use ds model for tokenizer

.github/workflows/gpu_test.yml

@@ -28,7 +28,7 @@ jobs:
           pip install -e .[test]
       - name: Running dataset tests
         run: |
-          rm -rf ~/verl-data && git clone --depth 1 https://github.com/eric-haibin-lin/verl-data ~/verl-data
+          [ ! -d "$HOME/verl-data" ] && git clone --depth 1 https://github.com/eric-haibin-lin/verl-data ~/verl-data
           pytest -s -x tests/verl
       - name: Running ray tests that need 2 GPUs
         run: |

examples/sft/gsm8k/run_gemma_2b.sh

@@ -8,7 +8,7 @@ if [ "$#" -lt 2 ]; then
 fi
 
 nproc_per_node=$1
-hdfs_path=$2
+save_path=$2
 
 # Shift the arguments so $@ refers to the rest
 shift 2
@@ -23,7 +23,7 @@ torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \
     +data.response_dict_keys=['answer'] \
     data.micro_batch_size=32 \
     model.partial_pretrain=google/gemma-2b-it \
-    trainer.default_hdfs_dir=$hdfs_path \
+    trainer.default_local_dir=$save_path \
     trainer.project_name=gsm8k-sft \
     trainer.experiment_name=gsm8k-sft-gemma-2b-it \
     trainer.total_epochs=3 \

tests/verl/utils/dataset/test_sft_dataset.py

@@ -28,7 +28,7 @@ def get_gsm8k_data():
 
 
 def test_sft_dataset():
-    tokenizer = AutoTokenizer.from_pretrained('Qwen/Qwen2.5-0.5B-Instruct')
+    tokenizer = AutoTokenizer.from_pretrained('deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct')
     set_pad_token_id(tokenizer)
     local_path = get_gsm8k_data()
     dataset = SFTDataset(parquet_files=local_path,
@@ -42,4 +42,4 @@ def test_sft_dataset():
     data = dataset[0]['input_ids']
     output = tokenizer.batch_decode([data])[0]
     assert len(output) > 1
-    assert type(output) == str
+    assert type(output) == str

verl/trainer/fsdp_sft_trainer.py

@@ -294,7 +294,7 @@ def save_checkpoint(self, step):
 
         path = os.path.join(self.config.trainer.default_local_dir, f'global_step_{step}')
         # save huggingface model
-        if self.device_mesh.get_rank() == 0:
+        if self.device_mesh.get_rank() == 0 and self.config.trainer.default_hdfs_dir:
             os.makedirs(path, exist_ok=True)
             hdfs_io.makedirs(self.config.trainer.default_hdfs_dir, exist_ok=True)
             self.model.save_pretrained(path, state_dict=state_dict)

verl/utils/logger/aggregate_logger.py

@@ -14,7 +14,6 @@
 """
 A Ray logger will receive logging info from different processes.
 """
-
 import numbers
 from typing import Dict
 
@@ -40,4 +39,4 @@ def flush(self):
 
     def log(self, data, step):
         if self.print_to_console:
-            print(concat_dict_to_str(data, step=step), flush=True)
+            print(concat_dict_to_str(data, step=step), flush=True)