Add metadata for wrap_fit when using GPU dataframes

[ChrisJar]

Enables using CREATE MODEL with gpu dataframes and wrap_fit=True. Depends on dask-ml/#862.

[charlesbluca]

@ChrisJar it looks like dask/dask-ml#862 is in - is this ready to review?

[VibhuJawa]

I think this PR is no longer needed as now with this state of dask-ml below should just work as the default now is infer for predict . See doc-string here .

That is below works now:

df = timeseries(freq="1d").reset_index(drop=True).compute()
df = cudf.from_pandas(df)
df['target'] = df['x']*df['y'] > 0

model = cuml.linear_model.LinearRegression()
model.fit(df[['x','y']], df['target'])

dask_model = dask_model = ParallelPostFit(model)
dask_cudf_df = dask_cudf.from_cudf(df, npartitions=2)
dask_model.predict(dask_cudf_df[['x','y']]).compute()

[VibhuJawa]

Can confirm below query with cuML and dask-cuDF now just works. :-D

model_query = '''
CREATE OR REPLACE MODEL my_model WITH (
    model_class = 'cuml.linear_model.LogisticRegression',
    wrap_predict = True,
    wrap_fit = False,
    target_column = 'target'
) AS (
    SELECT x, y, x*y > 0 AS target
    FROM timeseries
)
'''
predict_sql = """
        SELECT * FROM PREDICT(
            MODEL my_model,
            SELECT x, y FROM timeseries
        )
        """
df = timeseries(freq="1d").reset_index(drop=True).compute()
df = cudf.from_pandas(df)
df = dask_cudf.from_cudf(df, npartitions=2)
c.create_table('timeseries',input_table=df)


c.sql(model_query)
c.sql(predict_sql).compute()

[ChrisJar]

Awesome, thanks for letting me know. I'll close this PR