[v1] add models & accelerator (#9579)

src/llamafactory/v1/config/data_args.py

@@ -23,10 +23,6 @@ class DataArguments:
         default=None,
         metadata={"help": "Path to the dataset."},
     )
-    dataset_dir: str = field(
-        default="data",
-        metadata={"help": "Path to the folder containing the datasets."},
-    )
     cutoff_len: int = field(
         default=2048,
         metadata={"help": "Cutoff length for the dataset."},

src/llamafactory/v1/config/model_args.py

@@ -25,3 +25,11 @@ class ModelArguments:
         default=False,
         metadata={"help": "Trust remote code from Hugging Face."},
     )
+    use_fast_processor: bool = field(
+        default=True,
+        metadata={"help": "Use fast processor from Hugging Face."},
+    )
+    auto_model_class: str = field(
+        default="causallm",
+        metadata={"help": "Model class from Hugging Face."},
+    )

src/llamafactory/v1/config/sample_args.py

@@ -14,10 +14,20 @@
 
 
 from dataclasses import dataclass, field
+from enum import Enum
+
+
+class SampleBackend(Enum):
+    HF = "hf"
+    VLLM = "vllm"
 
 
 @dataclass
 class SampleArguments:
+    sample_backend: SampleBackend = field(
+        default=SampleBackend.HF,
+        metadata={"help": "Sampling backend, default to 'hf'."},
+    )
     max_new_tokens: int = field(
         default=128,
         metadata={"help": "Maximum number of new tokens to generate."},