first

ms-swift/examples/pytorch/llm/scripts/dbrx-instruct/lora_mp/sft.sh

+# Experimental environment: 4 * A100
+# 4 * 74GB GPU memory
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift sft \
+    --model_type dbrx-instruct \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype bf16 \
+    --output_dir output \
+    --dataset blossom-math-zh \
+    --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 \
+    --lora_dtype AUTO \
+    --gradient_checkpointing false \
+    --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/deepseek-v2-chat/lora_mp/deepseek2_device_map.json

+{
+    "model.embed_tokens": "cuda:0",
+    "model.layers.0": "cuda:0",
+    "model.layers.1": "cuda:0",
+    "model.layers.2": "cuda:0",
+    "model.layers.3": "cuda:0",
+    "model.layers.4": "cuda:0",
+    "model.layers.5": "cuda:0",
+    "model.layers.6": "cuda:0",
+    "model.layers.7": "cuda:1",
+    "model.layers.8": "cuda:1",
+    "model.layers.9": "cuda:1",
+    "model.layers.10": "cuda:1",
+    "model.layers.11": "cuda:1",
+    "model.layers.12": "cuda:1",
+    "model.layers.13": "cuda:1",
+    "model.layers.14": "cuda:2",
+    "model.layers.15": "cuda:2",
+    "model.layers.16": "cuda:2",
+    "model.layers.17": "cuda:2",
+    "model.layers.18": "cuda:2",
+    "model.layers.19": "cuda:2",
+    "model.layers.20": "cuda:2",
+    "model.layers.21": "cuda:3",
+    "model.layers.22": "cuda:3",
+    "model.layers.23": "cuda:3",
+    "model.layers.24": "cuda:3",
+    "model.layers.25": "cuda:3",
+    "model.layers.26": "cuda:3",
+    "model.layers.27": "cuda:3",
+    "model.layers.28": "cuda:4",
+    "model.layers.29": "cuda:4",
+    "model.layers.30": "cuda:4",
+    "model.layers.31": "cuda:4",
+    "model.layers.32": "cuda:4",
+    "model.layers.33": "cuda:4",
+    "model.layers.34": "cuda:4",
+    "model.layers.35": "cuda:4",
+    "model.layers.36": "cuda:5",
+    "model.layers.37": "cuda:5",
+    "model.layers.38": "cuda:5",
+    "model.layers.39": "cuda:5",
+    "model.layers.40": "cuda:5",
+    "model.layers.41": "cuda:5",
+    "model.layers.42": "cuda:5",
+    "model.layers.43": "cuda:5",
+    "model.layers.44": "cuda:6",
+    "model.layers.45": "cuda:6",
+    "model.layers.46": "cuda:6",
+    "model.layers.47": "cuda:6",
+    "model.layers.48": "cuda:6",
+    "model.layers.49": "cuda:6",
+    "model.layers.50": "cuda:6",
+    "model.layers.51": "cuda:6",
+    "model.layers.52": "cuda:7",
+    "model.layers.53": "cuda:7",
+    "model.layers.54": "cuda:7",
+    "model.layers.55": "cuda:7",
+    "model.layers.56": "cuda:7",
+    "model.layers.57": "cuda:7",
+    "model.layers.58": "cuda:7",
+    "model.layers.59": "cuda:7",
+    "model.norm": "cuda:7",
+    "lm_head": "cuda:7"
+}

ms-swift/examples/pytorch/llm/scripts/deepseek-v2-chat/lora_mp/infer.sh

+# Experimental environment: 8*A100
+# cd /path/to/swift/example/pytorch/llm
+
+PYTHONPATH=../../.. \
+
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+python llm_infer.py \
+    --ckpt_dir output/deepseek-v2-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 \
+    --device_map_config_path scripts/deepseek-v2-chat/lora_ddp_ds3/deepseek2_device_map.json

