1. 系统环境
硬件环境(Ascend/GPU/CPU): Ascend910
MindSpore版本: mindspore=2.2.0
执行模式（PyNative/ Graph）:不限
Python版本: Python=3.9
操作系统平台: 不限
2. 报错信息
2.1 问题描述
将qwen-7b全参微调时报错RuntimeError: Preprocess failed before run graph 1.
2.2 配置信息
seed: 0
output_dir: './output'
load_checkpoint: '/home/ma-user/work/infer/qwen_7b_base.ckpt'
src_strategy_path_or_dir: ''
auto_trans_ckpt: True # If true, auto transform load_checkpoint to load in distributed model
only_save_strategy: False
resume_training: False
use_parallel: True
run_mode: 'finetune'
# trainer config
trainer:
type: CausalLanguageModelingTrainer
model_name: 'qwen_7b'
# dataset
train_dataset: &train_dataset
data_loader:
type: MindDataset
dataset_dir: "/home/ma-user/work/infer/alpaca.mindrecord"
shuffle: True
input_columns: ["input_ids", "labels", "attention_mask"]
num_parallel_workers: 8
python_multiprocessing: False
drop_remainder: True
batch_size: 1
repeat: 1
numa_enable: False
prefetch_size: 1
train_dataset_task:
type: CausalLanguageModelDataset
dataset_config: *train_dataset
# runner config
runner_config:
epochs: 5
batch_size: 1
sink_mode: True
sink_size: 2
runner_wrapper:
type: MFTrainOneStepCell
scale_sense:
type: DynamicLossScaleUpdateCell
loss_scale_value: 65536
scale_factor: 2
scale_window: 1000
use_clip_grad: True
# optimizer
optimizer:
type: FP32StateAdamWeightDecay
beta1: 0.9
beta2: 0.95
eps: 1.e-6
weight_decay: 0.1
# lr sechdule
lr_schedule:
type: CosineWithWarmUpLR
learning_rate: 1.e-5
warmup_ratio: 0.01
total_steps: -1 # -1 means it will load the total steps of the dataset
# callbacks
callbacks:
- type: MFLossMonitor
- type: CheckpointMointor
prefix: "qwen"
save_checkpoint_steps: 10000
keep_checkpoint_max: 3
integrated_save: False
async_save: False
- type: ObsMonitor
# default parallel of device num = 8 910
parallel_config:
data_parallel: 8
model_parallel: 1
pipeline_stage: 1
micro_batch_num: 1
vocab_emb_dp: False
gradient_aggregation_group: 4
# when model parallel is greater than 1, we can set micro_batch_interleave_num=2, that may accelerate the train process.
micro_batch_interleave_num: 1
model:
model_config:
type: QwenConfig
batch_size: 1
seq_length: 1024
hidden_size: 4096
num_hidden_layers: 32
num_attention_heads: 32
vocab_size: 151936
intermediate_size: 11008
rms_norm_eps: 1.0e-6
emb_dropout_prob: 0.0
eos_token_id: 151643
pad_token_id: 151643
compute_dtype: "float16"
layernorm_compute_type: "float32"
softmax_compute_type: "float16"
rotary_dtype: "float16"
param_init_type: "float16"
use_past: True
use_flash_attention: False
use_past_shard: False
offset: 0
checkpoint_name_or_path: "/home/ma-user/work/infer/qwen_7b_base.ckpt"
repetition_penalty: 1
max_decode_length: 512
top_k: 0
top_p: 0.8
do_sample: False
# configuration items copied from Qwen
rotary_pct: 1.0
rotary_emb_base: 10000
kv_channels: 128
arch:
type: QwenForCausalLM
processor:
return_tensors: ms
tokenizer:
model_max_length: 8192
vocab_file: "/path/qwen.tiktoken"
pad_token: "<|endoftext|>"
type: QwenTokenizer
type: QwenProcessor
# mindspore context init config
context:
mode: 0 #0--Graph Mode; 1--Pynative Mode
device_target: "Ascend"
enable_graph_kernel: False
graph_kernel_flags: "--disable_expand_ops=Softmax,Dropout --enable_parallel_fusion=true --reduce_fuse_depth=8 --enable_auto_tensor_inplace=true"
ascend_config:
precision_mode: "must_keep_origin_dtype"
max_call_depth: 10000
max_device_memory: "58GB"
save_graphs: False
save_graphs_path: "./graph"
device_id: 0
# parallel context config
parallel:
parallel_mode: 1 # 0-data parallel, 1-semi-auto parallel, 2-auto parallel, 3-hybrid parallel
gradients_mean: False
enable_alltoall: False
full_batch: True
search_mode: "sharding_propagation"
enable_parallel_optimizer: True
strategy_ckpt_config:
save_file: "./ckpt_strategy.ckpt"
only_trainable_params: False
parallel_optimizer_config:
gradient_accumulation_shard: False
parallel_optimizer_threshold: 64
infer:
prefill_model_path: "/path/qwen_7b_prefill.mindir"
increment_model_path: "path/qwen_7b_inc.mindir"
infer_seq_length: 1024
2.3****报错信息
Traceback (most recent call last):
File "/home/ma-user/work/mindformers-dev/research/qwen/run_qwen.py", line 165, in <module>
