@@ -678,6 +678,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
678678 if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2" :
679679 # ref: https://huggingface.co/THUDM/glm-4-9b-hf
680680 res = "glm4"
681+ if chkhsh == "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902" :
682+ # ref: https://huggingface.co/zai-org/GLM-4.5-Air
683+ res = "glm4"
681684 if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35" :
682685 # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
683686 res = "minerva-7b"
@@ -6696,6 +6699,139 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
66966699 return super ().modify_tensors (data_torch , name , bid )
66976700
66986701
6702+ @ModelBase .register ("Glm4MoeForCausalLM" )
6703+ class Glm4MoeModel (TextModel ):
6704+ model_arch = gguf .MODEL_ARCH .GLM4_MOE
6705+
6706+ def __init__ (self , * args , ** kwargs ):
6707+ super ().__init__ (* args , ** kwargs )
6708+ # GLM4_MOE has num_hidden_layers + 1 actual layers (including NextN layer)
6709+ self .block_count = self .hparams ["num_hidden_layers" ] + self .hparams .get ("num_nextn_predict_layers" , 0 )
6710+ self .tensor_map = gguf .get_tensor_name_map (self .model_arch , self .block_count )
6711+
6712+ def set_vocab (self ):
6713+ from transformers import AutoTokenizer
6714+
6715+ tokenizer = AutoTokenizer .from_pretrained (self .dir_model )
6716+ special_vocab = gguf .SpecialVocab (self .dir_model , load_merges = True )
6717+ tokens , toktypes , tokpre = self .get_vocab_base ()
6718+ self .gguf_writer .add_tokenizer_model ("gpt2" )
6719+ self .gguf_writer .add_tokenizer_pre (tokpre )
6720+ self .gguf_writer .add_token_list (tokens )
6721+ self .gguf_writer .add_token_types (toktypes )
6722+
6723+ # Special tokens
6724+ # Note: Using <|endoftext|> (151329) for eot causes endless generation
6725+ special_vocab ._set_special_token ("bos" , tokenizer .get_added_vocab ()["[gMASK]" ]) # 151331
6726+ special_vocab ._set_special_token ("eot" , tokenizer .get_added_vocab ()["<|user|>" ]) # 151336
6727+ special_vocab ._set_special_token ("unk" , tokenizer .get_added_vocab ()["<|endoftext|>" ]) # 151329
6728+ special_vocab ._set_special_token ("eom" , tokenizer .get_added_vocab ()["<|observation|>" ]) # 151338
6729+
6730+ # Patch broken chat template
6731+ if isinstance (special_vocab .chat_template , str ) and "visible_text(m.content).endswith" in special_vocab .chat_template :
6732+ special_vocab .chat_template = special_vocab .chat_template .replace (
6733+ """{{ visible_text(m.content) }}\n {{- '/nothink' if (enable_thinking is defined and not enable_thinking and not visible_text(m.content).endswith("/nothink")) else '' -}}""" ,
6734+ """{% set content = visible_text(m.content) %}{{ content }}\n {{- '/nothink' if (enable_thinking is defined and not enable_thinking and not content.endswith("/nothink")) else '' -}}""" )
6735+
6736+ special_vocab .add_to_gguf (self .gguf_writer )
6737+
6738+ def set_gguf_parameters (self ):
6739+ super ().set_gguf_parameters ()
6740+ if (rope_dim := self .hparams .get ("head_dim" )) is None :
6741+ rope_dim = (
6742+ self .hparams ["hidden_size" ] // self .hparams ["num_attention_heads" ]
6743+ )
6744+ self .gguf_writer .add_rope_dimension_count (
6745+ int (rope_dim * self .hparams .get ("partial_rotary_factor" , 0.5 ))
6746+ )
6747+
6748+ # MoE parameters - Use only routed expert count (shared experts handled separately)
6749+ if (n_routed_experts := self .hparams .get ("n_routed_experts" )) is not None :
6750+ self .gguf_writer .add_expert_count (n_routed_experts )
6751+ if (moe_intermediate_size := self .hparams .get ("moe_intermediate_size" )) is not None :
