tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-26 19:55:04 +00:00
Code Issues Packages Projects Releases 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
6 changed files with 222 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
gguf-py/gguf/constants.py
View File

@ -291,6 +291,7 @@ class MODEL_ARCH(IntEnum):
BERT = auto()
NOMIC_BERT = auto()
NOMIC_BERT_MOE = auto()
NEO_BERT = auto()
JINA_BERT_V2 = auto()
BLOOM = auto()
STABLELM = auto()
@ -573,6 +574,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.BERT: "bert",
MODEL_ARCH.NOMIC_BERT: "nomic-bert",
MODEL_ARCH.NOMIC_BERT_MOE: "nomic-bert-moe",
MODEL_ARCH.NEO_BERT: "neo-bert",
MODEL_ARCH.JINA_BERT_V2: "jina-bert-v2",
MODEL_ARCH.BLOOM: "bloom",
MODEL_ARCH.STABLELM: "stablelm",
@ -1081,6 +1083,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.LAYER_OUT_NORM,
],
MODEL_ARCH.NEO_BERT: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.ENC_OUTPUT_NORM,
MODEL_TENSOR.CLS,
MODEL_TENSOR.CLS_OUT,
],
MODEL_ARCH.JINA_BERT_V2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.TOKEN_EMBD_NORM,

9
gguf-py/gguf/tensor_mapping.py
View File

@ -31,6 +31,7 @@ class TensorNameMap:
"model.embeddings", # rwkv7
"model.word_embeddings", # bailingmoe
"language_model.model.embed_tokens", # llama4
"encoder", # neobert
),
# Token type embeddings
@ -134,6 +135,7 @@ class TensorNameMap:
"rwkv.blocks.{bid}.ln1", # rwkv6
"model.layers.{bid}.ln1", # rwkv7
"model.layers.{bid}.input_layernorm", # llama4
"transformer_encoder.{bid}.attention_norm", # neobert
),
# Attention norm 2
@ -161,6 +163,7 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.qkv_proj", # phi3
"encoder.layers.{bid}.self_attention.query_key_value", # chatglm
"transformer.layers.{bid}.attn.qkv_proj", # openelm
"transformer_encoder.{bid}.qkv", # neobert
),
# Attention query
@ -236,6 +239,7 @@ class TensorNameMap:
"transformer.layers.{bid}.attn.out_proj", # openelm
"transformer.h.{bid}.attn.attention.out_proj", # exaone
"model.layers.{bid}.self_attn.o_proj", # llama4
"transformer_encoder.{bid}.wo", # neobert
),
# Attention output norm
@ -276,6 +280,7 @@ class TensorNameMap:
"encoder.layers.{bid}.post_attention_layernorm", # chatglm
"transformer.layers.{bid}.ffn_norm", # openelm
"model.layers.{bid}.post_attention_layernorm", # llama4
"transformer_encoder.{bid}.ffn_norm", # neobert
),
# Post feed-forward norm
@ -340,6 +345,7 @@ class TensorNameMap:
"encoder.layers.{bid}.mlp.dense_h_to_4h", # chatglm
"transformer.h.{bid}.mlp.c_fc_1", # exaone
"model.layers.{bid}.feed_forward.up_proj", # llama4
"transformer_encoder.{bid}.ffn.w12", # neobert
),
MODEL_TENSOR.FFN_UP_EXP: (
@ -422,6 +428,7 @@ class TensorNameMap:
"encoder.layers.{bid}.mlp.dense_4h_to_h", # chatglm
"model.layers.h.{bid}.mlp.c_proj", # exaone
"model.layers.{bid}.feed_forward.down_proj", # llama4
"transformer_encoder.{bid}.ffn.w3", # neobert
),
MODEL_TENSOR.FFN_DOWN_EXP: (
@ -832,12 +839,14 @@ class TensorNameMap:
# TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
MODEL_TENSOR.ENC_OUTPUT_NORM: (
"encoder.final_layer_norm", # t5
"layer_norm", # neobert
),
MODEL_TENSOR.CLS: (
"classifier", # jina
"classifier.dense", # roberta
"pre_classifier", # distillbert
"dense", # neobert
),
MODEL_TENSOR.CLS_OUT: (