Initial commit

swift-main/examples/pytorch/llm/scripts/qwen_vl_chat/qlora/sft.sh

+# Experimental environment: A10
+# 10GB GPU memory
+# Recommended to use `qwen_vl_chat_int4`
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+    --model_id_or_path qwen/Qwen-VL-Chat \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype AUTO \
+    --output_dir output \
+    --dataset coco-en-mini \
+    --train_dataset_sample -1 \
+    --num_train_epochs 1 \
+    --max_length 2048 \
+    --check_dataset_strategy warning \
+    --quantization_bit 4 \
+    --bnb_4bit_comp_dtype AUTO \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_target_modules c_attn attn.c_proj \
+    --gradient_checkpointing true \
+    --batch_size 1 \
+    --weight_decay 0.1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --max_grad_norm 0.5 \
+    --warmup_ratio 0.03 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 10 \
+    --use_flash_attn false \

swift-main/examples/pytorch/llm/scripts/qwen_vl_chat_int4/qlora/infer.sh

+# Experimental environment: A10
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/qwen-vl-chat-int4/vx_xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn false \
+    --max_new_tokens 2048 \
+    --temperature 0.7 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

swift-main/examples/pytorch/llm/scripts/qwen_vl_chat_int4/qlora/sft.sh

+# Experimental environment: A10
+# 11GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+    --model_id_or_path qwen/Qwen-VL-Chat-Int4 \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype fp16 \
+    --output_dir output \
+    --dataset coco-en-mini \
+    --train_dataset_sample -1 \
+    --num_train_epochs 1 \
+    --max_length 2048 \
+    --check_dataset_strategy warning \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_target_modules c_attn attn.c_proj \
+    --gradient_checkpointing true \
+    --batch_size 1 \
+    --weight_decay 0.1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --max_grad_norm 0.5 \
+    --warmup_ratio 0.03 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 10 \
+    --use_flash_attn false \
+    --push_to_hub false \
+    --hub_model_id qwen-vl-chat-int4-qlora \
+    --hub_private_repo true \
+    --hub_token 'your-sdk-token' \

swift-main/examples/pytorch/llm/scripts/qwen_vl_chat_int4/qlora_ddp_ds/infer.sh

+# Experimental environment: A10
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/qwen-vl-chat-int4/vx_xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn false \
+    --max_new_tokens 2048 \
+    --temperature 0.7 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

swift-main/examples/pytorch/llm/scripts/qwen_vl_chat_int4/qlora_ddp_ds/sft.sh

+# Experimental environment: 2 * A10
+# 2 * 13GB GPU memory
+nproc_per_node=2
+
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0,1 \
+torchrun \
+    --nproc_per_node=$nproc_per_node \
+    --master_port 29500 \
+    llm_sft.py \
+    --model_id_or_path qwen/Qwen-VL-Chat-Int4 \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype fp16 \
+    --output_dir output \
+    --ddp_backend nccl \
+    --dataset coco-en-mini \
+    --train_dataset_sample -1 \
+    --num_train_epochs 1 \
+    --max_length 2048 \
+    --check_dataset_strategy warning \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_target_modules c_attn attn.c_proj \
+    --gradient_checkpointing true \
+    --batch_size 1 \
+    --weight_decay 0.1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --max_grad_norm 0.5 \
+    --warmup_ratio 0.03 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 10 \
+    --use_flash_attn false \
+    --push_to_hub false \
+    --hub_model_id qwen-vl-chat-qlora \
+    --hub_private_repo true \
+    --hub_token 'your-sdk-token' \
+    --deepspeed default-zero2 \

swift-main/examples/pytorch/llm/scripts/seqgpt_560m/full/infer.sh

+# Experimental environment: A10
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/seqgpt-560m/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --max_new_tokens 2048 \
+    --temperature 0.3 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \

swift-main/examples/pytorch/llm/scripts/seqgpt_560m/full/sft.sh

+# Experimental environment: A10
+# 12GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+    --model_id_or_path damo/nlp_seqgpt-560m \
+    --model_revision master \
+    --sft_type full \
+    --template_type default-generation \
+    --dtype AUTO \
+    --output_dir output \
+    --dataset ner-jave-zh \
+    --train_dataset_sample -1 \
+    --num_train_epochs 3 \
+    --max_length 1024 \
+    --check_dataset_strategy warning \
+    --gradient_checkpointing true \
+    --batch_size 4 \
+    --weight_decay 0.1 \
+    --learning_rate 1e-5 \
+    --gradient_accumulation_steps 8 \
+    --max_grad_norm 0.5 \
+    --warmup_ratio 0.03 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_only_model false \
+    --save_total_limit 2 \
+    --logging_steps 10 \

