@@ -1216,6 +1216,55 @@ def _try_set_pooling_type(self) -> None:
12161216 raise NotImplementedError ("Only MEAN, CLS, and LAST pooling types supported" )
12171217 self .gguf_writer .add_pooling_type (pooling_type )
12181218
1219+ def _set_vocab_interns1 (self ):
1220+ tokens : list [str ] = []
1221+ toktypes : list [int ] = []
1222+
1223+ from transformers import AutoTokenizer
1224+ tokenizer = AutoTokenizer .from_pretrained (self .dir_model , trust_remote_code = True )
1225+ vocab = getattr (tokenizer , 'vocab' , tokenizer .get_vocab ())
1226+ vocab_size = self .hparams .get ("vocab_size" , len (vocab ))
1227+ assert max (vocab .values ()) < vocab_size
1228+
1229+ tokpre = self .get_vocab_base_pre (tokenizer )
1230+
1231+ reverse_vocab = {id_ : encoded_tok for encoded_tok , id_ in vocab .items ()}
1232+ added_vocab = tokenizer .get_added_vocab ()
1233+
1234+ added_tokens_decoder = tokenizer .added_tokens_decoder
1235+
1236+ for i in range (vocab_size ):
1237+ if i not in reverse_vocab :
1238+ tokens .append (f"[PAD{ i } )
1239+ toktypes .append (gguf .TokenType .UNUSED )
1240+ else :
1241+ token : str = reverse_vocab [i ]
1242+ if token in added_vocab :
1243+ # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
1244+ # To avoid unexpected issues - we make sure to normalize non-normalized tokens
1245+ if not added_tokens_decoder [i ].normalized :
1246+ previous_token = token
1247+ token = tokenizer .decode (tokenizer .encode (token , add_special_tokens = False ))
1248+ if previous_token != token :
1249+ logger .info (f"{ repr (previous_token )} { repr (token )} )
1250+
1251+ if added_tokens_decoder [i ].special or self .does_token_look_special (token ):
1252+ toktypes .append (gguf .TokenType .CONTROL )
1253+ else :
1254+ toktypes .append (gguf .TokenType .USER_DEFINED )
1255+ else :
1256+ toktypes .append (gguf .TokenType .NORMAL )
1257+ tokens .append (token )
1258+
1259+ self .gguf_writer .add_tokenizer_model ("gpt2" )
1260+ self .gguf_writer .add_tokenizer_pre (tokpre )
1261+ self .gguf_writer .add_token_list (tokens )
1262+ self .gguf_writer .add_token_types (toktypes )
1263+
1264+ special_vocab = gguf .SpecialVocab (self .dir_model , load_merges = True )
1265+ special_vocab ._set_special_token ("bos" , 151643 )
1266+ special_vocab .add_to_gguf (self .gguf_writer )
1267+
12191268
12201269class MmprojModel (ModelBase ):
12211270 model_type = ModelType .MMPROJ
@@ -2932,7 +2981,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
29322981 if "language_model." in name :
29332982 name = name .replace ("language_model." , "" ) # for InternVL
29342983 if name .startswith ("mlp" ) or name .startswith ("multi_modal_projector" ) \
2935- or name .startswith ("vision_model" ) or name .startswith ("audio_tower" ):
2984+ or name .startswith ("vision_model" ) or name .startswith ("audio_tower" ) \
2985+ or name .startswith ("model.vision_tower" ) or name .startswith ("model.multi_modal_projector" ):
29362986 # skip vision and audio tensors
29372987 return []
29382988 yield from super ().modify_tensors (data_torch , name , bid )
@@ -3109,7 +3159,7 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
31093159 yield from super ().modify_tensors (data_torch , name , bid )
31103160
31113161
3112- @ModelBase .register ("Ernie4_5_ForCausalLM" )
3162+ @ModelBase .register ("Ernie4_5_ForCausalLM" , "Ernie4_5ForCausalLM" )
31133163class Ernie4_5Model (TextModel ):
31143164 model_arch = gguf .MODEL_ARCH .ERNIE4_5
31153165
@@ -3604,6 +3654,19 @@ def prepare_tensors(self):
