diff --git a/llama.cpp/llama.cpp b/llama.cpp/llama.cpp index 78416fba34..b8ed12d0d0 100644 --- a/llama.cpp/llama.cpp +++ b/llama.cpp/llama.cpp @@ -162,6 +162,7 @@ enum llm_arch { LLM_ARCH_MINICPM, LLM_ARCH_GEMMA, LLM_ARCH_GEMMA2, + LLM_ARCH_GEMMA3, LLM_ARCH_STARCODER2, LLM_ARCH_MAMBA, LLM_ARCH_XVERSE, @@ -211,6 +212,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_MINICPM, "minicpm" }, { LLM_ARCH_GEMMA, "gemma" }, { LLM_ARCH_GEMMA2, "gemma2" }, + { LLM_ARCH_GEMMA3, "gemma3" }, { LLM_ARCH_STARCODER2, "starcoder2" }, { LLM_ARCH_MAMBA, "mamba" }, { LLM_ARCH_XVERSE, "xverse" }, @@ -1008,6 +1010,26 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, }, }, + { + LLM_ARCH_GEMMA3, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, + }, + }, { LLM_ARCH_STARCODER2, { @@ -1935,6 +1957,7 @@ struct llama_hparams { uint32_t n_layer; uint32_t n_rot; uint32_t n_swa = 0; // sliding window attention (SWA) + uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head uint32_t n_expert = 0; @@ -1962,7 +1985,9 @@ struct llama_hparams { float rope_attn_factor = 1.0f; float rope_freq_base_train; + float rope_freq_base_train_swa; float rope_freq_scale_train; + float rope_freq_scale_train_swa; uint32_t n_ctx_orig_yarn; float rope_yarn_log_mul; @@ -2001,6 +2026,7 @@ struct llama_hparams { if (this->n_layer != other.n_layer) return true; if (this->n_rot != other.n_rot) return true; if (this->n_swa != other.n_swa) return true; + if (this->n_swa_pattern != other.n_swa_pattern) return true; if (this->n_embd_head_k != other.n_embd_head_k) return true; if (this->n_embd_head_v != other.n_embd_head_v) return true; if (this->n_expert != other.n_expert) return true; @@ -2035,6 +2061,8 @@ struct llama_hparams { if (!is_float_close(this->rope_attn_factor, other.rope_attn_factor, EPSILON)) return true; if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true; if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true; + if (!is_float_close(this->rope_freq_base_train_swa, other.rope_freq_base_train_swa, EPSILON)) return true; + if (!is_float_close(this->rope_freq_scale_train_swa, other.rope_freq_scale_train_swa, EPSILON)) return true; if (!is_float_close(this->expert_weights_scale, other.expert_weights_scale, EPSILON)) return true; if (!is_float_close(this->rope_yarn_log_mul, other.rope_yarn_log_mul, EPSILON)) return true; if (!is_float_close(this->f_residual_scale, other.f_residual_scale, EPSILON)) return true; @@ -4446,6 +4474,10 @@ static void llm_load_hparams( } hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale; + // by default assume that the sliding-window layers use the same scaling type as the non-sliding-window layers + hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train; + hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train; + ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false); // non-transformer models do not have attention heads @@ -4779,6 +4811,8 @@ static void llm_load_hparams( case LLM_ARCH_GEMMA2: { hparams.n_swa = 4096; // default value of gemma 2 + hparams.n_swa_pattern = 2; + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping, false); @@ -4792,6 +4826,28 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_GEMMA3: + { + hparams.n_swa_pattern = 6; + + hparams.rope_freq_base_train_swa = 10000.0f; + hparams.rope_freq_scale_train_swa = 1.0f; + + 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); + + switch (hparams.n_layer) { + case 26: model.type = e_model::MODEL_1B; break; + case 34: model.type = e_model::MODEL_4B; break; + case 48: model.type = e_model::MODEL_12B; break; + case 62: model.type = e_model::MODEL_27B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + + hparams.f_attention_scale = model.type == e_model::MODEL_27B + ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0))) + : 1.0f / std::sqrt(float(hparams.n_embd_head_k)); + } break; case LLM_ARCH_STARCODER2: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -6844,6 +6900,38 @@ static bool llm_load_tensors( layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}); } } break; + case LLM_ARCH_GEMMA3: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + layer.attn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); + } + } break; case LLM_ARCH_STARCODER2: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -11727,7 +11815,8 @@ struct llm_build_context { for (int il = 0; il < n_layer; ++il) { // (il % 2) layers use SWA - struct ggml_tensor * KQ_mask_l = (il % 2 == 0) ? KQ_mask_swa : KQ_mask; + const bool is_swa = il % hparams.n_swa_pattern < (hparams.n_swa_pattern - 1); + struct ggml_tensor * KQ_mask_l = is_swa ? KQ_mask_swa : KQ_mask; // norm cur = llm_build_norm(ctx0, inpL, hparams, @@ -11839,6 +11928,142 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_gemma3() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); + + const int64_t n_embd_head_k = hparams.n_embd_head_k; + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // TODO: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(true); + struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(true); + + for (int il = 0; il < n_layer; ++il) { + const bool is_swa = il % hparams.n_swa_pattern < (hparams.n_swa_pattern - 1); + + const float freq_base_l = is_swa ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base; + const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale; + + struct ggml_tensor * KQ_mask_l = is_swa ? KQ_mask_swa : KQ_mask; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens); + Qcur = llm_build_norm(ctx0, Qcur, hparams, + model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Qcur, "Qcur", il); + + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens); + Kcur = llm_build_norm(ctx0, Kcur, hparams, + model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, lctx, kv_self, gf, + model.layers[il].wo, NULL, + Kcur, Vcur, Qcur, KQ_mask_l, n_tokens, kv_head, n_kv, hparams.f_attention_scale, cb, il); + } + + cur = llm_build_norm(ctx0, cur, hparams, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_post_norm", il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = llm_build_norm(ctx0, sa_out, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = llm_build_ffn(ctx0, lctx, cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = llm_build_norm(ctx0, cur, hparams, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, sa_out); + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + + cb(cur, "result_norm", -1); + + // lm_head + cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); + + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } struct ggml_cgraph * build_starcoder2() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); @@ -14136,6 +14361,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_gemma2(); } break; + case LLM_ARCH_GEMMA3: + { + result = llm.build_gemma3(); + } break; case LLM_ARCH_STARCODER2: { result = llm.build_starcoder2(); @@ -17315,6 +17544,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_PHI3: case LLM_ARCH_GEMMA: case LLM_ARCH_GEMMA2: + case LLM_ARCH_GEMMA3: case LLM_ARCH_STARCODER2: case LLM_ARCH_OPENELM: case LLM_ARCH_GPTNEOX: