Add support for GLM-4.5 models

convert_hf_to_gguf.py

@@ -618,6 +618,15 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
             # ref: https://huggingface.co/THUDM/glm-4-9b-chat
             res = "chatglm-bpe"
+        if chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
+            res = "glm4"
+        if chkhsh == "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902":
+            # ref: https://huggingface.co/zai-org/GLM-4.5-Air, https://huggingface.co/zai-org/GLM-4.5
+            res = "glm4"
         if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee":
             # ref: https://huggingface.co/LumiOpen/Viking-7B
             res = "viking"
@@ -3948,6 +3957,137 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
             return [(self.map_tensor_name(name), data_torch)]
         return super().modify_tensors(data_torch, name, bid)
 
+@Model.register("Glm4MoeForCausalLM")
+class Glm4MoeModel(Model):
+    model_arch = gguf.MODEL_ARCH.GLM4_MOE
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # GLM4_MOE has num_hidden_layers + 1 actual layers (including NextN layer)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        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 tokens
+        # Note: Using <|endoftext|> (151329) for eot causes endless generation
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
+        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
+
+        # Patch broken chat template
+        if isinstance(special_vocab.chat_template, str) and "visible_text(m.content).endswith" in special_vocab.chat_template:
+            special_vocab.chat_template = special_vocab.chat_template.replace(
+                """{{ visible_text(m.content) }}\n{{- '/nothink' if (enable_thinking is defined and not enable_thinking and not visible_text(m.content).endswith("/nothink")) else '' -}}""",
+                """{% set content = visible_text(m.content) %}{{ content }}\n{{- '/nothink' if (enable_thinking is defined and not enable_thinking and not content.endswith("/nothink")) else '' -}}""")
+
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if (rope_dim := self.hparams.get("head_dim")) is None:
+            rope_dim = (
+                self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+            )
+        self.gguf_writer.add_rope_dimension_count(
+            int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5))
+        )
+
+        # MoE parameters - Use only routed expert count (shared experts handled separately)
+        if (n_routed_experts := self.hparams.get("n_routed_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_routed_experts)
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+        if (n_shared_experts := self.hparams.get("n_shared_experts")) is not None:
+            self.gguf_writer.add_expert_shared_count(n_shared_experts)
+        if (first_k_dense_replace := self.hparams.get("first_k_dense_replace")) is not None:
+            self.gguf_writer.add_leading_dense_block_count(first_k_dense_replace)
+
+        # Expert gating function (sigmoid for GLM4_MOE)
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+
+        # Routed scaling factor
+        if (routed_scaling_factor := self.hparams.get("routed_scaling_factor")) is not None:
+            self.gguf_writer.add_expert_weights_scale(routed_scaling_factor)
+
+        # Normalise topk probabilities
+        if (norm_topk_prob := self.hparams.get("norm_topk_prob")) is not None:
+            self.gguf_writer.add_expert_weights_norm(norm_topk_prob)
+
+        # NextN/MTP prediction layers
+        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(
+        self, data_torch: Tensor, name: str, bid: int | None
+    ) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("model.visual."):  # ignore visual part
+            return []
+        elif name.startswith("model.language_model."):
+            name = name.replace("language_model.", "")  # for multimodal variants
+
+        # Handle main token embedding (but not layer-specific NextN embeddings)
+        if name == "model.embed_tokens.weight" and ".layers." not in name:
+            return [(self.map_tensor_name("token_embd.weight"), data_torch)]
+
+        # Handle routed experts
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["n_routed_experts"]
+            assert bid is not None
+
+            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:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # 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 []
+
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        new_name = self.map_tensor_name(name)
+
+        return [(new_name, data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
 
 @Model.register("ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(Model):

convert_hf_to_gguf_update.py

@@ -96,6 +96,8 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
     {"name": "deepseek-v3",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
     {"name": "seed-coder",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
+    {"name": "glm4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf",       "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2", },
+    {"name": "glm4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.5-Air",     "chkhsh": "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902", },
     {"name": "kimi-k2",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/moonshotai/Kimi-K2-Base", "chkhsh": "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890", },
 ]
 

