Merge remote-tracking branch 'origin/main' into if_eval

llama_stack/providers/inline/datasetio/localfs/datasetio.py

@@ -35,12 +35,12 @@ class PandasDataframeDataset:
         else:
             return self.df.iloc[idx].to_dict()
 
-    def load(self) -> None:
+    async def load(self) -> None:
         if self.df is not None:
             return
 
         if self.dataset_def.source.type == "uri":
-            self.df = get_dataframe_from_uri(self.dataset_def.source.uri)
+            self.df = await get_dataframe_from_uri(self.dataset_def.source.uri)
         elif self.dataset_def.source.type == "rows":
             self.df = pandas.DataFrame(self.dataset_def.source.rows)
         else:
@@ -95,7 +95,7 @@ class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
     ) -> IterrowsResponse:
         dataset_def = self.dataset_infos[dataset_id]
         dataset_impl = PandasDataframeDataset(dataset_def)
-        dataset_impl.load()
+        await dataset_impl.load()
 
         start_index = start_index or 0
 
@@ -114,7 +114,7 @@ class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
     async def append_rows(self, dataset_id: str, rows: List[Dict[str, Any]]) -> None:
         dataset_def = self.dataset_infos[dataset_id]
         dataset_impl = PandasDataframeDataset(dataset_def)
-        dataset_impl.load()
+        await dataset_impl.load()
 
         new_rows_df = pandas.DataFrame(rows)
         dataset_impl.df = pandas.concat([dataset_impl.df, new_rows_df], ignore_index=True)

llama_stack/providers/inline/inference/vllm/vllm.py

@@ -582,6 +582,7 @@ class VLLMInferenceImpl(Inference, ModelsProtocolPrivate):
                     tool_name=t.function.name,
                     # vLLM function args come back as a string. Llama Stack expects JSON.
                     arguments=json.loads(t.function.arguments),
+                    arguments_json=t.function.arguments,
                 )
                 for t in vllm_message.tool_calls
             ],

llama_stack/providers/remote/inference/sambanova/sambanova.py

@@ -42,9 +42,7 @@ from llama_stack.models.llama.datatypes import (
     TopKSamplingStrategy,
     TopPSamplingStrategy,
 )
-from llama_stack.providers.utils.inference.model_registry import (
-    ModelRegistryHelper,
-)
+from llama_stack.providers.utils.inference.model_registry import ModelRegistryHelper
 from llama_stack.providers.utils.inference.openai_compat import (
     process_chat_completion_stream_response,
 )
@@ -293,14 +291,12 @@ class SambaNovaInferenceAdapter(ModelRegistryHelper, Inference):
         if not tool_calls:
             return []
 
-        for call in tool_calls:
-            call_function_arguments = json.loads(call.function.arguments)
-
         compitable_tool_calls = [
             ToolCall(
                 call_id=call.id,
                 tool_name=call.function.name,
-                arguments=call_function_arguments,
+                arguments=json.loads(call.function.arguments),
+                arguments_json=call.function.arguments,
             )
             for call in tool_calls
         ]

llama_stack/providers/remote/inference/vllm/vllm.py

@@ -90,15 +90,12 @@ def _convert_to_vllm_tool_calls_in_response(
     if not tool_calls:
         return []
 
-    call_function_arguments = None
-    for call in tool_calls:
-        call_function_arguments = json.loads(call.function.arguments)
-
     return [
         ToolCall(
             call_id=call.id,
             tool_name=call.function.name,
-            arguments=call_function_arguments,
+            arguments=json.loads(call.function.arguments),
+            arguments_json=call.function.arguments,
         )
         for call in tool_calls
     ]
@@ -183,6 +180,7 @@ async def _process_vllm_chat_completion_stream_response(
                                 call_id=tool_call_buf.call_id,
                                 tool_name=tool_call_buf.tool_name,
                                 arguments=args,
+                                arguments_json=args_str,
                             ),
                             parse_status=ToolCallParseStatus.succeeded,
                         ),

llama_stack/providers/utils/datasetio/url_utils.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+import asyncio
 import base64
 import io
 from urllib.parse import unquote
@@ -13,12 +14,15 @@ import pandas
 from llama_stack.providers.utils.memory.vector_store import parse_data_url
 
 
-def get_dataframe_from_uri(uri: str):
+async def get_dataframe_from_uri(uri: str):
     df = None
     if uri.endswith(".csv"):
-        df = pandas.read_csv(uri)
+        # Moving to its own thread to avoid io from blocking the eventloop
+        # This isn't ideal as it moves more then just the IO to a new thread
+        # but it is as close as we can easly get
+        df = await asyncio.to_thread(pandas.read_csv, uri)
     elif uri.endswith(".xlsx"):
-        df = pandas.read_excel(uri)
+        df = await asyncio.to_thread(pandas.read_excel, uri)
     elif uri.startswith("data:"):
         parts = parse_data_url(uri)
         data = parts["data"]

llama_stack/providers/utils/inference/openai_compat.py

@@ -529,7 +529,11 @@ async def convert_message_to_openai_dict_new(
     ) -> Union[str, Iterable[OpenAIChatCompletionContentPartParam]]:
         async def impl(
             content_: InterleavedContent,
-        ) -> Union[str, OpenAIChatCompletionContentPartParam, List[OpenAIChatCompletionContentPartParam]]:
+        ) -> Union[
+            str,
+            OpenAIChatCompletionContentPartParam,
+            List[OpenAIChatCompletionContentPartParam],
+        ]:
             # Llama Stack and OpenAI spec match for str and text input
             if isinstance(content_, str):
                 return content_
@@ -570,7 +574,7 @@ async def convert_message_to_openai_dict_new(
                 OpenAIChatCompletionMessageToolCall(
                     id=tool.call_id,
                     function=OpenAIFunction(
-                        name=tool.tool_name if not isinstance(tool.tool_name, BuiltinTool) else tool.tool_name.value,
+                        name=(tool.tool_name if not isinstance(tool.tool_name, BuiltinTool) else tool.tool_name.value),
                         arguments=json.dumps(tool.arguments),
                     ),
                     type="function",
@@ -609,6 +613,7 @@ def convert_tool_call(
             call_id=tool_call.id,
             tool_name=tool_call.function.name,
             arguments=json.loads(tool_call.function.arguments),
+            arguments_json=tool_call.function.arguments,
         )
     except Exception:
         return UnparseableToolCall(
@@ -759,6 +764,7 @@ def _convert_openai_tool_calls(
             call_id=call.id,
             tool_name=call.function.name,
             arguments=json.loads(call.function.arguments),
+            arguments_json=call.function.arguments,
         )
         for call in tool_calls
     ]
@@ -890,7 +896,8 @@ async def convert_openai_chat_completion_stream(
             # ChatCompletionResponseEvent only supports one per stream
             if len(choice.delta.tool_calls) > 1:
                 warnings.warn(
-                    "multiple tool calls found in a single delta, using the first, ignoring the rest", stacklevel=2
+                    "multiple tool calls found in a single delta, using the first, ignoring the rest",
+                    stacklevel=2,
                 )
 
             if not enable_incremental_tool_calls:
@@ -971,6 +978,7 @@ async def convert_openai_chat_completion_stream(
                     call_id=buffer["call_id"],
                     tool_name=buffer["name"],
                     arguments=arguments,
+                    arguments_json=buffer["arguments"],
                 )
                 yield ChatCompletionResponseStreamChunk(
                     event=ChatCompletionResponseEvent(