#include "models.h" void llama_model_gemma_embedding2::load_arch_hparams(llama_model_loader & ml) { hparams.swa_type = LLAMA_SWA_TYPE_SYMMETRIC; load_swa_pattern(ml, 6); hparams.causal_attn = false; // embeddings do not use causal attention hparams.f_attention_scale = 1.0f; // same as Gemma4, q_norm makes the scaling unnecessary ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false); ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_EMBEDDING_LENGTH_PER_LAYER, hparams.n_embd_per_layer); switch (hparams.n_layer()) { case 24: type = LLM_TYPE_0_3B; break; default: type = LLM_TYPE_UNKNOWN; } } void llama_model_gemma_embedding2::load_arch_tensors(llama_model_loader &) { LLAMA_LOAD_LOCALS; const int64_t n_embd_per_layer = hparams.n_embd_per_layer; const int64_t n_embd_out = hparams.n_embd_out(); if (n_embd_head_k != n_embd_head_v) { throw std::runtime_error("EmbeddingGemma2 requires n_embd_head_k == n_embd_head_v"); } if (hparams.n_embd_head_k_swa != hparams.n_embd_head_v_swa) { throw std::runtime_error("EmbeddingGemma2 requires n_embd_head_k_swa == n_embd_head_v_swa"); } tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight", 0), {n_embd, n_embd_per_layer * n_layer}, 0); per_layer_proj_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM, "weight", 0), {n_embd_per_layer}, 0); output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); // projects the final hidden state to the embedding dimension output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_embd_out}, 0); for (int i = 0; i < n_layer; ++i) { auto & layer = layers[i]; const int64_t n_head = hparams.n_head(i); const int64_t n_embd_head = hparams.n_embd_head_k(i); const int64_t n_embd_k = hparams.n_embd_k_gqa(i); const int64_t n_embd_v = hparams.n_embd_v_gqa(i); layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head * n_head}, 0); layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k}, 0); layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v}, 0); layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head * n_head, n_embd}, 0); layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head}, 0); layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head}, 0); layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0); layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); layer.per_layer_inp_gate = create_tensor(tn(LLM_TENSOR_PER_LAYER_INP_GATE, "weight", i), {n_embd, n_embd_per_layer}, 0); layer.per_layer_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ, "weight", i), {n_embd_per_layer, n_embd}, 0); layer.per_layer_post_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_POST_NORM, "weight", i), {n_embd}, 0); layer.out_scale = create_tensor(tn(LLM_TENSOR_LAYER_OUT_SCALE, "weight", i), {1u}, 0); } } std::unique_ptr llama_model_gemma_embedding2::build_arch_graph(const llm_graph_params & params) const { return std::make_unique(*this, params); } llama_model_gemma_embedding2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) { ggml_tensor * cur; ggml_tensor * inpL; // important: do not normalize weights for raw embeddings input (i.e. encoded image embeddings) inpL = build_inp_embd(model.tok_embd, sqrtf(n_embd)); cb(inpL, "inp_scaled", -1); // inp_pos - contains the positions ggml_tensor * inp_pos = build_inp_pos(); auto * inp_attn = build_attn_inp_no_cache(); ggml_tensor * inp_out_ids = build_inp_out_ids(); // inp_per_layer shape: [n_embd_per_layer, n_tokens, n_layer] ggml_tensor * inp_per_layer = build_inp_per_layer(inpL); for (int il = 0; il < n_layer; ++il) { const int64_t n_embd_head = hparams.n_embd_head_k(il); const int64_t n_head = hparams.n_head(il); const int64_t n_head_kv = hparams.n_head_kv(il); const float freq_base_l = model.get_rope_freq_base(cparams, il); const float freq_scale_l = model.get_rope_freq_scale(cparams, il); const