关于swift框架微调

[jianmanlincjx]

感谢作者的工作！！
我想了解一下Intern-S1-mini的模型架构相较于Qwen3有什么区别吗，是否能通过swift框架直接sft微调呢

[Xuanaxx]

msswift的文档说已经支持了，可以尝试一下？
https://swift.readthedocs.io/zh-cn/latest/Instruction/%E6%94%AF%E6%8C%81%E7%9A%84%E6%A8%A1%E5%9E%8B%E5%92%8C%E6%95%B0%E6%8D%AE%E9%9B%86.html

[Xuanaxx]

msswift成功微调启动脚本

export NPROC_PER_NODE=8
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export MASTER_PORT=39312
export PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True'
export MAX_NUM=1

GPUS=8
BATCH_SIZE=32
PER_DEVICE_BATCH_SIZE=1
GRADIENT_ACC=$((BATCH_SIZE / PER_DEVICE_BATCH_SIZE / GPUS))
LR=0.00001
EPOCHS=3.0
SAVE_STEPS=250
MAX_LEN=8192
OUTPUR_DIR=''
DATASET_DIR=''
model_type='interns1'
dataset=''

swift sft \
    --model_type $model_type \
    --model models/Intern-S1-mini \
    --train_type full \
    --dataset $DATASET_DIR \
    --freeze_vit true \
    --freeze_aligner false \
    --freeze_llm false \
    --torch_dtype bfloat16 \
    --num_train_epochs $EPOCHS \
    --gradient_checkpointing true \
    --per_device_train_batch_size $PER_DEVICE_BATCH_SIZE \
    --per_device_eval_batch_size 1 \
    --learning_rate $LR \
    --gradient_accumulation_steps $GRADIENT_ACC \
    --save_strategy steps \
    --save_steps $SAVE_STEPS \
    --split_dataset_ratio 0.01 \
    --eval_strategy steps \
    --eval_steps $SAVE_STEPS \
    --save_total_limit 100 \
    --logging_steps 1 \
    --max_length $MAX_LEN \
    --output_dir $OUTPUR_DIR \
    --warmup_ratio 0.1 \
    --dataloader_num_workers 32 \
    --dataset_num_proc 16 \
    --deepspeed zero3 \
    --report_to tensorboard \
    --use_liger_kernel true \
    --attn_impl flash_attn \
    --truncation_strategy delete \
    2>&1 | tee ${OUTPUR_DIR}/training_log.txt

需要注意的是，MAX_NUM这个全局参数无法传参，因此max_patches默认为12，可能会导致图片token数量过长，目前通过修改Intern-S1-mini/preprocessor_config.json，硬编码max_patches解决，后续需要等swift更新适配。

[jianmanlincjx]

非常感谢您！！！

…

On Wed, Aug 27, 2025 at 7:02 PM Chen Zixuan ***@***.***> wrote: *Xuanaxx* left a comment (InternLM/Intern-S1#23) <#23 (comment)> msswift成功微调启动脚本 export NPROC_PER_NODE=8export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7export MASTER_PORT=39312export PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True'export MAX_NUM=1 GPUS=8 BATCH_SIZE=32 PER_DEVICE_BATCH_SIZE=1 GRADIENT_ACC=$((BATCH_SIZE / PER_DEVICE_BATCH_SIZE / GPUS)) LR=0.00001 EPOCHS=3.0 SAVE_STEPS=250 MAX_LEN=8192 OUTPUR_DIR='' DATASET_DIR='' model_type='interns1' dataset='' swift sft \ --model_type $model_type \ --model models/Intern-S1-mini \ --train_type full \ --dataset $DATASET_DIR \ --freeze_vit true \ --freeze_aligner false \ --freeze_llm false \ --torch_dtype bfloat16 \ --num_train_epochs $EPOCHS \ --gradient_checkpointing true \ --per_device_train_batch_size $PER_DEVICE_BATCH_SIZE \ --per_device_eval_batch_size 1 \ --learning_rate $LR \ --gradient_accumulation_steps $GRADIENT_ACC \ --save_strategy steps \ --save_steps $SAVE_STEPS \ --split_dataset_ratio 0.01 \ --eval_strategy steps \ --eval_steps $SAVE_STEPS \ --save_total_limit 100 \ --logging_steps 1 \ --max_length $MAX_LEN \ --output_dir $OUTPUR_DIR \ --warmup_ratio 0.1 \ --dataloader_num_workers 32 \ --dataset_num_proc 16 \ --deepspeed zero3 \ --report_to tensorboard \ --use_liger_kernel true \ --attn_impl flash_attn \ --truncation_strategy delete \ 2>&1 | tee ${OUTPUR_DIR}/training_log.txt 需要注意的是，MAX_NUM这个全局参数无法传参，因此max_patches默认为12，可能会导致图片token数量过长，目前通过修改Intern-S1-mini/preprocessor_config.json，硬编码max_patches解决，后续需要等swift更新适配。 — Reply to this email directly, view it on GitHub <#23 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BG3HDLKSY2RGERFDRCOCYK33PWF45AVCNFSM6AAAAACE3HG7T2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTEMRXG42TGMZXGE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[fak111]

