Sync StreamDelta directory

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

examples/train/rft/rft.py

@@ -0,0 +1,224 @@
+import os
+import shutil
+import subprocess
+import time
+from typing import List
+
+from swift.utils import get_device_count
+
+# NOTE: this script supports at most 8 GPUS in a node, if using multi node, please use custom logic.
+
+# Paste conda env
+# conda_prefix = 'source /root/miniconda3/etc/profile.d/conda.sh && conda activate py311 && '
+conda_prefix = ''
+
+
+def do_sample(model: str, model_type: str, dataset: List[str], iter: int):
+    device_count = get_device_count()
+    handlers = []
+    datasets = []
+    # Sampling cache, to avoid lmdeploy & PRM run at the same time
+    # Why lmdeploy not vllm? we found that the responses generated by lmdeploy are more similar than ones of vllm.
+    for device in range(device_count):
+        sample_cmd = (f'{conda_prefix} USE_OPENCOMPASS_EVALUATOR=True CUDA_VISIBLE_DEVICES={device} swift sample '
+                      f'--model {model} --model_type {model_type} '
+                      f'--dataset {" ".join(dataset)} '
+                      f'--data_range {device} {device_count} '
+                      f'--max_length 2048 '
+                      f'--system "You are a math model, you should **think step by step** carefully, '
+                      f'and always consider the basic math principles to avoid making calculating mistakes.'
+                      f'Give the final answer wrapped with \\boxed{{}}" '
+                      f'--load_args false '
+                      f'--sampler_engine vllm '
+                      f'--max_new_tokens 768 '
+                      f'--override_exist_file true '
+                      f'--num_sampling_per_gpu_batch_size 1 '
+                      f'--num_return_sequences 64 '
+                      f'--cache_files sample_output/iter_{iter}_proc_{device}_cache.jsonl '
+                      f'--output_file iter_{iter}_proc_{device}_cache.jsonl '
+                      f'--top_p 1.0 '
+                      f'--temperature 1.0 ')
+        print(f'Sampling caches of iter {iter}, part {device}.', flush=True)
+        env = os.environ.copy()
+        env['CUDA_VISIBLE_DEVICES'] = str(device)
+        handler = subprocess.Popen(
+            f'{sample_cmd}' + f' > logs/sample_iter_{iter}_proc_{device}_cache.log 2>&1',
+            env=os.environ.copy(),
+            shell=True,
+            executable='/bin/bash')
+        handlers.append(handler)
+
+    for proc, handler in enumerate(handlers):
+        handler.wait()
+        assert os.path.exists(os.path.join('sample_output', f'iter_{iter}_proc_{proc}_cache.jsonl'))
+
+    handlers = []
+    # Sample again, this time to filter with ORM & PRM
+    # Provide your PRM model or PRM name(add PRM in plugin/prm.py first)
+    # You can define your custom PRM logic in the plugin
+    # (like, split your steps, use the worst score/last score/avg score)
+    for device in range(device_count):
+        sample_cmd = (
+            f'{conda_prefix} USE_OPENCOMPASS_EVALUATOR=True CUDA_VISIBLE_DEVICES={device} swift sample '
+            f'--model {model} --model_type {model_type} '  # change to --resume_from_checkpoint to use the latest optimizer state # noqa
+            f'--dataset {" ".join(dataset)} '
+            f'--data_range {device} {device_count} '
+            f'--max_length 2048 '
+            f'--system "You are a math model, you should **think step by step** carefully, '
+            f'and always consider the basic math principles to avoid making calculating mistakes.'
+            f'Give the final answer wrapped with \\boxed{{}}" '
+            f'--load_args false '
+            f'--sampler_engine no '
+            f'--orm_model math '  # math defines in plugin/orm.py
+            f'--prm_model Qwen/Qwen2.5-Math-PRM-7B '
+            f'--prm_threshold {min(0.7 + 0.1*iter, 0.9)} '
+            f'--max_new_tokens 768 '
+            f'--override_exist_file true '  # no not override the existing sample files
+            f'--num_sampling_per_gpu_batch_size 1 '
+            f'--num_return_sequences 64 '
+            f'--output_file iter_{iter}_proc_{device}_sampling.jsonl '
+            f'--cache_files sample_output/iter_{iter}_proc_{device}_cache.jsonl ')
+        print(f'Sampling iter {iter}, part {device}.', flush=True)
+        env = os.environ.copy()
+        env['CUDA_VISIBLE_DEVICES'] = str(device)
+        handler = subprocess.Popen(
+            f'{sample_cmd}' + f' > logs/sample_iter_{iter}_proc_{device}.log 2>&1',
+            env=os.environ.copy(),
+            shell=True,
+            executable='/bin/bash')
+        handlers.append(handler)
+
+    for proc, handler in enumerate(handlers):
+        handler.wait()
+        assert os.path.exists(os.path.join('sample_output', f'iter_{iter}_proc_{proc}_sampling.jsonl')), (
+            f'{os.path.join("sample_output", f"iter_{iter}_proc_{proc}_sampling.jsonl")} not exists, '
+            'please check the sample logs to get the detail error.')
+        datasets.append(os.path.join('sample_output', f'iter_{iter}_proc_{proc}_sampling.jsonl'))
+    print(f'Sampling done, files:{datasets}', flush=True)
+    return datasets
+
+
+def do_train(model: str, model_type: str, datasets: List[str], iter, cmd='sft'):
+    gpu_prefix = ''
+    ds_config = ''
+    if get_device_count() > 1:
+        gpu_prefix = f'NPROC_PER_NODE={get_device_count()} '
+        ds_config = '--deepspeed zero3 '
+    extra_args = ''
+    if cmd == 'rlhf':
+        extra_args = '--rlhf_type dpo --beta 0.3 '  # use another reinforce learning method supported by swift
+    ga = 128 // get_device_count() // 2
+    train_cmd = (f'{conda_prefix} {gpu_prefix} swift {cmd} '
+                 f'--model {model} --model_type {model_type} '
+                 f'--dataset {" ".join(datasets)} '
+                 f'--max_length 2048 '
+                 f'--num_train_epochs 1 '
+                 f'--load_args false '
+                 f'--train_type full '
+                 f'{extra_args} '
+                 f'--eval_strategy no '
+                 f'--split_dataset_ratio 0 '
+                 f'--per_device_train_batch_size 2 '
+                 f'--gradient_accumulation_steps {ga} '
+                 f'--save_steps 1 '
+                 f'--save_strategy epoch '
+                 f'{ds_config} '
+                 f'--learning_rate 4e-6 ')
+
+    print(f'Training iter {iter}.', flush=True)
+    handler = subprocess.Popen(
+        f'{train_cmd}' + f' > logs/train_iter_{iter}.log 2>&1',
+        shell=True,
+        env=os.environ.copy(),
+        executable='/bin/bash')
+    handler.wait()
+    ckpt = None
+    with open(f'logs/train_iter_{iter}.log', 'r') as f:
+        for line in f.readlines():
+            if 'last_model_checkpoint: ' in line:
+                ckpt = line.split('last_model_checkpoint: ')[1]
+                break
+    assert ckpt is not None
+    print(f'Training done, ckpt: {ckpt.strip()}.', flush=True)
+    return ckpt.strip()
+
+
+def do_eval(model, model_type: str, iter):
+    eval_cmd = (
+        f'{conda_prefix} swift eval '
+        '--eval_dataset competition_math '  # eval another dataset
+        '--infer_backend vllm --eval_limit 500 '
+        f'--model {model} --model_type {model_type} '
+        '--system "You are a math model, you should **think step by step** carefully, '
+        'and always consider the basic math principles to avoid making calculating mistakes. '
+        'Give the final answer wrapped with \\boxed{}"')
+    print('Evaluating.', flush=True)
+    # Replace the original dataset to the math.json, this is for test, comment this if not need
+    replace_math_dataset()
+
+    if iter is None:
+        iter = 'origin'
+    env = os.environ.copy()
+    env['CUDA_VISIBLE_DEVICES'] = '0'
+    handler = subprocess.Popen(
+        f'{eval_cmd}' + f' > logs/eval_iter_{iter}.log 2>&1', shell=True, env=env, executable='/bin/bash')
+    handler.wait()
+
+    acc = None
+    # | math | 393424 | accuracy | gen | 39.00 |
+    with open(f'logs/eval_iter_{iter}.log', 'r') as f:
+        for line in f.readlines():
+            if 'Level 5' in line and 'AveragePass@1' in line:
+                parts = [p for p in line.split('|') if p.strip()]
+                acc = float(parts[-2])
+                break
+
+    print(f'Iter {iter} eval done with acc: {acc}.', flush=True)
+    return acc
+
+
+def replace_math_dataset():
+    # Note: This may run failed because this is special for math test,
+    # and one must run swift eval --eval_dataset math first to make sure opencompass has created
+    # the folder.
+    # You can use original math dataset either. just comment this call.
+    user_dir = os.path.expanduser('~')
+    if os.path.exists(os.path.join(user_dir, '.cache', 'opencompass', 'data', 'math', 'math.json')):
+        os.remove(os.path.join(user_dir, '.cache', 'opencompass', 'data', 'math', 'math.json'))
+    shutil.copy(
+        os.path.join('examples', 'train', 'rft', 'math.json'),
+        os.path.join(user_dir, '.cache', 'opencompass', 'data', 'math', 'math.json'))
+
+
+def main():
+    os.makedirs('logs', exist_ok=True)
+    max_acc = 0.
+    first_model = 'Qwen/Qwen2.5-Math-7B-Instruct'
+    model_type = 'qwen2_5_math'
+
+    if False:
+        # eval the original model
+        do_eval(first_model, None)
+
+    model = first_model
+    for i in range(5):
+        ts = time.time()
+        datasets = do_sample(model, model_type, ['tastelikefeet/competition_math'], i)
+        # add custom data filter here, for example: length or diversity control
+        print(f'do sample cost: {(time.time()-ts) / 60:.1f} minutes.', flush=True)
+        ts = time.time()
+        # if want to train the original dataset with datasets, add the original dataset here
+        # if want to train the original model everytime, change to first_model
+        ckpt = do_train(model, model_type, datasets, i)
+        print(f'do train cost: {(time.time() - ts) / 60:.1f} minutes.', flush=True)
+        ts = time.time()
+        acc = do_eval(ckpt, model_type, i)
+        print(f'do eval cost: {(time.time() - ts) / 60:.1f} minutes.', flush=True)
+        if acc > max_acc:
+            max_acc = acc
+        model = ckpt
+        print(f'acc: {acc}, upgrade model to : {model}', flush=True)
+
+
+if __name__ == '__main__':
+    main()

examples/train/rlhf/README.md

@@ -0,0 +1,3 @@
+# TIPS
+
+Multi-modal models' RLHF are also supported! Check the multimodal folder for details.

examples/train/rlhf/cpo.sh

@@ -0,0 +1,30 @@
+nproc_per_node=2
+
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=$nproc_per_node \
+swift rlhf \
+    --rlhf_type cpo \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset hjh0119/shareAI-Llama3-DPO-zh-en-emoji \
+    --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-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --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 zero2 \
+    --dataset_num_proc 4

examples/train/rlhf/dpo/full.sh

@@ -0,0 +1,31 @@
+# with padding_free: 4 * 47GiB, 1.90s/it
+# without padding_free: 4 * 57GiB 3.32s/it
+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 hjh0119/shareAI-Llama3-DPO-zh-en-emoji \
+    --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 \
+    --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 \
+    --save_only_model true \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --deepspeed zero3 \
+    --attn_impl flash_attn \
+    --rpo_alpha 0.1 \
+    --padding_free true

examples/train/rlhf/dpo/lora.sh

@@ -0,0 +1,30 @@
+# 24GiB
+# It is recommended to use padding_free. For more details, please refer to:
+# https://github.com/modelscope/ms-swift/blob/main/examples/train/padding_free/dpo.sh
+CUDA_VISIBLE_DEVICES=0 \
+swift rlhf \
+    --rlhf_type dpo \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset hjh0119/shareAI-Llama3-DPO-zh-en-emoji \
+    --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-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 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --rpo_alpha 0.1 \
+    --dataset_num_proc 4

examples/train/rlhf/gkd/fast.sh

@@ -0,0 +1,48 @@
+
+export teacher_model='Qwen/Qwen2.5-14B-Instruct'
+
+NPROC_PER_NODE=4 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift infer \
+    --model $teacher_model \
+    --infer_backend vllm \
+    --val_dataset 'AI-ModelScope/alpaca-gpt4-data-en#5000' 'AI-ModelScope/alpaca-gpt4-data-zh#5000' \
+    --vllm_gpu_memory_utilization 0.9 \
+    --vllm_max_model_len 8192 \
+    --max_new_tokens 2048 \
+    --write_batch_size 10000 \
+    --result_path new_dataset.jsonl
+
+
+# 4 * 56GiB, 1.55s/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 Qwen/Qwen2.5-7B \
+    --teacher_model $teacher_model \
+    --train_type full \
+    --dataset 'new_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 \
+    --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 zero3 \
+    --padding_free true \
+    --attn_impl flash_attn \
+    --lmbda 0

examples/train/rlhf/gkd/full.sh

@@ -0,0 +1,31 @@
+# 4 * 66GiB, 46s/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 Qwen/Qwen2.5-7B \
+    --teacher_model Qwen/Qwen2.5-14B-Instruct \
+    --train_type full \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-en#2000' 'AI-ModelScope/alpaca-gpt4-data-zh#2000' \
+    --split_dataset_ratio 0.01 \
+    --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 \
+    --gradient_accumulation_steps 1 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --max_length 2048 \
+    --max_completion_length 512 \
+    --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

examples/train/rlhf/gkd/think_model.sh

@@ -0,0 +1,50 @@
+
+export teacher_model='Qwen/Qwen3-8B'
+
+NPROC_PER_NODE=4 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift infer \
+    --model $teacher_model \
+    --infer_backend vllm \
+    --val_dataset 'AI-ModelScope/alpaca-gpt4-data-en#5000' 'AI-ModelScope/alpaca-gpt4-data-zh#5000' \
+    --vllm_gpu_memory_utilization 0.9 \
+    --vllm_max_model_len 8192 \
+    --max_new_tokens 2048 \
+    --write_batch_size 10000 \
+    --result_path new_dataset.jsonl
+
+
+# 4 * 67GiB, 2.50s/it
+# You need to additionally add sft_loss, because tokens like '<think>' have not been trained.
+NPROC_PER_NODE=4 \
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift rlhf \
+    --rlhf_type gkd \
+    --model Qwen/Qwen3-8B-Base \
+    --teacher_model $teacher_model \
+    --train_type full \
+    --dataset 'new_dataset.jsonl' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --learning_rate 1e-5 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --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 zero3 \
+    --packing true \
+    --attn_impl flash_attn \
+    --sft_alpha 0.1 \
+    --lmbda 0

examples/train/rlhf/kto.sh

@@ -0,0 +1,29 @@
+nproc_per_node=2
+
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=$nproc_per_node \
+swift rlhf \
+    --rlhf_type kto \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset 'AI-ModelScope/ultrafeedback-binarized-preferences-cleaned-kto#10000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --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 $(expr 16 / $nproc_per_node) \
+    --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 zero2 \
+    --dataset_num_proc 4

examples/train/rlhf/mpo.sh

@@ -0,0 +1,35 @@
+# https://arxiv.org/abs/2411.10442
+# modified from swift/examples/train/multimodal/rlhf/dpo/full.sh
+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 \
+    --loss_type sigmoid bco_pair sft \
+    --loss_weights 0.8 0.2 1.0 \
+    --rpo_alpha 0.0 \
+    --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 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --save_only_model true

examples/train/rlhf/orpo.sh

@@ -0,0 +1,30 @@
+nproc_per_node=2
+
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=$nproc_per_node \
+swift rlhf \
+    --rlhf_type orpo \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset hjh0119/shareAI-Llama3-DPO-zh-en-emoji \
+    --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-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --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 zero2 \
+    --dataset_num_proc 4

examples/train/rlhf/ppo/full.sh

@@ -0,0 +1,37 @@
+# 8 * 65 GiB
+# Currently, it only supports the case where the model and reward_model use the same template/tokenizer.
+# Currently, multimodal model PPO is not supported.
+
+# pip install "deepspeed==0.14.*"
+nproc_per_node=8
+
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+NPROC_PER_NODE=$nproc_per_node \
+swift rlhf \
+    --rlhf_type ppo \
+    --model LLM-Research/Meta-Llama-3.1-8B-Instruct \
+    --reward_model 'AI-ModelScope/Skywork-Reward-Llama-3.1-8B-v0.2' \
+    --train_type full \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#20000' 'AI-ModelScope/alpaca-gpt4-data-en#20000' \
+    --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-6 \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --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 \
+    --response_length 512 \
+    --temperature 0.7 \
+    --dataset_num_proc 4 \
+    --save_only_model true

examples/train/rlhf/ppo/lora.sh

@@ -0,0 +1,40 @@
+# 4 * 50GiB
+# Currently, it only supports the case where the model and reward_model use the same template/tokenizer.
+# Currently, multimodal model PPO is not supported.
+
+# pip install "deepspeed==0.14.*"
+nproc_per_node=4
+
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=$nproc_per_node \
+swift rlhf \
+    --rlhf_type ppo \
+    --model LLM-Research/Meta-Llama-3.1-8B-Instruct \
+    --reward_model 'AI-ModelScope/Skywork-Reward-Llama-3.1-8B-v0.2' \
+    --train_type lora \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#20000' 'AI-ModelScope/alpaca-gpt4-data-en#20000' \
+    --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 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --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 zero2 \
+    --response_length 512 \
+    --temperature 0.7 \
+    --dataset_num_proc 4 \
+    --save_only_model true

examples/train/rlhf/rm.sh

@@ -0,0 +1,30 @@
+nproc_per_node=2
+
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=$nproc_per_node \
+swift rlhf \
+    --rlhf_type rm \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset hjh0119/shareAI-Llama3-DPO-zh-en-emoji \
+    --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-4 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
+    --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 zero2 \
+    --dataset_num_proc 4

examples/train/rlhf/simpo.sh

@@ -0,0 +1,28 @@
+# 2*50GB
+nproc_per_node=2
+
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=$nproc_per_node \
+swift rlhf \
+    --rlhf_type simpo \
+    --model Qwen/Qwen2.5-3B-Instruct \
+    --train_type full \
+    --dataset hjh0119/shareAI-Llama3-DPO-zh-en-emoji \
+    --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 $(expr 16 / $nproc_per_node) \
+    --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 zero2 \
+    --dataset_num_proc 4

examples/train/seq_cls/bert/deploy.sh

@@ -0,0 +1,11 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift deploy \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --served_model_name bert-base-chinese \
+    --truncation_strategy right \
+    --max_length 512
+
+# curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
+# "model": "bert-base-chinese",
+# "messages": [{"role": "user", "content": "包装差，容易被调包。"}]
+# }'

examples/train/seq_cls/bert/infer.sh

@@ -0,0 +1,7 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --load_data_args true \
+    --max_batch_size 16 \
+    --truncation_strategy right \
+    --max_length 512

examples/train/seq_cls/bert/sft.sh

@@ -0,0 +1,30 @@
+# If `num_labels` is provided, it will be considered a classification task,
+# and AutoModelForSequenceClassification will be used to load the model.
+# The BERT model does not require templates, so it can usually be used without registration.
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model AI-ModelScope/bert-base-chinese \
+    --train_type lora \
+    --dataset 'DAMO_NLP/jd:cls#2000' \
+    --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-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 512 \
+    --truncation_strategy right \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --num_labels 2 \
+    --task_type seq_cls

examples/train/seq_cls/multi_label/sft.sh

@@ -0,0 +1,30 @@
+# Custom dataset format reference: https://swift.readthedocs.io/en/latest/Customization/Custom-dataset.html
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-0.5B \
+    --train_type lora \
+    --dataset '<your-dataset>' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 16 \
+    --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 2048 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --num_labels '<num-labels>' \
+    --task_type seq_cls \
+    --use_chat_template false \
+    --problem_type multi_label_classification

examples/train/seq_cls/qwen2_5/deploy.sh

@@ -0,0 +1,8 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift deploy \
+    --adapters output/vx-xxx/checkpoint-xxx
+
+# curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
+# "model": "Qwen2.5-0.5B",
+# "messages": [{"role": "user", "content": "包装差，容易被调包。"}]
+# }'

examples/train/seq_cls/qwen2_5/infer.sh

@@ -0,0 +1,5 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --load_data_args true \
+    --max_batch_size 16

examples/train/seq_cls/qwen2_5/sft.sh

@@ -0,0 +1,30 @@
+# If `num_labels` is provided, it will be considered a classification task,
+# and AutoModelForSequenceClassification will be used to load the model.
+# You can also specify `--model Qwen/Qwen2.5-0.5B-Instruct --use_chat_template true`.
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-0.5B \
+    --train_type lora \
+    --dataset 'DAMO_NLP/jd:cls#2000' \
+    --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-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 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --num_labels 2 \
+    --task_type seq_cls \
+    --use_chat_template false

examples/train/seq_cls/qwen2_5_omni/infer.py

@@ -0,0 +1,21 @@
+import os
+
+from swift.llm import BaseArguments, InferRequest, PtEngine, get_template
+
+os.environ['MAX_PIXELS'] = '1003520'
+
+infer_request = InferRequest(
+    messages=[{
+        'role': 'user',
+        'content': 'Task: Classify household waste.'
+    }], images=['xxx.jpg'])
+adapter_path = 'output/vx-xxx/checkpoint-xxx'
+args = BaseArguments.from_pretrained(adapter_path)
+
+engine = PtEngine(args.model, adapters=[adapter_path], task_type='seq_cls', num_labels=args.num_labels)
+template = get_template(args.template, engine.processor, args.system, use_chat_template=args.use_chat_template)
+engine.default_template = template
+
+resp_list = engine.infer([infer_request])
+response: int = resp_list[0].choices[0].message.content
+print(f'response: {response}')

examples/train/seq_cls/qwen2_5_omni/infer.sh

@@ -0,0 +1,5 @@
+CUDA_VISIBLE_DEVICES=0 \
+MAX_PIXELS=1003520 \
+swift infer \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --load_data_args true

examples/train/seq_cls/qwen2_5_omni/sft.sh

@@ -0,0 +1,28 @@
+CUDA_VISIBLE_DEVICES=0 \
+MAX_PIXELS=1003520 \
+swift sft \
+    --model Qwen/Qwen2.5-Omni-3B \
+    --train_type lora \
+    --dataset 'tany0699/garbage265#20000' \
+    --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-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 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --num_labels 265 \
+    --task_type seq_cls \
+    --use_chat_template true

examples/train/seq_cls/regression/deploy.sh

@@ -0,0 +1,8 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift deploy \
+    --adapters output/vx-xxx/checkpoint-xxx
+
+# curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
+# "model": "Qwen2.5-0.5B",
+# "messages": [{"role": "user", "content": "Task: Based on the given two sentences, provide a similarity score between 0.0 and 1.0.\nSentence 1: The animal is eating.\nSentence 2: A woman is dancing.\nSimilarity score: "}]
+# }'

examples/train/seq_cls/regression/infer.sh

@@ -0,0 +1,5 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --adapters output/vx-xxx/checkpoint-xxx \
+    --load_data_args true \
+    --max_batch_size 16

examples/train/seq_cls/regression/sft.sh

@@ -0,0 +1,30 @@
+# 2GB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-0.5B \
+    --train_type lora \
+    --dataset 'sentence-transformers/stsb:reg#20000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 16 \
+    --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 2048 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --num_labels 1 \
+    --task_type seq_cls \
+    --use_chat_template false \
+    --problem_type regression

examples/train/sequence_parallel/sequence_parallel.sh

@@ -0,0 +1,24 @@
+# Env: 8 * A100
+# Max Length: 65536
+# GPU Memory: 8 * 40GiB, Training Speed 26s/it
+NPROC_PER_NODE=8 \
+CELOSS_PARALLEL_SIZE=2048 \
+swift sft \
+    --model Qwen/Qwen2.5-3B-Instruct \
+    --dataset 'AI-ModelScope/LongAlpaca-12k' \
+    --load_from_cache_file true \
+    --train_type lora \
+    --torch_dtype bfloat16 \
+    --per_device_train_batch_size 4 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps 8 \
+    --save_total_limit 2 \
+    --save_only_model true \
+    --save_steps 50 \
+    --max_length 65536 \
+    --warmup_ratio 0.05 \
+    --attn_impl flash_attn \
+    --sequence_parallel_size 8 \
+    --logging_steps 1 \
+    --use_logits_to_keep false \
+    --padding_free true \

examples/train/sequence_parallel/sequence_parallel_512k.sh

@@ -0,0 +1,32 @@
+# Env: 8 * A100
+# Max Length: 512000
+# GPU Memory: 8 * 80GiB, Training Speed 150s/it
+NPROC_PER_NODE=8 \
+CELOSS_PARALLEL_SIZE=2048 \
+swift sft \
+    --model Qwen/QwQ-32B \
+    --train_type lora \
+    --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 \
+    --rope_scaling yarn \
+    --max_length 512000 \
+    --max_model_len 512000 \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --logging_steps 5 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 8 \
+    --dataset_num_proc 8 \
+    --save_total_limit 2 \
+    --use_liger_kernel true \
+    --save_only_model true \
+    --deepspeed zero3_offload \
+    --attn_impl flash_attn \
+    --sequence_parallel_size 8

examples/train/sequence_parallel/sequence_parallel_dpo.sh

@@ -0,0 +1,29 @@
+# Env: 4 * A100
+# GPU Memory: 4 * 25GiB, Training Speed 14s/it
+NPROC_PER_NODE=4 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift rlhf \
+    --rlhf_type dpo \
+    --model Qwen/Qwen2.5-VL-3B-Instruct \
+    --train_type full \
+    --dataset swift/RLAIF-V-Dataset \
+    --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 8 \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --logging_steps 5 \
+    --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-VL-3B-Instruct \
+    --deepspeed zero3 \
+    --attn_impl flash_attn \
+    --use_liger_kernel true \
+    --sequence_parallel_size 4

examples/train/sequence_parallel/sequence_parallel_grpo.sh

@@ -0,0 +1,44 @@
+NPROC_PER_NODE=8 \
+PYTORCH_CUDA_ALLOC_CONF='' \
+swift rlhf \
+    --rlhf_type grpo \
+    --model Qwen/Qwen2.5-7B \
+    --train_type full \
+    --use_vllm true \
+    --vllm_mode colocate \
+    --vllm_gpu_memory_utilization 0.5 \
+    --vllm_max_model_len 2048 \
+    --vllm_tensor_parallel_size 4 \
+    --dataset AI-MO/NuminaMath-TIR#5000 \
+    --load_from_cache_file true \
+    --torch_dtype bfloat16 \
+    --num_train_epochs 1 \
+    --max_length 2048 \
+    --per_device_train_batch_size 4 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 8 \
+    --eval_steps 1000 \
+    --save_steps 1000 \
+    --learning_rate 1e-6 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --output_dir output \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --max_completion_length 1024 \
+    --reward_funcs accuracy format \
+    --num_generations 32 \
+    --system examples/train/grpo/prompt.txt \
+    --deepspeed zero3 \
+    --temperature 1.0 \
+    --top_p 1.0 \
+    --top_k 80 \
+    --attn_impl flash_attn \
+    --log_completions true \
+    --async_generate false \
+    --offload_optimizer true \
+    --offload_model true \
+    --padding_free true \
+    --sequence_parallel_size 4 \
+    --dataloader_drop_last true \
+    --sleep_level 1

examples/train/streaming/lazy_tokenize.sh

@@ -0,0 +1,19 @@
+# 17GiB
+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 \
+    --lazy_tokenize true \
+    --model_author swift \
+    --model_name swift-robot

examples/train/streaming/streaming.sh

@@ -0,0 +1,17 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --dataset 'swift/self-cognition#1000' \
+    --streaming true \
+    --max_steps 1000 \
+    --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

examples/train/think_model/deepseek_r1.sh

@@ -0,0 +1,29 @@
+# 18GB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model deepseek-ai/DeepSeek-R1-0528-Qwen3-8B \
+    --train_type lora \
+    --dataset 'swift/DeepSeek-R1-Qwen3-8B-Distill#1800' \
+              'swift/self-cognition:empty_think#600' \
+    --loss_scale ignore_empty_think \
+    --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 \
+    --load_from_cache_file false \
+    --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 \
+    --use_liger_kernel true \
+    --model_author swift \
+    --model_name swift-robot

examples/train/think_model/qwen3_demo1.sh

@@ -0,0 +1,31 @@
+# use `--loss_scale ignore_empty_think`
+# Avoid losing the think capability by ignoring the loss of empty `<think>\n\n</think>\n\n`
+# This method is also applicable to the Deepseek-R1 series of models.
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen3-8B \
+    --train_type lora \
+    --dataset 'swift/Qwen3-SFT-Mixin#2000' \
+              'swift/self-cognition:empty_think#600' \
+    --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 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --use_liger_kernel true \
+    --load_from_cache_file false \
+    --loss_scale ignore_empty_think \
+    --model_author swift \
+    --model_name swift-robot

