model : add hunyuan moe

convert_hf_to_gguf.py

@@ -607,8 +607,7 @@
 
         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)
 
@@ -617,7 +616,7 @@
 
         added_tokens_decoder = tokenizer.added_tokens_decoder
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
                 tokens.append(f"[PAD{i}]")
                 toktypes.append(gguf.TokenType.UNUSED)
@@ -815,6 +814,9 @@
         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 +6392,114 @@
         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()
+
+    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 ######
 
 

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"},
 ]
 
 

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
 }
 

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

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 {

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()));
             }