echodict/llama.cpp

1#include "models.h"2 3#include "../llama-memory-hybrid-iswa.h"4#include "../llama-memory-hybrid.h"5 6template <bool iswa>7llm_build_lfm2<iswa>::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :8    llm_graph_context(params) {9    using inp_hybrid_type = std::conditional_t<iswa, llm_graph_input_mem_hybrid_iswa,  llm_graph_input_mem_hybrid>;10    using inp_attn_type   = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa,     llm_graph_input_attn_kv>;11    using mem_hybrid_ctx  = std::conditional_t<iswa, llama_memory_hybrid_iswa_context, llama_memory_hybrid_context>;12 13    // lambda helpers for readability14    auto build_dense_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {15        GGML_ASSERT(!model.layers[il].ffn_up_b);16        GGML_ASSERT(!model.layers[il].ffn_gate_b);17        GGML_ASSERT(!model.layers[il].ffn_down_b);18        return build_ffn(cur,19            model.layers[il].ffn_up, NULL, NULL,20            model.layers[il].ffn_gate, NULL, NULL,21            model.layers[il].ffn_down, NULL, NULL,22            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);23    };24    auto build_moe_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {25        return build_moe_ffn(cur,26                model.layers[il].ffn_gate_inp,27                model.layers[il].ffn_up_exps,28                model.layers[il].ffn_gate_exps,29                model.layers[il].ffn_down_exps,30                model.layers[il].ffn_exp_probs_b,31                n_expert, n_expert_used,32                LLM_FFN_SILU, true,33                hparams.expert_weights_scale,34                static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),35                il);36    };37    auto build_attn_block = [&model, this](ggml_tensor *   cur,38                                           ggml_tensor *   inp_pos,39                                           inp_attn_type * inp_attn,40                                           int             il) -> ggml_tensor * {41        GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));42        const auto n_embd_head = hparams.n_embd_head_v();43        const auto n_head_kv   = hparams.n_head_kv(il);44 45        auto [q, k, v] = build_qkv(model.layers[il], cur,46                n_embd_head, n_head, n_head_kv, il);47 48        // qk norm49        q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);50        cb(q, "model.layers.{}.self_attn.q_layernorm", il);51        k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);52        cb(k, "model.layers.{}.self_attn.k_layernorm", il);53 54        // RoPE55        q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,56                          attn_factor, beta_fast, beta_slow);57        k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,58                          attn_factor, beta_fast, beta_slow);59 60        cur = build_attn(inp_attn,61                model.layers[il].wo, NULL, model.layers[il].wo_s,62                q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);63 64        cb(cur, "model.layers.{}.self_attn.out_proj", il);65 66        return cur;67    };68    auto build_shortconv_block = [&model, this](ggml_tensor *        cur,69                                                llm_graph_input_rs * inp_recr,70                                                int                  il) -> ggml_tensor * {71        const auto * mctx_cur = static_cast<const mem_hybrid_ctx *>(mctx)->get_recr();72        const uint32_t kv_head      = mctx_cur->get_head();73        const int64_t  n_seq_tokens = ubatch.n_seq_tokens;74        const int64_t  n_seqs       = ubatch.n_seqs;75        GGML_ASSERT(n_seqs != 0);76        GGML_ASSERT(ubatch.equal_seqs());77        GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);78 79        GGML_ASSERT(hparams.n_shortconv_l_cache > 1);80        const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;81 82        // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}83        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);84 85        auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);86        cb(bcx, "model.layers.{}.conv.in_proj", il);87 88        constexpr auto n_chunks = 3;89        GGML_ASSERT(bcx->ne[0] % n_chunks == 0);90        const auto chunk_size = bcx->ne[0] / n_chunks;91        auto *     b          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],92                                             0 * chunk_size * ggml_element_size(bcx));93        auto *     c          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],94                                             1 * chunk_size * ggml_element_size(bcx));95        auto *     x          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],96                                             2 * chunk_size * ggml_element_size(bcx));97 98        auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));99 100        // read conv state101        auto * conv_state = mctx_cur->get_r_l(il);102        auto * conv_rs    = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);103        auto * conv       = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);104 105        bx = ggml_concat(ctx0, conv, bx, 0);106        GGML_ASSERT(bx->ne[0] > conv->ne[0]);107 108        // last d_conv columns is a new conv state109        auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],110                                       (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));111        GGML_ASSERT(ggml_are_same_shape(conv, new_conv));112 113        // write new conv conv state114        ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,115                                               ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),116                                                            kv_head * d_conv * n_embd * ggml_element_size(new_conv))));117 118        auto * conv_kernel = model.layers[il].shortconv.conv;119        auto * conv_out    = ggml_ssm_conv(ctx0, bx, conv_kernel);120        cb(conv_out, "model.layers.{}.conv.conv", il);121 122        auto * y = ggml_mul(ctx0, c, conv_out);123        y        = build_lora_mm(model.layers[il].shortconv.out_proj, y);124        cb(y, "model.layers.{}.conv.out_proj", il);125        // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}126        y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);127 128        return y;129    };130 131    // actual graph construction starts here132    ggml_tensor * cur = build_inp_embd(model.tok_embd);133    cb(cur, "model.embed_tokens", -1);134 135    ggml_build_forward_expand(gf, cur);136 137    inp_hybrid_type * inp_hybrid = nullptr;138    if constexpr (iswa) {139        inp_hybrid = build_inp_mem_hybrid_iswa();140    } else {141        inp_hybrid = build_inp_mem_hybrid();142    }143 144    ggml_tensor * inp_pos     = build_inp_pos();145    ggml_tensor * inp_out_ids = build_inp_out_ids();146 147    for (int il = 0; il < n_layer; ++il) {148        const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);149 150        auto * prev_cur = cur;151        cur             = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);152        cb(cur, "model.layers.{}.operator_norm", il);153 154        cur = hparams.is_recurrent(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) :155                                         build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);156 157        if (il == n_layer - 1 && inp_out_ids) {158            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);159            prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);160        }161 162        cur = ggml_add(ctx0, prev_cur, cur);163 164        auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);165        cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);166 167        ggml_tensor * ffn_out =168            is_moe_layer ? build_moe_feed_forward(ffn_norm_out, il) : build_dense_feed_forward(ffn_norm_out, il);169        cb(ffn_norm_out, "model.layers.{}.ffn_out", il);170 171        cur = ggml_add(ctx0, cur, ffn_out);172 173        cur = build_cvec(cur, il);174        cb(cur, "l_out", il);175    }176 177    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);178    cb(cur, "result_norm", -1);179    res->t_embd = cur;180 181    cur = build_lora_mm(model.output, cur);182    cb(cur, "result_output", -1);183 184    res->t_logits = cur;185 186    ggml_build_forward_expand(gf, cur);187}188 189// Explicit template instantiations190template struct llm_build_lfm2<true>;191template struct llm_build_lfm2<false>;192