swift-main/examples/pytorch/llm/scripts/seqgpt_560m/full_ddp/infer.sh

+# Experimental environment: A10
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/seqgpt-560m/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --max_new_tokens 2048 \
+    --temperature 0.3 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \

swift-main/examples/pytorch/llm/scripts/seqgpt_560m/full_ddp/sft.sh

+# Experimental environment: 2 * A10
+# 2 * 13GB GPU memory
+nproc_per_node=2
+
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0,1 \
+torchrun \
+    --nproc_per_node=$nproc_per_node \
+    --master_port 29500 \
+    llm_sft.py \
+    --model_id_or_path damo/nlp_seqgpt-560m \
+    --model_revision master \
+    --sft_type full \
+    --template_type default-generation \
+    --dtype AUTO \
+    --output_dir output \
+    --ddp_backend nccl \
+    --dataset ner-jave-zh \
+    --train_dataset_sample -1 \
+    --num_train_epochs 3 \
+    --max_length 1024 \
+    --check_dataset_strategy warning \
+    --gradient_checkpointing true \
+    --batch_size 4 \
+    --weight_decay 0.1 \
+    --learning_rate 1e-5 \
+    --gradient_accumulation_steps $(expr 32 / $nproc_per_node / 4) \
+    --max_grad_norm 0.5 \
+    --warmup_ratio 0.03 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_only_model false \
+    --save_total_limit 2 \
+    --logging_steps 10 \

swift-main/examples/pytorch/llm/scripts/skywork_13b/qlora/infer.sh

+# Experimental environment: A10, 3090
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/skywork-13b/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --max_new_tokens 2048 \
+    --temperature 0.7 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

swift-main/examples/pytorch/llm/scripts/skywork_13b/qlora/sft.sh

+# Experimental environment: A10, 3090
+# 16GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+    --model_id_or_path skywork/Skywork-13B-base \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type default-generation \
+    --dtype AUTO \
+    --output_dir output \
+    --dataset advertise-gen-zh \
+    --train_dataset_sample 20000 \
+    --num_train_epochs 1 \
+    --max_length 2048 \
+    --check_dataset_strategy warning \
+    --quantization_bit 4 \
+    --bnb_4bit_comp_dtype AUTO \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_target_modules DEFAULT \
+    --gradient_checkpointing true \
+    --batch_size 1 \
+    --weight_decay 0.1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --max_grad_norm 0.5 \
+    --warmup_ratio 0.03 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 10 \

swift-main/examples/pytorch/llm/scripts/sus_34b_chat/lora/infer.sh

+# Experimental environment: A100
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --ckpt_dir "output/sus-34b-chat/vx-xxx/checkpoint-xxx" \

swift-main/examples/pytorch/llm/scripts/sus_34b_chat/lora/sft.sh

+# ref: https://github.com/modelscope/swift/blob/main/docs/source/LLM/%E8%87%AA%E6%88%91%E8%AE%A4%E7%9F%A5%E5%BE%AE%E8%B0%83%E6%9C%80%E4%BD%B3%E5%AE%9E%E8%B7%B5.md
+# Experimental environment: A100
+# 70GB GPU memory
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model_id_or_path SUSTC/SUS-Chat-34B \
+    --dataset alpaca-zh alpaca-en \
+    --train_dataset_sample 500 \
+    --eval_steps 20 \
+    --logging_steps 5 \
+    --output_dir output \
+    --lora_target_modules ALL \
+    --self_cognition_sample 500 \
+    --model_name 小黄 'Xiao Huang' \
+    --model_author 魔搭 ModelScope \

swift-main/examples/pytorch/llm/scripts/telechat_12b/lora/infer.sh

+# Experiment env: A100
+# 1 * 26GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/telechat-12b/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --max_length 2048 \
+    --use_flash_attn true \
+    --max_new_tokens 2048 \
+    --temperature 0.5 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \
+    --dtype fp16 \
+    --stream false

swift-main/examples/pytorch/llm/scripts/telechat_12b/lora/sft.sh

+# Experiment env: A100
+# 1 * 30GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+  --model_type telechat-12b \
+  --dataset dureader-robust-zh \
+  --batch_size 1 \
+  --max_length 1024 \
+  --gradient_accumulation_steps 16 \
+  --learning_rate 5e-5 \
+  --use_flash_attn true \
+  --eval_steps 1000 \
+  --save_steps 1000 \
+  --train_dataset_sample -1 \
+  --num_train_epochs 2 \
+  --check_dataset_strategy none \
+  --gradient_checkpointing true \
+  --weight_decay 0.1 \
+  --max_grad_norm 1.0 \
+  --warmup_ratio 0.03 \
+  --save_total_limit 2 \
+  --logging_steps 10 \
+  --sft_type lora \
+  --lora_target_modules DEFAULT \
+  --lora_rank 8 \
+  --lora_alpha 32 \
+  --dtype fp16

