File size: 6,100 Bytes
b0c0df0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 |
# Copyright 2025 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import TYPE_CHECKING, Union
from transformers.integrations import is_deepspeed_zero3_enabled
from ...extras.misc import check_version
if TYPE_CHECKING:
from torch import nn
from transformers import PretrainedConfig, PreTrainedModel
from ...hparams import ModelArguments
def _set_z3_leaf_modules(model: "PreTrainedModel", leaf_modules: list[Union["nn.Module", str]]) -> None:
check_version("deepspeed>=0.13.0")
from deepspeed.utils import set_z3_leaf_modules # type: ignore
set_z3_leaf_modules(model, leaf_modules)
def add_z3_leaf_module(model: "PreTrainedModel") -> None:
r"""Set module as a leaf module to skip partitioning in deepspeed zero3."""
if not is_deepspeed_zero3_enabled():
return
model_type = getattr(model.config, "model_type", None)
text_config = getattr(model.config, "text_config", None)
text_model_type = getattr(text_config, "model_type", None)
if model_type == "dbrx":
from transformers.models.dbrx.modeling_dbrx import DbrxFFN
_set_z3_leaf_modules(model, [DbrxFFN])
if model_type == "deepseek_v2":
# deepseek v2 uses custom code
_set_z3_leaf_modules(model, ["DeepseekV2MoE"])
if model_type == "deepseek_v3" or model_type == "kimi_vl":
# deepseek v3 and kimi vl use custom code
_set_z3_leaf_modules(model, ["DeepseekV3MoE"])
if model_type == "ernie4_5_moe":
from transformers.models.ernie4_5_moe.modeling_ernie4_5_moe import Ernie4_5_MoeSparseMoeBlock
_set_z3_leaf_modules(model, [Ernie4_5_MoeSparseMoeBlock])
if model_type == "granitemoe":
from transformers.models.granitemoe.modeling_granitemoe import GraniteMoeMoE
_set_z3_leaf_modules(model, [GraniteMoeMoE])
if model_type == "glm4_moe":
from transformers.models.glm4_moe.modeling_glm4_moe import Glm4MoeMoE
_set_z3_leaf_modules(model, [Glm4MoeMoE])
if model_type == "glm4v_moe":
from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeTextMoE
_set_z3_leaf_modules(model, [Glm4vMoeTextMoE])
if model_type == "jamba":
from transformers.models.jamba.modeling_jamba import JambaSparseMoeBlock
_set_z3_leaf_modules(model, [JambaSparseMoeBlock])
if model_type == "jetmoe":
from transformers.models.jetmoe.modeling_jetmoe import JetMoeMoA, JetMoeMoE
_set_z3_leaf_modules(model, [JetMoeMoA, JetMoeMoE])
if model_type == "llama4":
from transformers.models.llama4.modeling_llama4 import Llama4TextMoe
_set_z3_leaf_modules(model, [Llama4TextMoe])
if model_type == "mixtral":
from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
_set_z3_leaf_modules(model, [MixtralSparseMoeBlock])
if model_type == "olmoe":
from transformers.models.olmoe.modeling_olmoe import OlmoeSparseMoeBlock
_set_z3_leaf_modules(model, [OlmoeSparseMoeBlock])
if model_type == "phimoe":
from transformers.models.phimoe.modeling_phimoe import PhimoeSparseMoeBlock
_set_z3_leaf_modules(model, [PhimoeSparseMoeBlock])
if model_type == "qwen2_moe":
from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
_set_z3_leaf_modules(model, [Qwen2MoeSparseMoeBlock])
if model_type == "qwen3_moe" or text_model_type == "qwen3_moe": # internvl 3.5
from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeSparseMoeBlock
_set_z3_leaf_modules(model, [Qwen3MoeSparseMoeBlock])
if model_type == "qwen3_vl_moe":
from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import Qwen3VLMoeTextSparseMoeBlock
_set_z3_leaf_modules(model, [Qwen3VLMoeTextSparseMoeBlock])
if model_type == "qwen3_omni_moe":
from transformers.models.qwen3_omni_moe.modeling_qwen3_omni_moe import Qwen3OmniMoeThinkerTextSparseMoeBlock
_set_z3_leaf_modules(model, [Qwen3OmniMoeThinkerTextSparseMoeBlock])
def configure_moe(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
if not is_trainable or not model_args.moe_aux_loss_coef:
return
model_type = getattr(config, "model_type", None)
text_config = getattr(config, "text_config", None) # for multimodal model
if model_type in [
"dbrx",
"ernie4_5_moe",
"granitemoe",
"jamba",
"jetmoe",
"llama4",
"mixtral",
"olmoe",
"phimoe",
"qwen2_moe",
"qwen3_moe",
]:
setattr(config, "output_router_logits", True)
if text_config and getattr(text_config, "model_type", None) in [
"glm4v_moe_text", # glmv4_5
"qwen3_moe", # internvl_3_5
]:
setattr(text_config, "output_router_logits", True)
if model_type in [
"ernie4_5_moe",
"granitemoe",
"jamba",
"llama4",
"mixtral",
"olmoe",
"phimoe",
"qwen2_moe",
"qwen3_moe",
]:
setattr(config, "router_aux_loss_coef", model_args.moe_aux_loss_coef)
elif text_config and getattr(text_config, "model_type", None) in ["qwen3_moe"]:
setattr(text_config, "router_aux_loss_coef", model_args.moe_aux_loss_coef)
elif model_type == "deepseek":
setattr(config, "aux_loss_alpha", model_args.moe_aux_loss_coef)
elif model_type == "jetmoe":
setattr(config, "aux_loss_coef", model_args.moe_aux_loss_coef)
|