From e0cb5c5cb8a61ac232130cf6bf878035f93824d9 Mon Sep 17 00:00:00 2001 From: lgai-exaone Date: Fri, 18 Jul 2025 17:45:49 +0900 Subject: [PATCH 1/5] model : add EXAONE 4.0 support (#14630) --- convert_hf_to_gguf.py | 72 +++++++++++++ convert_hf_to_gguf_update.py | 1 + gguf-py/gguf/constants.py | 19 ++++ src/llama-arch.cpp | 21 ++++ src/llama-arch.h | 1 + src/llama-chat.cpp | 20 ++++ src/llama-chat.h | 1 + src/llama-model.cpp | 195 +++++++++++++++++++++++++++++++++++ src/llama-vocab.cpp | 3 + 9 files changed, 333 insertions(+) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index d9185c80600..c8bf3c53830 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -843,6 +843,9 @@ def get_vocab_base_pre(self, tokenizer) -> str: if chkhsh == "169bf0296a13c4d9b7672313f749eb36501d931022de052aad6e36f2bf34dd51": # ref: https://huggingface.co/LiquidAI/LFM2-Tokenizer res = "lfm2" + if chkhsh == "2085e1638f6c377a0aa4ead21b27bb4cb941bf800df86ed391011769c1758dfb": + # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B + res = "exaone4" if res is None: logger.warning("\n") @@ -6780,6 +6783,75 @@ def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32)) +@ModelBase.register("Exaone4ForCausalLM") +class Exaone4Model(TextModel): + model_arch = gguf.MODEL_ARCH.EXAONE4 + + def set_vocab(self): + tokens, toktypes, tokpre = self.get_vocab_base() + self.gguf_writer.add_tokenizer_model("gpt2") + self.gguf_writer.add_tokenizer_pre(tokpre) + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True) + special_vocab.add_to_gguf(self.gguf_writer) + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + + if hparams.get("sliding_window") is not None: + self.gguf_writer.add_sliding_window(hparams["sliding_window"]) + if "layer_types" in hparams: + self.gguf_writer.add_sliding_window_pattern([t == "sliding_attention" for t in hparams["layer_types"]]) + elif "sliding_window_pattern" in hparams: + sliding_window_pattern = [] + if isinstance(hparams["sliding_window_pattern"], str): # e.g. LLLG + for i in range(hparams["num_hidden_layers"]): + sliding_window_pattern.append(hparams["sliding_window_pattern"][i % len(hparams["sliding_window_pattern"])] == "L") + if isinstance(hparams["sliding_window_pattern"], int): # e.g. 4 + for i in range(hparams["num_hidden_layers"]): + sliding_window_pattern.append((i + 1) % hparams["sliding_window_pattern"] != 0) + if len(sliding_window_pattern) == hparams["num_hidden_layers"]: + self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern) + + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) + + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): + if rope_scaling.get("rope_type", '').lower() == "llama3": + base = self.hparams.get("rope_theta", 10_000.0) + if (dim := self.hparams.get("head_dim")) is None: + dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"] + freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + + factor = rope_scaling.get("factor", 16.0) + low_freq_factor = rope_scaling.get("low_freq_factor", 1.0) + high_freq_factor = rope_scaling.get("high_freq_factor", 4.0) + old_context_len = self.hparams.get("original_max_position_embeddings", 8192) + + low_freq_wavelen = old_context_len / low_freq_factor + high_freq_wavelen = old_context_len / high_freq_factor + + rope_factors = [] + for freq in freqs: + wavelen = 2 * math.pi / freq + if wavelen < high_freq_wavelen: + rope_factors.append(1) + elif wavelen > low_freq_wavelen: + rope_factors.append(factor) + else: + smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) + rope_factors.append(1 / ((1 - smooth) / factor + smooth)) + + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32)) + + @ModelBase.register("GraniteForCausalLM") class GraniteModel(LlamaModel): """Conversion for IBM's GraniteForCausalLM""" diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index f7b6d97b19c..abaf2ea9a12 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -129,6 +129,7 @@ class TOKENIZER_TYPE(IntEnum): {"name": "a.x-4.0", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/skt/A.X-4.0", }, {"name": "midm-2.0", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", }, {"name": "lfm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"}, + {"name": "exaone4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", }, ] # some models are known to be broken upstream, so we will skip them as exceptions diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index a8f5947ac33..40e809f1ac8 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -354,6 +354,7 @@ class MODEL_ARCH(IntEnum): JAIS = auto() NEMOTRON = auto() EXAONE = auto() + EXAONE4 = auto() GRANITE = auto() GRANITE_MOE = auto() GRANITE_HYBRID = auto() @@ -671,6 +672,7 @@ class MODEL_TENSOR(IntEnum): MODEL_ARCH.JAIS: "jais", MODEL_ARCH.NEMOTRON: "nemotron", MODEL_ARCH.EXAONE: "exaone", + MODEL_ARCH.EXAONE4: "exaone4", MODEL_ARCH.GRANITE: "granite", MODEL_ARCH.GRANITE_MOE: "granitemoe", MODEL_ARCH.GRANITE_HYBRID: "granitehybrid", @@ -2197,6 +2199,23 @@ class MODEL_TENSOR(IntEnum): MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.EXAONE4: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_POST_NORM, + ], MODEL_ARCH.GRANITE: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index df3fc5d3e74..814ac93a6d8 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -68,6 +68,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_JAIS, "jais" }, { LLM_ARCH_NEMOTRON, "nemotron" }, { LLM_ARCH_EXAONE, "exaone" }, + { LLM_ARCH_EXAONE4, "exaone4" }, { LLM_ARCH_RWKV6, "rwkv6" }, { LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" }, { LLM_ARCH_RWKV7, "rwkv7" }, @@ -1510,6 +1511,26 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_EXAONE4, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { 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_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_RWKV6, { diff --git a/src/llama-arch.h b/src/llama-arch.h index 3bffe359eab..d09b7d7810b 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -72,6 +72,7 @@ enum llm_arch { LLM_ARCH_JAIS, LLM_ARCH_NEMOTRON, LLM_ARCH_EXAONE, + LLM_ARCH_EXAONE4, LLM_ARCH_RWKV6, LLM_ARCH_RWKV6QWEN2, LLM_ARCH_RWKV7, diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp index 240937eceee..80072ad2713 100644 --- a/src/llama-chat.cpp +++ b/src/llama-chat.cpp @@ -56,6 +56,7 @@ static const std::map LLM_CHAT_TEMPLATES = { { "glmedge", LLM_CHAT_TEMPLATE_GLMEDGE }, { "minicpm", LLM_CHAT_TEMPLATE_MINICPM }, { "exaone3", LLM_CHAT_TEMPLATE_EXAONE_3 }, + { "exaone4", LLM_CHAT_TEMPLATE_EXAONE_4 }, { "rwkv-world", LLM_CHAT_TEMPLATE_RWKV_WORLD }, { "granite", LLM_CHAT_TEMPLATE_GRANITE }, { "gigachat", LLM_CHAT_TEMPLATE_GIGACHAT }, @@ -168,6 +169,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) { } else if (tmpl_contains(LU8("<｜Assistant｜>")) && tmpl_contains(LU8("<｜User｜>")) && tmpl_contains(LU8("<｜end▁of▁sentence｜>"))) { return LLM_CHAT_TEMPLATE_DEEPSEEK_3; } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) { + if (tmpl_contains("[|tool|]")) { + return LLM_CHAT_TEMPLATE_EXAONE_4; + } // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb // EXAONE-3.0-7.8B-Instruct return LLM_CHAT_TEMPLATE_EXAONE_3; @@ -532,6 +536,22 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << "[|assistant|]"; } + } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_4) { + for (auto message : chat) { + std::string role(message->role); + if (role == "system") { + ss << "[|system|]" << trim(message->content) << "[|endofturn|]\n"; + } else if (role == "user") { + ss << "[|user|]" << trim(message->content) << "\n"; + } else if (role == "assistant") { + ss << "[|assistant|]" << trim(message->content) << "[|endofturn|]\n"; + } else if (role == "tool") { + ss << "[|tool|]" << trim(message->content) << "[|endofturn|]\n"; + } + } + if (add_ass) { + ss << "[|assistant|]"; + } } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) { // this template requires the model to have "\n\n" as EOT token for (size_t i = 0; i < chat.size(); i++) { diff --git a/src/llama-chat.h b/src/llama-chat.h index cab05334856..6968a19fbe1 100644 --- a/src/llama-chat.h +++ b/src/llama-chat.h @@ -35,6 +35,7 @@ enum llm_chat_template { LLM_CHAT_TEMPLATE_GLMEDGE, LLM_CHAT_TEMPLATE_MINICPM, LLM_CHAT_TEMPLATE_EXAONE_3, + LLM_CHAT_TEMPLATE_EXAONE_4, LLM_CHAT_TEMPLATE_RWKV_WORLD, LLM_CHAT_TEMPLATE_GRANITE, LLM_CHAT_TEMPLATE_GIGACHAT, diff --git a/src/llama-model.cpp b/src/llama-model.cpp index b88f4ebc5c0..cd3e456948c 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -1490,6 +1490,23 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_EXAONE4: + { + if (hparams.n_layer == 64) { // 32B + hparams.swa_type = LLAMA_SWA_TYPE_STANDARD; + hparams.n_swa = 4096; + hparams.set_swa_pattern(4); + } + + 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); + + switch (hparams.n_layer) { + case 30: type = LLM_TYPE_1_2B; break; + case 64: type = LLM_TYPE_32B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_RWKV6: case LLM_ARCH_RWKV6QWEN2: { @@ -4355,6 +4372,39 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); } } break; + case LLM_ARCH_EXAONE4: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); + + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "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_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); + } + } break; case