ms-swift/examples/pytorch/llm/scripts/deepseek-v2-chat/lora_mp/sft.sh

+# Experimental environment: 8*A100
+# 8*80GB GPU memory
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+swift sft \
+    --model_type deepseek-v2-chat \
+    --sft_type lora \
+    --tuner_backend peft \
+    --dtype bf16 \
+    --output_dir output \
+    --dataset alpaca-zh#5000 \
+    --num_train_epochs 1 \
+    --max_length 1024 \
+    --check_dataset_strategy warning \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_dtype AUTO \
+    --lora_target_modules DEFAULT \
+    --gradient_checkpointing false \
+    --use_flash_attn 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 10 \
+    --logging_steps 10 \
+    --device_map_config_path scripts/deepseek-v2-chat/lora_mp/deepseek2_device_map.json

ms-swift/examples/pytorch/llm/scripts/deepseek_moe_16b_chat/lora/infer.sh

+# Experimental environment: A100
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --ckpt_dir "output/deepseek-moe-16b-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 \

ms-swift/examples/pytorch/llm/scripts/deepseek_moe_16b_chat/lora/sft.sh

+# Experimental environment: A100
+# 52GB GPU memory
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model_type deepseek-moe-16b-chat \
+    --dataset damo-agent-mini-zh \
+    --train_dataset_sample 20000 \
+    --max_length 4096 \
+    --gradient_checkpointing true \
+    --eval_steps 100 \
+    --use_flash_attn true \
+    --output_dir output \

ms-swift/examples/pytorch/llm/scripts/dpo/lora/dpo.sh

+# Experimental environment: A100
+# Memory usage: 20G
+
+CUDA_VISIBLE_DEVICES=0 \
+swift dpo \
+    --model_type  yi-6b-chat \
+    --ref_model_type  yi-6b-chat \
+    --model_revision  master \
+    --sft_type  lora \
+    --tuner_backend  swift \
+    --dtype  AUTO  \
+    --output_dir  output  \
+    --dataset  hh-rlhf-cn:harmless_base_cn  \
+    --num_train_epochs  3  \
+    --max_length  1024  \
+    --max_prompt_length  512  \
+    --check_dataset_strategy  none  \
+    --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  5e-5  \
+    --gradient_accumulation_steps  16  \
+    --max_grad_norm  1.0  \
+    --warmup_ratio  0.03  \
+    --eval_steps  2000  \
+    --save_steps  2000  \
+    --save_total_limit  2  \
+    --logging_steps  10 \

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

+# Experimental environment: A10, 3090
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir output/mistral-7b/vx-xxx-xxx/checkpoint-xxx \
+    --load_dataset_config true \
+    --eval_human true \
+    --use_flash_attn false \
+    --max_new_tokens 1024 \
+    --temperature 0.3 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/dpo/lora_ddp_mp/dpo.sh

+# Experimental environment: 4*A100
+# Memory usage: 4 * 20G
+nproc_per_node=2
+
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=$nproc_per_node \
+MASTER_PORT=29500 \
+swift dpo \
+    --model_type  yi-6b-chat \
+    --ref_model_type  yi-6b-chat \
+    --model_revision  master \
+    --sft_type  lora \
+    --tuner_backend  swift \
+    --dtype  AUTO  \
+    --output_dir  output  \
+    --dataset  hh-rlhf-cn:harmless_base_cn  \
+    --num_train_epochs  3  \
+    --max_length  1024  \
+    --max_prompt_length  512  \
+    --check_dataset_strategy  none  \
+    --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  5e-5  \
+    --gradient_accumulation_steps  $(expr 16 / $nproc_per_node)  \
+    --max_grad_norm  1.0  \
+    --warmup_ratio  0.03  \
+    --eval_steps  2000  \
+    --save_steps  2000  \
+    --save_total_limit  2  \
+    --logging_steps  10 \

ms-swift/examples/pytorch/llm/scripts/dpo/lora_ddp_mp/infer.sh

+# Experimental environment: A10, 3090
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir output/mistral-7b/vx-xxx-xxx/checkpoint-xxx \
+    --load_dataset_config true \
+    --eval_human true \
+    --use_flash_attn false \
+    --max_new_tokens 1024 \
+    --temperature 0.3 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/gemma_2b_instruct/lora/infer.sh

