echodict/llama.cpp

1#include "models.h"2 3// JAIS-2 model graph builder4// Uses: LayerNorm (not RMSNorm), relu2 activation, separate Q/K/V, RoPE embeddings5llm_build_jais2::llm_build_jais2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {6    const int64_t n_embd_head = hparams.n_embd_head_v();7 8    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());9    GGML_ASSERT(n_embd_head == n_rot);10 11    ggml_tensor * cur;12    ggml_tensor * inpL;13 14    inpL = build_inp_embd(model.tok_embd);15 16    // inp_pos - contains the positions17    ggml_tensor * inp_pos = build_inp_pos();18 19    // KV input for attention20    auto * inp_attn = build_attn_inp_kv();21 22    ggml_tensor * inp_out_ids = build_inp_out_ids();23 24    for (int il = 0; il < n_layer; ++il) {25        // Pre-attention LayerNorm26        cur = build_norm(inpL,27                model.layers[il].attn_norm,28                model.layers[il].attn_norm_b,29                LLM_NORM, il);30        cb(cur, "attn_norm", il);31 32        // Self-attention with separate Q, K, V projections33        {34            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,35                    n_embd_head, n_head, n_head_kv, il);36 37            // Apply RoPE38            Qcur = ggml_rope_ext(39                ctx0, Qcur, inp_pos, nullptr,40                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,41                ext_factor, attn_factor, beta_fast, beta_slow42            );43 44            Kcur = ggml_rope_ext(45                ctx0, Kcur, inp_pos, nullptr,46                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,47                ext_factor, attn_factor, beta_fast, beta_slow48            );49 50            cb(Qcur, "Qcur_rope", il);51            cb(Kcur, "Kcur_rope", il);52 53            cur = build_attn(inp_attn,54                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,55                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);56        }57 58        if (il == n_layer - 1 && inp_out_ids) {59            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);60            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);61        }62 63        // Residual connection64        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);65        cb(ffn_inp, "ffn_inp", il);66 67        // Pre-FFN LayerNorm68        cur = build_norm(ffn_inp,69                model.layers[il].ffn_norm,70                model.layers[il].ffn_norm_b,71                LLM_NORM, il);72        cb(cur, "ffn_norm", il);73 74        // FFN with relu2 activation (ReLU squared) - no gate projection75        // up -> relu2 -> down76        cur = build_ffn(cur,77                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,78                NULL, NULL, NULL,  // no gate79                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,80                NULL,81                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);82        cb(cur, "ffn_out", il);83 84        // Residual connection85        inpL = ggml_add(ctx0, cur, ffn_inp);86        inpL = build_cvec(inpL, il);87        cb(inpL, "l_out", il);88    }89 90    // Final LayerNorm91    cur = build_norm(inpL,92            model.output_norm,93            model.output_norm_b,94            LLM_NORM, -1);95    cb(cur, "result_norm", -1);96 97    res->t_embd = cur;98 99    // Output projection100    cur = build_lora_mm(model.output, cur);101    cb(cur, "result_output", -1);102 103    res->t_logits = cur;104 105    ggml_build_forward_expand(gf, cur);106}107