diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index aed595e259ed5..12de60442a43e 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -607,7 +607,6 @@ def get_vocab_base(self) -> tuple[list[str], list[int], str]: from transformers import AutoTokenizer tokenizer = AutoTokenizer.from_pretrained(self.dir_model) - vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab)) assert max(tokenizer.vocab.values()) < vocab_size tokpre = self.get_vocab_base_pre(tokenizer) @@ -815,6 +814,9 @@ def get_vocab_base_pre(self, tokenizer) -> str: if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35": # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0 res = "minerva-7b" + if chkhsh == "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664": + # ref: https://huggingface.co/tencent/Hunyuan-A13B-Instruct + res = "hunyuan" if res is None: logger.warning("\n") @@ -6390,6 +6392,125 @@ def set_gguf_parameters(self): super().set_gguf_parameters() self.gguf_writer.add_audio_stack_factor(self.global_config["stack_factor"]) + +@ModelBase.register("HunYuanMoEV1ForCausalLM") +class HunYuanMoEModel(LlamaModel): + model_arch = gguf.MODEL_ARCH.HUNYUAN_MOE + undo_permute = False + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def set_vocab(self): + self._set_vocab_gpt2() + + def get_vocab_base(self) -> tuple[list[str], list[int], str]: + tokens: list[str] = [] + toktypes: list[int] = [] + + from transformers import AutoTokenizer + tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True) + + # merge logic is copied from QwenModel, maybe incorrect + merges = [] + vocab = {} + mergeable_ranks = tokenizer.mergeable_ranks + for token, rank in mergeable_ranks.items(): + vocab[QwenModel.token_bytes_to_string(token)] = rank + if len(token) == 1: + continue + merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank) + if len(merged) == 2: + merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged))) + self.gguf_writer.add_token_merges(merges) + + reverse_vocab = tokenizer.decoder + assert max(reverse_vocab.keys()) < tokenizer.vocab_size + + tokpre = self.get_vocab_base_pre(tokenizer) + added_vocab = tokenizer.get_added_vocab() + + added_tokens_decoder = tokenizer.added_tokens_decoder + + for i in range(tokenizer.vocab_size): + if i not in reverse_vocab: + tokens.append(f"[PAD{i}]") + toktypes.append(gguf.TokenType.UNUSED) + else: + token: str = reverse_vocab[i] + if token in added_vocab: + # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized. + # To avoid unexpected issues - we make sure to normalize non-normalized tokens + if not added_tokens_decoder[i].normalized: + previous_token = token + token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False)) + if previous_token != token: + logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer") + + if added_tokens_decoder[i].special or self.does_token_look_special(token): + toktypes.append(gguf.TokenType.CONTROL) + else: + # NOTE: this was added for Gemma. + # Encoding and decoding the tokens above isn't sufficient for this case. + token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ") # pre-normalize user-defined spaces + toktypes.append(gguf.TokenType.USER_DEFINED) + else: + toktypes.append(gguf.TokenType.NORMAL) + tokens.append(token) + + return tokens, toktypes, tokpre + + def set_gguf_parameters(self): + super().set_gguf_parameters() + + self.gguf_writer.add_expert_count(self.hparams["num_experts"]) + self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["intermediate_size"]) + + moe_intermediate_size = self.hparams["moe_intermediate_size"] + assert all(n == moe_intermediate_size[0] for n in moe_intermediate_size) + self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size[0]) + + moe_topk = self.hparams["moe_topk"] + assert all(topk == moe_topk[0] for topk in moe_topk) + self.gguf_writer.add_expert_used_count(moe_topk[0]) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # process the experts separately + if name.find("mlp.experts") != -1: + n_experts = self.hparams["num_experts"] + assert bid is not None + + tensors: list[tuple[str, Tensor]] = [] + + if self._experts