Sync StreamDelta directory

Co-authored-by: TRAE CLI <noreply@bytedance.com>

examples/train/multimodal/ocr.sh

@@ -0,0 +1,27 @@
+# 20GB
+CUDA_VISIBLE_DEVICES=0 \
+MAX_PIXELS=1003520 \
+swift sft \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --dataset 'AI-ModelScope/LaTeX_OCR:human_handwrite#20000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type lora \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --freeze_vit true \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 4096 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4

examples/train/multimodal/omni/infer.sh

@@ -0,0 +1,10 @@
+CUDA_VISIBLE_DEVICES=0 \
+VIDEO_MAX_PIXELS=50176 \
+FPS_MAX_FRAMES=12 \
+MAX_PIXELS=1003520 \
+ENABLE_AUDIO_OUTPUT=0 \
+swift infer \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --stream true \
+    --load_data_args true \
+    --max_new_tokens 2048

examples/train/multimodal/omni/sft.sh

@@ -0,0 +1,39 @@
+# 4*35GB
+# A demo for four modalities that can be run directly
+nproc_per_node=4
+
+# If using zero3, please set `ENABLE_AUDIO_OUTPUT=0`.
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+ENABLE_AUDIO_OUTPUT=1 \
+NPROC_PER_NODE=$nproc_per_node \
+VIDEO_MAX_PIXELS=50176 \
+FPS_MAX_FRAMES=12 \
+MAX_PIXELS=1003520 \
+swift sft \
+    --model Qwen/Qwen2.5-Omni-7B \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#2000' \
+              'AI-ModelScope/LaTeX_OCR:human_handwrite#2000' \
+              'speech_asr/speech_asr_aishell1_trainsets:validation#2000' \
+              'swift/VideoChatGPT:all#2000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type lora \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --freeze_vit true \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --deepspeed zero2

examples/train/multimodal/rlhf/dpo/full.sh

@@ -0,0 +1,32 @@
+# 4 * 50GiB
+nproc_per_node=4
+
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=$nproc_per_node \
+MAX_PIXELS=1003520 \
+swift rlhf \
+    --rlhf_type dpo \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --dataset 'swift/RLAIF-V-Dataset#20000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type full \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-5 \
+    --freeze_vit true \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --deepspeed zero3 \
+    --logging_steps 5 \
+    --max_length 4096 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --save_only_model true

examples/train/multimodal/rlhf/dpo/lora.sh

@@ -0,0 +1,35 @@
+# 4 * 50GiB
+# You can refer to `https://github.com/QwenLM/Qwen2.5-VL` for the meaning of the `MAX_PIXELS` parameter.
+# --rlhf_type cpo/orpo/simpo/rm are also supported
+nproc_per_node=2
+
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=$nproc_per_node \
+MAX_PIXELS=1003520 \
+swift rlhf \
+    --rlhf_type dpo \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --dataset 'swift/RLAIF-V-Dataset#20000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type lora \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --freeze_vit true \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --deepspeed zero2 \
+    --logging_steps 5 \
+    --max_length 4096 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4

examples/train/multimodal/rlhf/gkd/fast.sh

@@ -0,0 +1,49 @@
+
+export teacher_model='OpenGVLab/InternVL3-8B'
+
+NPROC_PER_NODE=4 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift infer \
+    --model $teacher_model \
+    --infer_backend vllm \
+    --val_dataset 'modelscope/coco_2014_caption:validation#5000' \
+    --vllm_gpu_memory_utilization 0.9 \
+    --vllm_max_model_len 8192 \
+    --max_new_tokens 2048 \
+    --write_batch_size 1000 \
+    --result_path new_coco_dataset.jsonl
+
+
+# 4 * 42GiB, 3.05s/it
+NPROC_PER_NODE=4 \
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift rlhf \
+    --rlhf_type gkd \
+    --model OpenGVLab/InternVL3-2B-Pretrained \
+    --teacher_model $teacher_model \
+    --train_type full \
+    --dataset 'new_coco_dataset.jsonl' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 4 \
+    --per_device_eval_batch_size 4 \
+    --learning_rate 1e-5 \
+    --freeze_vit true \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 4096 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --save_only_model true \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --deepspeed zero2 \
+    --padding_free true \
+    --attn_impl flash_attn \
+    --lmbda 0