msswift成功微调启动脚本

export NPROC_PER_NODE=8
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export MASTER_PORT=39312
export PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True'
export MAX_NUM=1

GPUS=8
BATCH_SIZE=32
PER_DEVICE_BATCH_SIZE=1
GRADIENT_ACC=$((BATCH_SIZE / PER_DEVICE_BATCH_SIZE / GPUS))
LR=0.00001
EPOCHS=3.0
SAVE_STEPS=250
MAX_LEN=8192
OUTPUR_DIR=''
DATASET_DIR=''
model_type='interns1'
dataset=''

swift sft
--model_type $model_type
--model models/Intern-S1-mini
--train_type full
--dataset $DATASET_DIR
--freeze_vit true
--freeze_aligner false
--freeze_llm false
--torch_dtype bfloat16
--num_train_epochs $EPOCHS
--gradient_checkpointing true
--per_device_train_batch_size $PER_DEVICE_BATCH_SIZE
--per_device_eval_batch_size 1
--learning_rate $LR
--gradient_accumulation_steps $GRADIENT_ACC
--save_strategy steps
--save_steps $SAVE_STEPS
--split_dataset_ratio 0.01
--eval_strategy steps
--eval_steps $SAVE_STEPS
--save_total_limit 100
--logging_steps 1
--max_length $MAX_LEN
--output_dir $OUTPUR_DIR
--warmup_ratio 0.1
--dataloader_num_workers 32
--dataset_num_proc 16
--deepspeed zero3
--report_to tensorboard
--use_liger_kernel true
--attn_impl flash_attn
--truncation_strategy delete
2>&1 | tee ${OUTPUR_DIR}/training_log.txt
需要注意的是，MAX_NUM这个全局参数无法传参，因此max_patches默认为12，可能会导致图片token数量过长，目前通过修改Intern-S1-mini/preprocessor_config.json，硬编码max_patches解决，后续需要等swift更新适配。

大佬我是这样启动的，会报错
model_type='interns1'

启动训练并获取PID

nohup swift sft
--model_type $model_type
--model models/Intern-S1-mini
xxx
> "$LOG_FILE" 2>&1 &

