基于PyTorch通信算子的分布式训练壅闭定位方法 - IT评测·应用市场-qidao123.com

import torch.distributed as dist
import torch.distributed
from collections import defaultdict
call_counts = defaultdict(int)
def recursive_tensor_processor(data, op_name, phase):
"""递归处理通信算子输入输出张量
Args:
data: 待处理数据(支持Tensor/List/Dict)
op_name: 通信算子名称
phase: 处理阶段(Input/Output)
"""
if torch.distributed.get_rank() != 0: # 仅主节点记录
return
if isinstance(data, torch.Tensor):
operation_stats[op_name] += 1
log_message = (
f"[{op_name}] {phase} #{operation_stats[op_name]} | "
f"Shape: {data.shape} | "
f"Mean: {data.float().mean().item():.4f} | "
f"Dtype: {data.dtype}"
)
print(log_message)
elif isinstance(data, (dict, list)):
container = data.items() if isinstance(data, dict) else enumerate(data)
for _, value in container:
recursive_tensor_processor(value, op_name, phase)
def create_debug_wrapper(native_func, op_name):
"""创建带调试功能的通信算子包装器
功能特性：
1. 设备同步保证时序准确性
2. 输入输出双向追踪
3. 异常处理扩展点
"""
def wrapped_function(tensor, *args, **kwargs):
# 前处理
torch.cuda.synchronize()
recursive_tensor_processor(tensor, op_name, "Input")
# 执行原生操作
result = native_func(tensor, *args, **kwargs)
# 后处理
torch.cuda.synchronize()
recursive_tensor_processor(tensor, op_name, "Output")
return result
return wrapped_function
import torch.distributed as dist
from collections import defaultdict
# 调试统计信息
operation_stats = defaultdict(int)
TRACKED_OPERATIONS = [
'all_reduce', 'reduce_scatter', 'reduce',
'all_gather', 'all_to_all', 'scatter',
'gather', 'broadcast', 'send', 'recv',
'all_to_all_single', 'batch_isend_irecv',
'isend', 'irecv'
]
def instrument_communication_ops():
"""注入通信算子调试功能"""
original_functions = {}
for op_name in TRACKED_OPERATIONS:
native_func = getattr(dist, op_name)
original_functions[op_name] = native_func
debug_wrapper = create_debug_wrapper(native_func, op_name)
setattr(dist, op_name, debug_wrapper)
return original_functions
def main():
pretrain(
train_valid_test_datasets_provider,
model_provider,
ModelType.encoder_or_decoder,
forward_step,
args_defaults={'tokenizer_type': 'GPT2BPETokenizer'},
)
if __name__ == "__main__":
# 注入调试功能
original_apis = instrument_communication_ops()
# 启动执行追踪
import sys
from trace import Trace
tracer = Trace(
count=False,
trace=True,
ignoredirs=[
sys.prefix,
sys.exec_prefix,
os.path.dirname(os.__file__)
]
)
tracer.run('main()')

复制代码