is None: + self._experts = [{} for _ in range(self.block_count)] + + self._experts[bid][name] = data_torch + + if len(self._experts[bid]) >= n_experts * 3: + # merge the experts into a single 3d tensor + for w_name in ["down_proj", "gate_proj", "up_proj"]: + datas: list[Tensor] = [] + + for xid in range(n_experts): + ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight" + datas.append(self._experts[bid][ename]) + del self._experts[bid][ename] + + data_torch = torch.stack(datas, dim=0) + + merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" + + new_name = self.map_tensor_name(merged_name) + + tensors.append((new_name, data_torch)) + + return tensors + else: + return [] + + return [(self.map_tensor_name(name), data_torch)] + ###### CONVERSION LOGIC ###### diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index 2f733f0973686..96a2b692a86c1 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -137,6 +137,7 @@ class TOKENIZER_TYPE(IntEnum): {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"}, {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"}, {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"}, + {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"}, ] diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index fb75143b0b545..e0a4f688ac7a8 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -354,6 +354,7 @@ class MODEL_ARCH(IntEnum): BAILINGMOE = auto() DOTS1 = auto() ARCEE = auto() + HUNYUAN_MOE = auto() class VISION_PROJECTOR_TYPE(IntEnum): @@ -654,6 +655,7 @@ class MODEL_TENSOR(IntEnum): MODEL_ARCH.BAILINGMOE: "bailingmoe", MODEL_ARCH.DOTS1: "dots1", MODEL_ARCH.ARCEE: "arcee", + MODEL_ARCH.HUNYUAN_MOE: "hunyuan-moe", } VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = { @@ -2177,6 +2179,27 @@ class MODEL_TENSOR(IntEnum): MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.HUNYUAN_MOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + 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.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + ], # TODO } diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index b30f77dbe3be7..87d95c776ed9a 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -303,6 +303,7 @@ class TensorNameMap: "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe "model.layers.{bid}.feed_forward.router", # llama4 "encoder.layers.{bid}.mlp.router.layer", # nomic-bert-moe + "model.layers.{bid}.mlp.gate.wg", # hunyuan ), MODEL_TENSOR.FFN_GATE_INP_SHEXP: ( @@ -362,6 +363,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe "model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek deepseek2 "model.layers.{bid}.feed_forward.shared_expert.up_proj", # llama4 + "model.layers.{bid}.mlp.shared_mlp.up_proj", # hunyuan ), # AWQ-activation gate @@ -398,6 +400,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe "model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek deepseek2 "model.layers.{bid}.feed_forward.shared_expert.gate_proj", # llama4 + "model.layers.{bid}.mlp.shared_mlp.gate_proj", # hunyuan ), # Feed-forward down @@ -447,11 +450,13 @@ class TensorNameMap: "model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek deepseek2 "model.layers.{bid}.feed_forward.shared_expert.down_proj", # llama4 "model.layers.{bid}.shared_mlp.output_linear", # granitemoe + "model.layers.{bid}.mlp.shared_mlp.down_proj", # hunyuan ), MODEL_TENSOR.ATTN_Q_NORM: ( "language_model.encoder.layers.{bid}.self_attention.q_layernorm", "model.layers.{bid}.self_attn.q_layernorm", # persimmon + "model.layers.{bid}.self_attn.query_layernorm", # hunyuan "model.layers.{bid}.self_attn.q_norm", # cohere olmoe chameleon olmo2 "transformer.blocks.{bid}.attn.q_ln", # sea-lion "encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2 @@ -461,6 +466,7 @@ class TensorNameMap: MODEL_TENSOR.ATTN_K_NORM: ( "language_model.encoder.layers.{bid}.self_attention.k_layernorm", "model.layers.{bid}.self_attn.k_layernorm", # persimmon + "model.layers.{bid}.self_attn.key_layernorm", # hunyuan "model.layers.{bid}.self_attn.k_norm", # cohere olmoe chameleon olmo2 "transformer.blocks.