LTSMConfig

module.py.LTSMConfig

Configuration dataclass for the LagLlama Transformer architecture.

__init__(self, feature_size: int = 9, block_size: int = 2048, n_layer: int = 32, n_head: int = 32, n_embd_per_head: int = 128, rope_scaling: Optional[dict] = None, dropout: float = 0.0)

Parameter	Type	Default	Description
feature_size	int	9	target + loc + scale + time features
block_size	int	2048	(undocumented)
n_layer	int	32	(undocumented)
n_head	int	32	(undocumented)
n_embd_per_head	int	128	(undocumented)
rope_scaling	Optional[dict]	None	(undocumented)
dropout	float	0.0	(undocumented)

RMSNorm

module.py.RMSNorm

Root Mean Square Layer Normalization.

__init__(self, size: int, dim: int = -1, eps: float = 1e-5) -> None

Parameter	Type	Default	Description
size	int	-	(undocumented)
dim	int	-1	(undocumented)
eps	float	1e-5	(undocumented)

LagLlamaModel

module.py.LagLlamaModel

(No prose description found in class docstring.)

__init__(self, context_length: int, max_context_length: int, scaling: str, input_size: int, n_layer: int, n_embd_per_head: int, n_head: int, lags_seq: List[int], distr_output: DistributionOutput, rope_scaling=None, num_parallel_samples: int = 100, time_feat: bool = True, dropout: float = 0.0) -> None

Parameter	Type	Default	Description
context_length	int	-	(undocumented)
max_context_length	int	-	(undocumented)
scaling	str	-	(undocumented)
input_size	int	-	(undocumented)
n_layer	int	-	(undocumented)
n_embd_per_head	int	-	(undocumented)
n_head	int	-	(undocumented)
lags_seq	List[int]	-	(undocumented)
distr_output	DistributionOutput	-	(undocumented)
rope_scaling	None	None	(undocumented)
num_parallel_samples	int	100	(undocumented)
time_feat	bool	True	(undocumented)
dropout	float	0.0	(undocumented)

prepare_input(self, past_target: torch.Tensor, past_observed_values: torch.Tensor, past_time_feat: Optional[torch.Tensor] = None, future_time_feat: Optional[torch.Tensor] = None, future_target: Optional[torch.Tensor] = None)

(No prose summary found.)

Parameters:

Parameter	Type	Default	Description
past_target	torch.Tensor	-	(undocumented)
past_observed_values	torch.Tensor	-	(undocumented)
past_time_feat	Optional[torch.Tensor]	None	(undocumented)
future_time_feat	Optional[torch.Tensor]	None	(undocumented)
future_target	Optional[torch.Tensor]	None	(undocumented)

forward(self, past_target: torch.Tensor, past_observed_values: torch.Tensor, past_time_feat: Optional[torch.Tensor] = None, future_time_feat: Optional[torch.Tensor] = None, future_target: Optional[torch.Tensor] = None, use_kv_cache: bool = False) -> torch.Tensor

(No prose summary found.)

Parameters:

Parameter	Type	Default	Description
past_target	torch.Tensor	-	(undocumented)
past_observed_values	torch.Tensor	-	(undocumented)
past_time_feat	Optional[torch.Tensor]	None	(undocumented)
future_time_feat	Optional[torch.Tensor]	None	(undocumented)
future_target	Optional[torch.Tensor]	None	(undocumented)
use_kv_cache	bool	False	(undocumented)

Returns: torch.Tensor.

reset_cache(self) -> None

Resets all cached key-values in attention. Has to be called after prediction loop in predictor

Returns: None.