[Tests] Fix oneshot + finetune test by passing splits to oneshot (#1316)

## Purpose ##
* Fix test which was failing because the calibration split was not
explicitly passed to the oneshot call. While normally this would not be
a problem, as `make_dataset_splits` will auto-infer which split is the
calibration split, the `make_dataset_splits` function does its inference
using the split name `train`, whereas ultrachat uses the spilt name
`train_sft`.

## Changes ##
* Pass `splits` argument to oneshot and train calls. This also reduces
the runtime of the test significantly.

---------

Signed-off-by: Kyle Sayers <[email protected]>

Author: kylesayrs

tests/llmcompressor/transformers/finetune/test_oneshot_and_finetune.py

@@ -25,6 +25,7 @@ def _test_oneshot_and_finetune(self):
 
         oneshot_args = dict(
             dataset=self.dataset,
+            splits=splits,
             output_dir=self.output,
             recipe=self.recipe,
             num_calibration_samples=64,
@@ -36,7 +37,6 @@ def _test_oneshot_and_finetune(self):
 
         train_args = dict(
             num_train_epochs=self.num_train_epochs,
-            splits=splits,
             precision="bfloat16",
             bf16=True,
         )