DeepSeek/ESFT
1
0
Fork 0
mirror of https://github.com/deepseek-ai/ESFT synced 2025-06-26 18:15:50 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

streamline code; add intermediate saving support for ep

Browse Source
This commit is contained in:
ZihanWang314
2025-05-22 07:21:52 +00:00
parent 98fd21ce21
commit f38f67706c
123 changed files with 710 additions and 5601 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
scripts/train_ep.sh
View File

@@ -1,11 +1,27 @@
export TOKENIZERS_PARALLELISM=false
exp_name="test/eval_translation"
base_model_path="deepseek-ai/esft-vanilla-lite"
exp_name="test/eval_intent"
base_model_path="deepseek-ai/ESFT-vanilla-lite"
torchrun --nproc-per-node=8 train_ep.py \
--base_model_path=${base_model_path} \
--expert_config=results/expert_configs/translation.json \
--train_dataset=translation \
--expert_config=results/expert_configs/intent.json \
--train_dataset=intent \
--train_config=configs/base.yaml \
--output_dir=results/checkpoints/${exp_name}
--output_dir=results/checkpoints/${exp_name}
cp results/expert_configs/intent.json results/checkpoints/${exp_name}/checkpoint-1/expert_cfg.json
python eval_multigpu.py \
--eval_dataset=intent \
--base_model_path=deepseek-ai/ESFT-vanilla-lite \
--adapter_dir=results/checkpoints/${exp_name}/checkpoint-1 \
--output_path=results/completions/token/intent.jsonl \
--max_new_tokens=512 \
--eval_batch_size=2 \
--world_size=4 \
--openai_api_key=REPLACE_WITH_YOUR_KEY \
--gpus_per_rank=2