fix examples (#2719)

Author: Jintao-Huang

examples/train/fsdp_qlora/train.sh

@@ -1,7 +1,6 @@
 # 14GiB * 2
 nproc_per_node=2
 
-PYTHONPATH=. \
 CUDA_VISIBLE_DEVICES=0,1 \
 accelerate launch --config_file "./examples/train/fsdp_qlora/fsdp_offload.json" \
     swift/cli/sft.py \

examples/train/multi-node/accelerate/train_node1.sh

@@ -10,7 +10,6 @@ accelerate launch --config_file ./examples/train/multi-node/accelerate/multi_nod
     --lora_alpha 32 \
     --learning_rate 1e-4 \
     --gradient_accumulation_steps 16 \
-    --gradient_checkpointing_kwargs '{"use_reentrant": false}' \
     --eval_steps 100 \
     --save_steps 100 \
     --save_total_limit 2 \

examples/train/multi-node/accelerate/train_node2.sh

@@ -10,7 +10,6 @@ accelerate launch --config_file ./examples/train/multi-node/accelerate/multi_nod
     --lora_alpha 32 \
     --learning_rate 1e-4 \
     --gradient_accumulation_steps 16 \
-    --gradient_checkpointing_kwargs '{"use_reentrant": false}' \
     --eval_steps 100 \
     --save_steps 100 \
     --save_total_limit 2 \

examples/train/multi-node/deepspeed/train.sh

@@ -11,7 +11,6 @@ deepspeed --hostfile=./examples/train/multi-node-deepspeed/host.txt \
     --lora_alpha 32 \
     --learning_rate 1e-4 \
     --gradient_accumulation_steps 16 \
-    --gradient_checkpointing_kwargs '{"use_reentrant": false}' \
     --eval_steps 100 \
     --save_steps 100 \
     --save_total_limit 2 \

examples/train/multimodal/dpo.sh

@@ -1,16 +1,18 @@
-# --rlhf_type cpo/orpo/simpo are also supported
+# 4*32GiB
+# You can refer to `https://github.com/QwenLM/Qwen2-VL` for the meaning of the `MAX_PIXELS` parameter.
+# --rlhf_type cpo/orpo/simpo/rm/kto are also supported
 nproc_per_node=4
 
 CUDA_VISIBLE_DEVICES=0,1,2,3 \
 NPROC_PER_NODE=$nproc_per_node \
+MAX_PIXELS=602112 \
 swift rlhf \
     --rlhf_type dpo \
     --model Qwen/Qwen2-VL-7B-Instruct \
     --train_type lora \
     --dataset swift/RLAIF-V-Dataset \
     --num_train_epochs 1 \
     --learning_rate 1e-4 \
-    --gradient_checkpointing_kwargs '{"use_reentrant": false}' \
     --lora_rank 8 \
     --lora_alpha 32 \
     --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \

examples/train/multimodal/train.sh

@@ -1,5 +1,7 @@
 # 22GiB
+# You can refer to `https://github.com/QwenLM/Qwen2-VL` for the meaning of the `MAX_PIXELS` parameter.
 CUDA_VISIBLE_DEVICES=0 \
+MAX_PIXELS=602112 \
 swift sft \
     --model Qwen/Qwen2-VL-7B-Instruct \
     --train_type lora \

examples/train/multimodal/train_grounding.sh

@@ -1,4 +1,6 @@
+# You can refer to `https://github.com/QwenLM/Qwen2-VL` for the meaning of the `MAX_PIXELS` parameter.
 CUDA_VISIBLE_DEVICES=0 \
+MAX_PIXELS=602112 \
 swift sft \
     --model Qwen/Qwen2-VL-7B-Instruct \
     --train_type lora \

examples/train/rlhf/simpo.sh

@@ -10,7 +10,6 @@ swift rlhf \
     --num_train_epochs 1 \
     --learning_rate 1e-5 \
     --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
-    --gradient_checkpointing_kwargs '{"use_reentrant": false}' \
     --warmup_ratio 0.03 \
     --eval_steps 100 \
     --save_steps 100 \