36043654class Qwen3Model (Qwen2Model ):
36053655 model_arch = gguf .MODEL_ARCH .QWEN3
36063656
3657+ def __init__ (self , * args , ** kwargs ):
3658+ super ().__init__ (* args , ** kwargs )
3659+ hparams = ModelBase .load_hparams (self .dir_model , is_mistral_format = False )
3660+ self .origin_hf_arch = hparams .get ('architectures' , [None ])[0 ]
3661+
3662+ def set_vocab (self ):
3663+ # deal with intern-s1-mini
3664+ if self .origin_hf_arch == 'InternS1ForConditionalGeneration' :
3665+ self ._set_vocab_interns1 ()
3666+ return
3667+
3668+ super ().set_vocab ()
3669+
36073670
36083671@ModelBase .register ("Qwen3MoeForCausalLM" )
36093672class Qwen3MoeModel (Qwen2MoeModel ):
@@ -3620,73 +3683,7 @@ def set_vocab(self):
36203683 self ._set_vocab_interns1 ()
36213684 return
36223685
3623- try :
3624- self ._set_vocab_sentencepiece ()
3625- except FileNotFoundError :
3626- self ._set_vocab_gpt2 ()
3627-
3628- def _set_vocab_interns1 (self ):
3629- tokens : list [str ] = []
3630- toktypes : list [int ] = []
3631-
3632- from transformers import AutoTokenizer
3633- tokenizer = AutoTokenizer .from_pretrained (self .dir_model , trust_remote_code = True )
3634- vocab = getattr (tokenizer , 'vocab' , tokenizer .get_vocab ())
3635- vocab_size = self .hparams .get ("vocab_size" , len (vocab ))
3636- assert max (vocab .values ()) < vocab_size
3637-
3638- tokpre = self .get_vocab_base_pre (tokenizer )
3639-
3640- reverse_vocab = {id_ : encoded_tok for encoded_tok , id_ in vocab .items ()}
3641- added_vocab = tokenizer .get_added_vocab ()
3642-
3643- added_tokens_decoder = tokenizer .added_tokens_decoder
3644-
3645- for i in range (vocab_size ):
3646- if i not in reverse_vocab :
3647- tokens .append (f"[PAD{ i } )
3648- toktypes .append (gguf .TokenType .UNUSED )
3649- else :
3650- token : str = reverse_vocab [i ]
3651- if token in added_vocab :
3652- # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
3653- # To avoid unexpected issues - we make sure to normalize non-normalized tokens
3654- if not added_tokens_decoder [i ].normalized :
3655- previous_token = token
3656- token = tokenizer .decode (tokenizer .encode (token , add_special_tokens = False ))
3657- if previous_token != token :
3658- logger .info (f"{ repr (previous_token )} { repr (token )} )
3659-
3660- if added_tokens_decoder [i ].special or self .does_token_look_special (token ):
3661- toktypes .append (gguf .TokenType .CONTROL )
3662- else :
3663- toktypes .append (gguf .TokenType .USER_DEFINED )
3664- else :
3665- toktypes .append (gguf .TokenType .NORMAL )
3666- tokens .append (token )
3667-
3668- self .gguf_writer .add_tokenizer_model ("gpt2" )
3669- self .gguf_writer .add_tokenizer_pre (tokpre )
3670- self .gguf_writer .add_token_list (tokens )
3671- self .gguf_writer .add_token_types (toktypes )
3672-
3673- special_vocab = gguf .SpecialVocab (self .dir_model , load_merges = True )
3674- special_tokens_map_file = self .dir_model / 'special_tokens_map.json'
3675- additional_special_tokens = []
3676- if special_tokens_map_file .is_file ():
3677- with open (special_tokens_map_file , encoding = 'utf-8' ) as f :
3678- additional_special_tokens = json .load (f ).get ('additional_special_tokens' , [])
3679- tokenizer_cfg_file = self .dir_model / 'special_tokens_map.json'
3680- if tokenizer_cfg_file .is_file ():
3681- with open (tokenizer_cfg_file , encoding = 'utf-8' ) as f :
3682- added_tokens_decoder = json .load (f ).get ('added_tokens_decoder' , {})
