first

ms-swift/examples/pytorch/llm/scripts/qwen1half_moe_a2_7b/lora/infer.sh

+# Experimental environment: A100
+# 36GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/qwen1half-moe-a2_7b/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn true \
+    --max_new_tokens 2048 \
+    --temperature 0.1 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/qwen1half_moe_a2_7b/lora/sft.sh

+# Experimental environment: A100
+# 42GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+    --model_type qwen1half-moe-a2_7b \
+    --sft_type lora \
+    --tuner_backend peft \
+    --dtype AUTO \
+    --output_dir output \
+    --dataset dureader-robust-zh \
+    --train_dataset_sample -1 \
+    --num_train_epochs 1 \
+    --max_length 1024 \
+    --check_dataset_strategy warning \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_target_modules ALL \
+    --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 true \

ms-swift/examples/pytorch/llm/scripts/qwen1half_moe_a2_7b_chat/lora/infer.sh

+# Experimental environment: A100
+# 36GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/qwen1half-moe-a2_7b-chat/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn true \
+    --max_new_tokens 2048 \
+    --temperature 0.1 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/qwen1half_moe_a2_7b_chat/lora/sft.sh

+# Experimental environment: A100
+# 42GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+    --model_type qwen1half-moe-a2_7b-chat \
+    --sft_type lora \
+    --tuner_backend peft \
+    --dtype AUTO \
+    --output_dir output \
+    --dataset blossom-math-zh \
+    --train_dataset_sample -1 \
+    --num_train_epochs 1 \
+    --max_length 1024 \
+    --check_dataset_strategy warning \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_target_modules ALL \
+    --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 true \

ms-swift/examples/pytorch/llm/scripts/qwen1half_moe_a2_7b_chat_int4/qlora/infer.sh

+# Experimental environment: A100
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --ckpt_dir "output/qwen1half-moe-a2_7b-chat-int4/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn true \
+    --max_new_tokens 2048 \
+    --temperature 0.1 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/qwen1half_moe_a2_7b_chat_int4/qlora/sft.sh

+# Experimental environment: A100
+# 17GB GPU memory
+
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model_type qwen1half-moe-a2_7b-chat-int4 \
+    --sft_type lora \
+    --output_dir output \
+    --dataset blossom-math-zh \
+    --train_dataset_sample -1 \
+    --num_train_epochs 3 \
+    --max_length 2048 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_target_modules ALL \
+    --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 true \

ms-swift/examples/pytorch/llm/scripts/qwen_14b/lora_ddp_ds/infer.sh

+# Experimental environment: A100
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/qwen-14b/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn true \
+    --max_new_tokens 2048 \
+    --temperature 0.7 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/qwen_14b/lora_ddp_ds/sft.sh

+# Experimental environment: 2 * A100
+# 2 * 32B GPU memory (use flash_attn)
+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-14B \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type default-generation \
+    --dtype AUTO \
+    --output_dir output \
+    --ddp_backend nccl \
+    --dataset dureader-robust-zh \
+    --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 ALL \
+    --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 true \
+    --deepspeed default-zero2 \

ms-swift/examples/pytorch/llm/scripts/qwen_14b/qlora/infer.sh

+# Experimental environment: A10
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/qwen-14b/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 \

ms-swift/examples/pytorch/llm/scripts/qwen_14b/qlora/sft.sh

+# Experimental environment: A10
+# 17GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+    --model_id_or_path qwen/Qwen-14B \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type default-generation \
+    --dtype AUTO \
+    --output_dir output \
+    --dataset dureader-robust-zh \
+    --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 ALL \
+    --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 \

ms-swift/examples/pytorch/llm/scripts/qwen_14b/qlora_ddp_ds/infer.sh

+# Experimental environment: A10
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/qwen-14b/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 \

