Data and request fixes for real data / chat_completions pathways (neuralmagic#131)

## Changes
- Enable config settings for controlling the route (text_completions or
chat_completions) used by the backend through GUIDELLM__PREFERRED_ROUTE
- Fix data files erroring out due to Path objects being passed into HF
load_dataset
- Fix silent bug where a batched request is treated as a single request
leading to issues in the output as well as slow times

Author: markurtz

src/guidellm/backend/openai.py

@@ -202,6 +202,13 @@ async def text_completions(  # type: ignore[override]
             and a ResponseSummary for the final response.
         """
         logger.debug("{} invocation with args: {}", self.__class__.__name__, locals())
+
+        if isinstance(prompt, list):
+            raise ValueError(
+                "List prompts (batching) is currently not supported for "
+                f"text_completions OpenAI pathways. Received: {prompt}"
+            )
+
         headers = self._headers()
         payload = self._completions_payload(
             orig_kwargs=kwargs,

src/guidellm/config.py

@@ -134,6 +134,9 @@ class Settings(BaseSettings):
         Literal["request", "response", "local"]
     ] = "response"
     preferred_backend: Literal["openai"] = "openai"
+    preferred_route: Literal["text_completions", "chat_completions"] = (
+        "text_completions"
+    )
     openai: OpenAISettings = OpenAISettings()
 
     # Output settings

src/guidellm/dataset/file.py

@@ -71,19 +71,21 @@ def load_dataset(
 
             dataset = Dataset.from_dict({"text": items}, **(data_args or {}))
         elif path.suffix.lower() == ".csv":
-            dataset = load_dataset("csv", data_files=path, **(data_args or {}))
+            dataset = load_dataset("csv", data_files=str(path), **(data_args or {}))
         elif path.suffix.lower() in {".json", ".jsonl"}:
-            dataset = load_dataset("json", data_files=path, **(data_args or {}))
+            dataset = load_dataset("json", data_files=str(path), **(data_args or {}))
         elif path.suffix.lower() == ".parquet":
-            dataset = load_dataset("parquet", data_files=path, **(data_args or {}))
+            dataset = load_dataset("parquet", data_files=str(path), **(data_args or {}))
         elif path.suffix.lower() == ".arrow":
-            dataset = load_dataset("arrow", data_files=path, **(data_args or {}))
+            dataset = load_dataset("arrow", data_files=str(path), **(data_args or {}))
         elif path.suffix.lower() == ".hdf5":
-            dataset = Dataset.from_pandas(pd.read_hdf(path), **(data_args or {}))
+            dataset = Dataset.from_pandas(pd.read_hdf(str(path)), **(data_args or {}))
         elif path.suffix.lower() == ".db":
-            dataset = Dataset.from_sql(con=path, **(data_args or {}))
+            dataset = Dataset.from_sql(con=str(path), **(data_args or {}))
         elif path.suffix.lower() == ".tar":
-            dataset = load_dataset("webdataset", data_files=path, **(data_args or {}))
+            dataset = load_dataset(
+                "webdataset", data_files=str(path), **(data_args or {})
+            )
         else:
             raise ValueError(f"Unsupported file type: {path.suffix} given for {path}. ")
 

src/guidellm/request/loader.py

@@ -11,6 +11,7 @@
 from datasets import Dataset, DatasetDict, IterableDataset, IterableDatasetDict
 from transformers import PreTrainedTokenizerBase  # type: ignore[import]
 
+from guidellm.config import settings
 from guidellm.dataset import ColumnInputTypes, load_dataset
 from guidellm.objects import StandardBaseModel
 from guidellm.request.request import GenerationRequest
@@ -61,6 +62,8 @@ class GenerativeRequestLoader(RequestLoader):
         "content",
         "conversation",
         "conversations",
+        "turn",
+        "turns",
         "text",
     ]
 
@@ -270,7 +273,7 @@ def _create_request(self, item: dict[str, Any]) -> GenerationRequest:
         )
 
         return GenerationRequest(
-            request_type="text_completions",
+            request_type=settings.preferred_route,
             content=item[self.column_mappings["prompt_column"]],
             stats=(
                 {"prompt_tokens": prompt_tokens} if prompt_tokens is not None else {}