Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2#include "../llama-memory-hybrid-iswa.h"3#include "../llama-memory-hybrid.h"4 5void llama_model_lfm2::load_arch_hparams(llama_model_loader & ml) {6    ml.get_key(LLM_KV_SHORTCONV_L_CACHE,           hparams.n_shortconv_l_cache);7    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);8 9    for (uint32_t il = 0; il < hparams.n_layer(); ++il) {10        hparams.is_recr_impl[il] = hparams.n_head_kv(il) == 0;11    }12 13    hparams.n_layer_dense_lead = hparams.n_layer();14 15    switch (hparams.n_ff()) {16        case  2560: type = LLM_TYPE_230M; break;17        case  4608: type = LLM_TYPE_350M; break;18        case  6912: type = LLM_TYPE_700M; break;19        case  8192: type = LLM_TYPE_1_2B; break;20        case 10752: type = LLM_TYPE_2_6B; break;21        default:    type = LLM_TYPE_UNKNOWN;22    }23 24    if (const auto is_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); is_swa && hparams.n_swa > 0) {25        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;26        for (uint32_t il = 0; il < hparams.n_layer(); ++il) {27            hparams.is_swa_impl[il] = !hparams.is_recr_impl[il];28        }29    }30}31 32void llama_model_lfm2::load_arch_tensors(llama_model_loader &) {33    LLAMA_LOAD_LOCALS;34 35    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);36 37    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM_LFM2, "weight"), {n_embd}, 0);38    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,           "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);39 40    if (output == NULL) {41        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);42    }43 44    for (int i = 0; i < n_layer; ++i) {45        auto & layer = layers[i];46 47        const bool is_moe_layer = i >= static_cast<int>(hparams.n_layer_dense_lead);48 49        // ffn/moe is same for transformer and conv layers50        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);51        if (is_moe_layer) {52            GGML_ASSERT(n_expert && n_expert_used);53            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i),  {n_embd, n_expert}, 0);54            layer.ffn_gate_exps   = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);55            layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp,   n_embd, n_expert}, 0);56            layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i),   {n_embd, hparams.n_ff_exp, n_expert}, 0);57            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);58        } else {  // dense59            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);60            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);61            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);62        }63 64        // for operator_norm65        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);66 67        if (!hparams.is_recr(i)) {68            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);69            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);70            GGML_ASSERT(n_embd_v_gqa == n_embd_k_gqa);71 72            create_tensor_qkv(layer, i, n_embd, n_embd, hparams.n_embd_k_gqa(i), hparams.n_embd_v_gqa(i), 0);73 74            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);75        } else {76            layer.shortconv.conv     = create_tensor(tn(LLM_TENSOR_SHORTCONV_CONV,    "weight", i), {hparams.n_shortconv_l_cache, n_embd}, 0);77            layer.shortconv.in_proj  = create_tensor(tn(LLM_TENSOR_SHORTCONV_INPROJ,  "weight", i), {n_embd, 3 * n_embd}, 0);78            layer.shortconv.out_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_OUTPROJ, "weight", i), {n_embd, n_embd}, 0);79        }80    }81 82    // for LFM2-ColBert-350M83    dense_2_out_layers   = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);84    dense_2_out_layers_b = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "bias"),   {hparams.n_embd_out()        }, TENSOR_NOT_REQUIRED);85}86 87std::unique_ptr<llm_graph_context> llama_model_lfm2::build_arch_graph(const llm_graph_params & params) const {88    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {89        return std::make_unique<graph<true>>(*this, params);90    } else {91        return std::make_unique<graph<false>>(*this, params);92    }93}94 95template <bool iswa>96llama_model_lfm2::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) :97    llm_graph_context(params) {98    using inp_hybrid_type = std::conditional_t<iswa, llm_graph_input_mem_hybrid_iswa,  llm_graph_input_mem_hybrid>;99    using inp_attn_type   = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa,     llm_graph_input_attn_kv>;100    using mem_hybrid_ctx  = std::conditional_t<iswa, llama_memory_hybrid_iswa_context, llama_memory_hybrid_context>;101 102    // lambda helpers for readability103    auto build_dense_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {104        GGML_ASSERT(!model.layers[il].ffn_up_b);105        GGML_ASSERT(!model.layers[il].ffn_gate_b);106        GGML_ASSERT(!model.layers[il].ffn_down_b);107        return build_ffn(cur,108            model.layers[il].ffn_up, NULL, NULL,109            model.layers[il].ffn_gate, NULL, NULL,110            model.layers[il].ffn_down, NULL, NULL,111            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);112    };113    auto build_moe_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {114        return build_moe_ffn(cur,115                model.layers[il].ffn_gate_inp,116                model.layers[il].ffn_up_exps,117                model.layers[il].ffn_gate_exps,118                model.layers[il].ffn_down_exps,119                model.layers[il].ffn_exp_probs_b,120                n_expert, n_expert_used,121                LLM_FFN_SILU, true,122                hparams.expert_weights_scale,123                static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),124                il);125    };126    auto build_attn_block = [&model, this](ggml_tensor *   cur,127                                           ggml_tensor *   inp_pos,128                                           inp_attn_type * inp_attn,129                                           int             il) -> ggml_tensor * {130        GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));131        const auto n_embd_head = hparams.n_embd_head_v();132        const auto n_head_kv   = hparams.n_head_kv(il);133 134        auto [q, k, v] = build_qkv(model.layers[il], cur,135                n_embd_head, n_head, n_head_kv, il);136 137        // qk norm138        q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);139        cb(q, "model.layers.{}.self_attn.q_layernorm", il);140        k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);141        cb(k, "model.layers.