feat: support GLM 4.5 family of models

Author: sammcj

convert_hf_to_gguf.py

@@ -6578,6 +6578,177 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Glm4MoeForCausalLM")
+class Glm4MoeModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.GLM4_MOE
+
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            self.dir_model, trust_remote_code=True
+        )
+        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_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab._set_special_token(
+            "eos", tokenizer.get_added_vocab()["<|endoftext|>"]
+        )
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
+        special_vocab._set_special_token(
+            "unk", tokenizer.get_added_vocab()["<|endoftext|>"]
+        )
+        special_vocab._set_special_token(
+            "bos", tokenizer.get_added_vocab()["<|endoftext|>"]
+        )
+        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
+        if (n_experts := self.hparams.get("n_routed_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+        # Note: expert_used_count is already set by parent class using num_experts_per_tok
+        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)
+
+    _experts: list[dict[str, Tensor]] | None = None
+    _shared_experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(
+        self, data_torch: Tensor, name: str, bid: int | None
+    ) -> Iterable[tuple[str, Tensor]]:
+        # Handle special GLM4_MOE layer 46 tensors (nextn prediction layer)
+        if bid is not None and bid == 46:
+            # Layer 46 is the nextn prediction layer - skip all tensors
+            return []
+
+        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
+        if name == "model.embed_tokens.weight":
+            return [(self.map_tensor_name("token_embd.weight"), data_torch)]
+
+        # Handle routed experts
+        if name.find("mlp.experts") != -1 and "shared_experts" not in name:
+            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)
+                    # Generate GGUF tensor names for merged experts
+                    if w_name == "down_proj":
+                        new_name = f"blk.{bid}.ffn_down_exps.weight"
+                    elif w_name == "gate_proj":
+                        new_name = f"blk.{bid}.ffn_gate_exps.weight"
+                    elif w_name == "up_proj":
+                        new_name = f"blk.{bid}.ffn_up_exps.weight"
+                    else:
+                        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 []
+
+        # Handle expert gating input (routing gate)
+        if ".mlp.gate.e_score_correction_bias" in name:
+            new_name = name.replace("model.layers.", "blk.").replace(
+                ".mlp.gate.e_score_correction_bias", ".ffn_gate_inp.bias"
+            )
+            return [(self.map_tensor_name(new_name), data_torch)]
+
+        # Handle shared expert tensors
+        if ".mlp.ffn_" in name and "_shexp" in name:
+            new_name = name.replace("model.layers.", "blk.")
+            return [(new_name, data_torch)]
+
+        # Handle regular dense FFN layers (for hybrid dense/MoE architecture)
+        if ".mlp." in name and "experts" not in name and "_shexp" not in name:
+            if "gate_proj" in name:
+                new_name = name.replace("model.layers.", "blk.").replace(
+                    ".mlp.gate_proj.weight", ".ffn_gate.weight"
+                )
+            elif "up_proj" in name:
+                new_name = name.replace("model.layers.", "blk.").replace(
+                    ".mlp.up_proj.weight", ".ffn_up.weight"
+                )
+            elif "down_proj" in name:
+                new_name = name.replace("model.layers.", "blk.").replace(
+                    ".mlp.down_proj.weight", ".ffn_down.weight"
+                )
+            else:
+                new_name = name
+            return [(self.map_tensor_name(new_name), data_torch)]
+
+        # Handle other special GLM4_MOE tensors (nextn prediction)
+        if (
+            ".embed_tokens." in name
+            or ".shared_head." in name
+            or ".eh_proj." in name
+            or ".enorm." in name
+            or ".hnorm." in name
+        ):
+            # Skip these special tensors - they are for nextn prediction
+            return []
+
+        return super().modify_tensors(data_torch, name, bid)
+
+    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}")
+
+
 @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CHATGLM