LLM_ARCH_RWKV6: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -13478,6 +13528,142 @@ struct llm_build_exaone : public llm_graph_context { } }; +template +struct llm_build_exaone4 : public llm_graph_context { + llm_build_exaone4(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_k; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_v); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + using inp_attn_type = std::conditional_t; + inp_attn_type * inp_attn = nullptr; + + if constexpr (iswa) { + inp_attn = build_attn_inp_kv_unified_iswa(); + } else { + inp_attn = build_attn_inp_kv_unified(); + } + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // use RoPE for SWA layers or non-SWA models + const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE; + + cur = inpL; + + // self-attention + { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + 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, NULL, LLM_NORM_RMS, il); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + cb(Kcur, "Kcur_normed", il); + + if (use_rope) { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "attn_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_ffn(ffn_inp, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + struct llm_build_rwkv6_base : public llm_graph_context { const llama_model & model; @@ -17163,6 +17349,14 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; + case LLM_ARCH_EXAONE4: + { + if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) { + llm = std::make_unique>(*this, params, gf); + } else { + llm = std::make_unique>(*this, params, gf); + } + } break; case LLM_ARCH_RWKV6: { llm = std::make_unique(*this, params); @@ -17430,6 +17624,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_ORION: case LLM_ARCH_NEMOTRON: case LLM_ARCH_EXAONE: + case LLM_ARCH_EXAONE4: case LLM_ARCH_MINICPM3: case LLM_ARCH_DOTS1: case LLM_ARCH_HUNYUAN_MOE: diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 2181c01e31a..e8bae645088 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -1925,6 +1925,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { } else if ( tokenizer_pre == "exaone") { pre_type = LLAMA_VOCAB_PRE_TYPE_EXAONE; + } else if ( + tokenizer_pre == "exaone4") { + pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2; } else if ( tokenizer_pre == "chameleon") { pre_type = LLAMA_VOCAB_PRE_TYPE_CHAMELEON; From eacdeb5bfcb6c6cd54461fd0e9f04cab78bf975b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 18 Jul 2025 11:53:55 +0300 Subject: [PATCH 2/5] model : fix build after merge conflict (#14754) --- src/llama-model.cpp | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/src/llama-model.cpp b/src/llama-model.cpp index cd3e456948c..2d90ec1ac68 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -13530,7 +13530,7 @@ struct llm_build_exaone : public llm_graph_context { template struct llm_build_exaone4 : public llm_graph_context { - llm_build_exaone4(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + llm_build_exaone4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { const int64_t n_embd_head = hparams.n_embd_head_k; GGML_ASSERT(n_embd_head == hparams.n_embd_head_v); @@ -13603,7 +13603,7 @@ struct llm_build_exaone4 : public llm_graph_context { cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); - cur = build_attn(inp_attn, gf, + cur = build_attn(inp_attn, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); cb(cur, "attn_out", il); @@ -17352,9 +17352,9 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { case LLM_ARCH_EXAONE4: { if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) { - llm = std::make_unique>(*this, params, gf); + llm = std::make_unique>(*this, params); } else { - llm = std::make_unique>(*this, params, gf); + llm = std::make_unique>(*this, params); } } break; case LLM_ARCH_RWKV6: From d498af3d5a00f96bdd37b534860f03a6d9e98d39 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 18 Jul 2025 14:31:15 +0300 Subject: [PATCH 3/5] graph : avoid huge warm-up graphs for MoE models (#14753) * graph : avoid huge warm-up graphs for MoE models ggml-ci * cont : bump max nodes to 8x model tensors --- src/llama-context.cpp | 2 +- src/llama-graph.cpp | 7 +++++-- 2 files changed, 6 insertions(+), 3 deletions(-) diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 1af19caa39d..6eb344736de 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -1312,7 +1312,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) { // uint32_t llama_context::graph_max_nodes() const { - return std::max(65536u, 5u*model.n_tensors()); + return std::max(1024u, 8u*model.n_tensors()); } llm_graph_result * llama_context::get_gf_res_reserve() const { diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 7ea7fd6156e..7cac3b98fa9 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -906,8 +906,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn( } // aggregate experts + // note: here we explicitly use hparams.n_expert_used instead of n_expert_used + // to avoid potentially a large number of add nodes during warmup + // ref: https://github.com/ggml-org/llama.cpp/pull/14753 ggml_tensor * moe_out = nullptr; - for (int i = 0; i < n_expert_used; ++i) { + for (uint32_t i = 0; i < hparams.n_expert_used; ++i) { ggml_tensor * cur_expert = ggml_view_2d(ctx0, experts, n_embd, n_tokens, experts->nb[2], i*experts->nb[1]); @@ -918,7 +921,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn( } } - if (n_expert_used == 1) { + if (hparams.n_expert_used == 1) { // avoid returning a non-contiguous tensor moe_out = ggml_cont(ctx0, moe_out); } From 021cc28bef4dd7d0bf9c91dbbd0803caa6cb15f2 Mon Sep 17 00:00:00 2001 From: Oliver Simons Date: Fri, 18 Jul 2025 13:35:32 +0200 Subject: [PATCH 4/5] cuda : Fix Gemma3n not executed as CUDA_GRAPH on NVGPUs (#14741) * Fix Gemma3n not executed as CUDA_GRAPH on NVGPUs Gemma3n uses Matrix-Matrix addition as part of their input processing, wrongly triggering CUDA_GRAPH disablement on NVGPUs even when batch-size of 1 is used. * Exclude `project_per_layer_input` by matching node names This ensures that all other graphs which don't exhibit this pattern do not have their behavior changed. * Revert unnecessary formatting changes --- ggml/src/ggml-cuda/ggml-cuda.cu | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 50a977c3076..dfc50ef0daf 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -2590,6 +2590,9 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud // Loop over nodes in GGML graph to obtain info needed for CUDA graph cuda_ctx->cuda_graph->cpy_dest_ptrs.clear(); + const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected"; + const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj"; + for (int i = 0; i < cgraph->n_nodes; i++) { ggml_tensor * node = cgraph->nodes[i]; @@ -2611,9 +2614,12 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud #endif } - if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) { - // disable CUDA graphs for batch size > 1 for now. - // Changes in batch size or context size can cause changes to the grid size of some kernels. + if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1 && (node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) && (node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true)) { + // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation + // by means of matching node names. See + // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and + // https://github.com/huggingface/transformers/blob/bda75b4011239d065de84aa3e744b67ebfa7b245/src/transformers/models/gemma3n/modeling_gemma3n.py#L1773, + // Generally, changes in batch size or context size can cause changes to the grid size of some kernels. use_cuda_graph = false; #ifndef NDEBUG GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]); From 2adf8d83acdb9b1bf58db6c9729ac9dc6847a58b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 18 Jul 2025 17:33:41 +0300 Subject: [PATCH 5/5] parallel : add option for different RNG seeds (#14757) ggml-ci --- examples/parallel/parallel.cpp | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp index 46fb451baa7..e48f48fc322 100644 --- a/examples/parallel/parallel.cpp +++ b/examples/parallel/parallel.cpp @@ -184,6 +184,9 @@ int main(int argc, char ** argv) { // extra text to insert in each client's prompt in order to make it larger const int32_t n_junk = std::max(1, params.n_junk); + // signed seed, use negative values to indicate different seeds for the different clients + const int32_t & sseed = params.sampling.seed; + // init llama.cpp llama_backend_init(); llama_numa_init(params.numa); @@ -219,12 +222,21 @@ int main(int argc, char ** argv) { const int n_ctx = llama_n_ctx(ctx); + if (sseed >= 0) { + LOG_INF("%s: initializing all samplers with the same RNG seed: %d (use a negative seed to have different seeds)\n", __func__, sseed); + } else { + LOG_INF("%s: initializing samplers with different RNG seeds, starting from %d\n", __func__, sseed); + } + std::vector clients(n_clients); for (size_t i = 0; i < clients.size(); ++i) { auto & client = clients[i]; client.id = i; client.smpl = common_sampler_init(model, params.sampling); - //params.sampling.seed++; + + if (sseed < 0) { + params.sampling.seed--; + } } std::vector tokens_system;