diff --git a/litellm/litellm_core_utils/prompt_templates/image_handling.py b/litellm/litellm_core_utils/prompt_templates/image_handling.py index 81132fa89a5..1890d4eb682 100644 --- a/litellm/litellm_core_utils/prompt_templates/image_handling.py +++ b/litellm/litellm_core_utils/prompt_templates/image_handling.py @@ -2,7 +2,11 @@ Helper functions to handle images passed in messages """ +import asyncio import base64 +from collections.abc import Mapping +from dataclasses import dataclass +from types import MappingProxyType from typing import Final from httpx import Response @@ -12,6 +16,7 @@ from litellm import verbose_logger from litellm.caching.caching import InMemoryCache from litellm.constants import MAX_IMAGE_URL_DOWNLOAD_SIZE_MB from litellm.litellm_core_utils.url_utils import async_safe_get, safe_get +from litellm.types.llms.openai import AllMessageValues MAX_IMGS_IN_MEMORY: Final = 10 @@ -124,3 +129,101 @@ def convert_url_to_base64(url: str) -> str: raise litellm.ImageFetchError( f"Error: Unable to fetch image from URL after 3 attempts. url={url}", ) + + +_REMOTE_URL_PREFIXES: Final = ("http://", "https://") + + +@dataclass(frozen=True, slots=True) +class _RemoteImage: + part: Mapping[str, object] + image_url: Mapping[str, object] | None + url: str + + +@dataclass(frozen=True, slots=True) +class _RemoteFile: + part: Mapping[str, object] + file: Mapping[str, object] + url: str + + +def _as_mapping(value: object) -> Mapping[str, object] | None: + return value if isinstance(value, Mapping) else None # pyright: ignore[reportUnknownVariableType] # fields are parsed one by one + + +def _remote_url(candidate: object) -> str | None: + return candidate if isinstance(candidate, str) and candidate.startswith(_REMOTE_URL_PREFIXES) else None + + +def _parse_remote_part(part: object) -> _RemoteImage | _RemoteFile | None: + fields: Final = _as_mapping(part) + if fields is None: + return None + if fields.get("type") == "image_url": + image_url: Final = fields.get("image_url") + image_url_fields: Final = _as_mapping(image_url) + url: Final = _remote_url(image_url_fields.get("url") if image_url_fields is not None else image_url) + return _RemoteImage(fields, image_url_fields, url) if url is not None else None + file: Final = _as_mapping(fields.get("file")) if fields.get("type") == "file" else None + file_url: Final = _remote_url(file.get("file_id")) if file is not None else None + return _RemoteFile(fields, file, file_url) if file is not None and file_url is not None else None + + +_PDF_FORMAT: Final = MappingProxyType({"format": "application/pdf"}) + + +def _inferred_format(file: Mapping[str, object], url: str) -> Mapping[str, str]: + return _PDF_FORMAT if "format" not in file and url.lower().endswith(".pdf") else MappingProxyType({}) + + +def _inlined_image_url(image_url: Mapping[str, object] | None, data_url: str) -> Mapping[str, object] | str: + return {**image_url, "url": data_url} if image_url is not None else data_url # mutable-ok: json-serialized part + + +def _inlined_file(file: Mapping[str, object], url: str, data_url: str) -> Mapping[str, object]: + kept: Final = {k: v for k, v in file.items() if k != "file_id"} # mutable-ok: json-serialized message part + return {**kept, **_inferred_format(file, url), "file_data": data_url} # mutable-ok: json-serialized part + + +def _inline(remote: _RemoteImage | _RemoteFile, data_url: str) -> Mapping[str, object]: + match remote: + case _RemoteImage(part, image_url, _): + return {**part, "image_url": _inlined_image_url(image_url, data_url)} # mutable-ok: json-serialized part + case _RemoteFile(part, file, url): + return {**part, "file": _inlined_file(file, url, data_url)} # mutable-ok: json-serialized message part + + +def _content_parts(message: Mapping[str, object]) -> tuple[object, ...]: + content: Final = message.get("content") + return tuple(content) if isinstance(content, list) else () # pyright: ignore[reportUnknownVariableType, reportUnknownArgumentType] # parts are parsed one by one + + +def _inline_message(message: AllMessageValues, data_urls: Mapping[str, str]) -> AllMessageValues: + parts: Final = _content_parts(message) + if not parts: + return message + inlined_parts: Final = [ # mutable-ok: content must stay a list for the transforms' isinstance checks + _inline(remote, data_urls[remote.url]) if (remote := _parse_remote_part(part)) is not None else part + for part in parts + ] + inlined_message: Final = {**message, "content": inlined_parts} # mutable-ok: json-serialized message + return inlined_message # pyright: ignore[reportReturnType] # the same message with its remote parts inlined + + +async def async_inline_remote_media( + messages: list[AllMessageValues], # mutable-ok: every transform_request takes list[AllMessageValues] +) -> list[AllMessageValues]: # mutable-ok: every transform_request takes list[AllMessageValues] + remote_urls: Final = tuple( + dict.fromkeys( + remote.url + for message in messages + for part in _content_parts(message) + if (remote := _parse_remote_part(part)) is not None + ) + ) + if not remote_urls: + return messages + data_urls: Final = await asyncio.gather(*(async_convert_url_to_base64(url) for url in remote_urls)) + inlined: Final = MappingProxyType(dict(zip(remote_urls, data_urls, strict=True))) + return [_inline_message(message, inlined) for message in messages] # mutable-ok: transform_request takes a list diff --git a/litellm/llms/base_llm/chat/transformation.py b/litellm/llms/base_llm/chat/transformation.py index bbe1cc85df1..7bfc87a30d6 100644 --- a/litellm/llms/base_llm/chat/transformation.py +++ b/litellm/llms/base_llm/chat/transformation.py @@ -411,6 +411,10 @@ class BaseConfig(ABC): def has_custom_stream_wrapper(self) -> bool: return False + @property + def uses_async_transform_request(self) -> bool: + return False + @property def supports_stream_param_in_request_body(self) -> bool: """ diff --git a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py index 67720451c00..38f280eef03 100644 --- a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py +++ b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py @@ -8,7 +8,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import ( convert_to_anthropic_image_obj, ) from litellm.litellm_core_utils.prompt_templates.image_handling import ( - async_convert_url_to_base64, + async_inline_remote_media, convert_url_to_base64, ) from litellm.llms.anthropic.chat.transformation import AnthropicConfig @@ -172,6 +172,10 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig): return _anthropic_request + @property + def uses_async_transform_request(self) -> bool: + return True + async def async_transform_request( self, model: str, @@ -180,26 +184,14 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig): litellm_params: dict, headers: dict, ) -> dict: - _anthropic_request: Final = self._build_bedrock_anthropic_request_base( + return self.transform_request( model=model, - messages=messages, + messages=await async_inline_remote_media(messages), optional_params=optional_params, litellm_params=litellm_params, headers=headers, ) - await self._async_convert_document_url_sources_to_base64(_anthropic_request) - beta_list: Final = self._compute_bedrock_invoke_beta_headers( - model=model, - messages=messages, - optional_params=optional_params, - headers=headers, - ) - if beta_list: - _anthropic_request["anthropic_beta"] = beta_list - - return _anthropic_request - def _build_bedrock_anthropic_request_base( self, model: str, @@ -321,45 +313,6 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig): "data": image_chunk["data"], } - async def _async_convert_document_url_sources_to_base64(self, anthropic_request: dict) -> None: - """ - Async version of document URL conversion for async completion paths. - """ - messages: Final = anthropic_request.get("messages") - if not isinstance(messages, list): - return - - for message in messages: - if not isinstance(message, dict): - continue - content = message.get("content") - if not isinstance(content, list): - continue - - for block in content: - if not isinstance(block, dict) or block.get("type") != "document": - continue - source = block.get("source") - if not isinstance(source, dict) or source.get("type") != "url": - continue - source_url = source.get("url") - if not isinstance(source_url, str): - continue - - inferred_format: str | None = None - if source_url.lower().endswith(".pdf"): - inferred_format = "application/pdf" - base64_url = await async_convert_url_to_base64(url=source_url) - image_chunk = convert_to_anthropic_image_obj( - openai_image_url=base64_url, - format=inferred_format, - ) - block["source"] = { - "type": "base64", - "media_type": image_chunk["media_type"], - "data": image_chunk["data"], - } - def _normalize_bedrock_tool_search_tools(self, optional_params: dict) -> dict: """ Convert tool search entries to the format supported by the Bedrock Invoke API. diff --git a/litellm/llms/bedrock/chat/mantle/transformation.py b/litellm/llms/bedrock/chat/mantle/transformation.py index 31fc079c0c9..7e2037c33f1 100644 --- a/litellm/llms/bedrock/chat/mantle/transformation.py +++ b/litellm/llms/bedrock/chat/mantle/transformation.py @@ -10,6 +10,7 @@ at a different endpoint (bedrock-mantle.