  = build_norm(B, model.layers[il].ssm_b_norm, NULL, LLM_NORM_RMS, il);251        C  = build_norm(C, model.layers[il].ssm_c_norm, NULL, LLM_NORM_RMS, il);252        dt = build_norm(dt, model.layers[il].ssm_dt_norm, NULL, LLM_NORM_RMS, il);253        cb(B, "mamba_B_normed", il);254        cb(C, "mamba_C_normed", il);255        cb(dt, "mamba_dt_normed", il);256 257        // dt_proj: {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}258        dt = build_lora_mm(model.layers[il].ssm_dt, dt);259        dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b);260        cb(dt, "mamba_dt_proj", il);261 262        ggml_tensor * A = ggml_reshape_2d(ctx0, model.layers[il].ssm_a, 1, n_heads);263        cb(A, "mamba_A", il);264 265        x = ggml_view_4d(ctx0, x, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x),266                         head_dim * n_heads * ggml_element_size(x),267                         head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);268        B = ggml_view_4d(ctx0, B, d_state, 1, n_seq_tokens, n_seqs, d_state * B->nb[0], B->nb[1], B->nb[2], 0);269        C = ggml_view_4d(ctx0, C, d_state, 1, n_seq_tokens, n_seqs, d_state * C->nb[0], C->nb[1], C->nb[2], 0);270 271        // use the states and the indices provided by build_recurrent_state272        // (this is necessary in order to properly use the states before they are overwritten,273        //  while avoiding to make unnecessary copies of the states)274        auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {275            ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_heads, mctx_cur->get_size());276 277            // Custom operator to optimize the parallel associative scan278            // as described in the Annex D of the Mamba paper.279            // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}280            return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);281        };282 283        ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);284        cb(y_ssm, "mamba_ssm_scan", il);285 286        // store last states287        ggml_build_forward_expand(288            gf, ggml_cpy(289                    ctx0,290                    ggml_view_1d(ctx0, y_ssm, n_heads * head_dim * d_state * n_seqs,291                                 n_heads * head_dim * n_seq_tokens * n_seqs * ggml_element_size(y_ssm)),292                    ggml_view_1d(ctx0, ssm_states_all, n_heads * head_dim * d_state * n_seqs,293                                 kv_head * n_seqs * n_heads * head_dim * d_state * ggml_element_size(ssm_states_all))));294        cb(ssm_states_all, "mamba_ssm_states", il);295 296        ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_heads, n_seq_tokens, n_seqs,297                                       head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x),298                                       head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);299        cb(y, "mamba_y_view", il);300 301        // Add D parameter and apply gating with z302        // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs}303        ggml_tensor * D = ggml_reshape_2d(ctx0, model.layers[il].ssm_d, 1, n_heads);304        y               = ggml_add(ctx0, y, ggml_mul(ctx0, x, D));305        cb(y, "mamba_y_add_d", il);306 307        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);308        cb(y, "mamba_y_swiglu_z", il);309 310        // out_proj: {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}311        y   = ggml_view_3d(ctx0, y, head_dim * n_heads, n_seq_tokens, n_seqs, y->nb[2], y->nb[3], 0);312        cur = build_lora_mm(model.layers[il].ssm_out, y);313        cb(cur, "mamba_out_proj", il);314    }315 316    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}317    cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);318    cb(cur, "mamba_out", il);319 320    return cur;321}322