Skip to contents

Control TabICL2 inference execution

Source: R/tab_icl2_control.R
tab_icl2_control.Rd

Control TabICL2 inference execution

Usage

tab_icl2_control(
  device = NULL,
  use_amp = TRUE,
  use_fa3 = TRUE,
  verbose = FALSE,
  min_batch_size = 1L,
  safety_factor = 0.8,
  offload = "auto",
  ignore_pretraining_limits = FALSE,
  col_config = list(),
  row_config = list(),
  icl_config = list(),
  ...
)

inference_config(
  device = NULL,
  use_amp = TRUE,
  use_fa3 = TRUE,
  verbose = FALSE,
  min_batch_size = 1L,
  safety_factor = 0.8,
  offload = "auto",
  ignore_pretraining_limits = FALSE,
  col_config = list(),
  row_config = list(),
  icl_config = list(),
  ...
)

Arguments

device

A character string or NULL for the torch device (e.g., "cpu", "cuda", "mps"). Applied to all three transformer stages.

use_amp

A logical. Enable automatic mixed precision. Applied to all stages.

use_fa3

A logical. Enable Flash Attention 3 kernels. Applied to all stages.

verbose

A logical. Enable verbose logging. Applied to all stages.

min_batch_size

A positive integer. Minimum batch size passed to the memory manager. Applied to all stages.

safety_factor

A number in [0, 1]. Memory safety margin for batch size estimation. Applied to all stages.

offload

A logical or one of "auto", "gpu", "cpu", "disk". Offload strategy for the column-embedding (COL) stage. Row-interaction (ROW) and in-context-learning (ICL) stages default to FALSE unless overridden via row_config/icl_config.

ignore_pretraining_limits

A logical. Bypass the default data size limits on training set rows (50,000) and predictors (2,000).

col_config

A named list of additional settings for the COL stage, overriding any of the defaults or parameters above.

row_config

A named list of additional settings for the ROW stage.

icl_config

A named list of additional settings for the ICL stage.

...

Additional named arguments stored in the returned control object. Use this for settings not covered above (e.g. auto_offload_threshold, disk_offload_dir, use_async, etc.) that apply uniformly. Per-stage values must go in col_config, row_config, or icl_config.

Value

A list with class "tab_icl2_control" containing resolved per-stage settings in $col, $row, and $icl, plus $ignore_pretraining_limits.

Examples

tab_icl2_control()
#> inference control for `tab_icl2()`
tab_icl2_control(use_amp = FALSE, offload = "cpu")
#> inference control for `tab_icl2()`
#> 
#> COL stage non-default arguments:
#> • `use_amp`: FALSE
#> • `offload`: "cpu"
#> ROW stage non-default arguments:
#> • `use_amp`: FALSE
#> ICL stage non-default arguments:
#> • `use_amp`: FALSE
tab_icl2_control(col_config = list(disk_offload_dir = tempdir()))
#> inference control for `tab_icl2()`
#> 
#> COL stage non-default arguments:
#> • `disk_offload_dir`:
#>   "/var/folders/df/djsxfhc17x95674wsm_g8s980000gn/T//RtmpaVcmdQ"
tab_icl2_control(ignore_pretraining_limits = TRUE)
#> inference control for `tab_icl2()`
#> 
#> non-default top-level arguments:
#> • `col`: TRUE, TRUE, FALSE, 1, 0.8, auto, 0.5, 0.85, 32768, NULL, 1024, 8192,
#>   TRUE, , NULL, 0.95, TRUE, and 4
#> • `row`: TRUE, TRUE, FALSE, 1, 0.8, FALSE, 0.5, 0.85, 32768, NULL, 1024, 8192,
#>   TRUE, , NULL, 0.95, TRUE, and 4
#> • `icl`: TRUE, TRUE, FALSE, 1, 0.8, FALSE, 0.5, 0.85, 32768, NULL, 1024, 8192,
#>   TRUE, , NULL, 0.95, TRUE, and 4
#> • `ignore_pretraining_limits`: TRUE