chore: import upstream snapshot with attribution

examples/train/packing/dpo.sh

@@ -0,0 +1,31 @@
+# 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 \
+    --tuner_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 \
+    --rpo_alpha 0.1 \
+    --packing true

examples/train/packing/dpo_vlm.sh

@@ -0,0 +1,34 @@
+# 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 \
+    --tuner_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 \
+    --freeze_aligner 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 \
+    --rpo_alpha 0.1 \
+    --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 \
+    --tuner_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 \
+    --tuner_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
+# 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 \
+    --tuner_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 \
+    --freeze_aligner 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 +1,35 @@
+# 4 * 36GB
+# 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 \
+    --tuner_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 \
+    --tuner_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