{region}.api.aws) with AWS SigV4 auth. from collections.abc import AsyncIterator, Iterator from typing import TYPE_CHECKING, Any, Final +from litellm.litellm_core_utils.prompt_templates.image_handling import async_inline_remote_media from litellm.llms.bedrock.chat.invoke_transformations.anthropic_claude3_transformation import ( AmazonAnthropicClaudeConfig, ) @@ -110,21 +111,13 @@ class AmazonMantleConfig(AmazonAnthropicClaudeConfig): litellm_params: dict, headers: dict, ) -> dict: - model_id: Final = model.replace("mantle/", "", 1) - - request: Final = self._build_bedrock_anthropic_request_base( - model=model_id, - messages=messages, + return self.transform_request( + model=model, + messages=await async_inline_remote_media(messages), optional_params=optional_params, litellm_params=litellm_params, headers=headers, ) - await self._async_convert_document_url_sources_to_base64(request) - return self._restore_mantle_body_fields( - request=request, - model_id=model_id, - optional_params=optional_params, - ) @staticmethod def _restore_mantle_body_fields(request: dict, model_id: str, optional_params: dict) -> dict: diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index f281c249c72..54c21f8d5b3 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -116,6 +116,7 @@ from litellm.types.llms.anthropic_skills import ( Skill, ) from litellm.types.llms.openai import ( + AllMessageValues, CreateBatchRequest, CreateFileRequest, FileContentRequest, @@ -488,7 +489,7 @@ class BaseLLMHTTPHandler: def completion( self, model: str, - messages: list, + messages: list[AllMessageValues], api_base: str | None, custom_llm_provider: str, model_response: ModelResponse, @@ -507,7 +508,7 @@ class BaseLLMHTTPHandler: shared_session: Optional["ClientSession"] = None, ): json_mode: Final[bool] = optional_params.pop("json_mode", False) - extra_body: Final[dict | None] = optional_params.pop("extra_body", None) + extra_body: Final[Mapping[str, object] | None] = optional_params.pop("extra_body", None) provider_config = provider_config or ProviderConfigManager.get_provider_chat_config( model=model, provider=litellm.LlmProviders(custom_llm_provider) @@ -522,14 +523,17 @@ class BaseLLMHTTPHandler: ) # get config from model, custom llm provider - headers = provider_config.validate_environment( - api_key=api_key, - headers=headers or {}, - model=model, - messages=messages, - optional_params=optional_params, - api_base=api_base, - litellm_params=litellm_params, + request_headers: Final = cast( # cast-ok: validate_environment is declared as a bare dict + "dict[str, object]", + provider_config.validate_environment( + api_key=api_key, + headers=headers or {}, + model=model, + messages=messages, + optional_params=optional_params, + api_base=api_base, + litellm_params=litellm_params, + ), ) api_base = provider_config.get_complete_url( @@ -541,93 +545,117 @@ class BaseLLMHTTPHandler: litellm_params=litellm_params, ) - data: dict[str, object] = provider_config.transform_request( - model=model, - messages=messages, - optional_params=optional_params, - litellm_params=litellm_params, - headers=headers, - ) - - if extra_body is not None: - data = {**data, **extra_body} - - headers, signed_json_body = provider_config.sign_request( - headers=headers, - optional_params={ - **optional_params, - **_aws_signing_overrides(optional_params, litellm_params), - }, - request_data=data, - api_base=api_base, - api_key=api_key, - stream=stream, - fake_stream=fake_stream, - model=model, - ) - - ## LOGGING - logging_obj.pre_call( - input=messages, - api_key=api_key, - additional_args={ - "complete_input_dict": data, - "api_base": api_base, - "headers": headers, - }, - ) - - # Check if stream was converted for WebSearch interception - # This is set by the async_pre_request_hook in WebSearchInterceptionLogger - if litellm_params.get("_websearch_interception_converted_stream", False): - logging_obj.model_call_details["websearch_interception_converted_stream"] = True - - if acompletion is True: - if stream is True: - data = self._add_stream_param_to_request_body( - data=data, - provider_config=provider_config, + def sign_and_log( + transformed: dict[str, object], # mutable-ok: async_completion takes dict + ) -> tuple[dict[str, object], dict[str, object], bytes | None]: # mutable-ok: async_completion takes dict + data: Final = {**transformed, **extra_body} if extra_body is not None else transformed + signed: Final = cast( # cast-ok: sign_request is declared as a bare dict + "tuple[dict[str, object], bytes | None]", + provider_config.sign_request( + headers=request_headers, + optional_params={ + **optional_params, + **_aws_signing_overrides(optional_params, litellm_params), + }, + request_data=data, + api_base=api_base, + api_key=api_key, + stream=stream, fake_stream=fake_stream, - ) + model=model, + ), + ) + logging_obj.pre_call( + input=messages, + api_key=api_key, + additional_args={ + "complete_input_dict": data, + "api_base": api_base, + "headers": signed[0], + }, + ) + if litellm_params.get("_websearch_interception_converted_stream", False): + logging_obj.model_call_details["websearch_interception_converted_stream"] = True + return data, signed[0], signed[1] + + def dispatch_async( + data: dict[str, object], # mutable-ok: async_completion takes dict + signed_headers: dict[str, object], # mutable-ok: async_completion takes dict + signed_json_body: bytes | None, + ): + async_client: Final = client if isinstance(client, AsyncHTTPHandler) else None + if stream is True: return