ValueError: model_type: 'interns1' not in ['emu3_gen', 'emu3_chat', 'baichuan', 'baichuan_m1', 'baichuan2', 'ernie', 'modern_bert', 'modern_bert_gte', 'modern_bert_gte_reranker', 'bert', 'chatglm2', 'chatglm3', 'glm4', 'glm4_0414', 'glm4_z1_rumination', 'longwriter_llama3_1', 'codegeex4', 'glm4v', 'glm4_1v', 'cogvlm', 'cogagent_chat', 'cogagent_vqa', 'cogvlm2', 'cogvlm2_video', 'glm_edge', 'glm_edge_v', 'glm4_5', 'qwen', 'modelscope_agent', 'qwen_audio', 'qwen_vl', 'qwen2', 'qwen2_5', 'qwen2_5_math', 'qwen2_moe', 'qwen3', 'qwen3_moe', 'qwen3_thinking', 'qwen3_moe_thinking', 'qwen2_vl', 'qvq', 'qwen2_5_vl', 'mimo_vl', 'qwen2_5_omni', 'midashenglm', 'qwen2_audio', 'marco_o1', 'qwq_preview', 'qwq', 'ovis1_6', 'ovis1_6_llama3', 'ovis2', 'ovis2_5', 'qwen2_reward', 'qwen2_5_prm', 'qwen2_5_math_reward', 'qwen3_emb', 'qwen3_reranker', 'codefuse_qwen', 'codefuse_codegeex2', 'codefuse_codellama', 'deepseek', 'deepseek_moe', 'deepseek_v2', 'deepseek_v2_5', 'deepseek_vl', 'deepseek_janus', 'deepseek_janus_pro', 'deepseek_vl2', 'deepseek_r1', 'deepseek_r1_distill', 'paligemma', 'gemma', 'gemma2', 'gemma3_text', 'gemma3_vision', 'gemma3n', 'internlm', 'internlm2', 'internlm3', 'internvl', 'internvl_phi3', 'internvl2', 'internvl2_phi3', 'internvl2_5', 'internvl3', 'xcomposer2_5', 'xcomposer2', 'xcomposer2_4khd', 'xcomposer2_5_ol_audio', 'internlm2_reward', 'llama', 'llama3', 'llama3_1', 'llama3_2', 'llama3_2_vision', 'llama4', 'llama3_1_omni', 'reflection', 'atom', 'mengzi3', 'numina', 'ziya', 'megrez', 'llava_llama3_hf', 'llava1_5_hf', 'llava_onevision_hf', 'llava_next_qwen_hf', 'llama3_llava_next_hf', 'llava1_6_vicuna_hf', 'llava1_6_mistral_hf', 'llava_llama3_1_hf', 'llava1_6_yi_hf', 'llava_next_video_hf', 'llava_next_video_yi_hf', 'llama3_llava_next', 'llava1_6_mistral', 'llava1_6_yi', 'llava_next_qwen', 'grok', 'polylm', 'yuan2', 'orion', 'dbrx', 'bluelm', 'seggpt', 'xverse', 'xverse_moe', 'c4ai', 'aya', 'ling', 'qwen2_gte', 'mimo', 'mimo_rl', 'dots1', 'hunyuan_moe', 'hunyuan', 'gpt_oss', 'mamba', 'phi3_vision', 'phi4_multimodal', 'florence', 'phi3_small', 'phi2', 'phi3', 'phi4', 'phi3_moe', 'minicpm_moe', 'minicpmv', 'minicpmv2_5', 'minicpmv2_6', 'minicpmo2_6', 'minicpmv4', 'minicpm', 'minicpm_chatml', 'minicpm3', 'minimax_vl', 'minimax', 'minimax_m1', 'mistral', 'mixtral', 'mistral_nemo', 'mistral_2501', 'zephyr', 'wizardlm2_moe', 'wizardlm2', 'devstral', 'mistral_2503', 'idefics3', 'pixtral', 'molmoe', 'molmo', 'megrez_omni', 'qwen2_gme', 'keye_vl', 'dots_ocr', 'moonlight', 'kimi_vl', 'mplug_owl2', 'mplug_owl2_1', 'mplug_owl3', 'mplug_owl3_241101', 'doc_owl2', 'openbuddy_llama', 'openbuddy_llama3', 'openbuddy_mistral', 'openbuddy_mixtral', 'skywork', 'skywork_o1', 'llama3_2_reward', 'gemma_reward', 'got_ocr2', 'got_ocr2_hf', 'step_audio', 'telechat', 'telechat2', 'valley', 'yi_vl', 'yi', 'yi_coder', 'sus']

[Xuanaxx]

msswift成功微调启动脚本
export NPROC_PER_NODE=8
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export MASTER_PORT=39312
export PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True'
export MAX_NUM=1
GPUS=8
BATCH_SIZE=32
PER_DEVICE_BATCH_SIZE=1
GRADIENT_ACC=$((BATCH_SIZE / PER_DEVICE_BATCH_SIZE / GPUS))
LR=0.00001
EPOCHS=3.0
SAVE_STEPS=250
MAX_LEN=8192
OUTPUR_DIR=''
DATASET_DIR=''
model_type='interns1'
dataset=''
swift sft
--model_type $model_type
--model models/Intern-S1-mini
--train_type full
--dataset $DATASET_DIR
--freeze_vit true
--freeze_aligner false
--freeze_llm false
--torch_dtype bfloat16
--num_train_epochs $EPOCHS
--gradient_checkpointing true
--per_device_train_batch_size $PER_DEVICE_BATCH_SIZE
--per_device_eval_batch_size 1
--learning_rate $LR
--gradient_accumulation_steps $GRADIENT_ACC
--save_strategy steps
--save_steps $SAVE_STEPS
--split_dataset_ratio 0.01
--eval_strategy steps
--eval_steps $SAVE_STEPS
--save_total_limit 100
--logging_steps 1
--max_length $MAX_LEN
--output_dir $OUTPUR_DIR
--warmup_ratio 0.1
--dataloader_num_workers 32
--dataset_num_proc 16
--deepspeed zero3
--report_to tensorboard
--use_liger_kernel true
--attn_impl flash_attn
--truncation_strategy delete
2>&1 | tee ${OUTPUR_DIR}/training_log.txt
需要注意的是，MAX_NUM这个全局参数无法传参，因此max_patches默认为12，可能会导致图片token数量过长，目前通过修改Intern-S1-mini/preprocessor_config.json，硬编码max_patches解决，后续需要等swift更新适配。

