Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_qwen3vl::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false);5    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);6    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);7 8    switch (hparams.n_layer()) {9        case 28: type = LLM_TYPE_1_7B; break;10        case 36: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_8B; break;11        case 64: type = LLM_TYPE_32B; break;12        default: type = LLM_TYPE_UNKNOWN;13    }14}15 16void llama_model_qwen3vl::load_arch_tensors(llama_model_loader &) {17    LLAMA_LOAD_LOCALS;18 19    int64_t n_vocab_out = n_vocab;20    if (arch == LLM_ARCH_QWEN3TTS) {21        n_vocab_out = 3072;22    }23 24    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);25 26    // output27    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);28    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab_out}, TENSOR_NOT_REQUIRED);29    // if output is NULL, init from the input tok embed30    if (output == NULL) {31        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);32    }33 34    // output rerank head35    cls_out = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);36 37    for (int i = 0; i < n_layer; ++i) {38        auto & layer = layers[i];39 40        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);41 42        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);43        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);44 45        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);46        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);47 48        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);49        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);50        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);51        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);52    }53}54 55std::unique_ptr<llm_graph_context> llama_model_qwen3vl::build_arch_graph(const llm_graph_params & params) const {56    return std::make_unique<graph>(*this, params);57}58 59llama_model_qwen3vl::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {60    const size_t n_deepstack_layers = hparams.n_deepstack_layers;61 62    const int64_t n_embd      = hparams.n_embd;63    const int64_t n_embd_head = hparams.n_embd_head_v();64 65    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());66    GGML_ASSERT(n_embd_head == n_rot);67 68    ggml_tensor * cur;69    ggml_tensor * inpL;70 71    inpL = build_inp_embd(model.tok_embd);72 73    int sections[4];74    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);75 76    // inp_pos - contains the positions77    ggml_tensor * inp_pos = build_inp_pos();78 79    auto * inp_attn = build_attn_inp_kv();80 81    ggml_tensor * inp_out_ids = build_inp_out_ids();82 83    for (int il = 0; il < n_layer; ++il) {84        ggml_tensor * inpSA = inpL;85 86        // norm87        cur = build_norm(inpL,88                model.layers[il].attn_norm, NULL,89                LLM_NORM_RMS, il);90        cb(cur, "attn_norm", il);91 92        // self-attention93        {94            // compute Q and K and RoPE them95            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,96                    n_embd_head, n_head, n_head_kv, il);97 98            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);99            cb(Qcur, "Qcur_normed", il);100 101            Qcur = ggml_rope_multi(102                    ctx0, Qcur, inp_pos, nullptr,103                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,104                    ext_factor, attn_factor, beta_fast, beta_slow105                    );106 107            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);108            cb(Kcur, "Kcur_normed", il);109 110            Kcur = ggml_rope_multi(111                    ctx0, Kcur, inp_pos, nullptr,112                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,113                    ext_factor, attn_factor, beta_fast, beta_slow114                    );115 116            cb(Qcur, "Qcur", il);117            cb(Kcur, "Kcur", il);118            cb(Vcur, "Vcur", il);119 120            cur = build_attn(inp_attn,121                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,122                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);123        }124 125        if (il == n_layer - 1 && inp_out_ids) {126            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);127            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);128        }129 130        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);131        cb(ffn_inp, "ffn_inp", il);132 133        // feed-forward network134        cur = build_norm(ffn_inp,135                model.layers[il].ffn_norm, NULL,136                LLM_NORM_RMS, il);137        cb(cur, "ffn_norm", il);138 139        cur = build_ffn(cur,140                model.layers[il].ffn_up,   NULL, NULL,141                model.layers[il].ffn_gate, NULL, NULL,142                model.layers[il].ffn_down, NULL, NULL,143                NULL,144                LLM_FFN_SILU, LLM_FFN_PAR, il);145        cb(cur, "ffn_out", il);146 147        cur = ggml_add(ctx0, cur, ffn_inp);148 149        cur = build_cvec(cur, il);150        cb(cur, "l_out", il);151 152        if (il < (int) n_deepstack_layers) {153            ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));154            cur = ggml_add(ctx0, cur, ds);155            cb(cur, "deepstack_out", il);156        }157 158        // input for next layer159        inpL = cur;160    }161 162    cur = inpL;163 164    cur = build_norm(cur,165            model.output_norm, NULL,166            LLM_NORM_RMS, -1);167 168    cb(cur, "result_norm", -1);169    res->t_embd = cur;170 171    // lm_head172    cur = build_lora_mm(model.output, cur, model.output_s);173 174    int64_t n_vocab_in  = model.tok_embd->ne[1];175    int64_t n_vocab_out = model.output->ne[1];176    if (n_vocab_in > n_vocab_out) {177        // case: Qwen3TTS model with codec_head as output178        GGML_ASSERT(model.output_norm);179        int64_t pad = n_vocab_in - n_vocab_out;180 181        // using this trick to get a scalar -inf tensor to pad the output182        ggml_tensor * neg_inf = ggml_scale_bias(ctx0,183                ggml_view_1d(ctx0, model.output_norm, 1, 0),184                0.0f, -INFINITY);185        neg_inf = ggml_repeat_4d(ctx0, neg_inf, pad, cur->ne[1], 1, 1);186        cur = ggml_concat(ctx0, neg_inf, cur, 0); // [padded .. n_vocab_out, n_stream]187 188    } else if (n_vocab_in < n_vocab_out) {189        GGML_ABORT("invalid case");190    }191 192    cb(cur, "result_output", -1);193    res->t_logits = cur;194 195    ggml_build_forward_expand(gf, cur);196}197