Remove the unreachable prepare_multimodal_messages_vllm

tests/test_data_utils.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import copy
 import textwrap
 from time import strftime
 
@@ -33,7 +32,6 @@
     maybe_unpair_preference_dataset,
     pack_dataset,
     prepare_multimodal_messages,
-    prepare_multimodal_messages_vllm,
     unpair_preference_dataset,
 )
 
@@ -268,98 +266,6 @@ def test_prepared_image_blocks_without_new_images(self):
         assert messages == expected
 
 
-@require_vision
-class TestPrepareMultimodalMessagesVLLM:
-    def test_single_image_conversion(self):
-        messages = [
-            {
-                "role": "user",
-                "content": [
-                    {"type": "image", "image": Image.new("RGB", (10, 10), color="blue")},
-                    {"type": "text", "text": "What color is the sky?"},
-                ],
-            }
-        ]
-
-        result = prepare_multimodal_messages_vllm(messages)
-
-        # Original should remain unchanged (deepcopy test)
-        assert messages[0]["content"][0]["type"] == "image"
-
-        # Converted version should have correct structure
-        assert result[0]["content"][0]["type"] == "image_pil"
-        assert "image_pil" in result[0]["content"][0]
-        assert "image" not in result[0]["content"][0]
-        assert isinstance(result[0]["content"][0]["image_pil"], Image.Image)
-        assert result[0]["content"][1]["type"] == "text"
-
-    def test_mixed_content_conversion(self):
-        messages = [
-            {
-                "role": "user",
-                "content": [
-                    {"type": "text", "text": "What color is the sky?"},
-                    {"type": "image", "image": Image.new("RGB", (10, 10), color="blue")},
-                ],
-            }
-        ]
-
-        result = prepare_multimodal_messages_vllm(messages)
-
-        # The image part should be converted, text should be unchanged
-        assert result[0]["content"][0]["type"] == "text"
-        assert result[0]["content"][1]["type"] == "image_pil"
-
-    def test_no_images(self):
-        messages = [{"role": "user", "content": [{"type": "text", "text": "What color is the sky?"}]}]
-
-        result = prepare_multimodal_messages_vllm(messages)
-
-        # Should be identical since there are no images
-        assert result == messages
-        # And a deepcopy — not the same object
-        assert result is not messages
-        assert result[0] is not messages[0]
-
-    def test_multiple_messages(self):
-        messages = [
-            {
-                "role": "user",
-                "content": [
-                    {"type": "text", "text": "What color is the sky?"},
-                    {"type": "image", "image": Image.new("RGB", (10, 10), color="blue")},
-                ],
-            },
-            {
-                "role": "assistant",
-                "content": [{"type": "text", "text": "It is blue."}],
-            },
-        ]
-
-        result = prepare_multimodal_messages_vllm(messages)
-
-        assert result[0]["content"][1]["type"] == "image_pil"
-        assert result[1]["content"][0]["type"] == "text"
-        assert result[1]["content"][0]["text"] == "It is blue."
-
-    def test_deepcopy_integrity(self):
-        messages = [
-            {
-                "role": "user",
-                "content": [
-                    {"type": "text", "text": "What color is the sky?"},
-                    {"type": "image", "image": Image.new("RGB", (10, 10), color="blue")},
-                ],
-            },
-        ]
-        original = copy.deepcopy(messages)
-
-        _ = prepare_multimodal_messages_vllm(messages)
-
-        # Original should not be mutated
-        assert messages == original
-
-
 class TestIsConversational(TrlTestCase):
     # fmt: off
     conversational_examples = [

trl/__init__.py

@@ -43,7 +43,6 @@
         "maybe_unpair_preference_dataset",
         "pack_dataset",
         "prepare_multimodal_messages",
-        "prepare_multimodal_messages_vllm",
         "unpair_preference_dataset",
     ],
     "models": ["create_reference_model"],
@@ -92,7 +91,6 @@
         maybe_unpair_preference_dataset,
         pack_dataset,
         prepare_multimodal_messages,
-        prepare_multimodal_messages_vllm,
         unpair_preference_dataset,
     )
     from .models import create_reference_model

trl/data_utils.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import copy
 from collections import defaultdict, deque
 from collections.abc import Callable, Sequence
 from itertools import takewhile
@@ -124,39 +123,6 @@ def prepare_multimodal_messages(messages: list[dict[str, Any]], images: list | N
     return new_messages
 
 
-def prepare_multimodal_messages_vllm(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
-    # docstyle-ignore  # because <Image> is not parsable in the code block
-    """
-    Convert structured multimodal messages into a format compatible with vLLM. Replaces `"type": "image"` blocks with
-    `"type": "image_pil"` blocks, and `"image": Image` with `"image_pil": Image`.
-
-    Args:
-        messages (`list[dict[str, Any]]`):
-            Messages with `"role"` and `"content"`. Content is expected to be a list of structured blocks.
-
-    Returns:
-        `list[dict[str, Any]]`:
-            A deep-copied list of messages compatible with vLLM's expected input format.
-
-    Example:
-    ```python
-    # Input
-    [{"role": "user", "content": [{"type": "image", "image": <PIL.Image.Image>}, {"type": "text", "text": "What's in this image?"}]}]
-
-    # Output
-    [{"role": "user", "content": [{"type": "image_pil", "image_pil": <PIL.Image.Image>}, {"type": "text", "text": "What's in this image?"}]}]
-    ```
-    """
-    messages = copy.deepcopy(messages)  # avoid modifying the original messages
-    for message in messages:
-        if isinstance(message["content"], list):
-            for part in message["content"]:
-                if part["type"] == "image":
-                    part["type"] = "image_pil"  # vLLM expects 'image_pil' key for images
-                    part["image_pil"] = part.pop("image")
-    return messages
-
-
 def is_conversational(example: dict[str, Any]) -> bool:
     r"""
     Check if the example is in a conversational format.