seed: 42
output_dir: './output'
load_checkpoint: ''
load_ckpt_format: 'safetensors'
src_strategy_path_or_dir: ''
auto_trans_ckpt: False
only_save_strategy: False
resume_training: False
use_parallel: True
run_mode: 'train'
train_precision_sync: False
trainer:
type: CausalLanguageModelingTrainer
model_name: 'deepseekV3'
runner_config:
epochs: 1
batch_size: 1
sink_mode: True
sink_size: 1
optimizer:
type: AdamW
betas: [0.9, 0.95]
eps: 1.e-8
weight_decay: 0.01
swap: True
lr_schedule:
type: ConstantWarmUpLR
learning_rate: 2.e-5
lr_end: 2.e-5
warmup_lr_init: 0
warmup_steps: 200
total_steps: -1
train_dataset: &train_dataset
data_loader:
type: BlendedMegatronDatasetDataLoader
datasets_type: "GPTDataset"
shuffle: False
sizes:
- 128000
- 0
- 0
config:
seed: 1234
split: "1, 0, 0"
seq_length: 4096
eod_mask_loss: False
reset_position_ids: False
create_attention_mask: False
reset_attention_mask: False
create_compressed_eod_mask: False
eod_pad_length: 128
eod: 1
pad: -1
num_dataset_builder_threads: 1
data_path:
- '1'
- "/path/megatraon_dataset/wiki103-megatron_text_document"
num_parallel_workers: 8
python_multiprocessing: False
drop_remainder: True
train_dataset_task:
type: CausalLanguageModelDataset
dataset_config: *train_dataset
context:
mode: 0
device_target: "Ascend"
max_call_depth: 10000
max_device_memory: "56GB"
mempool_block_size: "56GB"
memory_optimize_level: "O0"
save_graphs: False
save_graphs_path: "./output/graphs"
jit_config:
jit_level: "O0"
parallel_config:
data_parallel: 4
model_parallel: 8
pipeline_stage: 8
expert_parallel: 32
micro_batch_num: 32
use_seq_parallel: True
gradient_aggregation_group: 4
micro_batch_interleave_num: 1
parallel:
parallel_mode: 1
gradients_mean: False
enable_alltoall: True
search_mode: "sharding_propagation"
enable_parallel_optimizer: True
pipeline_config:
pipeline_interleave: True
pipeline_scheduler: seqpipe
strategy_ckpt_config:
save_file: "./ckpt_strategy.ckpt"
only_trainable_params: False
parallel_optimizer_config:
gradient_accumulation_shard: False
parallel_optimizer_threshold: 64
optimizer_weight_shard_size: 256
recompute_config:
recompute: True
select_recompute: False
mp_comm_recompute: True
use_legacy: False
model:
model_config:
model_type: "deepseek_v3"
architectures: ["DeepseekV3ForCausalLM"]
offset: [[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, -2]]
pp_interleave_num: 2
seq_length: 4096
num_hidden_layers: 61
vocab_size: 129280
hidden_size: 7168
intermediate_size: 18432
add_bias_linear: False
use_flash_attention: True
use_attn_mask_compression: True
multi_latent_attention: True
mla_qkv_concat: False
qk_layernorm: True
attention_dropout: 0.0
hidden_dropout: 0.0
params_dtype: "float32"
compute_dtype: "bfloat16"
layernorm_compute_dtype: "float32"
softmax_compute_dtype: "float32"
initializer_range: 0.01
num_nextn_predict_layers: 1
mtp_loss_factor: 0.3
apply_rope_fusion: False
bias_swiglu_fusion: False
router_dense_type: "float32"
gated_linear_unit: True
moe_grouped_gemm: True
use_fused_ops_topkrouter: True
n_group: 8
topk_group: 4
use_pad_tokens: False
moe_token_dispatcher_type: "alltoall"
moe_intermediate_size: 2048
moe_router_load_balancing_type: 'seq_aux_loss'
moe_aux_loss_coeff: 0.0001
moe_token_drop_policy: False
moe_router_enable_expert_bias: True
moe_router_bias_update_rate: 0.0001
callbacks:
- type: MFLossMonitor
- type: TopkBiasBalanceCallback
- type: CheckpointMonitor
prefix: "deepseekv3"
save_checkpoint_steps: 5000
keep_checkpoint_max: 1
integrated_save: False
async_save: False
checkpoint_format: "safetensors"
runner_wrapper:
type: MFTrainOneStepCell
scale_sense: 1.0
use_clip_grad: True
profile: False
profile_start_step: 10
profile_stop_step: 11
init_start_profile: True
profile_communication: False
profile_memory: False