ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

model : add NeoBERT (#14164)

Browse source 
* convert neobert model to gguf

* add inference graph

* fix flake8 lint

* followed reviewer suggestions

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* follow reviewers suggestions

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* override NeoBERT feed-forward length

---------

Co-authored-by: dinhhuy <huy.dinh@brains-tech.co.jp>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Đinh Trọng Huy 2025-06-16 21:53:41 +09:00 • committed by GitHub
parent 7d6d91babf
commit ad590be98c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 222 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

30
convert_hf_to_gguf.py
View file

@ -519,7 +519,7 @@ class TextModel(ModelBase):
def set_gguf_parameters(self):
self.gguf_writer.add_block_count(self.block_count)
if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None:
if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions", "max_length"], optional=True)) is not None:
self.gguf_writer.add_context_length(n_ctx)
logger.info(f"gguf: context length = {n_ctx}")
@ -4076,6 +4076,34 @@ class NomicBertModel(BertModel):
raise ValueError(f"unknown tokenizer: {toktyp}")
@ModelBase.register("NeoBERT", "NeoBERTLMHead", "NeoBERTForSequenceClassification")
class NeoBert(BertModel):
model_arch = gguf.MODEL_ARCH.NEO_BERT
def set_gguf_parameters(self):
super().set_gguf_parameters()
# NeoBERT uses 2/3 of the intermediate size as feed forward length
self.gguf_writer.add_feed_forward_length(int(2 * self.hparams["intermediate_size"] / 3))
self.gguf_writer.add_rope_freq_base(10000.0) # default value for NeoBERT
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
f_rms_eps = self.hparams.get("norm_eps", 1e-6) # default value for NeoBERT
self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
logger.info(f"gguf: rms norm epsilon = {f_rms_eps}")
self.gguf_writer.add_pooling_type(gguf.PoolingType.CLS) # https://huggingface.co/chandar-lab/NeoBERT#how-to-use
def modify_tensors(self, data_torch, name, bid):
if name.startswith("decoder."):
return []
if name.startswith("model."):
name = name[6:]
return super().modify_tensors(data_torch, name, bid)
@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
class XLMRobertaModel(BertModel):
model_arch = gguf.MODEL_ARCH.BERT