Skip to content

llama_factory_hyperparallel_sft_guide.md

Latest commit

 

History

History
194 lines (141 loc) · 4.74 KB

llama_factory_hyperparallel_sft_guide.md

File metadata and controls

194 lines (141 loc) · 4.74 KB
Raw

Llama Factory训练说明文档

版本及示例

请严格按照给出的版本进行实验，已发现torch/torch-npu/cann/机器有较严格的对应关系。

基础环境

项目 配置
计算设备 昇腾 910c
CANN 8.5.2
Python 3.11.10

框架与依赖

框架 版本 代码
LlamaFactory master https://github.com/hiyouga/LlamaFactory
HyperParallel master https://gitcode.com/mindspore/hyper-parallel.git
Transformers 4.57.1
Accelerate 1.11.0
torch 2.7.1
torch_npu 2.7.1.post4 torch_npu按对应版本安装，此处仅为参考
torchvision 0.22.1
torchaudio 2.7.1
trl 0.24.0
datasets 4.0.0
peft 0.18.0

yaml配置

qwen3vlmoe_full_sft_fsdp2.yaml

# Start FSDP2 fine-tuning
# accelerate launch \
#     --config_file examples/accelerate/fsdp2_config.yaml \
#     src/train.py examples/ascend/qwen3vlmoe_full_sft_fsdp2.yaml
# Change `num_processes` in fsdp2_config.yaml to 16 in A3

### model
model_name_or_path: /path/to/Qwen3-VL-30B-A3B-Instruct
image_max_pixels: 262144
video_max_pixels: 16384
trust_remote_code: true
use_v1_kernels: true
flash_attn: fa2

### enable hyper parallel FSDP
use_hyper_parallel: true
### enable hyper parallel CP
hyper_parallel_cp_size: 2

### method
stage: sft
do_train: true
finetuning_type: full
disable_gradient_checkpointing: false

### dataset
dataset: llava_1k_en, llava_1k_zh
template: qwen3_vl
cutoff_len: 16384
overwrite_cache: true
preprocessing_num_workers: 16
dataloader_num_workers: 4

### output
output_dir: saves/Qwen3-VL-30B-A3B-Instruct/full/sft
logging_steps: 1
save_steps: 1000
max_steps: 100
plot_loss: true
overwrite_output_dir: true
save_only_model: true
report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 1
learning_rate: 1.0e-4
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
resume_from_checkpoint: null
seed: 1234

启动命令

先下载模型权重，将yaml文件中model_path替换为上述文件夹路径。

accelerate config:

compute_environment: LOCAL_MACHINE
debug: false
distributed_type: FSDP
downcast_bf16: 'no'
fsdp_config:
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_cpu_ram_efficient_loading: true
  fsdp_offload_params: false
  fsdp_reshard_after_forward: true
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_version: 2
machine_rank: 0
main_training_function: main
mixed_precision: bf16  # or fp16
num_machines: 1  # the number of nodes
num_processes: 16  # the number of GPUs in all nodes
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

Hyper Parallel FSDP启动命令：

accelerate launch --config_file examples/accelerate/fsdp2_config.yaml src/train.py examples/ascend/qwen3vlmoe_full_sft_fsdp2.yaml

即可跑通Hyper Parallel FSDP的训练流程。

精度验证

Torch FSDP/ Hyper FSDP精度对比

Model: Qwen3-VL-30B-A3B-Instruct

Configuration:

  • Basline torch FSDP, ours hyper FSDP

  • 16 cards

  • BF16

  • per_device_batch_size=1

  • cutoff_len=16384

Torch FSDP 与 Hyper FSDP 精度对比

Hyper FSDP + CP精度

Model: Qwen3-VL-30B-A3B-Instruct

Configuration:

  • Hyper FSDP + CP

  • 16 cards

  • CP size 2

  • BF16

  • per_device_batch_size=1

  • cutoff_len=16384

Hyper FSDP + CP 精度

Hyper FSDP/Hyper FSDP + CP精度对比

Model: Qwen3-VL-30B-A3B-Instruct(num_hidden_layers=24)

Configuration:

  • Baseline 8 NPU FSDP, ours 8NPU FSDP + CP_SIZE 2

  • 16 cards

  • BF16

  • per_device_batch_size=1

  • cutoff_len=16384

Hyper FSDP 与 Hyper FSDP + CP 精度对比

性能

方案 性能
torch FSDP 4.4 s/iter
hyper FSDP 3.54 s/iter
hyper FSDP + CP 2.96s/iter