[feat, SFT] Support LigerKernel for SFT

.github/workflows/e2e_sft.yml

@@ -54,3 +54,7 @@ jobs:
         run: |
           ray stop --force
           bash tests/sft/run_sft_sp_loss_match.sh
+      - name: Running gsm8k e2e training tests on 8 L20 GPUs with sequence parallism and liger
+        run: |
+          ray stop --force
+          bash tests/sft/run_sft_qwen05_sp2_liger.sh 8 $HOME/ckpts/

examples/sft/gsm8k/run_qwen_05_sp2_liger.sh

@@ -0,0 +1,32 @@
+set -x
+
+if [ "$#" -lt 2 ]; then
+    echo "Usage: run_qwen_05_sp2.sh <nproc_per_node> <save_path> [other_configs...]"
+    exit 1
+fi
+
+nproc_per_node=$1
+save_path=$2
+
+# Shift the arguments so $@ refers to the rest
+shift 2
+
+torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \
+     -m verl.trainer.fsdp_sft_trainer \
+    data.train_files=$HOME/data/gsm8k/train.parquet \
+    data.val_files=$HOME/data/gsm8k/test.parquet \
+    data.prompt_key=extra_info \
+    data.response_key=extra_info \
+    optim.lr=1e-4 \
+    +data.prompt_dict_keys=['question'] \
+    +data.response_dict_keys=['answer'] \
+    data.micro_batch_size=4 \
+    model.partial_pretrain=Qwen/Qwen2.5-0.5B-Instruct \
+    model.use_liger=True \
+    trainer.default_local_dir=$save_path \
+    trainer.project_name=gsm8k-sft \
+    trainer.experiment_name=gsm8k-sft-qwen-2.5-0.5b-instruct-sp2-liger \
+    trainer.logger=['console'] \
+    trainer.default_hdfs_dir=null $@ \
+    ulysses_sequence_parallel_size=2 \
+    use_remove_padding=true

tests/sft/run_sft_qwen05_sp2_liger.sh

@@ -0,0 +1,33 @@
+set -x
+
+if [ "$#" -lt 2 ]; then
+    echo "Usage: run_qwen_05_sp2.sh <nproc_per_node> <save_path> [other_configs...]"
+    exit 1
+fi
+
+nproc_per_node=$1
+save_path=$2
+
+# Shift the arguments so $@ refers to the rest
+shift 2
+
+torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \
+     -m verl.trainer.fsdp_sft_trainer \
+    data.train_files=$HOME/data/gsm8k/train.parquet \
+    data.val_files=$HOME/data/gsm8k/test.parquet \
+    data.prompt_key=extra_info \
+    data.response_key=extra_info \
+    optim.lr=1e-4 \
+    +data.prompt_dict_keys=['question'] \
+    +data.response_dict_keys=['answer'] \
+    data.micro_batch_size=4 \
+    model.partial_pretrain=Qwen/Qwen2.5-0.5B-Instruct \
+    model.use_liger=True \
+    trainer.default_local_dir=$save_path \
+    trainer.project_name=gsm8k-sft \
+    trainer.experiment_name=gsm8k-sft-qwen-2.5-0.5b-instruct-sp2-liger \
+    trainer.logger=['console'] \
+    trainer.total_training_steps=1 \
+    trainer.default_hdfs_dir=null $@ \
+    ulysses_sequence_parallel_size=2 \
+    use_remove_padding=true

verl/trainer/config/sft_trainer.yaml

@@ -23,6 +23,7 @@ model:
   lora_rank: 0  # Set to positive value to enable LoRA (e.g., 32)
   lora_alpha: 16  # LoRA scaling factor
   target_modules: all-linear  # Target modules for LoRA adaptation
+  use_liger: False
 optim:
   lr: 1e-5
   betas: [0.9, 0.95]

verl/trainer/fsdp_sft_trainer.py

@@ -213,6 +213,11 @@ def _build_model_optimizer(self):
                                                                                attn_implementation='flash_attention_2',
                                                                                trust_remote_code=trust_remote_code)
 
+            # Apply Liger kernel if use_liger is enabled
+            if self.config.model.get('use_liger', False):
+                from liger_kernel.transformers.monkey_patch import _apply_liger_kernel_to_instance
+                _apply_liger_kernel_to_instance(model=self.model)
+
             if self.config.model.get('lora_rank', 0) > 0:
                 self.model.enable_input_require_grads()
                 # Convert config to regular Python types before creating PEFT model