+# Experimental environment: V100, A10, 3090
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --ckpt_dir "output/gemma-2b-instruct/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --max_new_tokens 2048 \
+    --temperature 0.1 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --do_sample true \

ms-swift/examples/pytorch/llm/scripts/gemma_2b_instruct/lora/sft.sh

+# Experimental environment: V100, A10, 3090
+# 12GB GPU memory
+
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model_id_or_path AI-ModelScope/gemma-2b-it \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype AUTO \
+    --output_dir output \
+    --dataset hc3-zh \
+    --train_dataset_sample 5000 \
+    --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 16 \
+    --max_grad_norm 0.5 \
+    --warmup_ratio 0.1 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 10 \

ms-swift/examples/pytorch/llm/scripts/grok-1/lora_ddp_ds/infer.sh

+# Experimental environment: 8 * A100
+# Memory cost: 8 * 80G
+PYTHONPATH=../../.. \
+python llm_infer.py \
+    --ckpt_dir output/grok-1/vxx-xxxx-xxxx/checkpoint-xxx \
+    --dtype bf16 \
+    --load_dataset_config true \
+    --max_new_tokens 64 \
+    --do_sample true \
+    --dtype bf16 \
+    --eval_human false \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/grok-1/lora_ddp_ds/sft.sh

+# Experimental environment: 8 * A100
+# Memory cost: 8 * 21G
+nproc_per_node=8
+
+PYTHONPATH=../../.. \
+torchrun \
+    --nproc_per_node=$nproc_per_node \
+    --master_port 29500 \
+    llm_sft.py \
+    --model_type grok-1 \
+    --sft_type lora \
+    --tuner_backend peft \
+    --dtype bf16 \
+    --output_dir output \
+    --ddp_backend nccl \
+    --dataset dureader-robust-zh \
+    --train_dataset_sample -1 \
+    --num_train_epochs 1 \
+    --max_length 512 \
+    --check_dataset_strategy warning \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --lora_dropout_p 0.05 \
+    --lora_dtype AUTO \
+    --lora_target_modules DEFAULT \
+    --gradient_checkpointing true \
+    --batch_size 2 \
+    --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 \
+    --deepspeed zero3-offload \

ms-swift/examples/pytorch/llm/scripts/internlm2_7b_sft_chat/lora/infer.sh

+# Experimental environment: A10
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --ckpt_dir "output/internlm2-7b-sft-chat/vx-xxx/checkpoint-xxx" \
+    --load_dataset_config true \
+    --max_new_tokens 2048 \
+    --temperature 0.5 \
+    --top_p 0.7 \
+    --repetition_penalty 1. \
+    --stream false \
+    --do_sample true \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/internlm2_7b_sft_chat/lora/sft.sh

+# Experimental environment: A10
+# 22GB GPU memory
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model_type internlm2-7b-sft-chat \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type AUTO \
+    --dtype AUTO \
+    --output_dir output \
+    --dataset dureader-robust-zh \
+    --train_dataset_sample 20000 \
+    --num_train_epochs 1 \
+    --max_length 2048 \
+    --system 'You are a helpful assistant.' \
+    --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 \
+    --neftune_noise_alpha 5 \
+    --use_flash_attn false \

ms-swift/examples/pytorch/llm/scripts/internlm_20b/lora_ddp/infer.sh

+# Experimental environment: A100
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/internlm-20b/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 \
+    --merge_lora false \

ms-swift/examples/pytorch/llm/scripts/internlm_20b/lora_ddp/sft.sh

+# Experimental environment: 2 * A100
+# 2 * 56GB 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 Shanghai_AI_Laboratory/internlm-20b \
+    --model_revision master \
+    --sft_type lora \
+    --tuner_backend peft \
+    --template_type default-generation \
+    --dtype AUTO \
+    --output_dir output \
+    --ddp_backend nccl \
+    --dataset jd-sentiment-zh \
+    --train_dataset_sample -1 \
+    --val_dataset_sample 1000 \
+    --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 DEFAULT \
+    --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 \

ms-swift/examples/pytorch/llm/scripts/internlm_20b/qlora/infer.sh

+# Experimental environment: A10
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_infer.py \
+    --ckpt_dir "output/internlm-20b/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 \

ms-swift/examples/pytorch/llm/scripts/internlm_20b/qlora/sft.sh

+# Experimental environment: A10
+# 14GB GPU memory
+PYTHONPATH=../../.. \
+CUDA_VISIBLE_DEVICES=0 \
+python llm_sft.py \
+    --model_id_or_path Shanghai_AI_Laboratory/internlm-20b \
+    --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 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 \