self.acompletion_stream_function( model=model, messages=messages, api_base=api_base, - headers=headers, + headers=signed_headers, custom_llm_provider=custom_llm_provider, provider_config=provider_config, timeout=timeout, logging_obj=logging_obj, - data=data, + data=self._add_stream_param_to_request_body( + data=data, + provider_config=provider_config, + fake_stream=fake_stream, + ), fake_stream=fake_stream, - client=(client if client is not None and isinstance(client, AsyncHTTPHandler) else None), + client=async_client, litellm_params=litellm_params, json_mode=json_mode, optional_params=optional_params, signed_json_body=signed_json_body, ) + return self.async_completion( + custom_llm_provider=custom_llm_provider, + provider_config=provider_config, + api_base=api_base, + headers=signed_headers, + data=data, + timeout=timeout, + model=model, + model_response=model_response, + logging_obj=logging_obj, + api_key=api_key, + messages=messages, + optional_params=optional_params, + litellm_params=litellm_params, + encoding=encoding, + client=async_client, + json_mode=json_mode, + signed_json_body=signed_json_body, + shared_session=shared_session, + ) - else: - return self.async_completion( - custom_llm_provider=custom_llm_provider, - provider_config=provider_config, - api_base=api_base, - headers=headers, - data=data, - timeout=timeout, - model=model, - model_response=model_response, - logging_obj=logging_obj, - api_key=api_key, - messages=messages, - optional_params=optional_params, - litellm_params=litellm_params, - encoding=encoding, - client=(client if client is not None and isinstance(client, AsyncHTTPHandler) else None), - json_mode=json_mode, - signed_json_body=signed_json_body, - shared_session=shared_session, + if acompletion is True and provider_config.uses_async_transform_request: + + async def transform_then_dispatch(): + transformed: Final = cast( # cast-ok: async_transform_request is declared as a bare dict + "dict[str, object]", + await provider_config.async_transform_request( + model=model, + messages=messages, + optional_params=optional_params, + litellm_params=litellm_params, + headers=request_headers, + ), ) + return await dispatch_async(*sign_and_log(transformed)) + + return transform_then_dispatch() + + data, signed_headers, signed_json_body = sign_and_log( + provider_config.transform_request( + model=model, + messages=messages, + optional_params=optional_params, + litellm_params=litellm_params, + headers=request_headers, + ) + ) + + if acompletion is True: + return dispatch_async(data, signed_headers, signed_json_body) if stream is True: data = self._add_stream_param_to_request_body( @@ -641,7 +669,7 @@ class BaseLLMHTTPHandler: custom_llm_provider=custom_llm_provider, logging_obj=logging_obj, api_base=api_base, - headers=headers, + headers=signed_headers, data=data, signed_json_body=signed_json_body, messages=messages, @@ -651,7 +679,7 @@ class BaseLLMHTTPHandler: completion_stream, headers = self.make_sync_call( provider_config=provider_config, api_base=api_base, - headers=headers, + headers=signed_headers, data=data, signed_json_body=signed_json_body, original_data=data, @@ -684,7 +712,7 @@ class BaseLLMHTTPHandler: sync_httpx_client=sync_httpx_client, provider_config=provider_config, api_base=api_base, - headers=headers, + headers=signed_headers, data=data, signed_json_body=signed_json_body, timeout=timeout, diff --git a/litellm/llms/snowflake/chat/transformation.py b/litellm/llms/snowflake/chat/transformation.py index c64fc583edc..f65b0876202 100644 --- a/litellm/llms/snowflake/chat/transformation.py +++ b/litellm/llms/snowflake/chat/transformation.py @@ -23,6 +23,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import ( create_anthropic_image_param, select_anthropic_content_block_type_for_file, ) +from litellm.litellm_core_utils.prompt_templates.image_handling import async_inline_remote_media from litellm.llms.anthropic.chat.handler import ModelResponseIterator as AnthropicStreamParser from litellm.llms.anthropic.chat.transformation import AnthropicConfig from litellm.llms.anthropic.common_utils import normalize_cache_control_in_anthropic_payload @@ -421,6 +422,21 @@ class SnowflakeConfig(SnowflakeBaseConfig, OpenAIGPTConfig): return self._transform_request_anthropic(model, messages, optional_params, stream, extra_body) return self._transform_request_openai(model, messages, optional_params, stream, extra_body) + @property + def uses_async_transform_request(self) -> bool: + return True + + async def async_transform_request( + self, + model: str, + messages: list[AllMessageValues], # mutable-ok: BaseConfig signature + optional_params: dict[str, object], # mutable-ok: BaseConfig signature + litellm_params: dict[str, object], # mutable-ok: BaseConfig signature + headers: dict[str, object], # mutable-ok: BaseConfig signature + ) -> dict[str, object]: # mutable-ok: BaseConfig signature + inlined_messages: Final = await