import asyncio
import os
import re
import subprocess
import sys
import time
import jsonlines as jl
from data_juicer.utils.file_utils import follow_read
from data_juicer.utils.lazy_loader import LazyLoader
from data_juicer.utils.model_utils import (
get_model,
prepare_model,
update_sampling_params,
)
from jsonargparse import namespace_to_dict
from loguru import logger
from data_juicer_sandbox.data_pool_manipulators import check_io_paths
from data_juicer_sandbox.helper_funcs import ALL_FUNCS
[文档]
class BaseModelExecutor(object):
"""
Base abstraction for model executor within the DataJuicer's sandbox
"""
[文档]
def __init__(self, model_config: dict, watcher=None):
self.model_config = model_config
self.executor = None
self.watcher = watcher
# log to tell the end of model execution
self.END_OF_MODEL_EXEC = "<DJ-Sandbox> End of ModelExecutor's running <DJ-Sandbox>"
[文档]
async def run(self, run_type, run_obj=None, **kwargs):
"""
conduct some model-related execution tasks
given specified run_type and run_obj
"""
watch_task = asyncio.create_task(self.watch_run(run_type, run_obj, **kwargs))
if self.watcher is None:
run_task = asyncio.create_task(self._run(run_type, run_obj, **kwargs))
ret = await run_task
return ret
else:
original_stdout = sys.stdout
original_stderr = sys.stderr
try:
timestamp = time.strftime("%Y%m%d%H%M%S", time.localtime(time.time()))
log_f_name = os.path.join(self.watcher.sandbox_cfg.work_dir, f"model_exe_{run_type}_{timestamp}.log")
self.watcher.model_exe_log_file = log_f_name
with open(log_f_name, "w") as log_f:
sys.stdout = log_f
sys.stderr = log_f
run_task = asyncio.create_task(self._run(run_type, run_obj, **kwargs))
ret = await run_task
print(self.END_OF_MODEL_EXEC, flush=True)
await watch_task
finally:
sys.stdout = original_stdout
sys.stderr = original_stderr
return ret
[文档]
def run_subprocess(self, script_path, run_args, working_dir, cmd="bash"):
run_args = [str(arg) for arg in run_args]
args = [cmd, script_path] + run_args
subprocess.run(args, cwd=working_dir)
async def _run(self, run_type, run_obj=None, **kwargs):
raise NotImplementedError
[文档]
async def watch_run(self, run_type, run_obj=None, **kwargs):
"""
watch the running process in an online manner, and
return the summarized results
"""
met_eof = False
while not met_eof:
if os.path.exists(self.watcher.model_exe_log_file):
async for line in follow_read(self.watcher.model_exe_log_file):
if self.END_OF_MODEL_EXEC in line:
met_eof = True
break
else:
self._watch_run(line, **kwargs)
else:
await asyncio.sleep(0.1)
def _watch_run(self, line, **kwargs):
"""
customized log watcher, depending on specific model executor
"""
raise NotImplementedError
[文档]
def data_connector(self, input_data, **kwargs):
"""
convert input_data (usually in Data-Juicer's Dataset format) into
the appropriate format for specific model executor
"""
raise NotImplementedError
[文档]
class ModelScopeExecutor(BaseModelExecutor):
[文档]
def data_connector(self, input_data, split="train", key_remapping=None, **kwargs):
try:
from modelscope.msdatasets import MsDataset
if isinstance(input_data, str):
# dataset path
ds = MsDataset.load(input_data, split=split).ds_instance
else:
ds = input_data
if key_remapping:
features = ds.features
for key in key_remapping:
if key in features:
ds = ds.rename_column(key, key_remapping[key])
return ds
except ModuleNotFoundError:
raise ModuleNotFoundError("modelscope package not installed")
def _watch_run(self, line, **kwargs):
# e.g., "2023-07-02 17:26:50,324 - modelscope - INFO - epoch
# [1][100/4953]\tlr: 4.825e-05, memory: 8742, loss: 5.6125\n"
pattern = r"loss:\s*([0-9.]+)"
if self.watcher is not None:
match = re.search(pattern, line)
if match:
loss_value = float(match.group(1))
self.watcher.watch(loss_value, "loss")
[文档]
class ModelscopeInferProbeExecutor(ModelScopeExecutor):
[文档]
def __init__(self, model_config: dict):
super().__init__(model_config)
try:
from modelscope.pipelines import pipeline
self.executor = pipeline(**model_config)
except ModuleNotFoundError:
raise ModuleNotFoundError(
"`modelscope` package not installed. Please refer to the doc: "
"https://datajuicer.github.io/data-juicer/en/main/docs/Sandbox.html#requirements"
)
async def _run(self, run_type, run_obj=None, **kwargs):
if run_type == "infer_on_data":
return self.executor(self.data_connector(run_obj), **kwargs)
else:
raise ValueError(f"run_type {run_type} not supported")
[文档]
class ModelscopeTrainExecutor(ModelScopeExecutor):
