diff --git a/PR_DESCRIPTION.md b/PR_DESCRIPTION.md
deleted file mode 100644
index 1c8ff8f..0000000
--- a/PR_DESCRIPTION.md
+++ /dev/null
@@ -1,9 +0,0 @@
-[bug fix] Require arctic-inference 0.3.0+ for the [rl] extra
-
-## Summary
-
-`[rl]` asks for `arctic-inference[server,vllm]>=0.3.0` and does not declare `vllm`. That extra pulls vLLM and applies that release's pin. `tensordict` lives on `[sft]`. `[testing]` is pytest plugins only. Unit Tests Setup environment is `uv pip install ".[sft,testing]"` after CPU torch, so the CPU runner does not pull arctic-inference.
-
-## Testing
-
-`tests/test_dependency_groups.py`. GitHub Unit Tests Setup environment is unrun on this SHA.
diff --git a/arctic_platform/common/utils/batch.py b/arctic_platform/common/utils/batch.py
index a4110fd..933c390 100644
--- a/arctic_platform/common/utils/batch.py
+++ b/arctic_platform/common/utils/batch.py
@@ -79,7 +79,7 @@ def _split_value(val, num_chunks: int):
                 f"{num_chunks}. The client must send at least one sample per "
                 "DP worker."
             )
-        return list(torch.chunk(val, num_chunks, dim=0))
+        return list(torch.tensor_split(val, num_chunks, dim=0))
     if isinstance(val, list):
         if len(val) < num_chunks:
             raise ValueError(
diff --git a/tests/common/test_split_dict.py b/tests/common/test_split_dict.py
new file mode 100644
index 0000000..0bd012f
--- /dev/null
+++ b/tests/common/test_split_dict.py
@@ -0,0 +1,40 @@
+# Copyright 2025 Snowflake Inc.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""``split_dict`` must return one shard per DP rank for any batch size ``B >= n``."""
+
+from __future__ import annotations
+
+import torch
+
+from arctic_platform.common.utils.batch import split_dict
+from arctic_platform.testing_utils import TestCasePlus
+from arctic_platform.testing_utils import torch_assert_equal
+
+
+class TestSplitDictRemainder(TestCasePlus):
+    def test_tensor_split_returns_one_shard_per_rank(self):
+        # torch.chunk(B, n) can return fewer than n tensors (B=6 n=4 → 3 chunks).
+        for batch_size, num_chunks in ((6, 4), (5, 4), (9, 4), (13, 8), (4, 4), (7, 4)):
+            ids = torch.arange(batch_size * 3).view(batch_size, 3)
+            shards = split_dict({"input_ids": ids}, num_chunks)
+            self.assertEqual(len(shards), num_chunks, msg=f"B={batch_size} n={num_chunks}")
+            rows = [int(s["input_ids"].shape[0]) for s in shards]
+            self.assertEqual(sum(rows), batch_size, msg=f"B={batch_size} n={num_chunks} rows={rows}")
+            torch_assert_equal(torch.cat([s["input_ids"] for s in shards], dim=0), ids)
+
+    def test_batch_smaller_than_ranks_is_rejected(self):
+        ids = torch.arange(6).view(3, 2)
+        with self.assertRaises(ValueError):
+            split_dict({"input_ids": ids}, 4)