examples/train/multimodal/rlhf/gkd/full.sh

@@ -0,0 +1,34 @@
+# 4 * 45GiB, 10.29s/it
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+MASTER_PORT=29501 \
+NPROC_PER_NODE=4 \
+swift rlhf \
+    --rlhf_type gkd \
+    --model OpenGVLab/InternVL3-2B-Pretrained \
+    --teacher_model OpenGVLab/InternVL3-8B \
+    --dataset 'modelscope/coco_2014_caption:validation#2000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type full \
+    --seq_kd true \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 4 \
+    --per_device_eval_batch_size 4 \
+    --learning_rate 1e-5 \
+    --freeze_vit true \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --deepspeed zero2 \
+    --attn_impl flash_attn \
+    --logging_steps 5 \
+    --max_length 4096 \
+    --max_completion_length 512 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --save_only_model true

examples/train/multimodal/rlhf/kto.sh

@@ -0,0 +1,34 @@
+# Due to the absence of a multi-modal open-source dataset for kto,
+# we will use a pure text kto dataset as an example here.
+nproc_per_node=2
+
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=$nproc_per_node \
+MAX_PIXELS=1003520 \
+swift rlhf \
+    --rlhf_type kto \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --dataset 'AI-ModelScope/ultrafeedback-binarized-preferences-cleaned-kto#10000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type lora \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --freeze_vit true \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --deepspeed zero2 \
+    --logging_steps 5 \
+    --max_length 4096 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4

examples/train/multimodal/video.sh

@@ -0,0 +1,33 @@
+# 2*24GB
+# You can refer to `https://github.com/QwenLM/Qwen2.5-VL` for the meaning of the `VIDEO_MAX_PIXELS` parameter.
+nproc_per_node=2
+
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=$nproc_per_node \
+VIDEO_MAX_PIXELS=50176 \
+FPS_MAX_FRAMES=12 \
+swift sft \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --dataset swift/VideoChatGPT:all \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type lora \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --freeze_vit true \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --deepspeed zero2

examples/train/multimodal/vit_gradient_checkpointing.sh

@@ -0,0 +1,37 @@
+# gc true, vgc true: 48GiB, 2.45s/it
+# gc true, vgc false: 62GiB 2.32s/it
+# gc false, vgc true: 56GiB 2.16s/it
+# gc false, vgc false: 77GiB 1.95s/it
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=4 \
+VIDEO_MAX_PIXELS=50176 \
+FPS_MAX_FRAMES=12 \
+swift sft \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --dataset swift/VideoChatGPT:all \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type full \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 4 \
+    --per_device_eval_batch_size 4 \
+    --learning_rate 1e-5 \
+    --freeze_vit false \
+    --freeze_aligner false \
+    --gradient_accumulation_steps 1 \
+    --gradient_checkpointing true \
+    --vit_gradient_checkpointing true \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --deepspeed zero3 \
+    --use_liger_kernel true \
+    --attn_impl flash_attn \
+    --padding_free true \
+    --save_only_model true

examples/train/new_special_tokens/infer.sh

@@ -0,0 +1,6 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --max_batch_size 16 \
+    --load_data_args true \
+    --temperature 0

examples/train/new_special_tokens/merge_lora.sh

@@ -0,0 +1,13 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift export \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --merge_lora true
+
+
+# infer
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --adapters output/vx-xxx/checkpoint-xxx-merged \
+    --max_batch_size 16 \
+    --load_data_args true \
+    --temperature 0

examples/train/new_special_tokens/tokens.txt

@@ -0,0 +1,100 @@
+<|0|>
+<|1|>
+<|2|>
+<|3|>
+<|4|>
+<|5|>
+<|6|>
+<|7|>
+<|8|>
+<|9|>
+<|10|>
+<|11|>
+<|12|>
+<|13|>
+<|14|>
+<|15|>
+<|16|>
+<|17|>
+<|18|>
+<|19|>
+<|20|>
+<|21|>
+<|22|>
+<|23|>
+<|24|>
+<|25|>
+<|26|>
+<|27|>
+<|28|>
+<|29|>
+<|30|>
+<|31|>
+<|32|>
+<|33|>
+<|34|>
+<|35|>
+<|36|>
+<|37|>
+<|38|>
+<|39|>
+<|40|>
+<|41|>
+<|42|>
+<|43|>
+<|44|>
+<|45|>
+<|46|>
+<|47|>
+<|48|>
+<|49|>
+<|50|>
+<|51|>
+<|52|>
+<|53|>
+<|54|>
+<|55|>
+<|56|>
+<|57|>
+<|58|>
+<|59|>
+<|60|>
+<|61|>
+<|62|>
+<|63|>
+<|64|>
+<|65|>
+<|66|>
+<|67|>
+<|68|>
+<|69|>
+<|70|>
+<|71|>
+<|72|>
+<|73|>
+<|74|>
+<|75|>
+<|76|>
+<|77|>
+<|78|>
+<|79|>
+<|80|>
+<|81|>
+<|82|>
+<|83|>
+<|84|>
+<|85|>
+<|86|>
+<|87|>
+<|88|>
+<|89|>
+<|90|>
+<|91|>
+<|92|>
+<|93|>
+<|94|>
+<|95|>
+<|96|>
+<|97|>
+<|98|>
+<|99|>