[文档]
def __init__(self, model_config, watcher=None):
super().__init__(model_config, watcher)
self.model_config = namespace_to_dict(model_config)
self.executor = None
[文档]
def cfg_modify_fn(self, cfg):
cfg.merge_from_dict(self.model_config)
return cfg
[文档]
def build_executor(self, model_name, trainer_name, work_dir, train_dataset=None, eval_dataset=None):
try:
from modelscope.trainers import build_trainer
kwargs = dict(
model=model_name,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
work_dir=work_dir,
cfg_modify_fn=self.cfg_modify_fn,
)
self.executor = build_trainer(
name=trainer_name,
default_args=kwargs,
)
except ModuleNotFoundError:
raise ModuleNotFoundError("modelscope package not installed")
async def _run(self, run_type, run_obj=None, **kwargs):
# training cfg updated, such as datasets and training parameters
builder_kwargs = {
"model_name": self.model_config["model_name"],
"trainer_name": self.model_config["trainer_name"],
}
if "key_remapping" in self.model_config:
key_remapping = self.model_config["key_remapping"]
else:
key_remapping = None
if "train_dataset" in self.model_config:
builder_kwargs["train_dataset"] = self.data_connector(
self.model_config["train_dataset"], split="train", key_remapping=key_remapping
)
if "eval_dataset" in self.model_config:
builder_kwargs["eval_dataset"] = self.data_connector(
self.model_config["eval_dataset"], split="val", key_remapping=key_remapping
)
if "work_dir" in self.model_config:
builder_kwargs["work_dir"] = self.model_config["work_dir"]
self.work_dir = builder_kwargs["work_dir"]
self.build_executor(**builder_kwargs)
self.executor.train()
return self.work_dir
[文档]
class LLMInferExecutor(BaseModelExecutor):
"""
A inference executor for LLM inference.
The model preparation method should be implemented by the subclass for specific type of model.
The config file for this type of executor should at least include the following items:
1. `type`: model type.
2. `build_messages_func`: the helper func to build the messages.
3. `parse_output_func`: the helper func to build the messages.
4. `dataset_path`: the input datasets or data pools use to construct the input messages for LLM inference.
Only support jsonl files for now.
5. `export_path`: the output dir to store the inference results.
6. `infer_res_key`: the key name to store the inference results. It's "response" in default.
"""
[文档]
def __init__(self, model_config: dict, watcher=None):
super().__init__(model_config, watcher)
# model related params
self.executor, self.sampling_params = self.prepare_executor()
# inference format related params
self.build_messages_func = model_config.get("build_messages_func", "build_input")
self.parse_output_func = model_config.get("parse_output_func", "parse_output")
self.func_kwargs = model_config.get("func_kwargs", {})
self.build_messages_func = ALL_FUNCS.get(self.build_messages_func)
self.parse_output_func = ALL_FUNCS.get(self.parse_output_func)
# inference dataset related
self.dataset_path = model_config.get("dataset_path", [])
self.export_path = model_config.get("export_path", None)
self.infer_res_key = model_config.get("infer_res_key", "response")
if isinstance(self.dataset_path, str):
self.dataset_path = [self.dataset_path]
[文档]
def prepare_executor(self):
raise NotImplementedError
[文档]
def executor_infer(self, messages):
raise NotImplementedError
async def _run(self, run_type, run_obj=None, **kwargs):
# check I/O paths
existing_input_paths, export_path = check_io_paths(self.dataset_path, self.export_path)
output_paths = []
for input_path in existing_input_paths:
input_basename = os.path.splitext(os.path.basename(input_path))[0]
output_path = os.path.join(export_path, f"{input_basename}_inference.jsonl")
with jl.open(output_path, "w") as writer:
with jl.open(input_path) as reader:
for item in reader:
non_batch = False
messages_list = self.build_messages_func(item, **self.func_kwargs)
if len(messages_list) > 0 and not isinstance(messages_list[0], list):
messages_list = [messages_list]
non_batch = True
results = []
for messages in messages_list:
output = self.executor_infer(messages)
results.append(self.parse_output_func(output, item, **self.func_kwargs))
if non_batch:
item[self.infer_res_key] = results[0]
else:
item[self.infer_res_key] = results
writer.write(item)
output_paths.append(output_path)
return output_paths
[文档]
class VLLMInferExecutor(LLMInferExecutor):
"""
A inference executor for model inference with vLLM.