async_inline_remote_media(messages) if _is_claude_model(model) else messages + return self.transform_request(model, inlined_messages, optional_params, litellm_params, headers) + def _transform_request_openai( self, model: str, diff --git a/litellm/llms/vertex_ai/gemini/transformation.py b/litellm/llms/vertex_ai/gemini/transformation.py index e2d62be6a69..c9480f07150 100644 --- a/litellm/llms/vertex_ai/gemini/transformation.py +++ b/litellm/llms/vertex_ai/gemini/transformation.py @@ -27,6 +27,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import ( convert_to_gemini_tool_call_result, response_schema_prompt, ) +from litellm.litellm_core_utils.prompt_templates.image_handling import async_inline_remote_media from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler from litellm.llms.vertex_ai.common_utils import pop_vertex_request_labels from litellm.types.files import ( @@ -1348,13 +1349,15 @@ async def async_transform_request_body( vertex_auth_header=vertex_auth_header, ) - if _openai_messages_may_need_sync_gcs_metadata_fetch(messages): + inlined_messages: Final = await async_inline_remote_media(messages) if custom_llm_provider == "gemini" else messages + + if _openai_messages_may_need_sync_gcs_metadata_fetch(inlined_messages): # _transform_request_body may issue a sync httpx.get (up to 5s timeout) # via _get_gcs_object_content_type to fetch GCS object metadata. Run the # whole sync transformation on a worker thread so it does not block the # async event loop. return await asyncify(_transform_request_body)( - messages=messages, + messages=inlined_messages, model=model, custom_llm_provider=custom_llm_provider, litellm_params=litellm_params, @@ -1363,7 +1366,7 @@ async def async_transform_request_body( ) return _transform_request_body( - messages=messages, + messages=inlined_messages, model=model, custom_llm_provider=custom_llm_provider, litellm_params=litellm_params, diff --git a/litellm/main.py b/litellm/main.py index 2929790f2bd..9970c203e03 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -4474,7 +4474,7 @@ def _complete_snowflake(ctx: _CompletionDispatchContext) -> _CompletionDispatchR acompletion: Final = ctx.acompletion api_base: Final = ctx.api_base api_key: Final = ctx.api_key - client = _dispatch_client_http(ctx) + injected_client: Final = _dispatch_client_http(ctx) custom_llm_provider: Final = ctx.custom_llm_provider headers: Final = ctx.headers litellm_params: Final = ctx.litellm_params @@ -4486,11 +4486,11 @@ def _complete_snowflake(ctx: _CompletionDispatchContext) -> _CompletionDispatchR shared_session: Final = ctx.shared_session stream: Final = ctx.stream timeout: Final = ctx.timeout + client: Final = ( + injected_client if injected_client is not None else (HTTPHandler(timeout=timeout) if stream is False else None) + ) # Keep this here, otherwise, the httpx.client closes and streaming is impossible try: - client = ( - HTTPHandler(timeout=timeout) if stream is False else None - ) # Keep this here, otherwise, the httpx.client closes and streaming is impossible response: Final = base_llm_http_handler.completion( model=model, messages=messages, diff --git a/tests/test_litellm/conftest.py b/tests/test_litellm/conftest.py index 62c95cb100b..aa8ba168ecc 100644 --- a/tests/test_litellm/conftest.py +++ b/tests/test_litellm/conftest.py @@ -7,9 +7,12 @@ # 4. Added proper cleanup in fixtures # 5. Added worker-specific isolation for parallel execution +import base64 import importlib import os from pathlib import Path +from types import SimpleNamespace +import httpx import pytest import asyncio @@ -595,3 +598,35 @@ def pytest_sessionfinish(session, exitstatus): _close_handler_if_needed(getattr(litellm, "aclient", None)) _close_handler_if_needed(getattr(litellm, "client", None)) _run_coroutine_if_needed(close_litellm_async_clients()) + + +ONE_PIXEL_PNG = base64.b64decode( + "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNkYPhfDwAChwGA60e6kgAAAABJRU5ErkJggg==" +) + + +@pytest.fixture +def async_only_image_fetch(monkeypatch): + from litellm.litellm_core_utils.prompt_templates import image_handling + + fetch = SimpleNamespace( + fetched=[], + base64_png=base64.b64encode(ONE_PIXEL_PNG).decode(), + data_url="data:image/png;base64," + base64.b64encode(ONE_PIXEL_PNG).decode(), + ) + + def forbid_sync_fetch(client, url, **kwargs): + raise litellm.ImageFetchError(f"sync image fetch ran on the event loop: {url}") + + async def serve_png(client, url, **kwargs): + fetch.fetched.append(url) + return httpx.Response( + 200, + content=ONE_PIXEL_PNG, + headers={"content-type": "image/png"}, + request=httpx.Request("GET", url), + ) + + monkeypatch.setattr(image_handling, "safe_get", forbid_sync_fetch) + monkeypatch.setattr(image_handling, "async_safe_get", serve_png) + return fetch diff --git a/tests/test_litellm/litellm_core_utils/test_image_handling.py b/tests/test_litellm/litellm_core_utils/test_image_handling.py index 893472d63ae..5b6d403fa21 