swift-main/examples/pytorch/llm/scripts/telechat_7b/lora/infer.sh

+# Experiment env: A100
+# 1 * 16GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/telechat-7b/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --max_length 2048 \
+    --use_flash_attn true \
+    --max_new_tokens 2048 \
+    --temperature 0.5 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \
+    --dtype fp16 \
+    --stream false

swift-main/examples/pytorch/llm/scripts/telechat_7b/lora/sft.sh

+# Experiment env: A100
+# 1 * 18GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+  --model_type telechat-7b \
+  --dataset dureader-robust-zh \
+  --batch_size 1 \
+  --max_length 1024 \
+  --gradient_accumulation_steps 16 \
+  --learning_rate 5e-5 \
+  --use_flash_attn true \
+  --eval_steps 1000 \
+  --save_steps 1000 \
+  --train_dataset_sample -1 \
+  --num_train_epochs 2 \
+  --check_dataset_strategy none \
+  --gradient_checkpointing true \
+  --weight_decay 0.1 \
+  --max_grad_norm 1.0 \
+  --warmup_ratio 0.03 \
+  --save_total_limit 2 \
+  --logging_steps 10 \
+  --sft_type lora \
+  --lora_target_modules DEFAULT \
+  --lora_rank 8 \
+  --lora_alpha 32 \
+  --dtype fp16

swift-main/examples/pytorch/llm/scripts/tongyi_finance_14b_chat_int4/qlora/infer.sh

+# Experimental environment: V100, A10, 3090
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/tongyi-finance-14b-chat-int4/vx_xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn false \
+    --max_new_tokens 2048 \
+    --temperature 0.3 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

swift-main/examples/pytorch/llm/scripts/tongyi_finance_14b_chat_int4/qlora/sft.sh

+# Experimental environment: V100, A10, 3090
+# 18GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+    --model_type tongyi-finance-14b-chat-int4 \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype fp16 \
+    --output_dir output \
+    --dataset xxx.jsonl \
+    --val_dataset yyy.jsonl \
+    --train_dataset_sample -1 \
+    --num_train_epochs 1 \
+    --max_length 4096 \
+    --check_dataset_strategy warning \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_target_modules DEFAULT \
+    --gradient_checkpointing true \
+    --batch_size 1 \
+    --weight_decay 0.1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --max_grad_norm 0.5 \
+    --warmup_ratio 0.03 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 10 \
+    --use_flash_attn false \
+    --push_to_hub false \
+    --hub_model_id tongyi-finance-14b-chat-int4-qlora \
+    --hub_private_repo true \
+    --hub_token 'your-sdk-token' \

swift-main/examples/pytorch/llm/scripts/torchacc/baichuan2_13b_chat/acc_lora_dp_sft.sh

+# Experimental environment: 2 * A100
+# 80GB GPU memory
+# Note: TorchAcc is currently only available internally.
+# torchacc dp
+export USE_TORCHACC=1
+export XLA_FLAGS='--xla_gpu_force_compilation_parallelism=32 --xla_multiheap_size_constraint_per_heap=4831838208 --xla_disable_hlo_passes=all-gather-combiner,all-reduce-combiner,reduce-scatter-combiner,gpu-convert-async-collectives-to-sync,rematerialization'
+export XLA_IR_SHAPE_CACHE_SIZE=100000000
+export XLA_ALLOCATOR_FRACTION=0.95
+export XLA_EXPERIMENTAL=nonzero:masked_select
+
+NPROC_PER_NODE=2 \
+CUDA_VISIBLE_DEVICES=0,1 \
+MASTER_PORT=27829 \
+swift sft \
+  --model_id_or_path baichuan-inc/Baichuan2-13B-Chat \
+  --model_layer_cls_name BaichuanLayer \
+  --dataset codefuse-python-en \
+  --sft_type lora \
+  --output_dir output \
+  --num_train_epochs 1 \
+  --max_length 2048 \
+  --batch_size 12 \
+  --use_flash_attn true \
+  --gradient_accumulation_steps 1 \
+  --gradient_checkpointing no \
+  --tuner_backend 'peft' \
+  --dataset_test_ratio 0 \
+  --save_strategy no \
+  --eval_steps 2000000 \
+  --save_steps 2000000 \
+  --logging_steps 100 \
+  --preprocess_num_proc 1 \
+  --metric_warmup_step 0.1 \
+  --report_to 'none'