{bid}.attn.k_ln", # sea-lion "encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2 diff --git a/include/llama.h b/include/llama.h index 3eda9bc68608c..dc86aea41dcbd 100644 --- a/include/llama.h +++ b/include/llama.h @@ -117,6 +117,7 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33, LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34, LLAMA_VOCAB_PRE_TYPE_SEED_CODER = 35, + LLAMA_VOCAB_PRE_TYPE_HUNYUAN = 36, }; enum llama_rope_type { diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 435e3b9ba3db8..403dfe8fcc028 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -76,6 +76,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_BAILINGMOE, "bailingmoe" }, { LLM_ARCH_DOTS1, "dots1" }, { LLM_ARCH_ARCEE, "arcee" }, + { LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -1658,6 +1659,30 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" }, } }, + { + LLM_ARCH_HUNYUAN_MOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { 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_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, { LLM_ARCH_UNKNOWN, { diff --git a/src/llama-arch.h b/src/llama-arch.h index 9181ad053f6b3..14617d2caf061 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -80,6 +80,7 @@ enum llm_arch { LLM_ARCH_BAILINGMOE, LLM_ARCH_DOTS1, LLM_ARCH_ARCEE, + LLM_ARCH_HUNYUAN_MOE, LLM_ARCH_UNKNOWN, }; diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp index 5d317f4ee62eb..d05335a685b49 100644 --- a/src/llama-chat.cpp +++ b/src/llama-chat.cpp @@ -64,6 +64,7 @@ static const std::map LLM_CHAT_TEMPLATES = { { "bailing", LLM_CHAT_TEMPLATE_BAILING }, { "llama4", LLM_CHAT_TEMPLATE_LLAMA4 }, { "smolvlm", LLM_CHAT_TEMPLATE_SMOLVLM }, + { "hunyuan-moe", LLM_CHAT_TEMPLATE_HUNYUAN_MOE }, }; llm_chat_template llm_chat_template_from_str(const std::string & name) { @@ -185,6 +186,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) { return LLM_CHAT_TEMPLATE_LLAMA4; } else if (tmpl_contains("<|endofuserprompt|>")) { return LLM_CHAT_TEMPLATE_DOTS1; + } else if (tmpl_contains("<|startoftext|>") && tmpl_contains("<|extra_4|>")) { + return LLM_CHAT_TEMPLATE_HUNYUAN_MOE; } return LLM_CHAT_TEMPLATE_UNKNOWN; } @@ -665,6 +668,21 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << "<|response|>"; } + } else if (tmpl == LLM_CHAT_TEMPLATE_HUNYUAN_MOE) { + // tencent/Hunyuan-A13B-Instruct + for (auto message : chat) { + std::string role(message->role); + if (role == "system") { + ss << "<|startoftext|>" << message->content << "<|extra_4|>"; + } else if (role == "assistant") { + ss << "<|startoftext|>" << message->content << "<|eos|>"; + } else { + ss << "<|startoftext|>" << message->content << "<|extra_0|>"; + } + } + if (add_ass) { + ss << "<|startoftext|>"; + } } else { // template not supported return -1; diff --git a/src/llama-chat.h b/src/llama-chat.h index 38800010ae48b..b621fda281669 100644 --- a/src/llama-chat.h +++ b/src/llama-chat.h @@ -44,6 +44,7 @@ enum llm_chat_template { LLM_CHAT_TEMPLATE_LLAMA4, LLM_CHAT_TEMPLATE_SMOLVLM, LLM_CHAT_TEMPLATE_DOTS1, + LLM_CHAT_TEMPLATE_HUNYUAN_MOE, LLM_CHAT_TEMPLATE_UNKNOWN, }; diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 71ee431a977ba..a299c89ecc7e4 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -705,6 +705,13 @@ ggml_tensor * llm_graph_context::build_moe_ffn( ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens] cb(weights, "ffn_moe_weights", il); + if (arch == LLM_ARCH_HUNYUAN_MOE) { + weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_expert_used, n_tokens] + weights = ggml_div(ctx0, weights, ggml_sum_rows(ctx0, weights)); // [1, n_tokens] + weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens); // [1, n_expert_used, n_tokens] + cb(weights, "ffn_moe_weights_scaled", il); + } + if (norm_w) { weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); diff --git a/src/llama-model.cpp b/src/llama-model.cpp index fc39195ed5177..172a1d7409caa 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -101,6 +101,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_57B_A14B: return "57B.A14B"; case LLM_TYPE_17B_16E: return "17Bx16E (Scout)"; case LLM_TYPE_17B_128E: return "17Bx128E (Maverick)"; + case