File size: 3,840 Bytes
eafbe80 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 |
import math
import warnings
from transformers.configuration_utils import PretrainedConfig
class SambaConfig(PretrainedConfig):
model_type = "samba"
def __init__(
self,
hidden_size: int = 2304,
state_size: int = 16,
num_hidden_layers: int = 18,
norm_eps=1e-5,
pad_token_id: int = 0,
bos_token_id: int = 1,
eos_token_id: int = 2,
expand: int = 2,
conv_kernel: int = 4,
use_bias: bool = False,
use_conv_bias: bool = True,
hidden_act: str = "swish",
initializer_range: float = 0.02,
residual_in_fp32: bool = False,
time_step_rank: str = "auto",
time_step_scale: float = 1.0,
time_step_min: float = 0.001,
time_step_max: float = 0.1,
time_step_init_scheme: str = "random",
time_step_floor: float = 1e-4,
max_position_embeddings: int = 2048,
attn: dict | None = {
'layers': (1, 3, 5, 7, 9, 11, 13, 15, 17),
'num_heads': 18,
'num_kv_heads': 18,
'qkv_bias': False,
'window_size': 2048,
'rope_theta': 10000.,
},
hidden_ratio: int | None = 4,
rescale_prenorm_residual: bool = False,
use_cache: bool = True,
fuse_norm: bool = True,
fuse_swiglu: bool = True,
fuse_cross_entropy: bool = True,
fuse_linear_cross_entropy: bool = False,
use_l2warp: bool = False,
vocab_size: int = 32000,
tie_word_embeddings: bool = False,
**kwargs,
):
self.hidden_size = hidden_size
self.state_size = state_size
self.num_hidden_layers = num_hidden_layers
self.norm_eps = norm_eps
self.conv_kernel = conv_kernel
self.expand = expand
self.intermediate_size = int(expand * self.hidden_size)
self.bos_token_id = bos_token_id
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.use_bias = use_bias
self.use_conv_bias = use_conv_bias
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.time_step_rank = math.ceil(self.hidden_size / 16) if time_step_rank == "auto" else time_step_rank
self.time_step_scale = time_step_scale
self.time_step_min = time_step_min
self.time_step_max = time_step_max
self.time_step_init_scheme = time_step_init_scheme
self.time_step_floor = time_step_floor
self.max_position_embeddings = max_position_embeddings
self.attn = attn
self.hidden_ratio = hidden_ratio
self.rescale_prenorm_residual = rescale_prenorm_residual
self.residual_in_fp32 = residual_in_fp32
self.use_cache = use_cache
self.fuse_norm = fuse_norm
self.fuse_swiglu = fuse_swiglu
self.fuse_cross_entropy = fuse_cross_entropy
self.fuse_linear_cross_entropy = fuse_linear_cross_entropy
self.use_l2warp = use_l2warp
self.vocab_size = vocab_size
if fuse_cross_entropy and fuse_linear_cross_entropy:
raise ValueError(
"`fuse_cross_entropy` and `fuse_linear_cross_entropy` cannot be True at the same time.",
)
if fuse_linear_cross_entropy:
warnings.warn(
"`fuse_linear_cross_entropy` is enabled, which can improves memory efficiency "
"at the potential cost of reduced precision. "
"If you observe issues like loss divergence, consider disabling this setting.",
)
super().__init__(
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
pad_token_id=pad_token_id,
tie_word_embeddings=tie_word_embeddings,
**kwargs,
)
|