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

llama : Add support for DeepSeek V3 (#11049)

Browse source 
* convert : extend DEEPSEEK2 model architecture to support DeepseekV3ForCausalLM by adding EXPERT_WEIGHTS_NORM and EXPERT_GATING_FUNC model parameters and FFN_EXP_PROBS_B tensor type

* vocab : add DeepSeek V3 pre-tokenizer regexes

* unicode : handle ACCENT_MARK and SYMBOL categories in regex

* llama : add DeepSeek V3 chat template, handle new model parameters and tensor types

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
This commit is contained in:
fairydreaming 2025-01-04 21:06:11 +01:00 • committed by GitHub
parent f922a9c542
commit 9394bbd484
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
16 changed files with 162 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

10
gguf-py/gguf/constants.py
View file

@ -102,6 +102,8 @@ class Keys:
EXPERT_USED_COUNT = "{arch}.expert_used_count"
EXPERT_SHARED_COUNT = "{arch}.expert_shared_count"
EXPERT_WEIGHTS_SCALE = "{arch}.expert_weights_scale"
EXPERT_WEIGHTS_NORM = "{arch}.expert_weights_norm"
EXPERT_GATING_FUNC = "{arch}.expert_gating_func"
POOLING_TYPE = "{arch}.pooling_type"
LOGIT_SCALE = "{arch}.logit_scale"
DECODER_START_TOKEN_ID = "{arch}.decoder_start_token_id"
@ -313,6 +315,7 @@ class MODEL_TENSOR(IntEnum):
FFN_GATE_SHEXP = auto()
FFN_DOWN_SHEXP = auto()
FFN_UP_SHEXP = auto()
FFN_EXP_PROBS_B = auto()
ATTN_Q_NORM = auto()
ATTN_K_NORM = auto()
LAYER_OUT_NORM = auto()
@ -498,6 +501,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.FFN_GATE_EXP: "blk.{bid}.ffn_gate_exps",
MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down_exps",
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up_exps",
MODEL_TENSOR.FFN_EXP_PROBS_B: "blk.{bid}.exp_probs_b",
MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm",
MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in",
MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d",
@ -1290,6 +1294,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_EXP_PROBS_B,
],
MODEL_ARCH.CHATGLM : [
MODEL_TENSOR.TOKEN_EMBD,
@ -1590,6 +1595,11 @@ class GGMLQuantizationType(IntEnum):
TQ2_0 = 35
class ExpertGatingFuncType(IntEnum):
SOFTMAX = 1
SIGMOID = 2
# TODO: add GGMLFileType from ggml_ftype in ggml.h

7
gguf-py/gguf/gguf_writer.py
View file

@ -26,6 +26,7 @@ from .constants import (
RopeScalingType,
PoolingType,
TokenType,
ExpertGatingFuncType,
)
from .quants import quant_shape_from_byte_shape
@ -715,6 +716,12 @@ class GGUFWriter:
def add_expert_weights_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
def add_expert_weights_norm(self, value: bool) -> None:
self.add_bool(Keys.LLM.EXPERT_WEIGHTS_NORM.format(arch=self.arch), value)
def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None:
self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value)
def add_swin_norm(self, value: bool) -> None:
self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)

4
gguf-py/gguf/tensor_mapping.py
View file

@ -276,6 +276,10 @@ class TensorNameMap:
"model.layers.{bid}.mlp.shared_expert_gate", # qwen2moe
),
MODEL_TENSOR.FFN_EXP_PROBS_B: (
"model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
),
# Feed-forward up
MODEL_TENSOR.FFN_UP: (
"gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox