chore: import upstream snapshot with attribution

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 \
+    --tuner_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 \
+    --tuner_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 +1,30 @@
+# 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 \
+    --tuner_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 \
+    --rpo_alpha 0.1 \
+    --sequence_parallel_size 4

examples/train/sequence_parallel/sequence_parallel_emb.sh

@@ -0,0 +1,28 @@
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+INFONCE_TEMPERATURE=0.1 \
+INFONCE_MASK_FAKE_NEGATIVE=true \
+INFONCE_INCLUDE_QQ=true \
+INFONCE_INCLUDE_DD=false \
+NPROC_PER_NODE=4 \
+swift sft \
+    --model Qwen/Qwen3-Embedding-0.6B \
+    --task_type embedding \
+    --model_type qwen3_emb \
+    --tuner_type full \
+    --dataset sentence-transformers/stsb \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.05 \
+    --eval_strategy steps \
+    --output_dir output \
+    --save_steps 50 \
+    --eval_steps 50 \
+    --num_train_epochs 5 \
+    --per_device_train_batch_size 256 \
+    --per_device_eval_batch_size 256 \
+    --gradient_accumulation_steps 1 \
+    --learning_rate 6e-6 \
+    --loss_type infonce \
+    --dataloader_drop_last true \
+    --sequence_parallel_size 4 \
+    --padding_free true \
+    --attn_impl flash_attn

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 \
+    --tuner_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/sequence_parallel/sequence_parallel_qwen3_5.sh

@@ -0,0 +1,35 @@
+# Env: 8 * H800
+# Max Length: 65536
+# GPU Memory: 8 * 80GiB, Training Speed 18.97s/it
+NPROC_PER_NODE=8 \
+CELOSS_PARALLEL_SIZE=2048 \
+swift sft \
+    --model Qwen/Qwen3.5-4B \
+    --dataset 'AI-ModelScope/LongAlpaca-12k' \
+    --load_from_cache_file true \
+    --tuner_type lora \
+    --torch_dtype bfloat16 \
+    --per_device_train_batch_size 2 \
+    --target_modules all-linear \
+    --gradient_accumulation_steps 8 \
+    --save_total_limit 2 \
+    --save_only_model true \
+    --save_steps 50 \
+    --max_length 65535 \
+    --warmup_ratio 0.05 \
+    --attn_impl flash_attn \
+    --logging_steps 1 \
+    --use_logits_to_keep false \
+    --sequence_parallel_size 4 \
+    --padding_free true
+# Train:   1%|          | 1/189 [02:41<8:25:06, 161.21s/it]
+
+# {'loss': '1.461', 'grad_norm': '1.705', 'learning_rate': '1e-05', 'token_acc': '0.6399', 'epoch': '0.016', 'global_step/max_steps': '1/189', 'elapsed_time': '2m 41s', 'remaining_time': '8h 25m 20s', 'memory(GiB)': '67.95', 'train_speed(s/it)': '161.3'}
+# {'loss': '1.484', 'grad_norm': '1.666', 'learning_rate': '2e-05', 'token_acc': '0.6324', 'epoch': '0.032', 'global_step/max_steps': '2/189', 'elapsed_time': '3m 55s', 'remaining_time': '6h 6m 35s', 'memory(GiB)': '68.05', 'train_speed(s/it)': '117.6'}
+# {'loss': '1.477', 'grad_norm': '1.767', 'learning_rate': '3e-05', 'token_acc': '0.6311', 'epoch': '0.048', 'global_step/max_steps': '3/189', 'elapsed_time': '4m 48s', 'remaining_time': '4h 57m 21s', 'memory(GiB)': '68.05', 'train_speed(s/it)': '95.92'}
+# {'loss': '1.498', 'grad_norm': '1.683', 'learning_rate': '4e-05', 'token_acc': '0.6355', 'epoch': '0.064', 'global_step/max_steps': '4/189', 'elapsed_time': '5m 36s', 'remaining_time': '4h 18m 56s', 'memory(GiB)': '68.05', 'train_speed(s/it)': '83.98'}
+# {'loss': '1.422', 'grad_norm': '1.528', 'learning_rate': '5e-05', 'token_acc': '0.6329', 'epoch': '0.08', 'global_step/max_steps': '5/189', 'elapsed_time': '6m 17s', 'remaining_time': '3h 51m 17s', 'memory(GiB)': '68.05', 'train_speed(s/it)': '75.42'}
+# {'loss': '1.322', 'grad_norm': '1.203', 'learning_rate': '6e-05', 'token_acc': '0.6457', 'epoch': '0.096', 'global_step/max_steps': '6/189', 'elapsed_time': '6m 53s', 'remaining_time': '3h 30m 1s', 'memory(GiB)': '68.05', 'train_speed(s/it)': '68.86'}
+# ...
+# Train:  24%|██▍       | 46/189 [17:53<45:13, 18.97s/it]
+# {'loss': '1.049', 'grad_norm': '0.6095', 'learning_rate': '9.035e-05', 'token_acc': '0.7018', 'epoch': '0.736', 'global_step/max_steps': '46/189', 'elapsed_time': '17m 54s', 'remaining_time': '55m 38s', 'memory(GiB)': '68.06', 'train_speed(s/it)': '23.34'}

examples/train/sequence_parallel/sequence_parallel_reranker.sh

@@ -0,0 +1,25 @@
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=4 \
+swift sft \
+    --model Qwen/Qwen3-Reranker-0.6B \
+    --task_type generative_reranker \
+    --loss_type pointwise_reranker \
+    --tuner_type full \
+    --dataset MTEB/scidocs-reranking \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.05 \
+    --eval_strategy steps \
+    --output_dir output \
+    --eval_steps 100 \
+    --num_train_epochs 1 \
+    --save_steps 200 \
+    --per_device_train_batch_size 2 \
+    --per_device_eval_batch_size 2 \
+    --gradient_accumulation_steps 8 \
+    --dataset_num_proc 8 \
+    --learning_rate 6e-6 \
+    --label_names labels \
+    --dataloader_drop_last true \
+    --sequence_parallel_size 4 \
+    --padding_free true \
+    --attn_impl flash_attn

examples/train/sequence_parallel/sequence_parallel_seq_cls.sh

@@ -0,0 +1,31 @@
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=2 \
+swift sft \
+    --model Qwen/Qwen2.5-0.5B \
+    --tuner_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 \
+    --sequence_parallel_size 2 \
+    --padding_free true \
+    --attn_impl flash_attn