examples/train/new_special_tokens/train.sh

@@ -0,0 +1,31 @@
+# 4 * 26GB
+# This example is just a demo showing how to add new_special_tokens.
+NPROC_PER_NODE=4 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset 'swift/new_special_tokens' \
+    --split_dataset_ratio 0.01 \
+    --new_special_tokens 'examples/train/new_special_tokens/tokens.txt' \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 5 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 16 \
+    --padding_free true \
+    --attn_impl flash_attn \
+    --learning_rate 1e-4 \
+    --lora_rank 16 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --modules_to_save embed_tokens lm_head \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 500 \
+    --save_steps 500 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --deepspeed zero2

examples/train/optimizer/muon.sh

@@ -0,0 +1,31 @@
+# 17GB
+# ref: https://github.com/MoonshotAI/Moonlight/blob/master/examples/toy_train.py
+# `moonshotai/Moonlight-16B-A3B-Instruct` does not support training; here we use `Qwen/Qwen2.5-7B-Instruct` as an example.
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
+              'AI-ModelScope/alpaca-gpt4-data-en#500' \
+              'swift/self-cognition#500' \
+    --optimizer muon \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --system 'You are a helpful assistant.' \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/packing/dpo.sh

@@ -0,0 +1,30 @@
+# 4 * 50GiB; 2h
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+NPROC_PER_NODE=4 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift rlhf \
+    --rlhf_type dpo \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type full \
+    --dataset AI-ModelScope/orpo-dpo-mix-40k \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-5 \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 8192 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --save_only_model true \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --deepspeed zero3 \
+    --attn_impl flash_attn \
+    --packing true

examples/train/packing/dpo_vlm.sh

@@ -0,0 +1,32 @@
+# 50GiB; 6h
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=4 \
+MAX_PIXELS=1003520 \
+swift rlhf \
+    --rlhf_type dpo \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --dataset 'swift/RLAIF-V-Dataset' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type full \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-5 \
+    --freeze_vit true \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --save_total_limit 2 \
+    --deepspeed zero3 \
+    --logging_steps 5 \
+    --max_length 8192 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 64 \
+    --attn_impl flash_attn \
+    --save_only_model true \
+    --packing true

examples/train/packing/liger_kernel.sh

@@ -0,0 +1,31 @@
+# Env: 4 * A100
+# https://github.com/modelscope/ms-swift/blob/main/examples/megatron/long_text.sh
+# Max Length: 16K
+# GPU Memory: 4 * 42GB, Training Speed 10s/it
+NPROC_PER_NODE=4 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift sft \
+    --model Qwen/Qwen2.5-7B \
+    --train_type full \
+    --dataset 'AI-ModelScope/LongAlpaca-12k' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --torch_dtype bfloat16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-5 \
+    --gradient_accumulation_steps 2 \
+    --packing true \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --logging_steps 5 \
+    --max_length 16384 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 8 \
+    --dataset_num_proc 8 \
+    --save_total_limit 2 \
+    --save_only_model true \
+    --output_dir output/Qwen2.5-7B \
+    --deepspeed zero3 \
+    --use_liger_kernel true \
+    --attn_impl flash_attn

examples/train/packing/llm.sh

@@ -0,0 +1,32 @@
+# 22GB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --packing true \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
+              'AI-ModelScope/alpaca-gpt4-data-en#500' \
+              'swift/self-cognition#500' \
+    --split_dataset_ratio 0.01 \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 3 \
+    --attn_impl flash_attn \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps 4 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --system 'You are a helpful assistant.' \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/packing/qwen2_5_omni.sh

