wassname/alignment-handbook
1
0
Fork 0
mirror of https://github.com/wassname/alignment-handbook.git synced 2026-06-27 17:14:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add doc

Browse Source
This commit is contained in:
Lewis Tunstall
2023-11-09 07:32:24 +00:00
parent e2c19a0252
commit 2de17f5ba1
3 changed files with 32 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
recipes/accelerate_configs/deepspeed_zero1.yaml
-19
View File
@@ -1,19 +0,0 @@
compute_environment: LOCAL_MACHINE
debug: false
deepspeed_config:
deepspeed_multinode_launcher: standard
zero3_init_flag: false
zero_stage: 1
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 8
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
recipes/accelerate_configs/deepspeed_zero2.yaml
-21
View File
@@ -1,21 +0,0 @@
compute_environment: LOCAL_MACHINE
debug: false
deepspeed_config:
deepspeed_multinode_launcher: standard
offload_optimizer_device: none
offload_param_device: none
zero3_init_flag: false
zero_stage: 2
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 8
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
scripts/README.md
+32
View File
@@ -1,6 +1,38 @@
## Supervised Fine-Tuning (SFT)
We provide 3 main ways to train SFT models:
* Distributed fine-tuning of all model weights with ZeRO-3
* Fine-tuning with LoRA adapters and ZeRO-3
* Fine-tuning with QLoRA adapters and DDP
```shell
# Full training with ZeRO-3
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/deepspeed_zero3.yaml scripts/run_sft.py recipes/{model_name}/sft/config_full.yaml
# LoRA training with ZeRO-3
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/deepspeed_zero3.yaml scripts/run_sft.py recipes/{model_name}/sft/config_16bit.yaml
# QLoRA training with DDP
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/multi_gpu.yaml scripts/run_sft.py recipes/{model_name}/sft/config_8bit.yaml
```
You can override the parameters in each YAML config by appending them to the command as follows:
```shell
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/deepspeed_zero3.yaml scripts/run_sft.py recipes/{model_name}/sft/config_full.yaml --per_device_train_batch_size=2 --num_train_epochs=3
```
## Direct Preference Optimisation (DPO)
```shell
# Full training with ZeRO-3
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/deepspeed_zero3.yaml scripts/run_dpo.py recipes/{model_name}/dpo/config_full.yaml
# LoRA training with ZeRO-3
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/deepspeed_zero3.yaml scripts/run_dpo.py recipes/{model_name}/dpo/config_16bit.yaml
# QLoRA training with DDP
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/multi_gpu.yaml scripts/run_dpo.py recipes/{model_name}/dpo/config_8bit.yaml
```