gguf-py/gguf/constants.py

@@ -354,6 +354,7 @@ class MODEL_ARCH(IntEnum):
     DEEPSEEK2        = auto()
     CHATGLM          = auto()
     GLM4             = auto()
+    GLM4_MOE         = auto()
     BITNET           = auto()
     T5               = auto()
     T5ENCODER        = auto()
@@ -422,6 +423,9 @@ class MODEL_TENSOR(IntEnum):
     FFN_GATE_EXP         = auto()
     FFN_DOWN_EXP         = auto()
     FFN_UP_EXP           = auto()
+    FFN_GATE_EXPS        = auto()  # merged experts
+    FFN_DOWN_EXPS        = auto()  # merged experts  
+    FFN_UP_EXPS          = auto()  # merged experts
     FFN_GATE_SHEXP       = auto()
     FFN_DOWN_SHEXP       = auto()
     FFN_UP_SHEXP         = auto()
@@ -673,6 +677,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.DEEPSEEK2:        "deepseek2",
     MODEL_ARCH.CHATGLM:          "chatglm",
     MODEL_ARCH.GLM4:             "glm4",
+    MODEL_ARCH.GLM4_MOE:         "glm4moe",
     MODEL_ARCH.BITNET:           "bitnet",
     MODEL_ARCH.T5:               "t5",
     MODEL_ARCH.T5ENCODER:        "t5encoder",
@@ -747,6 +752,9 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.FFN_GATE_EXP:              "blk.{bid}.ffn_gate_exps",
     MODEL_TENSOR.FFN_DOWN_EXP:              "blk.{bid}.ffn_down_exps",
     MODEL_TENSOR.FFN_UP_EXP:                "blk.{bid}.ffn_up_exps",
+    MODEL_TENSOR.FFN_GATE_EXPS:             "blk.{bid}.ffn_gate_exps",  # merged experts
+    MODEL_TENSOR.FFN_DOWN_EXPS:             "blk.{bid}.ffn_down_exps",  # merged experts
+    MODEL_TENSOR.FFN_UP_EXPS:               "blk.{bid}.ffn_up_exps",    # merged experts
     MODEL_TENSOR.FFN_EXP_PROBS_B:           "blk.{bid}.exp_probs_b",
     MODEL_TENSOR.LAYER_OUT_NORM:            "blk.{bid}.layer_output_norm",
     MODEL_TENSOR.PER_LAYER_TOKEN_EMBD:      "per_layer_token_embd",           # gemma3n
@@ -2102,6 +2110,30 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.ATTN_POST_NORM,
         MODEL_TENSOR.FFN_POST_NORM,
     ],
+    MODEL_ARCH.GLM4_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_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_NORM,
+        MODEL_TENSOR.FFN_GATE,          # dense layers
+        MODEL_TENSOR.FFN_DOWN,          # dense layers  
+        MODEL_TENSOR.FFN_UP,            # dense layers
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXPS,
+        MODEL_TENSOR.FFN_DOWN_EXPS,
+        MODEL_TENSOR.FFN_UP_EXPS,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.ATTN_POST_NORM,
+    ],
     MODEL_ARCH.BITNET: [
         MODEL_TENSOR.ATTN_Q,
         MODEL_TENSOR.ATTN_K,

src/llama-arch.cpp

@@ -62,6 +62,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DEEPSEEK2,        "deepseek2"        },
     { LLM_ARCH_CHATGLM,          "chatglm"          },
     { LLM_ARCH_GLM4,             "glm4"             },
+    { LLM_ARCH_GLM4_MOE,         "glm4moe"          },
     { LLM_ARCH_BITNET,           "bitnet"           },
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -1389,6 +1390,33 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
         },
     },
+    {
+        LLM_ARCH_GLM4_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_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },        // dense layers
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },        // dense layers  
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },          // dense layers
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { 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_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_ATTN_POST_NORM,     "blk.%d.post_attention_norm" },
+        },
+    },
     {
         LLM_ARCH_BITNET,
         {

src/llama-arch.h

@@ -66,6 +66,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
     LLM_ARCH_GLM4,
+    LLM_ARCH_GLM4_MOE,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,