happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

update

This commit is contained in:
Xuan Son Nguyen 2026-02-12 16:24:55 +01:00
parent 9bfafada7f
commit 825540135f
3 changed files with 22 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

50
convert_hf_to_gguf.py
View File

@ -1605,6 +1605,23 @@ class TextModel(ModelBase):
special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])
special_vocab.add_to_gguf(self.gguf_writer)
def _set_vocab_glm(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
special_vocab.add_to_gguf(self.gguf_writer)
def _set_vocab_interns1(self):
tokens: list[str] = []
toktypes: list[int] = []
@ -8562,27 +8579,6 @@ class Glm4MoeModel(TextModel):
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)
# using staticmethod here to allow re-using it in other classes
@staticmethod
def set_vocab_glm(model: TextModel):
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(model.dir_model)
special_vocab = gguf.SpecialVocab(model.dir_model, load_merges=True)
tokens, toktypes, tokpre = model.get_vocab_base()
model.gguf_writer.add_tokenizer_model("gpt2")
model.gguf_writer.add_tokenizer_pre(tokpre)
model.gguf_writer.add_token_list(tokens)
model.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
special_vocab.add_to_gguf(model.gguf_writer)
def set_vocab(self):
Glm4MoeModel.set_vocab_glm(self)
@ -8685,7 +8681,7 @@ class Glm4MoeLiteModel(DeepseekV2Model):
model_arch = gguf.MODEL_ARCH.DEEPSEEK2
def set_vocab(self):
Glm4MoeModel.set_vocab_glm(self)
return self._set_vocab_glm()
@ModelBase.register("GlmMoeDsaForCausalLM")
@ -8700,20 +8696,15 @@ class GlmMoeDsaModel(DeepseekV2Model):
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
def set_vocab(self):
Glm4MoeModel.set_vocab_glm(self)
return self._set_vocab_glm()
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_leading_dense_block_count(3) # TODO: not to hard-code this for future models
rope_dim = self.hparams["qk_rope_head_dim"]
partial_rotary_factor = self.hparams.get("partial_rotary_factor", 1.0)
self.gguf_writer.add_rope_dimension_count(int(rope_dim * partial_rotary_factor))
# Expert gating function (sigmoid for GLM4_MOE)
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
# 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)
@ -8723,9 +8714,6 @@ class GlmMoeDsaModel(DeepseekV2Model):
self.gguf_writer.add_indexer_key_length(self.hparams["index_head_dim"])
self.gguf_writer.add_indexer_top_k(self.hparams["index_topk"])
def modify_tensors(self, data_torch, name, bid):
yield from super().modify_tensors(data_torch, name, bid)
@ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
class ChatGLMModel(TextModel):

3
src/llama-model.cpp
View File

@ -136,6 +136,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_300B_A47B: return "300B.A47B";
case LLM_TYPE_310B_A15B: return "310B.A15B";
case LLM_TYPE_355B_A32B: return "355B.A32B";
case LLM_TYPE_744B_A40B: return "744B.A40B";
case LLM_TYPE_E2B: return "E2B";
case LLM_TYPE_E4B: return "E4B";
default: return "?B";
@ -1860,7 +1861,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
switch (hparams.n_layer) {
// TODO
case 78: type = LLM_TYPE_744B_A40B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;

1
src/llama-model.h
View File

@ -129,6 +129,7 @@ enum llm_type {
LLM_TYPE_300B_A47B, // Ernie MoE big
LLM_TYPE_310B_A15B, // /MiMo-V2-Flash
LLM_TYPE_355B_A32B, // GLM-4.5
LLM_TYPE_744B_A40B, // GLM-5
LLM_TYPE_E2B,
LLM_TYPE_E4B,
};