echodict/llama.cpp

1#include "models.h"2 3ggml_cgraph * clip_graph_internvl::build() {4    GGML_ASSERT(model.class_embedding != nullptr);5    GGML_ASSERT(model.position_embeddings != nullptr);6 7    const int n_pos = n_patches + 1;8    ggml_tensor * inp = build_inp();9 10    // add CLS token11    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);12 13    // The larger models use a different ViT, which uses RMS norm instead of layer norm14    // ref: https://github.com/ggml-org/llama.cpp/pull/13443#issuecomment-286978618815    norm_type norm_t = (hparams.n_embd == 3200 && hparams.n_layer == 45)16        ? NORM_TYPE_RMS // 6B ViT (Used by InternVL 2.5/3 - 26B, 38B, 78B)17        : NORM_TYPE_NORMAL; // 300M ViT (Used by all smaller InternVL models)18 19    ggml_tensor * cur = build_vit(20                            inp, n_pos,21                            norm_t,22                            hparams.ffn_op,23                            model.position_embeddings,24                            nullptr);25 26    // remove CLS token27    cur = ggml_view_2d(ctx0, cur,28        n_embd, n_patches,29        ggml_row_size(cur->type, n_embd), 0);30 31    // pixel shuffle32    {33        const int scale_factor = model.hparams.n_merge;34        const int bsz    = 1; // batch size, always 1 for now since we don't support batching35        const int height = n_patches_y;36        const int width  = n_patches_x;37        GGML_ASSERT(scale_factor > 0);38        cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, height / scale_factor, width, bsz);39        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);40        cur = ggml_cont_4d(ctx0, cur,41            n_embd * scale_factor * scale_factor,42            height / scale_factor,43            width / scale_factor,44            bsz);45        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);46        // flatten to 2D47        cur = ggml_cont_2d(ctx0, cur,48            n_embd * scale_factor * scale_factor,49            cur->ne[1] * cur->ne[2]);50    }51 52    // projector (always using GELU activation)53    {54        // projector LayerNorm uses pytorch's default eps = 1e-555        // ref: https://huggingface.co/OpenGVLab/InternVL3-8B-Instruct/blob/a34d3e4e129a5856abfd6aa6de79776484caa14e/modeling_internvl_chat.py#L7956        cur = build_norm(cur, model.mm_0_w, model.mm_0_b, NORM_TYPE_NORMAL, 1e-5, -1);57        cur = build_ffn(cur,58            model.mm_1_w, model.mm_1_b,59            nullptr, nullptr,60            model.mm_3_w, model.mm_3_b,61            FFN_GELU,62            -1);63    }64 65    // build the graph66    ggml_build_forward_expand(gf, cur);67 68    return gf;69}70