@@ -0,0 +1,40 @@
+# 4 * 32GB
+# Multimodal packing currently only supports qwen2_vl, qwen2_5_vl, qwen2_5_omni, internvl2_5/3
+# A demo for four modalities that can be run directly
+# For local datasets, it is recommended to use streaming: `--streaming true` (save memory)
+NPROC_PER_NODE=4 \
+ENABLE_AUDIO_OUTPUT=1 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+VIDEO_MAX_PIXELS=50176 \
+FPS_MAX_FRAMES=12 \
+MAX_PIXELS=1003520 \
+swift sft \
+    --model Qwen/Qwen2.5-Omni-7B \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#10000' \
+              'AI-ModelScope/LaTeX_OCR#2000' \
+              'speech_asr/speech_asr_aishell1_trainsets:validation#2000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type lora \
+    --torch_dtype bfloat16 \
+    --attn_impl flash_attn \
+    --packing true \
+    --num_train_epochs 3 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --freeze_vit true \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 4096 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 8 \
+    --deepspeed zero2

examples/train/packing/qwen2_5_vl.sh

@@ -0,0 +1,36 @@
+# 4 * 36GB
+# Multimodal packing currently only supports qwen2_vl, qwen2_5_vl, qwen2_5_omni, internvl2_5/3
+# Efficiency: With packing: 10 minutes; Without packing: >=1 hour
+# For local datasets, it is recommended to use streaming: `--streaming true` (save memory)
+# You can also use padding_free to avoid the space/time cost caused by multi-modal packing:
+# https://github.com/modelscope/ms-swift/blob/main/examples/train/padding_free/sft.sh
+NPROC_PER_NODE=4 \
+MAX_PIXELS=1003520 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift sft \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --train_type lora \
+    --dataset 'AI-ModelScope/LaTeX_OCR#20000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --torch_dtype bfloat16 \
+    --attn_impl flash_attn \
+    --packing true \
+    --num_train_epochs 3 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 8192 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 8 \
+    --deepspeed zero2

examples/train/packing/streaming.sh

@@ -0,0 +1,34 @@
+# 4 * 36GB
+# A demo using the Hugging Face dataset
+# The first model weights will be saved around step 70.
+NPROC_PER_NODE=4 \
+MAX_PIXELS=1003520 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+HF_ENDPOINT=https://hf-mirror.com \
+swift sft \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --train_type lora \
+    --dataset 'HF::linxy/LaTeX_OCR:full#20000' \
+    --torch_dtype bfloat16 \
+    --attn_impl flash_attn \
+    --streaming true \
+    --shuffle_buffer_size 1000 \
+    --packing true \
+    --save_strategy epoch \
+    --max_steps 1000 \
+    --max_epochs 5 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps 1 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 8192 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 1 \
+    --dataset_num_proc 8 \
+    --deepspeed zero2

examples/train/padding_free/dpo_vlm.sh

@@ -0,0 +1,33 @@
+# with padding_free: 4 * 53GiB, 3.55s/it
+# without padding_free: 4 * 62GiB 4.41s/it
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=4 \
+MAX_PIXELS=1003520 \
+swift rlhf \
+    --rlhf_type dpo \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --dataset 'swift/RLAIF-V-Dataset#20000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type full \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 4 \
+    --per_device_eval_batch_size 4 \
+    --learning_rate 1e-5 \
+    --freeze_vit true \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --deepspeed zero3 \
+    --logging_steps 5 \
+    --max_length 4096 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --attn_impl flash_attn \
+    --save_only_model true \
+    --padding_free true

examples/train/padding_free/sft.sh

@@ -0,0 +1,31 @@
+# Supported multimodal models reference:
+# https://github.com/modelscope/ms-swift/blob/main/examples/train/packing/qwen2_5_vl.sh
+# without padding_free: 4 * 60GiB, 26h
+# padding_free: 4 * 44GiB, 13h
+NPROC_PER_NODE=4 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift sft \
+    --model Qwen/Qwen2.5-7B \
+    --train_type full \
+    --dataset 'liucong/Chinese-DeepSeek-R1-Distill-data-110k-SFT' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --torch_dtype bfloat16 \
+    --per_device_train_batch_size 8 \
+    --per_device_eval_batch_size 8 \
+    --learning_rate 1e-5 \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --logging_steps 5 \
+    --max_length 8192 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 8 \
+    --dataset_num_proc 8 \
+    --save_total_limit 2 \
+    --save_only_model true \
+    --output_dir output/Qwen2.5-7B \
+    --deepspeed zero3 \
+    --use_liger_kernel true \
+    --attn_impl flash_attn \
+    --padding_free true

