update default_lr; fix do_sample in vllm (#336)

Author: Jintao-Huang

docs/source/LLM/命令行参数.md

@@ -55,7 +55,7 @@
 - `--num_train_epochs`: 训练的epoch数, 默认为`1`. 如果`max_steps >= 0`, 则覆盖`num_train_epochs`.
 - `--max_steps`: 训练的max_steps数, 默认为`-1`. 如果`max_steps >= 0`, 则覆盖`num_train_epochs`.
 - `--optim`: 默认为`'adamw_torch'`.
-- `--learning_rate`: 默认值为`None`, 即如果`sft_type`为lora, 则设置为1e-4, 如果`sft_type`为full, 则设置为2e-5.
+- `--learning_rate`: 默认值为`None`, 即如果`sft_type`为lora, 则设置为1e-4, 如果`sft_type`为full, 则设置为1e-5.
 - `--weight_decay`: 默认值为`0.01`.
 - `--gradient_accumulation_steps`: 梯度累加, 默认值为`None`, 设置为`math.ceil(16 / self.batch_size / world_size)`. `total_batch_size =  batch_size * gradient_accumulation_steps * world_size`.
 - `--max_grad_norm`: 梯度裁剪, 默认值为`0.5`.

examples/pytorch/llm/scripts/qwen_1_8b_chat/full/sft.sh

@@ -17,7 +17,7 @@ python llm_sft.py \
     --gradient_checkpointing true \
     --batch_size 1 \
     --weight_decay 0.01 \
-    --learning_rate 2e-5 \
+    --learning_rate 1e-5 \
     --gradient_accumulation_steps 16 \
     --max_grad_norm 0.5 \
     --warmup_ratio 0.03 \

examples/pytorch/llm/scripts/qwen_1_8b_chat/full_ddp/sft.sh

@@ -24,7 +24,7 @@ torchrun \
     --gradient_checkpointing true \
     --batch_size 1 \
     --weight_decay 0.01 \
-    --learning_rate 2e-5 \
+    --learning_rate 1e-5 \
     --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
     --max_grad_norm 0.5 \
     --warmup_ratio 0.03 \

examples/pytorch/llm/scripts/qwen_7b_chat/full/sft.sh

@@ -9,7 +9,7 @@ swift sft \
     --output_dir output \
     --num_train_epochs 1 \
     --max_length 4096 \
-    --learning_rate 2e-5 \
+    --learning_rate 1e-5 \
     --use_flash_attn true \
     --save_only_model true \
     --dataset codefuse-evol-instruction-zh \

examples/pytorch/llm/scripts/qwen_7b_chat/full_freeze_ddp/sft.sh

@@ -10,7 +10,7 @@ swift sft \
     --output_dir output \
     --num_train_epochs 1 \
     --max_length 4096 \
-    --learning_rate 2e-5 \
+    --learning_rate 1e-5 \
     --use_flash_attn true \
     --save_only_model true \
     --dataset codefuse-evol-instruction-zh \

examples/pytorch/llm/scripts/qwen_7b_chat/full_mp/sft.sh

@@ -17,7 +17,7 @@ python llm_sft.py \
     --gradient_checkpointing true \
     --batch_size 1 \
     --weight_decay 0.01 \
-    --learning_rate 2e-5 \
+    --learning_rate 1e-5 \
     --gradient_accumulation_steps 16 \
     --max_grad_norm 0.5 \
     --warmup_ratio 0.03 \

examples/pytorch/llm/scripts/qwen_7b_chat/full_mp_ddp/sft.sh

@@ -22,7 +22,7 @@ torchrun \
     --gradient_checkpointing true \
     --batch_size 1 \
     --weight_decay 0.01 \
-    --learning_rate 2e-5 \
+    --learning_rate 1e-5 \
     --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
     --max_grad_norm 0.5 \
     --warmup_ratio 0.03 \

examples/pytorch/llm/scripts/qwen_audio_chat/full_mp/sft.sh

@@ -9,7 +9,7 @@ swift sft \
     --output_dir output \
     --num_train_epochs 1 \
     --max_length 2048 \
-    --learning_rate 2e-5 \
+    --learning_rate 1e-5 \
     --use_flash_attn true \
     --save_only_model true \
     --dataset aishell1-mini-zh \

examples/pytorch/llm/scripts/qwen_audio_chat/full_mp_ddp/sft.sh

@@ -10,7 +10,7 @@ swift sft \
     --output_dir output \
     --num_train_epochs 1 \
     --max_length 2048 \
-    --learning_rate 2e-5 \
+    --learning_rate 1e-5 \
     --use_flash_attn true \
     --save_only_model true \
     --dataset aishell1-mini-zh \

examples/pytorch/llm/scripts/qwen_vl_chat/full_mp/sft.sh

@@ -9,7 +9,7 @@ swift sft \
     --output_dir output \
     --num_train_epochs 1 \
     --max_length 2048 \
-    --learning_rate 2e-5 \
+    --learning_rate 1e-5 \
     --use_flash_attn true \
     --save_only_model true \
     --dataset coco-mini-en \