Error with gpu-enabled xgboost

[mitchellmanware]

The problem

I am getting an error message when implementing tidymodels functions for xgboost boosted tree model with GPU acceleration. Error reports problem with evals parameter in xgb.train

Reproducible example

Utilizing GPU-enabled xgboost version 2.1.3, accessed from Releases.
Package installation:

# Build and install xgboost with GPU support
wget https://github.com/dmlc/xgboost/releases/download/v2.1.3/xgboost_r_gpu_linux_2.1.3.tar.gz    
Rscript -e "install.packages(c('data.table', 'jsonlite'))"
R CMD INSTALL ./xgboost_r_gpu_linux_2.1.3.tar.gz

Reprex with credit_data example data:

> library(xgboost); library(tibble); library(rsample); library(dplyr); library(purrr); library(parsnip); library(recipes); library(workflows); library(dials); library(tune); library(yardstick); library(tidymodels)
> data(credit_data, package = "modeldata")
> credit <- credit_data |> tibble::as_tibble()
> credit_split <- rsample::initial_split(
          credit, prop = 0.8, strata = "Status"
        )
> train <- rsample::training(credit_split)
        test <- rsample::testing(credit_split)
> cv_split <- rsample::vfold_cv(
          train, v = 5, repeats = 2, strata = "Status"
        )
> scaler <- train |>
          dplyr::count(Status) |>
          dplyr::pull(n) |>
          purrr::reduce(`/`)
> xg_spec <- parsnip::boost_tree(
          mode = "classification",
          trees = parsnip::tune(),
          tree_depth = parsnip::tune(),
          sample_size = parsnip::tune()
        ) |>
          parsnip::set_engine(
            "xgboost",
            params = list(tree_method = "gpu_hist"),
            scale_pos_weight = scaler
          )
> rec_xg_flow <- recipes::recipe(Status ~ ., data = train) |>
          recipes::step_unknown(
            all_nominal_predictors(), new_level = "missing"
          ) |>
          recipes::step_novel(all_nominal_predictors(), new_level = "Unseen") |>
          recipes::step_other(all_nominal_predictors(), other = "Misc") |>
          recipes::step_dummy(all_nominal_predictors(), one_hot = TRUE)
> xg_flow <- workflows::workflow(
          preprocessor = rec_xg_flow, spec = xg_spec
        )
> xg_trees <- dials::trees(range = c(10, 300))
        xg_tree_depth <- dials::tree_depth(range = c(2, 8))
        xg_sample_size <- dials::sample_prop(range = c(0.3, 1.0))
        xg_params <- dials::parameters(xg_trees, xg_tree_depth, xg_sample_size)
> xg_tune <- tune::tune_grid(
          xg_flow,
          resamples = cv_split,
          param_info = xg_params,
          metrics = yardstick::metric_set(
            yardstick::roc_auc, yardstick::accuracy
          ),
          grid = 30,
          control = tune::control_grid(allow_par = TRUE, save_pred = TRUE)
        )
→ A | error:   <text>:1:10: unexpected '<'
               1: evals <- <
                            ^
There were issues with some computations   A: x300
Warning message:
All models failed. Run `show_notes(.Last.tune.result)` for more information.
> tune::show_notes(.Last.tune.result)
unique notes:
───────────────────────────
<text>:1:10: unexpected '<'
1: evals <- <
             ^

Reprex derived from GPU Computing with R presentation (slides 61 - 75)
@sigmafelix @kyle-messier

[EmilHvitfeldt]

Hello @mitchellmanware 👋

This appears to be related to #1227 and/or #1087

[mitchellmanware]

Thanks for the prompt response @EmilHvitfeldt

I thought it was related to those issues as well since the info=list() param has been dropped from xgb.DMatrix.

However, I cloned the parsnip repo and excluded this parameter from xgb.DMatrix and the evals <- - error persists. I am installing this custom version of parsnip in my Apptainer definition file.

# Build and install xgboost with GPU support
wget https://github.com/dmlc/xgboost/releases/download/v2.1.3/xgboost_r_gpu_linux_2.1.3.tar.gz    
Rscript -e "install.packages(c('data.table', 'jsonlite'))"
R CMD INSTALL ./xgboost_r_gpu_linux_2.1.3.tar.gz

# Install R packages (mimicking rocker/ml-verse setup)
Rscript -e "install.packages(c('pak', 'devtools'))"
Rscript -e "devtools::install_version('targets', version = '1.10.0')"
Rscript -e "pak::pak('mitchellmanware/parsnip@beethoven')"
Rscript -e "install.packages(c('tarchetypes', \
    'testthat', 'tidymodels', 'bonsai', 'qs2', \
    'torch', 'tidyverse', 'glmnet'))"
Rscript -e "options(timeout = 1800); torch::install_torch(version = 'latest', type = 'gpu')"
Rscript -e "pak::pak('tidymodels/brulee@gpu')"
Rscript -e "pak::pak('shikokuchuo/nanonext'); pak::pak('shikokuchuo/mirai')"
Rscript -e "pak::pak('wlandau/crew'); pak::pak('wlandau/crew.cluster')"
Rscript -e "pak::pak('NIEHS/beethoven@dev-mm-0108')"

[EmilHvitfeldt]

thank you! that will be helpful. we are working on making sure that parsnip work with CRAN versions of xgboost once it comes out