ms-swift/examples/pytorch/llm/scripts/qwen_14b/qlora_ddp_ds/sft.sh

+# Experimental environment: 2 * A10
+# 2 * 14GB 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-14B \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type default-generation \
+    --dtype AUTO \
+    --output_dir output \
+    --ddp_backend nccl \
+    --dataset dureader-robust-zh \
+    --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 ALL \
+    --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 \
+    --deepspeed default-zero2 \

ms-swift/examples/pytorch/llm/scripts/qwen_14b_chat/full_ddp_zero3/infer.sh

+# Experimental environment: A100
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --ckpt_dir "output/qwen-14b-chat/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn true \
+    --max_new_tokens 2048 \
+    --temperature 0.1 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/qwen_14b_chat/full_ddp_zero3/sft.sh

+# Experimental environment: 4 * A100
+# 4 * 78GB GPU memory
+nproc_per_node=4
+
+NPROC_PER_NODE=$nproc_per_node \
+MASTER_PORT=29500 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift sft \
+    --model_id_or_path qwen/Qwen-14B-Chat \
+    --model_revision master \
+    --sft_type full \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype AUTO \
+    --output_dir output \
+    --ddp_backend nccl \
+    --dataset blossom-math-zh \
+    --train_dataset_sample -1 \
+    --num_train_epochs 5 \
+    --max_length 2048 \
+    --check_dataset_strategy warning \
+    --gradient_checkpointing true \
+    --batch_size 1 \
+    --weight_decay 0.1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps $(expr 64 / $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 true \
+    --deepspeed 'default-zero3' \
+    --save_only_model true \

ms-swift/examples/pytorch/llm/scripts/qwen_14b_chat/lora_ddp_ds/infer.sh

+# Experimental environment: A100
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/qwen-14b-chat/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn true \
+    --max_new_tokens 2048 \
+    --temperature 0.1 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/qwen_14b_chat/lora_ddp_ds/sft.sh

+# Experimental environment: 2 * A100
+# 2 * 30GB GPU memory (use flash_attn)
+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-14B-Chat \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype AUTO \
+    --output_dir output \
+    --ddp_backend nccl \
+    --dataset blossom-math-zh \
+    --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 ALL \
+    --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 true \
+    --deepspeed default-zero2 \

ms-swift/examples/pytorch/llm/scripts/qwen_14b_chat/lora_ddp_zero3/infer.sh

+# Experimental environment: 2 * 3090
+
+CUDA_VISIBLE_DEVICES=0,1 \
+swift infer \
+    --ckpt_dir "output/qwen-14b-chat/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn false \
+    --max_new_tokens 2048 \
+    --temperature 0.1 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/qwen_14b_chat/lora_ddp_zero3/sft.sh

+# Experimental environment: 4 * 3090
+# 4 * 24GB GPU memory
+nproc_per_node=4
+
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=$nproc_per_node \
+MASTER_PORT=29500 \
+swift sft \
+    --model_id_or_path qwen/Qwen-14B-Chat \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype AUTO \
+    --output_dir output \
+    --ddp_backend nccl \
+    --dataset blossom-math-zh \
+    --train_dataset_sample -1 \
+    --num_train_epochs 5 \
+    --max_length 2048 \
+    --check_dataset_strategy warning \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_target_modules ALL \
+    --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 \
+    --deepspeed default-zero3 \

ms-swift/examples/pytorch/llm/scripts/qwen_14b_chat/qlora/infer.sh

+# Experimental environment: A10, 3090
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --ckpt_dir "output/qwen-14b-chat/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --use_flash_attn false \
+    --max_new_tokens 2048 \
+    --temperature 0.1 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/qwen_14b_chat/qlora/sft.sh

+# Experimental environment: A10, 3090
+# 16GB GPU memory
+# Recommended to use `qwen_14b_chat_int4`
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model_id_or_path qwen/Qwen-14B-Chat \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype AUTO \
+    --output_dir output \
+    --dataset blossom-math-zh \
+    --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 ALL \
+    --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 \