echodict/llama.cpp

1#include "models.h"2 3ggml_cgraph * clip_graph_glm4v::build() {4    GGML_ASSERT(model.patch_bias != nullptr);5    GGML_ASSERT(model.class_embedding == nullptr);6 7    const int batch_size = 1;8 9    norm_type norm_t = NORM_TYPE_RMS;10 11    ggml_tensor * inp_raw = build_inp_raw();12    ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);13 14    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};15    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches * 4);16    ggml_set_name(positions, "positions");17    ggml_set_input(positions);18 19    GGML_ASSERT(img.nx % (patch_size * 2) == 0);20    GGML_ASSERT(img.ny % (patch_size * 2) == 0);21 22    // second conv dimension23    {24        auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);25        inp = ggml_add(ctx0, inp, inp_1);26 27        inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]28        inp = ggml_cont_4d(29            ctx0, inp,30            n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);31        inp = ggml_reshape_4d(32            ctx0, inp,33            n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));34        inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);35        inp = ggml_cont_3d(36            ctx0, inp,37            n_embd, n_patches_x * n_patches_y, batch_size);38    }39 40    // add patch bias41    inp = ggml_add(ctx0, inp, model.patch_bias);42    cb(inp, "patch_bias", -1);43 44    // pos-conv norm45    inp = build_norm(inp, model.norm_embd_w, model.norm_embd_b, norm_t, eps, -1);46 47    ggml_tensor * learned_pos_embd = nullptr;48    // Note: GLM-OCR does not have learned position embeddings49    if (model.position_embeddings != nullptr) {50        learned_pos_embd = resize_position_embeddings(GGML_SCALE_MODE_BICUBIC);51        learned_pos_embd = ggml_cont_4d(52            ctx0, learned_pos_embd,53            n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);54        learned_pos_embd = ggml_reshape_4d(55            ctx0, learned_pos_embd,56            n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));57        learned_pos_embd = ggml_permute(ctx0, learned_pos_embd, 0, 2, 1, 3);58        learned_pos_embd = ggml_cont_3d(59            ctx0, learned_pos_embd,60            n_embd, n_patches_x * n_patches_y, batch_size);61        cb(learned_pos_embd, "learned_pos_embd", -1);62    }63 64    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {65        return ggml_rope_multi(66                    ctx0, cur, positions, nullptr,67                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION,68                    32768, hparams.rope_theta, 1, 0, 1, 32, 1);69    };70 71    ggml_tensor * cur = build_vit(72                            inp, n_patches,73                            norm_t,74                            hparams.ffn_op,75                            learned_pos_embd,76                            add_pos);77 78    cb(cur, "vit_out", -1);79    // cb(ggml_sum(ctx0, cur), "vit_out_sum", -1);80 81    // GLM4V projector82    // ref: https://github.com/huggingface/transformers/blob/40dc11cd3eb4126652aa41ef8272525affd4a636/src/transformers/models/glm4v/modeling_glm4v.py#L116-L13083 84    // patch merger (downsample)85    {86        int n_merge = hparams.n_merge;87        GGML_ASSERT(n_merge > 0);88 89        int n_token_out = n_patches / n_merge / n_merge;90        cur = ggml_reshape_4d(ctx0, cur, n_embd, n_merge, n_merge, n_token_out);91        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 2, 0, 1, 3)); // [n_merge, n_merge, n_embd, n_token_out]92        cur = ggml_conv_2d(ctx0, model.mm_patch_merger_w, cur, n_merge, n_merge, 0, 0, 1, 1);93        cur = ggml_reshape_2d(ctx0, cur, cur->ne[2], n_token_out); // [n_embd_out, n_token_out]94 95        cur = ggml_add(ctx0, cur, model.mm_patch_merger_b);96    }97 98    // FC projector99    {100        cur = build_mm(model.mm_fc_w, cur);101        // default LayerNorm (post_projection_norm)102        cur = build_norm(cur, model.mm_post_norm_w, model.mm_post_norm_b, NORM_TYPE_NORMAL, 1e-5, -1);103        cur = ggml_gelu_erf(ctx0, cur);104        cb(cur, "after_fc_proj", -1);105    }106 107    // FFN projector108    {109        cur = build_ffn(cur,110            model.mm_ffn_up_w, model.mm_ffn_up_b,111            model.mm_ffn_gate_w, model.mm_ffn_gate_b,112            model.mm_ffn_down_w, model.mm_ffn_down_b,113            hparams.ffn_op, -1);114        cb(cur, "after_ffn_proj", -1);115        // cb(ggml_sum(ctx0, cur), "merged_sum", -1);116    }117 118    // build the graph119    ggml_build_forward_expand(gf, cur);120 121    return gf;122}123