100644 --- a/tests/test_litellm/litellm_core_utils/test_image_handling.py +++ b/tests/test_litellm/litellm_core_utils/test_image_handling.py @@ -1,3 +1,5 @@ +import copy +import uuid from unittest.mock import patch import pytest @@ -8,6 +10,7 @@ from litellm import constants from litellm.litellm_core_utils.prompt_templates import image_handling from litellm.litellm_core_utils.prompt_templates.image_handling import ( async_convert_url_to_base64, + async_inline_remote_media, convert_url_to_base64, ) @@ -268,3 +271,62 @@ def test_image_size_limit_disabled(monkeypatch): assert "Image URL download is disabled" in str(excinfo.value) assert "MAX_IMAGE_URL_DOWNLOAD_SIZE_MB=0" in str(excinfo.value) + + +async def test_async_inline_remote_media_inlines_every_remote_part_shape(async_only_image_fetch): + image_url = f"http://img.example/{uuid.uuid4()}.png" + pdf_url = f"http://docs.example/{uuid.uuid4()}.pdf" + messages = [ + {"role": "system", "content": "be terse"}, + { + "role": "user", + "content": [ + {"type": "text", "text": "what is this?"}, + {"type": "image_url", "image_url": {"url": image_url, "detail": "low"}}, + {"type": "image_url", "image_url": image_url}, + {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo="}}, + {"type": "file", "file": {"file_id": pdf_url}}, + {"type": "file", "file": {"file_id": image_url, "format": "image/png"}}, + ], + }, + ] + snapshot = copy.deepcopy(messages) + + inlined = await async_inline_remote_media(messages) + + data_url = async_only_image_fetch.data_url + assert inlined[0] == {"role": "system", "content": "be terse"} + assert inlined[1]["content"] == [ + {"type": "text", "text": "what is this?"}, + {"type": "image_url", "image_url": {"url": data_url, "detail": "low"}}, + {"type": "image_url", "image_url": data_url}, + {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo="}}, + {"type": "file", "file": {"format": "application/pdf", "file_data": data_url}}, + {"type": "file", "file": {"format": "image/png", "file_data": data_url}}, + ] + assert sorted(async_only_image_fetch.fetched) == sorted([image_url, pdf_url]) + assert messages == snapshot + + +async def test_async_inline_remote_media_leaves_messages_without_remote_parts_alone(async_only_image_fetch): + messages = [ + {"role": "user", "content": [{"type": "text", "text": "hi"}]}, + { + "role": "user", + "content": [{"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo="}}], + }, + ] + + assert await async_inline_remote_media(messages) is messages + assert async_only_image_fetch.fetched == [] + + +async def test_async_inline_remote_media_raises_image_fetch_error_when_the_fetch_fails(monkeypatch): + async def serve_404(client, url, **kwargs): + return Response(404, request=Request("GET", url)) + + monkeypatch.setattr(image_handling, "async_safe_get", serve_404) + url = f"http://img.example/{uuid.uuid4()}.png" + + with pytest.raises(litellm.ImageFetchError, match="Status code: 404"): + await async_inline_remote_media([{"role": "user", "content": [{"type": "image_url", "image_url": url}]}]) diff --git a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py index 0d7573a2536..e808a087e3d 100644 --- a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py +++ b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py @@ -1,15 +1,19 @@ import asyncio import json +import uuid from unittest.mock import patch +import httpx import pytest # Ensure the project root is on the import path so `litellm` can be imported when # tests are executed from any working directory. +import litellm from litellm.llms.bedrock.chat.invoke_transformations.anthropic_claude3_transformation import ( AmazonAnthropicClaudeConfig, ) +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler def test_get_supported_params_thinking(): @@ -714,3 +718,48 @@ def test_bedrock_chat_invoke_response_format_stub_still_upgrades_legacy_thinking assert result["thinking"] == {"type": "adaptive"} assert result["output_config"] == {"effort": "high"} + + +async def test_bedrock_invoke_claude_async_completion_inlines_remote_images_off_the_event_loop(async_only_image_fetch): + image_url = f"http://img.example/{uuid.uuid4()}.png" + captured = {} + + def handle(request): + captured["body"] = request.content.decode() + return httpx.Response( + 200, + json={ + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "us.anthropic.claude-sonnet-5", + "content": [{"type": "text", "text": "Green"}], + "stop_reason": "end_turn", + "usage": {"input_tokens": 1, "output_tokens": 1}, + }, + ) + + client = AsyncHTTPHandler() + client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle)) + + response = await litellm.acompletion( + model="bedrock/invoke/us.anthropic.claude-sonnet-5", + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "What colour is this?"