CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-09 10:17:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/mamba2

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-06-18 23:59:21 -04:00
parent 0b6f6becb4 8d94713654
commit a42f239418
91 changed files with 3604 additions and 1088 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
gguf-py/gguf/constants.py
View File

@@ -292,6 +292,7 @@ class MODEL_ARCH(IntEnum):
BERT = auto()
NOMIC_BERT = auto()
NOMIC_BERT_MOE = auto()
NEO_BERT = auto()
JINA_BERT_V2 = auto()
BLOOM = auto()
STABLELM = auto()
@@ -345,6 +346,8 @@ class MODEL_ARCH(IntEnum):
WAVTOKENIZER_DEC = auto()
PLM = auto()
BAILINGMOE = auto()
DOTS1 = auto()
ARCEE = auto()
class VISION_PROJECTOR_TYPE(IntEnum):
@@ -574,6 +577,7 @@ 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",
@@ -627,6 +631,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
MODEL_ARCH.PLM: "plm",
MODEL_ARCH.BAILINGMOE: "bailingmoe",
MODEL_ARCH.DOTS1: "dots1",
MODEL_ARCH.ARCEE: "arcee",
}
VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1082,6 +1088,18 @@ 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,
@@ -2062,6 +2080,45 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
],
MODEL_ARCH.DOTS1: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_EXP_PROBS_B,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_UP_SHEXP,
],
MODEL_ARCH.ARCEE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
# TODO
}

2
gguf-py/gguf/gguf_writer.py
View File

@@ -271,7 +271,7 @@ class GGUFWriter:
def add_key_value(self, key: str, val: Any, vtype: GGUFValueType, sub_type: GGUFValueType | None = None) -> None:
if any(key in kv_data for kv_data in self.kv_data):
raise ValueError(f'Duplicated key name {key!r}')
logger.warning(f'Duplicated key name {key!r}, overwriting it with new value {val!r} of type {vtype.name}')
self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type)

11
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
@@ -305,7 +310,7 @@ class TensorNameMap:
),
MODEL_TENSOR.FFN_EXP_PROBS_B: (
"model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
"model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
),
# Feed-forward up
@@ -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: (
@@ -836,12 +843,14 @@ 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: (