The config file for this executor should at least include the following items:
1. `type`: must be "vllm".
2. `model_path`: the path to the vLLM model.
3. `model_params`: extra parameters for the model.
4. `sampling_params`: extra sampling parameters for the model.
5. other parameters can be referred to the class LLMInferExecutor
"""
[文档]
def __init__(self, model_config: dict, watcher=None):
super().__init__(model_config, watcher)
[文档]
def prepare_executor(self):
# model related params
torch = LazyLoader("torch")
vllm = LazyLoader("vllm")
model = self.model_config.get("model", None)
model_params = self.model_config.get("model_params", {})
sampling_params = self.model_config.get("sampling_params", {})
if model_params.get("tensor_parallel_size") is None:
tensor_parallel_size = torch.cuda.device_count()
logger.info(
f"Set tensor_parallel_size to \
{tensor_parallel_size} for vllm."
)
model_params["tensor_parallel_size"] = tensor_parallel_size
executor, _ = get_model(
prepare_model(
model_type="vllm",
pretrained_model_name_or_path=model,
**model_params,
),
use_cuda=True,
)
sampling_params = vllm.SamplingParams(**update_sampling_params(sampling_params, model, False))
return executor, sampling_params
[文档]
def executor_infer(self, messages):
response = self.executor.chat(messages, self.sampling_params)
output = response[0].outputs[0].text
return output
[文档]
class APIModelInferExecutor(LLMInferExecutor):
"""
A inference executor for model inference with OpenAI API.
The config file for this executor should at least include the following items:
1. `type`: must be "api".
2. `model`: the API model used to inference.
3. `model_params`: extra parameters for the model.
4. `sampling_params`: extra sampling parameters for the model.
5. `api_endpoint`: URL endpoint for the API.
6. `response_path`: Path to extract content from the API response. Defaults to 'choices.0.message.content'.
7. `max_retry_num`: the max number of retries when the API request fails.
8. other parameters can be referred to the class LLMInferExecutor
"""
[文档]
def __init__(self, model_config: dict, watcher=None):
super().__init__(model_config, watcher)
# model related params
self.max_retry_num = self.model_config.get("max_retry_num", 5)
[文档]
def prepare_executor(self):
# model related params
api_endpoint = self.model_config.get("api_endpoint", None)
response_path = self.model_config.get("response_path", None)
model = self.model_config.get("model", None)
model_params = self.model_config.get("model_params", {})
sampling_params = self.model_config.get("sampling_params", {})
executor = get_model(
prepare_model(
model_type="api",
model=model,
endpoint=api_endpoint,
response_path=response_path,
**model_params,
)
)
sampling_params = sampling_params
return executor, sampling_params
[文档]
def executor_infer(self, messages):
try_count = 0
while try_count <= self.max_retry_num:
try:
output = self.executor(messages, **self.sampling_params)
return output
except Exception as e:
logger.warning(f"error: {e} -- retries: {try_count}")
try_count += 1
if try_count > self.max_retry_num:
logger.error("Retried too many times. Abort!")
raise
time.sleep(try_count * 1)
[文档]
class LLaVAExecutor(BaseModelExecutor):
[文档]
def __init__(self, model_config: dict):
super().__init__(model_config)
raise NotImplementedError("To be refactored, from DJ's refinement for " "LLaVA related experiments.")
[文档]
class LLaMAFactoryExecutor(BaseModelExecutor):
[文档]
def __init__(self, model_config: dict):
super().__init__(model_config)
raise NotImplementedError("To be refactored, which is used in " "data-juicer competition.")
[文档]
class MegatronExecutor(BaseModelExecutor):
[文档]
def __init__(self, model_config: dict):
super().__init__(model_config)
raise NotImplementedError("To be refactored from dj's `thirdparty`.")
