Spaces:
Running
Running
File size: 8,372 Bytes
e81015c |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 |
# 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.
import os
import shutil
from typing import TYPE_CHECKING, Any, Optional
import torch
import torch.distributed as dist
from transformers import EarlyStoppingCallback, PreTrainedModel
from ..data import get_template_and_fix_tokenizer
from ..extras import logging
from ..extras.constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
from ..extras.misc import infer_optim_dtype
from ..extras.packages import is_ray_available
from ..hparams import get_infer_args, get_ray_args, get_train_args, read_args
from ..model import load_model, load_tokenizer
from .callbacks import LogCallback, PissaConvertCallback, ReporterCallback
from .dpo import run_dpo
from .kto import run_kto
from .ppo import run_ppo
from .pt import run_pt
from .rm import run_rm
from .sft import run_sft
from .trainer_utils import get_ray_trainer, get_swanlab_callback
if is_ray_available():
import ray
from ray.train.huggingface.transformers import RayTrainReportCallback
if TYPE_CHECKING:
from transformers import TrainerCallback
logger = logging.get_logger(__name__)
def _training_function(config: dict[str, Any]) -> None:
args = config.get("args")
callbacks: list[Any] = config.get("callbacks")
model_args, data_args, training_args, finetuning_args, generating_args = get_train_args(args)
callbacks.append(LogCallback())
if finetuning_args.pissa_convert:
callbacks.append(PissaConvertCallback())
if finetuning_args.use_swanlab:
callbacks.append(get_swanlab_callback(finetuning_args))
if finetuning_args.early_stopping_steps is not None:
callbacks.append(EarlyStoppingCallback(early_stopping_patience=finetuning_args.early_stopping_steps))
callbacks.append(ReporterCallback(model_args, data_args, finetuning_args, generating_args)) # add to last
if finetuning_args.stage == "pt":
run_pt(model_args, data_args, training_args, finetuning_args, callbacks)
elif finetuning_args.stage == "sft":
run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
elif finetuning_args.stage == "rm":
run_rm(model_args, data_args, training_args, finetuning_args, callbacks)
elif finetuning_args.stage == "ppo":
run_ppo(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
elif finetuning_args.stage == "dpo":
run_dpo(model_args, data_args, training_args, finetuning_args, callbacks)
elif finetuning_args.stage == "kto":
run_kto(model_args, data_args, training_args, finetuning_args, callbacks)
else:
raise ValueError(f"Unknown task: {finetuning_args.stage}.")
if is_ray_available() and ray.is_initialized():
return # if ray is intialized it will destroy the process group on return
try:
if dist.is_initialized():
dist.destroy_process_group()
except Exception as e:
logger.warning(f"Failed to destroy process group: {e}.")
def run_exp(args: Optional[dict[str, Any]] = None, callbacks: Optional[list["TrainerCallback"]] = None) -> None:
args = read_args(args)
if "-h" in args or "--help" in args:
get_train_args(args)
ray_args = get_ray_args(args)
callbacks = callbacks or []
if ray_args.use_ray:
callbacks.append(RayTrainReportCallback())
trainer = get_ray_trainer(
training_function=_training_function,
train_loop_config={"args": args, "callbacks": callbacks},
ray_args=ray_args,
)
trainer.fit()
else:
_training_function(config={"args": args, "callbacks": callbacks})
def export_model(args: Optional[dict[str, Any]] = None) -> None:
model_args, data_args, finetuning_args, _ = get_infer_args(args)
if model_args.export_dir is None:
raise ValueError("Please specify `export_dir` to save model.")
if model_args.adapter_name_or_path is not None and model_args.export_quantization_bit is not None:
raise ValueError("Please merge adapters before quantizing the model.")
tokenizer_module = load_tokenizer(model_args)
tokenizer = tokenizer_module["tokenizer"]
processor = tokenizer_module["processor"]
template = get_template_and_fix_tokenizer(tokenizer, data_args)
model = load_model(tokenizer, model_args, finetuning_args) # must after fixing tokenizer to resize vocab
if getattr(model, "quantization_method", None) is not None and model_args.adapter_name_or_path is not None:
raise ValueError("Cannot merge adapters to a quantized model.")
if not isinstance(model, PreTrainedModel):
raise ValueError("The model is not a `PreTrainedModel`, export aborted.")
if getattr(model, "quantization_method", None) is not None: # quantized model adopts float16 type
setattr(model.config, "torch_dtype", torch.float16)
else:
if model_args.infer_dtype == "auto":
output_dtype = getattr(model.config, "torch_dtype", torch.float32)
if output_dtype == torch.float32: # if infer_dtype is auto, try using half precision first
output_dtype = infer_optim_dtype(torch.bfloat16)
else:
output_dtype = getattr(torch, model_args.infer_dtype)
setattr(model.config, "torch_dtype", output_dtype)
model = model.to(output_dtype)
logger.info_rank0(f"Convert model dtype to: {output_dtype}.")
model.save_pretrained(
save_directory=model_args.export_dir,
max_shard_size=f"{model_args.export_size}GB",
safe_serialization=(not model_args.export_legacy_format),
)
if model_args.export_hub_model_id is not None:
model.push_to_hub(
model_args.export_hub_model_id,
token=model_args.hf_hub_token,
max_shard_size=f"{model_args.export_size}GB",
safe_serialization=(not model_args.export_legacy_format),
)
if finetuning_args.stage == "rm":
if model_args.adapter_name_or_path is not None:
vhead_path = model_args.adapter_name_or_path[-1]
else:
vhead_path = model_args.model_name_or_path
if os.path.exists(os.path.join(vhead_path, V_HEAD_SAFE_WEIGHTS_NAME)):
shutil.copy(
os.path.join(vhead_path, V_HEAD_SAFE_WEIGHTS_NAME),
os.path.join(model_args.export_dir, V_HEAD_SAFE_WEIGHTS_NAME),
)
logger.info_rank0(f"Copied valuehead to {model_args.export_dir}.")
elif os.path.exists(os.path.join(vhead_path, V_HEAD_WEIGHTS_NAME)):
shutil.copy(
os.path.join(vhead_path, V_HEAD_WEIGHTS_NAME),
os.path.join(model_args.export_dir, V_HEAD_WEIGHTS_NAME),
)
logger.info_rank0(f"Copied valuehead to {model_args.export_dir}.")
try:
tokenizer.padding_side = "left" # restore padding side
tokenizer.init_kwargs["padding_side"] = "left"
tokenizer.save_pretrained(model_args.export_dir)
if model_args.export_hub_model_id is not None:
tokenizer.push_to_hub(model_args.export_hub_model_id, token=model_args.hf_hub_token)
if processor is not None:
processor.save_pretrained(model_args.export_dir)
if model_args.export_hub_model_id is not None:
processor.push_to_hub(model_args.export_hub_model_id, token=model_args.hf_hub_token)
except Exception as e:
logger.warning_rank0(f"Cannot save tokenizer, please copy the files manually: {e}.")
ollama_modelfile = os.path.join(model_args.export_dir, "Modelfile")
with open(ollama_modelfile, "w", encoding="utf-8") as f:
f.write(template.get_ollama_modelfile(tokenizer))
logger.info_rank0(f"Ollama modelfile saved in {ollama_modelfile}")
|