Categorical support in multi GPU training with dask_cudf

[alexHeu]

Hi,

currently I am working on a multi-gpu workflow using dask_cudf together with xgboost.
I am getting the following error:

File /databricks/python/lib/python3.8/site-packages/xgboost/data.py:603, in _cudf_array_interfaces()
601 for i, col in enumerate(data):
602 if is_categorical_dtype(data[col].dtype):
--> 603 codes = cat_codes[i]
604 interface = codes.cuda_array_interface
605 else:

IndexError: list index out of range

Library Versions:
dask_cudf 22.06.00
xgboost 1.6.1

My training code basically looks as follows:

cluster = LocalCUDACluster(n_workers=4)
client = Client(cluster)

df = dask_cudf.read_parquet("path_to_parquet")

for cat_feature in categorical_columns:
    df[cat_feature] = df[cat_feature].astype("category")

df = df.categorize()

X_train = df[df["CalDay"] < validation_start_date][features]
y_train = df[df["CalDay"] < validation_start_date][label]

X_val = df[df["CalDay"] >= validation_start_date][features]
y_val = df[df["CalDay"] >= validation_start_date][label]

params = {
  'tree_method': 'gpu_hist'
}

es = xgb.callback.EarlyStopping(rounds=early_stopping_rounds, save_best=False)

Xy = dxgb.DaskDeviceQuantileDMatrix(client, X_train, y_train, enable_categorical=True)
Xy_valid = dxgb.DaskDMatrix(client, X_val, y_val, enable_categorical=True)

booster = xgb.dask.train(
    client,
    params,
    Xy,
    evals=[(Xy_valid, "Valid")],
    num_boost_round=1000,
    callbacks=[es],
    obj=assym_loss,  # pass the custom objective
    verbose_eval=True
)

Are dask_cudf categoricals currently not supported?

Thanks!

[trivialfis]

Could you please use 1.7? I believe it's fixed by #8280

[alexHeu]

@trivialfis You are right, 1.7 solved this. Thanks!

[trivialfis]

Great news!