{}.self_attn.k_layernorm", il);142 143        // RoPE144        q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,145                          attn_factor, beta_fast, beta_slow);146        k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,147                          attn_factor, beta_fast, beta_slow);148 149        cur = build_attn(inp_attn,150                model.layers[il].wo, NULL, model.layers[il].wo_s,151                q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);152 153        cb(cur, "model.layers.{}.self_attn.out_proj", il);154 155        return cur;156    };157    auto build_shortconv_block = [&model, this](ggml_tensor *        cur,158                                                llm_graph_input_rs * inp_recr,159                                                int                  il) -> ggml_tensor * {160        const auto * mctx_cur = static_cast<const mem_hybrid_ctx *>(mctx)->get_recr();161        const uint32_t kv_head      = mctx_cur->get_head();162        const int64_t  n_seq_tokens = ubatch.n_seq_tokens;163        const int64_t  n_seqs       = ubatch.n_seqs;164        GGML_ASSERT(n_seqs != 0);165        GGML_ASSERT(ubatch.equal_seqs());166        GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);167 168        GGML_ASSERT(hparams.n_shortconv_l_cache > 1);169        const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;170 171        // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}172        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);173 174        auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);175        cb(bcx, "model.layers.{}.conv.in_proj", il);176 177        constexpr auto n_chunks = 3;178        GGML_ASSERT(bcx->ne[0] % n_chunks == 0);179        const auto chunk_size = bcx->ne[0] / n_chunks;180        auto *     b          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],181                                             0 * chunk_size * ggml_element_size(bcx));182        auto *     c          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],183                                             1 * chunk_size * ggml_element_size(bcx));184        auto *     x          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],185                                             2 * chunk_size * ggml_element_size(bcx));186 187        auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));188 189        // read conv state190        auto * conv_state = mctx_cur->get_r_l(il);191        auto * conv_rs    = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);192        auto * conv       = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);193 194        // causal prepends the state, non-causal pads symmetrically for a centered window195        if (hparams.causal_attn) {196            bx = ggml_concat(ctx0, conv, bx, 0);197        } else {198            const int64_t pad = (hparams.n_shortconv_l_cache - 1) / 2;199            auto * left = ggml_cont(ctx0,200                ggml_view_3d(ctx0, conv, pad, hparams.n_embd, n_seqs, conv->nb[1], conv->nb[2], (d_conv - pad) * conv->nb[0]));201            bx = ggml_pad_ext(ctx0, ggml_concat(ctx0, left, bx, 0), 0, pad, 0, 0, 0, 0, 0, 0);202        }203        GGML_ASSERT(bx->ne[0] > conv->ne[0]);204 205        // last d_conv columns is a new conv state206        auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],207                                       (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));208        GGML_ASSERT(ggml_are_same_shape(conv, new_conv));209 210        // write new conv conv state211        ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,212                                               ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),213                                                            kv_head * d_conv * n_embd * ggml_element_size(new_conv))));214 215        auto * conv_kernel = model.layers[il].shortconv.conv;216        auto * conv_out    = ggml_ssm_conv(ctx0, bx, conv_kernel);217        cb(conv_out, "model.layers.{}.conv.conv", il);218 219        auto * y = ggml_mul(ctx0, c, conv_out);220        y        = build_lora_mm(model.layers[il].shortconv.out_proj, y);221        cb(y, "model.layers.{}.conv.out_proj", il);222        // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}223        y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);224 225        return y;226    };227 228    // actual graph construction starts here229    ggml_tensor * cur = build_inp_embd(model.tok_embd);230    cb(cur, "model.embed_tokens", -1);231 232    ggml_build_forward_expand(gf, cur);233 234    inp_hybrid_type * inp_hybrid = nullptr;235    if constexpr (iswa) {236        inp_hybrid = build_inp_mem_hybrid_iswa();237    } else {238        inp_hybrid = build_inp_mem_hybrid();239    }240 241    ggml_tensor * inp_pos     = build_inp_pos();242    ggml_tensor * inp_out_ids = build_inp_out_ids();243 244    for (int il = 0; il < n_layer; ++il) {245        const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);246 247        auto * prev_cur = cur;248        cur             = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);249        cb(cur, "model.layers.{}.operator_norm", il);250 251        cur = hparams.is_recr(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) :252                                    build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);253 254        if (il == n_layer - 1 && inp_out_ids) {255            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);256            prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);257        }258 259        cur = ggml_add(ctx0, prev_cur, cur);260 261        auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);262        cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);263 264        ggml_tensor * ffn_out =265            is_moe_layer ? build_moe_feed_forward(ffn_norm_out, il) : build_dense_feed_forward(ffn_norm_out, il);266        cb(ffn_norm_out, "model.layers.{}.ffn_out", il);267 268        cur = ggml_add(ctx0, cur, ffn_out);269 270        cur = build_cvec(cur, il);271        cb(cur, "l_out", il);272    }273 274    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);275    cb(cur, "result_norm", -1);276    res->t_embd = cur;277 278    if (!cparams.embeddings) {279        cur = build_lora_mm(model.output, cur, model.output_s);280        cb(cur, "result_output", -1);281 282        res->t_logits = cur;283    }284 285    ggml_build_forward_expand(gf, cur);286}287 288// Explicit template instantiations289template struct llama_model_lfm2::graph<true>;290template struct llama_model_lfm2::graph<false>;291