@@ -27,7 +27,7 @@ class DataArgument:
27
27
)
28
28
intokens : bool = field (default = False , metadata = {"help" : "Whether to use InTokens data stream" })
29
29
intokens_max_length : int = field (
30
- default = 1024 ,
30
+ default = 2048 ,
31
31
metadata = {"help" : "The max length for InTokens data stream. Only effective when intokens is True" },
32
32
)
33
33
@@ -43,9 +43,6 @@ class ModelArgument:
43
43
lora : bool = field (default = False , metadata = {"help" : "Whether to use LoRA technique" })
44
44
lora_path : str = field (default = None , metadata = {"help" : "Initialize lora state dict." })
45
45
lora_rank : int = field (default = 8 , metadata = {"help" : "Lora attention dimension" })
46
- lora_merge_weights : bool = field (
47
- default = False , metadata = {"help" : "Merge weights of the original model and the Lora model" }
48
- )
49
46
50
47
# prefix tuning related parameters
51
48
prefix_tuning : bool = field (default = False , metadata = {"help" : "Whether to use Prefix technique" })
@@ -54,38 +51,41 @@ class ModelArgument:
54
51
55
52
@dataclass
56
53
class QuantArgument :
57
- quant_type : str = field (default = "A8W8" , metadata = {"help" : "Quantization type. Supported values: A8W8, W4,A8W4" })
54
+ quant_type : str = field (
55
+ default = "A8W8" , metadata = {"help" : "Quantization type. Supported values: A8W8, WINT4,WINT8" }
56
+ )
58
57
59
58
# QAT related parameters
59
+ # Not Yet support
60
60
do_qat : bool = field (default = False , metadata = {"help" : "Whether to use QAT technique" })
61
61
62
- # GPTQ related parameters
63
- do_gptq : bool = field (default = False , metadata = {"help" : "Whether to use GPTQ" })
64
- gptq_step : int = field (default = 8 , metadata = {"help" : "Step for GPTQ" })
65
-
66
62
# PTQ related parameters
67
63
do_ptq : bool = field (default = False , metadata = {"help" : "Whether to use PTQ" })
68
- ptq_step : int = field (default = 8 , metadata = {"help" : "Step for PTQ" })
64
+ ptq_step : int = field (default = 32 , metadata = {"help" : "Step for PTQ" })
69
65
70
66
shift : bool = field (default = False , metadata = {"help" : "Whether to use Shift" })
71
67
shift_all_linears : bool = field (default = False , metadata = {"help" : "Whether to shift all linears" })
72
68
shift_sampler : str = field (
73
69
default = "ema" , metadata = {"help" : "The name of shift sampler, choosen from ['ema', 'none']" }
74
70
)
75
- shift_step : int = field (default = 8 , metadata = {"help" : "Sample steps when shift" })
71
+ shift_step : int = field (default = 32 , metadata = {"help" : "Sample steps when shift" })
76
72
77
73
smooth : bool = field (default = False , metadata = {"help" : "Whether to use Smooth" })
78
74
smooth_all_linears : bool = field (default = False , metadata = {"help" : "Whether to smooth all linears" })
79
75
smooth_sampler : str = field (
80
76
default = "none" , metadata = {"help" : "The name of smooth sampler, choosen from ['multi_step','none']" }
81
77
)
82
- smooth_step : int = field (default = 8 , metadata = {"help" : "Sample steps when smooth" })
78
+ smooth_step : int = field (default = 32 , metadata = {"help" : "Sample steps when smooth" })
83
79
smooth_piecewise_search : bool = field (
84
80
default = False , metadata = {"help" : "The number of piece in piecewise search for smooth strategy." }
85
81
)
86
- smooth_k_piece : int = field (default = 6 , metadata = {"help" : "Number of pieces for K-search" })
82
+ smooth_k_piece : int = field (default = 3 , metadata = {"help" : "Number of pieces for K-search" })
87
83
smooth_search_piece : bool = field (default = False , metadata = {"help" : "Whether search k_piece when piecewise search" })
88
84
85
+ # GPTQ related parameters
86
+ do_gptq : bool = field (default = False , metadata = {"help" : "Whether to use GPTQ" })
87
+ gptq_step : int = field (default = 8 , metadata = {"help" : "Step for GPTQ" })
88
+
89
89
90
90
@dataclass
91
91
class GenerateArgument :
@@ -98,8 +98,3 @@ class GenerateArgument:
98
98
top_p : float = field (
99
99
default = 1.0 , metadata = {"help" : "The cumulative probability for top-p-filtering in the sampling strategy." }
100
100
)
101
- num_beams : int = field (default = 1 , metadata = {"help" : "The number of beams in the beam_search strategy." })
102
- decode_strategy : str = field (default = "sampling" , metadata = {"help" : "The decoding strategy in generation." })
103
- repetition_penalty : float = field (
104
- default = 1.0 , metadata = {"help" : "The parameter for repetition penalty. 1.0 means no penalty." }
105
- )
0 commit comments