ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 11:21:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Deepseek V3 support added (#176)

Browse Source 
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
This commit is contained in:
saood06
2025-01-23 10:24:10 -06:00
committed by GitHub
parent 6d1b4adaac
commit 5c0a01bdaf
9 changed files with 136 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -89,6 +89,8 @@ 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"
@@ -257,6 +259,7 @@ 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()
@@ -387,6 +390,7 @@ 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",
@@ -978,6 +982,7 @@ 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,
@@ -1177,6 +1182,10 @@ class GGMLQuantizationType(IntEnum):
IQ2_TN = 42,
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
@@ -670,6 +671,12 @@ 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_layer_norm_eps(self, value: float) -> None:
self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)

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

@@ -251,6 +251,10 @@ 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