int n_rot_l = hparams.n_rot(il); // norm cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); cb(cur, "attn_norm", il); // self-attention { ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s); ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s); ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s); cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps); cb(Qcur, "Qcur_normed", il); cb(Kcur, "Kcur_normed", il); cb(Vcur, "Vcur_normed", il); Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, ext_factor, attn_factor, beta_fast, beta_slow); Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, ext_factor, attn_factor, beta_fast, beta_slow); cb(Qcur, "Qcur_pos", il); cb(Kcur, "Kcur_pos", il); cur = build_attn(inp_attn, model.layers[il].wo, nullptr, model.layers[il].wo_s, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, hparams.f_attention_scale, il); } if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } cur = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", il); ggml_tensor * attn_out = ggml_add(ctx0, cur, inpL); cb(attn_out, "attn_out", il); // feed-forward network cur = build_norm(attn_out, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il); cb(cur, "ffn_norm", il); cur = build_ffn(cur, model.layers[il].ffn_up, nullptr, model.layers[il].ffn_up_s, model.layers[il].ffn_gate, nullptr, model.layers[il].ffn_gate_s, model.layers[il].ffn_down, nullptr, model.layers[il].ffn_down_s, nullptr, LLM_FFN_GELU, LLM_FFN_PAR, il); cb(cur, "ffn_out", il); cur = build_norm(cur, model.layers[il].ffn_post_norm, nullptr, LLM_NORM_RMS, -1); cb(cur, "ffn_post_norm", il); cur = ggml_add(ctx0, cur, attn_out); // per-layer embedding { ggml_tensor * pe_in = cur; cb(cur, "pe_in", il); cur = build_lora_mm(model.layers[il].per_layer_inp_gate, cur); // [n_embd_per_layer, n_tokens] cur = ggml_gelu(ctx0, cur); ggml_tensor * inp_this_layer = ggml_view_2d(ctx0, inp_per_layer, inp_per_layer->ne[0], inp_per_layer->ne[1], inp_per_layer->nb[1], il * inp_per_layer->nb[2]); if (il == n_layer - 1 && inp_out_ids) { inp_this_layer = ggml_get_rows(ctx0, inp_this_layer, inp_out_ids); } cur = ggml_mul(ctx0, cur, inp_this_layer); cur = build_lora_mm(model.layers[il].per_layer_proj, cur); // [n_embd, n_tokens] cur = build_norm(cur, model.layers[il].per_layer_post_norm, nullptr, LLM_NORM_RMS, il); cb(cur, "per_layer_embd_out", il); cur = ggml_add(ctx0, pe_in, cur); } // layer_scalar cur = ggml_mul(ctx0, cur, model.layers[il].out_scale); cb(cur, "out_scaled", il); cur = build_cvec(cur, il); cb(cur, "l_out", il); // input for next layer inpL = cur; } cur = build_norm(inpL, model.output_norm, nullptr, LLM_NORM_RMS, -1); cb(cur, "result_norm", -1); // projecting per token is equivalent to projecting after mean pooling cur = build_lora_mm(model.output, cur, model.output_s); cb(cur, "result_embd", -1); res->t_embd = cur; ggml_build_forward_expand(gf, cur); } // equivalent to EmbeddingGemma2TextPLE in python code // this model has no per-layer token embeddings, the per-layer inputs come only from the projection // inpL shape: [n_embd, n_tokens] // output shape: [n_embd_per_layer, n_tokens, n_layer] ggml_tensor * llama_model_gemma_embedding2::graph::build_inp_per_layer(ggml_tensor * inpL) { const int64_t n_embd_per_layer = hparams.n_embd_per_layer; ggml_tensor * cur = ggml_mul_mat(ctx0, model.per_layer_model_proj, inpL); // [n_embd_per_layer * n_layer, n_tokens] cur = ggml_scale(ctx0, cur, 1.0f / sqrtf((float) n_embd)); cur = ggml_reshape_3d(ctx0, cur, n_embd_per_layer, n_layer, n_tokens); cur = build_norm(cur, model.per_layer_proj_norm, nullptr, LLM_NORM_RMS, -1); cb(cur, "inp_per_layer", -1); // permute to shape: [n_embd_per_layer, n_tokens, n_layer] cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 2, 1, 3)); return cur; }