大佬我是这样启动的，会报错 model_type='interns1'

启动训练并获取PID

nohup swift sft --model_type $model_type --model models/Intern-S1-mini xxx > "$LOG_FILE" 2>&1 &

ValueError: model_type: 'interns1' not in ['emu3_gen', 'emu3_chat', 'baichuan', 'baichuan_m1', 'baichuan2', 'ernie', 'modern_bert', 'modern_bert_gte', 'modern_bert_gte_reranker', 'bert', 'chatglm2', 'chatglm3', 'glm4', 'glm4_0414', 'glm4_z1_rumination', 'longwriter_llama3_1', 'codegeex4', 'glm4v', 'glm4_1v', 'cogvlm', 'cogagent_chat', 'cogagent_vqa', 'cogvlm2', 'cogvlm2_video', 'glm_edge', 'glm_edge_v', 'glm4_5', 'qwen', 'modelscope_agent', 'qwen_audio', 'qwen_vl', 'qwen2', 'qwen2_5', 'qwen2_5_math', 'qwen2_moe', 'qwen3', 'qwen3_moe', 'qwen3_thinking', 'qwen3_moe_thinking', 'qwen2_vl', 'qvq', 'qwen2_5_vl', 'mimo_vl', 'qwen2_5_omni', 'midashenglm', 'qwen2_audio', 'marco_o1', 'qwq_preview', 'qwq', 'ovis1_6', 'ovis1_6_llama3', 'ovis2', 'ovis2_5', 'qwen2_reward', 'qwen2_5_prm', 'qwen2_5_math_reward', 'qwen3_emb', 'qwen3_reranker', 'codefuse_qwen', 'codefuse_codegeex2', 'codefuse_codellama', 'deepseek', 'deepseek_moe', 'deepseek_v2', 'deepseek_v2_5', 'deepseek_vl', 'deepseek_janus', 'deepseek_janus_pro', 'deepseek_vl2', 'deepseek_r1', 'deepseek_r1_distill', 'paligemma', 'gemma', 'gemma2', 'gemma3_text', 'gemma3_vision', 'gemma3n', 'internlm', 'internlm2', 'internlm3', 'internvl', 'internvl_phi3', 'internvl2', 'internvl2_phi3', 'internvl2_5', 'internvl3', 'xcomposer2_5', 'xcomposer2', 'xcomposer2_4khd', 'xcomposer2_5_ol_audio', 'internlm2_reward', 'llama', 'llama3', 'llama3_1', 'llama3_2', 'llama3_2_vision', 'llama4', 'llama3_1_omni', 'reflection', 'atom', 'mengzi3', 'numina', 'ziya', 'megrez', 'llava_llama3_hf', 'llava1_5_hf', 'llava_onevision_hf', 'llava_next_qwen_hf', 'llama3_llava_next_hf', 'llava1_6_vicuna_hf', 'llava1_6_mistral_hf', 'llava_llama3_1_hf', 'llava1_6_yi_hf', 'llava_next_video_hf', 'llava_next_video_yi_hf', 'llama3_llava_next', 'llava1_6_mistral', 'llava1_6_yi', 'llava_next_qwen', 'grok', 'polylm', 'yuan2', 'orion', 'dbrx', 'bluelm', 'seggpt', 'xverse', 'xverse_moe', 'c4ai', 'aya', 'ling', 'qwen2_gte', 'mimo', 'mimo_rl', 'dots1', 'hunyuan_moe', 'hunyuan', 'gpt_oss', 'mamba', 'phi3_vision', 'phi4_multimodal', 'florence', 'phi3_small', 'phi2', 'phi3', 'phi4', 'phi3_moe', 'minicpm_moe', 'minicpmv', 'minicpmv2_5', 'minicpmv2_6', 'minicpmo2_6', 'minicpmv4', 'minicpm', 'minicpm_chatml', 'minicpm3', 'minimax_vl', 'minimax', 'minimax_m1', 'mistral', 'mixtral', 'mistral_nemo', 'mistral_2501', 'zephyr', 'wizardlm2_moe', 'wizardlm2', 'devstral', 'mistral_2503', 'idefics3', 'pixtral', 'molmoe', 'molmo', 'megrez_omni', 'qwen2_gme', 'keye_vl', 'dots_ocr', 'moonlight', 'kimi_vl', 'mplug_owl2', 'mplug_owl2_1', 'mplug_owl3', 'mplug_owl3_241101', 'doc_owl2', 'openbuddy_llama', 'openbuddy_llama3', 'openbuddy_mistral', 'openbuddy_mixtral', 'skywork', 'skywork_o1', 'llama3_2_reward', 'gemma_reward', 'got_ocr2', 'got_ocr2_hf', 'step_audio', 'telechat', 'telechat2', 'valley', 'yi_vl', 'yi', 'yi_coder', 'sus']

