seed: 0
run_mode: 'train'
output_dir: './output'
load_checkpoint: ""
src_strategy_path_or_dir: ''
auto_trans_ckpt: False
only_save_strategy: False
resume_training: False
context:
mode: 0
device_target: "Ascend"
device_id: 0
remote_save_url: "Please input obs url on AICC platform."
runner_config:
epochs: 3
batch_size: 4
sink_mode: True
sink_size: 2
runner_wrapper:
type: MFTrainOneStepCell
scale_sense:
type: DynamicLossScaleUpdateCell
loss_scale_value: 4294967296
scale_factor: 2
scale_window: 1000
use_clip_grad: True
use_parallel: False
parallel:
parallel_mode: 0
gradients_mean: True
search_mode: "sharding_propagation"
enable_parallel_optimizer: False
parallel_config:
data_parallel: 1
model_parallel: 1
pipeline_stage: 1
use_seq_parallel: False
micro_batch_num: 1
vocab_emb_dp: True
gradient_aggregation_group: 4
micro_batch_interleave_num: 1
moe_config:
expert_num: 1
capacity_factor: 1.05
aux_loss_factor: 0.05
num_experts_chosen: 1
recompute_config:
recompute: False
select_recompute: False
parallel_optimizer_comm_recompute: False
mp_comm_recompute: True
recompute_slice_activation: False
auto_tune: False
filepath_prefix: './autotune'
autotune_per_step: 10
profile: False
profile_start_step: 1
profile_stop_step: 10
init_start_profile: False
profile_communication: False
profile_memory: True
trainer:
type: CausalLanguageModelingTrainer
model_name: 'gpt2'
do_eval: False
train_dataset: &train_dataset
data_loader:
type: MindDataset
dataset_dir: ""
shuffle: True
input_columns: ["input_ids", "attention_mask"]
num_parallel_workers: 8
python_multiprocessing: False
drop_remainder: True
batch_size: 8
repeat: 1
numa_enable: False
prefetch_size: 1
train_dataset_task:
type: CausalLanguageModelDataset
dataset_config: *train_dataset
eval_dataset: &eval_dataset
data_loader:
type: MindDataset
dataset_dir: ""
shuffle: False
input_columns: ["input_ids", "attention_mask"]
num_parallel_workers: 8
python_multiprocessing: False
drop_remainder: False
repeat: 1
numa_enable: False
prefetch_size: 1
eval_dataset_task:
type: CausalLanguageModelDataset
dataset_config: *eval_dataset
model:
model_config:
type: GPT2Config
seq_length: 1024
vocab_size: 50257
hidden_size: 768
num_layers: 12
num_heads: 12
expand_ratio: 4
hidden_act: "gelu"
use_flash_attention: False
use_prompt_flash_attention: False
hidden_dropout_rate: 0.1
attention_dropout_rate: 0.1
param_init_type: "float32"
layernorm_compute_type: "float32"
softmax_compute_type: "float32"
compute_dtype: "float16"
checkpoint_name_or_path: "gpt2"
eos_token_id: 50256
repetition_penalty: 1
max_decode_length: 1024
top_k: 5
top_p: 1
do_sample: True
use_past: False
arch:
type: GPT2LMHeadModel
lr_schedule:
type: polynomial
learning_rate: 0.0001
lr_end: 0.00001
warmup_steps: 0
total_steps: -1
layer_scale: False
layer_decay: 0.65
optimizer:
type: FusedAdamWeightDecay
beta1: 0.9
beta2: 0.95
eps: 0.00000001
weight_decay: 0.1
lr_scale: False
lr_scale_factor: 256
callbacks:
- type: MFLossMonitor
- type: CheckpointMonitor
prefix: "gpt"
save_checkpoint_steps: 1
integrated_save: True
save_network_params: True
save_trainable_params: False
async_save: False
- type: ObsMonitor
eval_callbacks:
- type: ObsMonitor
metric:
type: PerplexityMetric
processor:
return_tensors: ms
tokenizer:
unk_token: '<|endoftext|>'
bos_token: '<|endoftext|>'
eos_token: '<|endoftext|>'
pad_token: '<|endoftext|>'
type: GPT2Tokenizer
type: GPT2Processor