6752+ self .gguf_writer .add_expert_feed_forward_length (moe_intermediate_size )
6753+ if (n_shared_experts := self .hparams .get ("n_shared_experts" )) is not None :
6754+ self .gguf_writer .add_expert_shared_count (n_shared_experts )
6755+ if (first_k_dense_replace := self .hparams .get ("first_k_dense_replace" )) is not None :
6756+ self .gguf_writer .add_leading_dense_block_count (first_k_dense_replace )
6757+
6758+ # Expert gating function (sigmoid for GLM4_MOE)
6759+ self .gguf_writer .add_expert_gating_func (gguf .ExpertGatingFuncType .SIGMOID )
6760+
6761+ # Routed scaling factor
6762+ if (routed_scaling_factor := self .hparams .get ("routed_scaling_factor" )) is not None :
6763+ self .gguf_writer .add_expert_weights_scale (routed_scaling_factor )
6764+
6765+ # Normalise topk probabilities
6766+ if (norm_topk_prob := self .hparams .get ("norm_topk_prob" )) is not None :
6767+ self .gguf_writer .add_expert_weights_norm (norm_topk_prob )
6768+
6769+ # NextN/MTP prediction layers
6770+ if (num_nextn_predict_layers := self .hparams .get ("num_nextn_predict_layers" )) is not None :
6771+ self .gguf_writer .add_nextn_predict_layers (num_nextn_predict_layers )
6772+
6773+ _experts : list [dict [str , Tensor ]] | None = None
6774+
6775+ def modify_tensors (
6776+ self , data_torch : Tensor , name : str , bid : int | None
6777+ ) -> Iterable [tuple [str , Tensor ]]:
6778+ if name .startswith ("model.visual." ): # ignore visual part
6779+ return []
6780+ elif name .startswith ("model.language_model." ):
6781+ name = name .replace ("language_model." , "" ) # for multimodal variants
6782+
6783+ # Handle main token embedding (but not layer-specific NextN embeddings)
6784+ if name == "model.embed_tokens.weight" and ".layers." not in name :
6785+ return [(self .map_tensor_name ("token_embd.weight" ), data_torch )]
6786+
6787+ # Handle routed experts
6788+ if name .find ("mlp.experts" ) != - 1 :
6789+ n_experts = self .hparams ["n_routed_experts" ]
6790+ assert bid is not None
6791+
6792+ if self ._experts is None :
6793+ self ._experts = [{} for _ in range (self .block_count )]
6794+
6795+ self ._experts [bid ][name ] = data_torch
6796+
6797+ if len (self ._experts [bid ]) >= n_experts * 3 :
6798+ tensors : list [tuple [str , Tensor ]] = []
6799+
6800+ # merge the experts into a single 3d tensor
6801+ for w_name in ["down_proj" , "gate_proj" , "up_proj" ]:
6802+ datas : list [Tensor ] = []
6803+
6804+ for xid in range (n_experts ):
6805+ ename = f"model.layers.{ bid } { xid } { w_name }
6806+ datas .append (self ._experts [bid ][ename ])
6807+ del self ._experts [bid ][ename ]
6808+
6809+ data_torch = torch .stack (datas , dim = 0 )
6810+
6811+ merged_name = f"model.layers.{ bid } { w_name }
6812+
6813+ new_name = self .map_tensor_name (merged_name )
6814+ tensors .append ((new_name , data_torch ))
6815+ return tensors
6816+ else :
6817+ return []
6818+
6819+ if name .endswith ("e_score_correction_bias" ):
6820+ name = name .replace ("e_score_correction_bias" , "e_score_correction.bias" )
6821+
6822+ new_name = self .map_tensor_name (name )
6823+
6824+ return [(new_name , data_torch )]
6825+
6826+ def prepare_tensors (self ):
6827+ super ().prepare_tensors ()
6828+ if self ._experts is not None :
6829+ # flatten `list[dict[str, Tensor]]` into `list[str]`
6830+ experts = [k for d in self ._experts for k in d .keys ()]
6831+ if len (experts ) > 0 :
6832+ raise ValueError (f"Unprocessed experts: { experts } )
6833+
6834+
66996835@ModelBase .register ("GlmForCausalLM" , "ChatGLMModel" , "ChatGLMForConditionalGeneration" )
67006836class ChatGLMModel (TextModel ):
67016837 model_arch = gguf .MODEL_ARCH .CHATGLM
0 commit comments