Spaces:
Sleeping
Sleeping
Update configs.py
Browse files- configs.py +79 -0
configs.py
CHANGED
@@ -18,6 +18,85 @@ from dataclasses import dataclass, field
|
|
18 |
from typing import Any, Dict, List, NewType, Optional, Tuple
|
19 |
|
20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
@dataclass
|
22 |
class DataArguments:
|
23 |
"""
|
|
|
18 |
from typing import Any, Dict, List, NewType, Optional, Tuple
|
19 |
|
20 |
|
21 |
+
@dataclass
|
22 |
+
class ModelArguments:
|
23 |
+
"""
|
24 |
+
Arguments pertaining to which model/config/tokenizer we are going to fine-tune.
|
25 |
+
"""
|
26 |
+
|
27 |
+
base_model_revision: Optional[str] = field(
|
28 |
+
default=None,
|
29 |
+
metadata={"help": ("The base model checkpoint for weights initialization with PEFT adatpers.")},
|
30 |
+
)
|
31 |
+
model_name_or_path: Optional[str] = field(
|
32 |
+
default=None,
|
33 |
+
metadata={
|
34 |
+
"help": (
|
35 |
+
"The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
|
36 |
+
)
|
37 |
+
},
|
38 |
+
)
|
39 |
+
model_revision: str = field(
|
40 |
+
default="main",
|
41 |
+
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
|
42 |
+
)
|
43 |
+
model_code_revision: str = field(default=None, metadata={"help": "The branch of the IFT model"})
|
44 |
+
torch_dtype: Optional[str] = field(
|
45 |
+
default=None,
|
46 |
+
metadata={
|
47 |
+
"help": (
|
48 |
+
"Override the default `torch.dtype` and load the model under this dtype. If `auto` is passed, the "
|
49 |
+
"dtype will be automatically derived from the model's weights."
|
50 |
+
),
|
51 |
+
"choices": ["auto", "bfloat16", "float16", "float32"],
|
52 |
+
},
|
53 |
+
)
|
54 |
+
trust_remote_code: bool = field(default=False, metadata={"help": "Trust remote code when loading a model."})
|
55 |
+
use_flash_attention_2: bool = field(
|
56 |
+
default=False,
|
57 |
+
metadata={
|
58 |
+
"help": (
|
59 |
+
"Whether to use flash attention 2. You must install this manually by running `pip install flash-attn --no-build-isolation`"
|
60 |
+
)
|
61 |
+
},
|
62 |
+
)
|
63 |
+
use_peft: bool = field(
|
64 |
+
default=False,
|
65 |
+
metadata={"help": ("Whether to use PEFT or not for training.")},
|
66 |
+
)
|
67 |
+
lora_r: Optional[int] = field(
|
68 |
+
default=16,
|
69 |
+
metadata={"help": ("LoRA R value.")},
|
70 |
+
)
|
71 |
+
lora_alpha: Optional[int] = field(
|
72 |
+
default=32,
|
73 |
+
metadata={"help": ("LoRA alpha.")},
|
74 |
+
)
|
75 |
+
lora_dropout: Optional[float] = field(
|
76 |
+
default=0.05,
|
77 |
+
metadata={"help": ("LoRA dropout.")},
|
78 |
+
)
|
79 |
+
lora_target_modules: Optional[List[str]] = field(
|
80 |
+
default=None,
|
81 |
+
metadata={"help": ("LoRA target modules.")},
|
82 |
+
)
|
83 |
+
lora_modules_to_save: Optional[List[str]] = field(
|
84 |
+
default=None,
|
85 |
+
metadata={"help": ("Model layers to unfreeze & train")},
|
86 |
+
)
|
87 |
+
load_in_8bit: bool = field(default=False, metadata={"help": "use 8 bit precision"})
|
88 |
+
load_in_4bit: bool = field(default=False, metadata={"help": "use 4 bit precision"})
|
89 |
+
|
90 |
+
bnb_4bit_quant_type: Optional[str] = field(
|
91 |
+
default="nf4", metadata={"help": "precise the quantization type (fp4 or nf4)"}
|
92 |
+
)
|
93 |
+
use_bnb_nested_quant: bool = field(default=False, metadata={"help": "use nested quantization"})
|
94 |
+
|
95 |
+
def __post_init__(self):
|
96 |
+
if self.load_in_8bit and self.load_in_4bit:
|
97 |
+
raise ValueError("You can't use 8 bit and 4 bit precision at the same time")
|
98 |
+
|
99 |
+
|
100 |
@dataclass
|
101 |
class DataArguments:
|
102 |
"""
|