examples/train/think_model/qwen3_demo2.sh

@@ -0,0 +1,30 @@
+# use `swift/self-cognition:qwen3`
+# Avoid losing the thinking capability by appending `/no_think` to the dataset query.
+# https://github.com/modelscope/ms-swift/blob/77985c2ccdac8ed4037174ee222e79d1f1d5059d/swift/llm/dataset/dataset/llm.py#L835
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen3-8B \
+    --train_type lora \
+    --dataset 'swift/Qwen3-SFT-Mixin#2000' \
+              'swift/self-cognition:qwen3#600' \
+    --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 \
+    --warmup_ratio 0.05 \
+    --dataloader_num_workers 4 \
+    --use_liger_kernel true \
+    --load_from_cache_file false \
+    --model_author swift \
+    --model_name swift-robot

examples/train/tuners/adalora/train.sh

@@ -0,0 +1,16 @@
+# 17GiB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type adalora \
+    --dataset 'swift/self-cognition#1000' \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/tuners/adapter/train.sh

@@ -0,0 +1,16 @@
+# 17GiB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type adapter \
+    --dataset 'swift/self-cognition#1000' \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/tuners/boft/train.sh

@@ -0,0 +1,16 @@
+# 17GiB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type boft \
+    --dataset 'swift/self-cognition#1000' \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/tuners/bone/train.sh

@@ -0,0 +1,16 @@
+# 17.3GiB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type bone \
+    --dataset 'swift/self-cognition#1000' \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/tuners/dora/train.sh

@@ -0,0 +1,19 @@
+# 17.2GiB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type lora \
+    --use_dora true \
+    --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

examples/train/tuners/galore/train_galore.sh

@@ -0,0 +1,18 @@
+# 38GiB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type full \
+    --dataset 'swift/self-cognition#1000' \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --learning_rate 1e-5 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot \
+    --use_galore true \
+    --galore_optim_per_parameter true

examples/train/tuners/galore/train_qgalore.sh

@@ -0,0 +1,22 @@
+# 35GiB
+# pip install bitsandbytes==0.40.0
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type full \
+    --torch_dtype bfloat16 \
+    --dataset 'lvjianjin/AdvertiseGen#1000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --learning_rate 1e-5 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot \
+    --use_galore true \
+    --galore_quantization true

examples/train/tuners/lisa/train.sh

@@ -0,0 +1,17 @@
+# 29GiB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type full \
+    --dataset 'swift/self-cognition#1000' \
+    --lisa_activated_layers 2 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --learning_rate 1e-5 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/tuners/llamapro/train.sh

@@ -0,0 +1,17 @@
+# 25.4GiB
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2.5-7B-Instruct \
+    --train_type llamapro \
+    --dataset 'swift/self-cognition#1000' \
+    --llamapro_num_new_blocks 4 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --learning_rate 1e-4 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot

examples/train/tuners/longlora/train.sh

@@ -0,0 +1,18 @@
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model LLM-Research/Meta-Llama-3.1-8B-Instruct \
+    --train_type longlora \
+    --dataset 'AI-ModelScope/LongAlpaca-12k#1000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --num_train_epochs 1 \
+    --learning_rate 1e-4 \
+    --attn_impl flash_attn \
+    --gradient_accumulation_steps 16 \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --max_length 10000 \
+    --save_total_limit 2 \
+    --logging_steps 5

examples/train/tuners/lora-ga/train.sh

@@ -0,0 +1,33 @@
+# Train
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model Qwen/Qwen2-1.5B-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 \
+    --init_weights lora-ga \
+    --lora_ga_batch_size 2 \
+    --lora_ga_iters 2 \
+    --lora_ga_max_length 1024 \
+    --lora_ga_direction ArB2r \
+    --lora_ga_scale stable \
+    --lora_ga_stable_gamma 16 \
+    --gradient_accumulation_steps 16 \
+    --eval_steps 100 \
+    --save_steps 100 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --model_author swift \
+    --model_name swift-robot
+
+# Infer
+# swift infer \
+#     --model Qwen/Qwen2-1.5B-Instruct \
+#     --ckpt_dir ./output/Qwen2-1.5B-Instruct/v0-20241214-191235/checkpoint-62/converted/default \
+#     --infer_backend pt \
+#     --stream true \
+#     --max_new_tokens 2048

examples/train/tuners/lora/train.sh

@@ -0,0 +1,18 @@
+# 17.2GiB
+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