main(task=args.task,
File "/home/ma-user/work/mindformers-dev/research/qwen/run_qwen.py", line 118, in main
trainer.finetune(finetune_checkpoint=ckpt, auto_trans_ckpt=auto_trans_ckpt)
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/_checkparam.py", line 1313, in wrapper
return func(*args, **kwargs)
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindformers/trainer/trainer.py", line 498, in finetune
self.trainer.train(
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindformers/trainer/causal_language_modeling/causal_language_modeling.py", line 97, in train
self.training_process(
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindformers/trainer/base_trainer.py", line 696, in training_process
transform_and_load_checkpoint(config, model, network, dataset)
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindformers/trainer/utils.py", line 311, in transform_and_load_checkpoint
build_model(config, model, dataset, do_eval=do_eval, do_predict=do_predict)
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindformers/trainer/utils.py", line 436, in build_model
model.build(train_dataset=dataset, epoch=config.runner_config.epochs,
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/train/model.py", line 1274, in build
self._init(train_dataset, valid_dataset, sink_size, epoch)
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/train/model.py", line 529, in _init
train_network.compile(*inputs)
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/nn/cell.py", line 997, in compile
_cell_graph_executor.compile(self, phase=self.phase,
File "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/common/api.py", line 1547, in compile
result = self._graph_executor.compile(obj, args, kwargs, phase, self._use_vm_mode())
RuntimeError: Preprocess failed before run graph 1.
----------------------------------------------------
- Framework Error Message:
----------------------------------------------------
Out of Memory!!! Request memory size: 35051914240B, Memory Statistic:
Device HBM memory size: 32768M
MindSpore Used memory size: 30678M
MindSpore memory base address: 0x124180000000
Total Static Memory size: 12040M
Total Dynamic memory size: 0M
Dynamic memory size of this graph: 0M
Please try to reduce 'batch_size' or check whether exists extra large shape. For more details, please refer to 'Out of Memory' at https://www.mindspore.cn .
----------------------------------------------------
- C++ Call Stack: (For framework developers)
----------------------------------------------------
mindspore/ccsrc/plugin/device/ascend/hal/hardware/ascend_kernel_executor.cc:286 PreprocessBeforeRunGraph
mindspore/ccsrc/plugin/device/ascend/hal/device/ascend_memory_adapter.cc:185 MallocDynamicDevMem
[WARNING] MD(123569,ffffa9748010,python):2024-01-04-14:24:06.999.824 [mindspore/ccsrc/minddata/dataset/engine/datasetops/data_queue_op.cc:163] ~DataQueueOp]
preprocess_batch: 100;
batch_queue: 1, 1, 1, 1, 1, 1, 1, 1, 1, 1;
push_start_time -> push_end_time
2024-01-04-14:18:18.185.953 -> 2024-01-04-14:18:18.186.203
2024-01-04-14:18:18.205.302 -> 2024-01-04-14:18:18.220.553
2024-01-04-14:18:18.220.798 -> 2024-01-04-14:18:18.221.154
2024-01-04-14:18:18.221.300 -> 2024-01-04-14:18:18.221.999
2024-01-04-14:18:18.222.250 -> 2024-01-04-14:18:18.222.646
2024-01-04-14:18:18.223.158 -> 2024-01-04-14:18:18.223.419
2024-01-04-14:18:18.223.665 -> 2024-01-04-14:18:18.223.945
2024-01-04-14:18:18.224.214 -> 2024-01-04-14:18:18.224.818
2024-01-04-14:18:18.225.096 -> 2024-01-04-14:18:18.225.442
2024-01-04-14:18:18.225.901 -> 2024-01-04-14:18:18.226.317
For more details, please refer to the FAQ at https://www.mindspore.cn/docs/en/master/faq/data_processing.html.
3. 根因分析
单从报错上看是卡内存不足
Out of Memory!!! Request memory size: 35051914240B, Memory Statistic: Device HBM memory size: 32768M MindSpore Used memory size: 30678M
但是看yaml配置文件
batch_size: 1
seq_length: 1024
这个配置设置并不高，按理说910的32g是可以支持的。
然后再配置文件里面有看到另外一个参数
max_device_memory: "58GB"
这个明显不对了，这个应该是给设备内存64G用的。910只有32G，所以设置超限制了。
4. 解决方案
修改max_device_memory
910最大支持32G，去除默认占用的，可以设置为"29G"