[trainer] update config (#7174)

Former-commit-id: b4b89b4ff3bc03aa388569e253d62580755a77a5

examples/train_full/llama3_full_sft.yaml

@@ -15,6 +15,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/llama3-8b/full/sft
@@ -22,6 +23,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -32,6 +34,7 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
 # eval_dataset: alpaca_en_demo

examples/train_full/qwen2vl_full_sft.yaml

@@ -20,6 +20,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/qwen2_vl-7b/full/sft
@@ -27,6 +28,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -37,6 +39,7 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
 # val_size: 0.1

examples/train_lora/llama3_lora_dpo.yaml

@@ -18,6 +18,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/llama3-8b/lora/dpo
@@ -25,6 +26,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -35,6 +37,7 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
 # eval_dataset: dpo_en_demo

examples/train_lora/llama3_lora_pretrain.yaml

@@ -15,6 +15,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/llama3-8b/lora/pretrain
@@ -22,6 +23,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -32,6 +34,7 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
 # eval_dataset: c4_demo

examples/train_lora/llama3_lora_reward.yaml

@@ -16,6 +16,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/llama3-8b/lora/reward
@@ -23,6 +24,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -33,6 +35,7 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
 # eval_dataset: dpo_en_demo

examples/train_lora/llama3_lora_sft.yaml

@@ -16,6 +16,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -23,6 +24,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -33,6 +35,7 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
 # eval_dataset: alpaca_en_demo

examples/train_lora/llama3_lora_sft_ds3.yaml

@@ -17,6 +17,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/llama3-8b/lora/sft
@@ -24,6 +25,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -34,8 +36,10 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
+# eval_dataset: alpaca_en_demo
 # val_size: 0.1
 # per_device_eval_batch_size: 1
 # eval_strategy: steps

examples/train_lora/llama3_lora_sft_ray.yaml

@@ -17,6 +17,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: tmp_dir
@@ -24,9 +25,11 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### ray
 ray_run_name: llama3_8b_sft_lora
+ray_storage_path: ./saves
 ray_num_workers: 4  # number of GPUs to use
 resources_per_worker:
   GPU: 1
@@ -41,8 +44,10 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
+# eval_dataset: alpaca_en_demo
 # val_size: 0.1
 # per_device_eval_batch_size: 1
 # eval_strategy: steps

examples/train_lora/llava1_5_lora_sft.yaml

@@ -16,6 +16,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/llava1_5-7b/lora/sft
@@ -23,6 +24,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -33,6 +35,7 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
 # val_size: 0.1

examples/train_lora/qwen2vl_lora_dpo.yaml

@@ -20,6 +20,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/qwen2_vl-7b/lora/dpo
@@ -27,6 +28,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -37,6 +39,7 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
 # val_size: 0.1

examples/train_lora/qwen2vl_lora_sft.yaml

@@ -18,6 +18,7 @@ cutoff_len: 2048
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
+dataloader_num_workers: 4
 
 ### output
 output_dir: saves/qwen2_vl-7b/lora/sft
@@ -25,6 +26,7 @@ logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
+save_only_model: false
 
 ### train
 per_device_train_batch_size: 1
@@ -35,6 +37,7 @@ lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
 ddp_timeout: 180000000
+resume_from_checkpoint: null
 
 ### eval
 # val_size: 0.1

src/llamafactory/data/mm_plugin.py

@@ -521,9 +521,7 @@ class MiniCPMVPlugin(BasePlugin):
         processor: Optional["ProcessorMixin"],
     ) -> List[Dict[str, str]]:
         self._validate_input(processor, images, videos, audios)
-        num_image_tokens = 0
+        num_image_tokens, num_video_tokens, num_audio_tokens = 0, 0, 0
-        num_video_tokens = 0
-        num_audio_tokens = 0
         messages = deepcopy(messages)
         image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
         mm_inputs = {}
@@ -1038,7 +1036,7 @@ class Qwen2AudioPlugin(BasePlugin):
 
 
 @dataclass
-class Qwen2vlPlugin(BasePlugin):
+class Qwen2VLPlugin(BasePlugin):
     @override
     def _preprocess_image(self, image: "ImageObject", **kwargs) -> "ImageObject":
         image = super()._preprocess_image(image, **kwargs)
@@ -1124,7 +1122,10 @@ class Qwen2vlPlugin(BasePlugin):
         processor: Optional["ProcessorMixin"],
     ) -> List[Dict[str, str]]:
         self._validate_input(processor, images, videos, audios)
+        num_image_tokens, num_video_tokens = 0, 0
+        messages = deepcopy(messages)
         image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
+
         merge_length: int = getattr(image_processor, "merge_size") ** 2
         if self.expand_mm_tokens:
             mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
@@ -1134,8 +1135,6 @@ class Qwen2vlPlugin(BasePlugin):
             image_grid_thw = [None] * len(images)
             video_grid_thw = [None] * len(videos)
 
-        num_image_tokens, num_video_tokens = 0, 0
-        messages = deepcopy(messages)
         for message in messages:
             content = message["content"]
             while IMAGE_PLACEHOLDER in content:
@@ -1273,7 +1272,7 @@ PLUGINS = {
     "paligemma": PaliGemmaPlugin,
     "pixtral": PixtralPlugin,
     "qwen2_audio": Qwen2AudioPlugin,
-    "qwen2_vl": Qwen2vlPlugin,
+    "qwen2_vl": Qwen2VLPlugin,
     "video_llava": VideoLlavaPlugin,
 }
 

src/llamafactory/train/tuner.py

@@ -17,6 +17,7 @@ import shutil
 from typing import TYPE_CHECKING, Any, Dict, List, Optional
 
 import torch
+import torch.distributed as dist
 from transformers import PreTrainedModel
 
 from ..data import get_template_and_fix_tokenizer
@@ -76,6 +77,12 @@ def _training_function(config: Dict[str, Any]) -> None:
     else:
         raise ValueError(f"Unknown task: {finetuning_args.stage}.")
 
+    try:
+        if dist.is_initialized():
+            dist.destroy_process_group()
+    except Exception as e:
+        logger.warning(f"Failed to destroy process group: {e}.")
+
 
 def run_exp(args: Optional[Dict[str, Any]] = None, callbacks: Optional[List["TrainerCallback"]] = None) -> None:
     args = read_args(args)