Ascend910A Lora微调Qwen1.5 7B 报错Cast failed, original value: 0.0883883, type: FP32Imm 解决方法

环境

• MS：mindspore2.2.13
• CANN:7.0
• mindformers:
  • version:1.0
  • branch: 1.0
  • tag:ac544ba7b4346700142c0cf4d0d73391d8548504
• npu: Ascend910A * 8
• config:

seed: 0
output_dir: './output' # path to save checkpoint/strategy
load_checkpoint: '/mnt/dataset/mindformers/research/pretrain/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: 'qwen2_7b'

# dataset
train_dataset: &train_dataset
  data_loader:
    type: MindDataset
    dataset_dir: "/mnt/dataset/alpaca_ds/mindrecord/alpaca-messages.mindrecord"
    shuffle: True
  input_columns: ["input_ids", "target_ids", "attention_mask"]
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: True
  batch_size: 2
  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: 2
  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: 8.e-4
  lr_end: 3.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: "qwen2"
    save_checkpoint_steps: 10000
    keep_checkpoint_max: 3
    integrated_save: False
    async_save: False
  - type: ObsMonitor

# default parallel of device num = 8 for Atlas 800T A2
parallel_config:
  data_parallel: 2
  model_parallel: 4
  pipeline_stage: 1
  micro_batch_num: 4
  vocab_emb_dp: True
  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

# recompute config
recompute_config:
  recompute: False
  select_recompute: False
  parallel_optimizer_comm_recompute: False
  mp_comm_recompute: False
  recompute_slice_activation: False

model:
  model_config:
    type: LlamaConfig
    batch_size: 1
    seq_length: 2048
    hidden_size: 4096
    num_layers: 32
    num_heads: 32
    vocab_size: 151936
    intermediate_size: 11008
    qkv_has_bias: True
    rms_norm_eps: 1.0e-6
    theta: 1000000.0
    emb_dropout_prob: 0.0
    eos_token_id: 151643
    pad_token_id: 151643
    bos_token_id: 151643
    compute_dtype: "float16"
    layernorm_compute_type: "float16"
    softmax_compute_type: "float16"
    rotary_dtype: "float16"
    param_init_type: "float16"
    use_past: True
    use_flash_attention: False
    use_paged_attention: False
    # block_size: 16
    # num_blocks: 512
    use_past_shard: False
    offset: 0
    checkpoint_name_or_path: ""
    repetition_penalty: 1
    max_decode_length: 512
    top_k: 0
    top_p: 0.8
    do_sample: False

    pet_config:
      pet_type: lora
      # configuration of lora
      lora_rank: 64
      lora_alpha: 16
      lora_dropout: 0.05
      target_modules: '.*wq|.*wk|.*wv|.*wo|.*w1|.*w2|.*w3'
      freeze_exclude: ["*wte*", "*lm_head*"]

    # configuration items copied from Qwen
    rotary_pct: 1.0
    rotary_emb_base: 10000
    kv_channels: 128

  arch:
    type: LlamaForCausalLM

processor:
  return_tensors: ms
  tokenizer:
    model_max_length: 32768
    vocab_file: "/mnt/dataset/mindformers/research/pretrain/vocab.json"
    merges_file: "/mnt/dataset/mindformers/research/pretrain/merges.txt"
    unk_token: "<|endoftext|>"
    eos_token: "<|endoftext|>"
    pad_token: "<|endoftext|>"
    type: Qwen2Tokenizer
  type: Qwen2Processor

# 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: "31GB"
  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

、

问题描述

910A使用MindFormers1.0 跑qwen1_5 7b的lora微调报错，Cast failed, original value: 0.0883883, type: FP32Imm，如下：

  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: Cast failed, original value: 0.0883883, type: FP32Imm

----------------------------------------------------
- Framework Unexpected Exception Raised:
----------------------------------------------------
This exception is caused by framework's unexpected error. Please create an issue at https://gitee.com/mindspore/mindspore/issues to get help.

----------------------------------------------------
- C++ Call Stack: (For framework developers)
----------------------------------------------------
mindspore/core/ir/anf.h:910 GetValue

[WARNING] MD(67780,ffffb77d0010,python):2024-05-11-03:04:21.629.364 [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;

解决办法

如下，配置文件修改项：

    use_flash_attention: False
    use_paged_attention: False