gguf-py/gguf/constants.py

@@ -91,6 +91,7 @@ class LLM:
         EXPERT_WEIGHTS_SCALE              = "{arch}.expert_weights_scale"
         EXPERT_WEIGHTS_NORM               = "{arch}.expert_weights_norm"
         EXPERT_GATING_FUNC                = "{arch}.expert_gating_func"
+        NEXTN_PREDICT_LAYERS              = "{arch}.nextn_predict_layers"
         POOLING_TYPE                      = "{arch}.pooling_type"
         LOGIT_SCALE                       = "{arch}.logit_scale"
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
@@ -159,6 +160,13 @@ class Tokenizer:
         CHAT_TEMPLATE_N      = "tokenizer.chat_template.{name}"
         CHAT_TEMPLATES       = "tokenizer.chat_templates"
         # FIM/Infill special tokens constants
+        FIM_PRE_ID           = "tokenizer.ggml.fim_pre_token_id"
+        FIM_SUF_ID           = "tokenizer.ggml.fim_suf_token_id"
+        FIM_MID_ID           = "tokenizer.ggml.fim_mid_token_id"
+        FIM_PAD_ID           = "tokenizer.ggml.fim_pad_token_id"
+        FIM_REP_ID           = "tokenizer.ggml.fim_rep_token_id"
+        FIM_SEP_ID           = "tokenizer.ggml.fim_sep_token_id"
+        # FIM/Infill special tokens constants
         PREFIX_ID            = "tokenizer.ggml.prefix_token_id"
         SUFFIX_ID            = "tokenizer.ggml.suffix_token_id"
         MIDDLE_ID            = "tokenizer.ggml.middle_token_id"
@@ -220,6 +228,7 @@ class MODEL_ARCH(IntEnum):
     OPENELM      = auto()
     ARCTIC       = auto()
     DEEPSEEK2    = auto()
+    GLM4_MOE     = auto()
     CHATGLM      = auto()
     BITNET       = auto()
     BITNET_25    = auto()
@@ -314,6 +323,12 @@ class MODEL_TENSOR(IntEnum):
     ENC_FFN_DOWN         = auto()
     ENC_FFN_UP           = auto()
     ENC_OUTPUT_NORM      = auto()
+    NEXTN_EH_PROJ        = auto()   # nextn tensors (glm4moe)
+    NEXTN_EMBED_TOKENS   = auto()   # nextn tensors (glm4moe)
+    NEXTN_ENORM          = auto()   # nextn tensors (glm4moe)
+    NEXTN_HNORM          = auto()   # nextn tensors (glm4moe)
+    NEXTN_SHARED_HEAD_HEAD = auto() # nextn tensors (glm4moe)
+    NEXTN_SHARED_HEAD_NORM = auto() # nextn tensors (glm4moe)
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -358,6 +373,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.ARCTIC:         "arctic",
     MODEL_ARCH.DEEPSEEK2:      "deepseek2",
     MODEL_ARCH.CHATGLM:        "chatglm",
+    MODEL_ARCH.GLM4_MOE:         "glm4moe",
     MODEL_ARCH.BITNET:         "bitnet",
     MODEL_ARCH.BITNET_25:      "bitnet-25",
     MODEL_ARCH.T5:             "t5",
@@ -451,6 +467,13 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.ENC_FFN_DOWN:         "enc.blk.{bid}.ffn_down",
     MODEL_TENSOR.ENC_FFN_UP:           "enc.blk.{bid}.ffn_up",
     MODEL_TENSOR.ENC_OUTPUT_NORM:      "enc.output_norm",
+    # NextN/MTP
+    MODEL_TENSOR.NEXTN_EH_PROJ:             "blk.{bid}.nextn.eh_proj",
+    MODEL_TENSOR.NEXTN_EMBED_TOKENS:        "blk.{bid}.nextn.embed_tokens",
+    MODEL_TENSOR.NEXTN_ENORM:               "blk.{bid}.nextn.enorm",
+    MODEL_TENSOR.NEXTN_HNORM:               "blk.{bid}.nextn.hnorm",
+    MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD:    "blk.{bid}.nextn.shared_head_head",
+    MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM:    "blk.{bid}.nextn.shared_head_norm",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -1070,6 +1093,37 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.GLM4_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        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,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        # NextN/MTP tensors - preserved but unused
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
+    ],
     MODEL_ARCH.BITNET: [
         MODEL_TENSOR.ATTN_Q,
         MODEL_TENSOR.ATTN_K,
@@ -1633,6 +1687,14 @@ def get_type(val: Any) -> GGUFValueType:
 KEY_TOKENIZER_MASK_ID    = Keys.Tokenizer.MASK_ID
 KEY_TOKENIZER_HF_JSON    = Keys.Tokenizer.HF_JSON
 KEY_TOKENIZER_RWKV       = Keys.Tokenizer.RWKV
+
+KEY_TOKENIZER_FIM_PRE_ID = Keys.Tokenizer.FIM_PRE_ID
+KEY_TOKENIZER_FIM_SUF_ID = Keys.Tokenizer.FIM_SUF_ID
+KEY_TOKENIZER_FIM_MID_ID = Keys.Tokenizer.FIM_MID_ID
+KEY_TOKENIZER_FIM_PAD_ID = Keys.Tokenizer.FIM_PAD_ID
+KEY_TOKENIZER_FIM_REP_ID = Keys.Tokenizer.FIM_REP_ID
+KEY_TOKENIZER_FIM_SEP_ID = Keys.Tokenizer.FIM_SEP_ID
+
 KEY_TOKENIZER_PREFIX_ID  = Keys.Tokenizer.PREFIX_ID
 KEY_TOKENIZER_SUFFIX_ID  = Keys.Tokenizer.SUFFIX_ID
 KEY_TOKENIZER_MIDDLE_ID  = Keys.Tokenizer.MIDDLE_ID

gguf-py/gguf/gguf_writer.py

@@ -677,6 +677,9 @@ def add_expert_weights_norm(self, value: bool) -> None:
     def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None:
         self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value)
 
+    def add_nextn_predict_layers(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.NEXTN_PREDICT_LAYERS.format(arch=self.arch), count)
+
     def add_layer_norm_eps(self, value: float) -> None:
         self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)
 

gguf-py/gguf/tensor_mapping.py

@@ -592,6 +592,31 @@ class TensorNameMap:
         MODEL_TENSOR.ENC_OUTPUT_NORM: (
             "encoder.final_layer_norm", # t5
         ),
+
+        # NextN/MTP tensors for GLM4_MOE
+        MODEL_TENSOR.NEXTN_EH_PROJ: (
+            "model.layers.{bid}.eh_proj",
+        ),
+
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS: (
+            "model.layers.{bid}.embed_tokens",
+        ),
+
+        MODEL_TENSOR.NEXTN_ENORM: (
+            "model.layers.{bid}.enorm",
+        ),
+
+        MODEL_TENSOR.NEXTN_HNORM: (
+            "model.layers.{bid}.hnorm",
+        ),
+
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD: (
+            "model.layers.{bid}.shared_head.head",
+        ),
+
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM: (
+            "model.layers.{bid}.shared_head.norm",
+        ),
     }
 
     # architecture-specific block mappings