}, + {"type": "image_url", "image_url": {"url": image_url}}, + ], + } + ], + aws_access_key_id="AKIAEXAMPLE", + aws_secret_access_key="fake-secret", + aws_region_name="us-east-1", + client=client, + ) + + assert response.choices[0].message.content == "Green" + assert async_only_image_fetch.fetched == [image_url] + assert image_url not in captured["body"] + assert async_only_image_fetch.base64_png in captured["body"] diff --git a/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py b/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py new file mode 100644 index 00000000000..9b491d305c7 --- /dev/null +++ b/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py @@ -0,0 +1,51 @@ +import uuid + +import httpx + +import litellm +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler + + +async def test_bedrock_mantle_claude_async_completion_inlines_remote_images_off_the_event_loop(async_only_image_fetch): + image_url = f"http://img.example/{uuid.uuid4()}.png" + captured = {} + + def handle(request): + captured["body"] = request.content.decode() + return httpx.Response( + 200, + json={ + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "us.anthropic.claude-sonnet-5", + "content": [{"type": "text", "text": "Green"}], + "stop_reason": "end_turn", + "usage": {"input_tokens": 1, "output_tokens": 1}, + }, + ) + + client = AsyncHTTPHandler() + client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle)) + + response = await litellm.acompletion( + model="bedrock/mantle/us.anthropic.claude-sonnet-5", + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "What colour is this?"}, + {"type": "image_url", "image_url": {"url": image_url}}, + ], + } + ], + aws_access_key_id="AKIAEXAMPLE", + aws_secret_access_key="fake-secret", + aws_region_name="us-east-1", + client=client, + ) + + assert response.choices[0].message.content == "Green" + assert async_only_image_fetch.fetched == [image_url] + assert image_url not in captured["body"] + assert async_only_image_fetch.base64_png in captured["body"] diff --git a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py index 023e2d8843f..e86e671b939 100644 --- a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py +++ b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py @@ -17,7 +17,7 @@ from litellm.llms.base_llm.audio_transcription.transformation import ( AudioTranscriptionRequestData, BaseAudioTranscriptionConfig, ) -from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler from litellm.llms.custom_httpx.llm_http_handler import ( BaseLLMHTTPHandler, @@ -30,7 +30,7 @@ from litellm.llms.azure.videos.transformation import AzureVideoConfig from litellm.llms.openai.videos.transformation import OpenAIVideoConfig from litellm.types.llms.openai import ResponsesAPIResponse from litellm.types.router import GenericLiteLLMParams -from litellm.types.utils import TranscriptionResponse +from litellm.types.utils import ModelResponse, TranscriptionResponse _ACTIVE_KEY = "_code_interpreter_interception_active" _SANDBOX_KEY = "_code_interpreter_interception_sandbox_key" @@ -3186,3 +3186,105 @@ async def test_async_container_list_handler_transforms_success_response(): assert [container.id for container in response.data] == ["cntr_a"] assert response.has_more is True + + +class _TransformRecordingConfig(BaseConfig): + def __init__(self, transform_async: bool): + self.transform_async = transform_async + self.transform_calls = [] + + @property + def uses_async_transform_request(self) -> bool: + return self.transform_async + + def get_supported_openai_params(self, model): + return [] + + def map_openai_params(self, non_default_params, optional_params, model, drop_params): + return optional_params + + def validate_environment( + self, headers, model, messages, optional_params, litellm_params, api_key=None, api_base=None + ): + return {} + + def transform_request(self, model, messages, optional_params, litellm_params, headers): + self.transform_calls.append("sync") + return {"transformed_by": "sync"} + + async def async_transform_request(self, model, messages, optional_params, litellm_params, headers): + self.transform_calls.append("async") + return {"transformed_by": "async"} + + def transform_response( + self, + model, + raw_response, + model_response, + logging_obj, + request_data, + messages, + optional_params, + litellm_params, + encoding, + api_key=None, + json_mode=None, + ): + model_response.choices[0].message.content = raw_response.json()["transformed_by"] + return model_response + + def get_error_class(self, error_message, status_code, headers): + return BaseLLMException(status_code=status_code, message=error_message, headers=headers) + + +def _start_async_completion(config): + captured = {} + + def handle(request): + captured["body"] = json.loads(request.content) + return httpx.Response(200, json=captured["body"]) + + client = AsyncHTTPHandler() + client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle)) + pending = BaseLLMHTTPHandler().completion( + model="stub-model", + messages=[{"role": "user", "content": "hi"}], + api_base="https://llm.example/v1/chat", + custom_llm_provider="openai", + model_response=ModelResponse(), + encoding=None, + logging_obj=Mock(dynamic_success_callbacks=None, model_call_details={}), + optional_params={}, + timeout=10.0, + litellm_params={}, + acompletion=True, + client=client, + provider_config=config, + ) + return pending, captured + + +async def test_completion_awaits_async_transform_request_when_config_opts_in(): + config = _TransformRecordingConfig(transform_async=True) + + pending, captured = _start_async_completion(config) + assert config.transform_calls == [] + + response = await pending + + assert config.transform_calls == ["async"] + assert captured["body"] == {"transformed_by": "async"} + assert response.choices[0].message.content == "async" + + +async def test_completion_keeps_sync_transform_request_before_returning_by_default(): + config = _TransformRecordingConfig(transform_async=False) + + pending, captured = _start_async_completion(config) + assert config.transform_calls == ["sync"] + + response = await pending + + assert config.transform_calls == ["sync"] + assert captured["body"] == {"transformed_by": "sync"} + assert response.choices[0].message.content == "sync" diff --git a/tests/test_litellm/llms/snowflake/chat/test_snowflake_chat_transformation.py b/tests/test_litellm/llms/snowflake/chat/test_snowflake_chat_transformation.py index 25a961c3413..5687a319f06 100644 --- a/tests/test_litellm/llms/snowflake/chat/test_snowflake_chat_transformation.py +++ b/tests/test_litellm/llms/snowflake/chat/test_snowflake_chat_transformation.py @@ -6,6 +6,7 @@ Tests tool calling request/response transformations and chat completions import asyncio import os import copy +import uuid import json from typing import Any, Dict, List @@ -945,3 +946,48 @@ class TestSnowflakeChatCompletion: assert len(chunks_received) > 0 content = "".join(c.choices[0].delta.content for c in chunks_received if c.choices[0].delta.content) + + +async def test_snowflake_claude_async_completion_inlines_remote_images_off_the_event_loop(async_only_image_fetch): + image_url = f"http://img.example/{uuid.uuid4()}.png" + captured = {} + + def handle(request): + captured["body"] = request.content.decode() + return httpx.Response( + 200, + json={ + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "claude-sonnet-4-6", + "content": [{"type": "text", "text": "Green"}], + "stop_reason": "end_turn", + "usage": {"input_tokens": 1, "output_tokens": 1}, + }, + ) + + client = AsyncHTTPHandler() + client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle)) + + response = await litellm.acompletion( + model="snowflake/claude-sonnet-4-6", + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "What colour is this?"}, + {"type": "image_url", "image_url": {"url": image_url}}, + ], + } + ], + api_key="fake-jwt", + account_id="FAKE-ACCOUNT", + api_base=FAKE_API_BASE, + client=client, + ) + + assert response.choices[0].message.content == "Green" + assert async_only_image_fetch.fetched == [image_url] + assert image_url not in captured["body"] + assert async_only_image_fetch.base64_png in captured["body"] diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py b/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py index fad310fc5c0..ec445342523 100644 --- a/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py +++ b/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py @@ -1,6 +1,10 @@ +import uuid +import httpx import pytest +import litellm +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler from litellm.llms.vertex_ai.gemini import transformation from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import ( VertexGeminiConfig, @@ -338,3 +342,41 @@ def test_map_function_enterprise_web_search_snake_case(): assert len(result) == 1 assert "enterpriseWebSearch" in result[0] + + +async def test_gemini_ai_studio_async_completion_inlines_remote_images_off_the_event_loop(async_only_image_fetch): + image_url = f"http://img.example/{uuid.uuid4()}.png" + captured = {} + + def handle(request): + captured["body"] = request.content.decode() + return httpx.Response( + 200, + json={ + "candidates": [{"content": {"parts": [{"text": "Green"}], "role": "model"}, "finishReason": "STOP"}], + "usageMetadata": {"promptTokenCount": 1, "candidatesTokenCount": 1, "totalTokenCount": 2}, + }, + ) + + client = AsyncHTTPHandler() + client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle)) + + response = await litellm.acompletion( + model="gemini/gemini-3.8-flash", + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "What colour is this?"}, + {"type": "image_url", "image_url": {"url": image_url}}, + ], + } + ], + api_key="fake-gemini-key", + client=client, + ) + + assert response.choices[0].message.content == "Green" + assert async_only_image_fetch.fetched == [image_url] + assert image_url not in captured["body"] + assert async_only_image_fetch.base64_png in captured["body"]