fetch一下swift的最新代码吧，看报错像是model_type没更新，msswift在8月22日才commit了对interns1的支持

[fak111]

感谢您的回复。
我试了下没有指定 model_type ,反而启动成果了。我的版本Name: ms_swift
Version: 3.7.2。

没有指定 model_type

启动训练并获取PID

nohup swift sft
--model '/root/share/new_models/Intern-S1-mini'
--dataset '/root/user/xxl'
--eval_steps 1000
--train_type lora
--lora_rank 8
--lora_dropout 0.01
--lora_alpha 16 \

[jianmanlincjx]

感谢您的回复。 我试了下没有指定 model_type ,反而启动成果了。我的版本Name: ms_swift Version: 3.7.2。

没有指定 model_type

启动训练并获取PID

nohup swift sft --model '/root/share/new_models/Intern-S1-mini' --dataset '/root/user/xxl' --eval_steps 1000 --train_type lora --lora_rank 8 --lora_dropout 0.01 --lora_alpha 16 \

你好，请问你训练完成后可以推理吗，我运行推理代码，需要我输入model_type，但是interns1不在预设定的列表里面

[fak111]

感谢您的回复。 我试了下没有指定 model_type ,反而启动成果了。我的版本Name: ms_swift Version: 3.7.2。

没有指定 model_type

启动训练并获取PID

nohup swift sft --model '/root/share/new_models/Intern-S1-mini' --dataset '/root/user/xxl' --eval_steps 1000 --train_type lora --lora_rank 8 --lora_dropout 0.01 --lora_alpha 16

你好，请问你训练完成后可以推理吗，我运行推理代码，需要我输入model_type，但是interns1不在预设定的列表里面

我这边可以推理，我发现可能大概率是 ms-swift版本问题。源码下载 3.8.0 好像就好了。
git clone https://xget.xi-xu.me/gh/modelscope/ms-swift.git cd ms-swift pip install -e .

然后第二个 bug，发现微调后可以推理但读不到图。
然后我在微调厚的权重目录下补了缺失的文件

我的微调脚本:
'''

没有指定 model_type

nohup swift sft
--model '/root/share/new_models/Intern-S1-mini'
--dataset '/root/share/datasets/VLM-formula-recognition-dataset_intern_camp/train/train_mini.jsonl'
--eval_steps 1000
--train_type lora
--lora_rank 8
--lora_dropout 0.01
--lora_alpha 16
--torch_dtype bfloat16
--num_train_epochs 5
--per_device_train_batch_size 1
--per_device_eval_batch_size 1
--learning_rate 1e-4
--warmup_ratio 0.05
--gradient_accumulation_steps 4
--save_steps 2000
--save_total_limit 10
--gradient_checkpointing_kwargs '{"use_reentrant": false}'
--logging_steps 1
--max_length 8000
--output_dir ./swift_output/SFT-Interns1mini
--dataset_num_proc 16
--dataloader_num_workers 16
--model_author JimmyMa99
--model_name SFT-camp6
--metric acc
--freeze_vit true
> "$LOG_FILE" 2>&1 &
'''