examples/train/plugins/loss_scale.sh

@@ -0,0 +1,22 @@
+# loss_scale all to train all tokens
+# use loss_type loss_scale
+# This is just an example
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset 'swift/self-cognition#1000' \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot \
+    --loss_scale all \
+    --loss_type loss_scale

examples/train/plugins/tuner_phi4_mm.sh

@@ -0,0 +1,22 @@
+# `--train_type dummy`
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model LLM-Research/Phi-4-multimodal-instruct \
+    --dataset 'AI-ModelScope/LaTeX_OCR:human_handwrite#20000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type dummy \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4

examples/train/predict_with_generate/train.sh

@@ -0,0 +1,30 @@
+# 20GiB
+CUDA_VISIBLE_DEVICES=0 \
+MAX_PIXELS=1003520 \
+swift sft \
+    --model Qwen/Qwen2.5-VL-7B-Instruct \
+    --dataset 'AI-ModelScope/LaTeX_OCR:human_handwrite#20000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type lora \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 2 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --freeze_vit true \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --predict_with_generate true \
+    --metric_for_best_model rouge-l \
+    --greater_is_better true

examples/train/pretrain/train.sh

@@ -0,0 +1,31 @@
+# If not using flash_attn, or transformers<4.44,
+# or encountering an abnormally large loss (i.e., the model does not support packing),
+# please remove `--packing true`.
+nproc_per_node=4
+
+NPROC_PER_NODE=$nproc_per_node \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift pt \
+    --model Qwen/Qwen2.5-7B \
+    --train_type full \
+    --dataset swift/chinese-c4 \
+    --torch_dtype bfloat16 \
+    --streaming true \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-5 \
+    --gradient_accumulation_steps $(expr 64 / $nproc_per_node) \
+    --packing true \
+    --eval_steps 500 \
+    --save_steps 500 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --deepspeed zero3 \
+    --max_length 8192 \
+    --max_steps 10000 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 8 \
+    --save_only_model true \
+    --output_dir output/Qwen2.5-7B \
+    --attn_impl flash_attn

examples/train/qlora/awq/merge_lora.sh

@@ -0,0 +1,5 @@
+# This will store the full, unquantized weights.
+swift export \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --merge_lora true

examples/train/qlora/awq/train.sh

@@ -0,0 +1,28 @@
+# 10GB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct-AWQ \
+    --train_type lora \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
+              'AI-ModelScope/alpaca-gpt4-data-en#500' \
+              'swift/self-cognition#500' \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --system 'You are a helpful assistant.' \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/qlora/bnb/merge_lora.sh

@@ -0,0 +1,4 @@
+# This will store the full, unquantized weights.
+swift export \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --merge_lora true

examples/train/qlora/bnb/train.sh

@@ -0,0 +1,34 @@
+# 10GB
+# pip install bitsandbytes
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
+              'AI-ModelScope/alpaca-gpt4-data-en#500' \
+              'swift/self-cognition#500' \
+    --torch_dtype bfloat16 \
+    --bnb_4bit_compute_dtype bfloat16 \
+    --bnb_4bit_quant_type nf4 \
+    --bnb_4bit_use_double_quant true \
+    --quant_method bnb \
+    --quant_bits 4 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --system 'You are a helpful assistant.' \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/qlora/gptq.sh

@@ -0,0 +1,27 @@
+# 2 * 30GiB
+CUDA_VISIBLE_DEVICES=0,1 \
+MAX_PIXELS=1003520 \
+swift sft \
+    --model Qwen/Qwen2.5-VL-72B-Instruct-GPTQ-Int4 \
+    --dataset 'modelscope/coco_2014_caption:validation#20000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --train_type lora \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --freeze_vit true \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4

examples/train/qlora/hqq.sh

@@ -0,0 +1,31 @@
+# 10GB
+# pip install hqq
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
+              'AI-ModelScope/alpaca-gpt4-data-en#500' \
+              'swift/self-cognition#500' \
+    --torch_dtype bfloat16 \
+    --quant_method hqq \
+    --quant_bits 4 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --learning_rate 1e-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --output_dir output \
+    --system 'You are a helpful assistant.' \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --model_author swift \
+    --model_name swift-robot