3683- token2ids_map = {data ['content' ] : int (token ) for token , data in added_tokens_decoder .items () if data ['special' ]}
3684- for token in additional_special_tokens :
3685- if token in token2ids_map :
3686- special_vocab ._set_special_token (token , token2ids_map [token ])
3687- special_vocab ._set_special_token ('eos' , 151645 )
3688- special_vocab ._set_special_token ("bos" , 151643 )
3689- special_vocab .add_to_gguf (self .gguf_writer )
3686+ super ().set_vocab ()
36903687
36913688
36923689@ModelBase .register ("GPT2LMHeadModel" )
@@ -5854,6 +5851,11 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
58545851 return [(self .map_tensor_name (name ), data_torch )]
58555852
58565853
5854+ @ModelBase .register ("SeedOssForCausalLM" )
5855+ class SeedOssModel (TextModel ):
5856+ model_arch = gguf .MODEL_ARCH .SEED_OSS
5857+
5858+
58575859@ModelBase .register ("Olmo2ForCausalLM" )
58585860class Olmo2Model (TextModel ):
58595861 model_arch = gguf .MODEL_ARCH .OLMO2
@@ -6252,9 +6254,11 @@ def prepare_tensors(self):
62526254 raise ValueError (f"Unprocessed experts: { experts } )
62536255
62546256
6255- @ModelBase .register ("DeepseekV2ForCausalLM" )
6256- @ModelBase .register ("DeepseekV3ForCausalLM" )
6257- @ModelBase .register ("KimiVLForConditionalGeneration" )
6257+ @ModelBase .register (
6258+ "DeepseekV2ForCausalLM" ,
6259+ "DeepseekV3ForCausalLM" ,
6260+ "KimiVLForConditionalGeneration" ,
6261+ )
62586262class DeepseekV2Model (TextModel ):
62596263 model_arch = gguf .MODEL_ARCH .DEEPSEEK2
62606264
@@ -8505,6 +8509,43 @@ def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", "
85058509 return "mm.2.weight"
85068510 return super ().map_tensor_name (name , try_suffixes )
85078511
8512+
8513+ @ModelBase .register ("KimiVLForConditionalGeneration" )
8514+ class KimiVLModel (MmprojModel ):
8515+ def __init__ (self , * args , ** kwargs ):
8516+ super ().__init__ (* args , ** kwargs )
8517+ assert self .hparams_vision is not None
8518+ self .hparams_vision ["image_size" ] = 64 * 14 # for compatibility
8519+
8520+ def set_gguf_parameters (self ):
8521+ super ().set_gguf_parameters ()
8522+ self .gguf_writer .add_clip_projector_type (gguf .VisionProjectorType .KIMIVL )
8523+ self .gguf_writer .add_vision_use_gelu (True )
8524+ self .gguf_writer .add_vision_projector_scale_factor (2 )
8525+ # eps is the same as pytorch's default value
8526+ assert self .hparams_vision is not None
8527+ self .gguf_writer .add_vision_attention_layernorm_eps (self .hparams_vision .get ("layer_norm_eps" , 1e-5 ))
8528+
8529+ def modify_tensors (self , data_torch : Tensor , name : str , bid : int | None ) -> Iterable [tuple [str , Tensor ]]:
8530+ del bid # unused
8531+ is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
8532+
8533+ if is_vision_tensor :
8534+ if "pos_emb.weight" in name :
8535+ data_torch = data_torch .view (data_torch .shape [0 ] * data_torch .shape [1 ], data_torch .shape [2 ])
8536+ elif "wqkv" in name :
8537+ split_dim = 0 if "weight" in name else - 1
8538+ wq , wk , wv = data_torch .chunk (3 , dim = split_dim )
8539+ return [
8540+ (self .map_tensor_name (name .replace ("wqkv" , "wq" )), wq ),
8541+ (self .map_tensor_name (name .replace ("wqkv" , "wk" )), wk ),
8542+ (self .map_tensor_name (name .replace ("wqkv" , "wv" )), wv )
8543+ ]
8544+
8545+ return [(self .map_tensor_name (name ), data_torch )]
8546+
8547+ return [] # skip other tensors
8548+
85088549###### CONVERSION LOGIC ######
85098550
85108551
0 commit comments