LLM_TYPE_A13B: return "A13B"; case LLM_TYPE_30B_A3B: return "30B.A3B"; case LLM_TYPE_235B_A22B: return "235B.A22B"; case LLM_TYPE_E2B: return "E2B"; @@ -1504,6 +1505,22 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_HUNYUAN_MOE: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp); + + // TODO: read from gguf + float n_dim = hparams.n_embd_head_k; + float alpha = 1000.0f; // NTK-Aware + hparams.rope_freq_base_train = 10000.0f * std::powf(alpha, n_dim / (n_dim - 2.0f)); + + switch (hparams.n_layer) { + case 32: type = LLM_TYPE_A13B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; default: throw std::runtime_error("unsupported model architecture"); } @@ -4348,6 +4365,43 @@ 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_HUNYUAN_MOE: + { + 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.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_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_head_k * n_head, n_embd}, 0); + + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -14260,6 +14314,166 @@ struct llm_build_arcee : public llm_graph_context { } }; +struct llm_build_hunyuan_moe : public llm_graph_context { + llm_build_hunyuan_moe(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_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + 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(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + 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 = 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 + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + 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 + ); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, nullptr, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_norm", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, nullptr, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_norm", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, 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); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network (non-MoE) + ggml_tensor * cur_mlp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur_mlp, "ffn_mlp", il); + + // MoE branch + ggml_tensor * cur_moe = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur_moe, "ffn_moe_out", il); + + ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp); + cb(ffn_out, "ffn_out", il); + + cur = ggml_add(ctx0, ffn_out, 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); + } +}; + llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const { llama_memory_i * res; @@ -14635,6 +14849,10 @@ llm_graph_result_ptr llama_model::build_graph( { llm = std::make_unique(*this, params, gf); } break; + case LLM_ARCH_HUNYUAN_MOE: + { + llm = std::make_unique(*this, params, gf); + } break; default: GGML_ABORT("fatal error"); } @@ -14821,6 +15039,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_EXAONE: case LLM_ARCH_MINICPM3: case LLM_ARCH_DOTS1: + case LLM_ARCH_HUNYUAN_MOE: return LLAMA_ROPE_TYPE_NEOX; case LLM_ARCH_QWEN2VL: diff --git a/src/llama-model.h b/src/llama-model.h index 40063b790d434..5bbfa108fc9e0 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -93,6 +93,7 @@ enum llm_type { LLM_TYPE_57B_A14B, LLM_TYPE_17B_16E, // llama4 Scout LLM_TYPE_17B_128E, // llama4 Maverick + LLM_TYPE_A13B, LLM_TYPE_30B_A3B, LLM_TYPE_235B_A22B, LLM_TYPE_E2B, diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 5c9eb87566dde..551bba171c0e0 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -351,6 +351,7 @@ struct llm_tokenizer_bpe : llm_tokenizer { break; case LLAMA_VOCAB_PRE_TYPE_STABLELM2: case LLAMA_VOCAB_PRE_TYPE_QWEN2: + case LLAMA_VOCAB_PRE_TYPE_HUNYUAN: regex_exprs = { // original regex from tokenizer.json // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" @@ -1656,6 +1657,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "seed-coder") { pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER; clean_spaces = false; + } else if ( + tokenizer_pre == "hunyuan") { + pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN; + clean_spaces = false; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); }