forked from phoenix/litellm-mirror
LiteLLM Minor Fixes & Improvements (10/30/2024) (#6519)
* refactor: move gemini translation logic inside the transformation.py file easier to isolate the gemini translation logic * fix(gemini-transformation): support multiple tool calls in message body Merges https://github.com/BerriAI/litellm/pull/6487/files * test(test_vertex.py): add remaining tests from https://github.com/BerriAI/litellm/pull/6487 * fix(gemini-transformation): return tool calls for multiple tool calls * fix: support passing logprobs param for vertex + gemini * feat(vertex_ai): add logprobs support for gemini calls * fix(anthropic/chat/transformation.py): fix disable parallel tool use flag * fix: fix linting error * fix(_logging.py): log stacktrace information in json logs Closes https://github.com/BerriAI/litellm/issues/6497 * fix(utils.py): fix mem leak for async stream + completion Uses a global executor pool instead of creating a new thread on each request Fixes https://github.com/BerriAI/litellm/issues/6404 * fix(factory.py): handle tool call + content in assistant message for bedrock * fix: fix import * fix(factory.py): maintain support for content as a str in assistant response * fix: fix import * test: cleanup test * fix(vertex_and_google_ai_studio/): return none for content if no str value * test: retry flaky tests * (UI) Fix viewing members, keys in a team + added testing (#6514) * fix listing teams on ui * LiteLLM Minor Fixes & Improvements (10/28/2024) (#6475) * fix(anthropic/chat/transformation.py): support anthropic disable_parallel_tool_use param Fixes https://github.com/BerriAI/litellm/issues/6456 * feat(anthropic/chat/transformation.py): support anthropic computer tool use Closes https://github.com/BerriAI/litellm/issues/6427 * fix(vertex_ai/common_utils.py): parse out '$schema' when calling vertex ai Fixes issue when trying to call vertex from vercel sdk * fix(main.py): add 'extra_headers' support for azure on all translation endpoints Fixes https://github.com/BerriAI/litellm/issues/6465 * fix: fix linting errors * fix(transformation.py): handle no beta headers for anthropic * test: cleanup test * fix: fix linting error * fix: fix linting errors * fix: fix linting errors * fix(transformation.py): handle dummy tool call * fix(main.py): fix linting error * fix(azure.py): pass required param * LiteLLM Minor Fixes & Improvements (10/24/2024) (#6441) * fix(azure.py): handle /openai/deployment in azure api base * fix(factory.py): fix faulty anthropic tool result translation check Fixes https://github.com/BerriAI/litellm/issues/6422 * fix(gpt_transformation.py): add support for parallel_tool_calls to azure Fixes https://github.com/BerriAI/litellm/issues/6440 * fix(factory.py): support anthropic prompt caching for tool results * fix(vertex_ai/common_utils): don't pop non-null required field Fixes https://github.com/BerriAI/litellm/issues/6426 * feat(vertex_ai.py): support code_execution tool call for vertex ai + gemini Closes https://github.com/BerriAI/litellm/issues/6434 * build(model_prices_and_context_window.json): Add 'supports_assistant_prefill' for bedrock claude-3-5-sonnet v2 models Closes https://github.com/BerriAI/litellm/issues/6437 * fix(types/utils.py): fix linting * test: update test to include required fields * test: fix test * test: handle flaky test * test: remove e2e test - hitting gemini rate limits * Litellm dev 10 26 2024 (#6472) * docs(exception_mapping.md): add missing exception types Fixes https://github.com/Aider-AI/aider/issues/2120#issuecomment-2438971183 * fix(main.py): register custom model pricing with specific key Ensure custom model pricing is registered to the specific model+provider key combination * test: make testing more robust for custom pricing * fix(redis_cache.py): instrument otel logging for sync redis calls ensures complete coverage for all redis cache calls * (Testing) Add unit testing for DualCache - ensure in memory cache is used when expected (#6471) * test test_dual_cache_get_set * unit testing for dual cache * fix async_set_cache_sadd * test_dual_cache_local_only * redis otel tracing + async support for latency routing (#6452) * docs(exception_mapping.md): add missing exception types Fixes https://github.com/Aider-AI/aider/issues/2120#issuecomment-2438971183 * fix(main.py): register custom model pricing with specific key Ensure custom model pricing is registered to the specific model+provider key combination * test: make testing more robust for custom pricing * fix(redis_cache.py): instrument otel logging for sync redis calls ensures complete coverage for all redis cache calls * refactor: pass parent_otel_span for redis caching calls in router allows for more observability into what calls are causing latency issues * test: update tests with new params * refactor: ensure e2e otel tracing for router * refactor(router.py): add more otel tracing acrosss router catch all latency issues for router requests * fix: fix linting error * fix(router.py): fix linting error * fix: fix test * test: fix tests * fix(dual_cache.py): pass ttl to redis cache * fix: fix param * fix(dual_cache.py): set default value for parent_otel_span * fix(transformation.py): support 'response_format' for anthropic calls * fix(transformation.py): check for cache_control inside 'function' block * fix: fix linting error * fix: fix linting errors --------- Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com> --------- Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com> * ui new build * Add retry strat (#6520) Signed-off-by: dbczumar <corey.zumar@databricks.com> * (fix) slack alerting - don't spam the failed cost tracking alert for the same model (#6543) * fix use failing_model as cache key for failed_tracking_alert * fix use standard logging payload for getting response cost * fix kwargs.get("response_cost") * fix getting response cost * (feat) add XAI ChatCompletion Support (#6373) * init commit for XAI * add full logic for xai chat completion * test_completion_xai * docs xAI * add xai/grok-beta * test_xai_chat_config_get_openai_compatible_provider_info * test_xai_chat_config_map_openai_params * add xai streaming test --------- Signed-off-by: dbczumar <corey.zumar@databricks.com> Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com> Co-authored-by: Corey Zumar <39497902+dbczumar@users.noreply.github.com>
This commit is contained in:
Krish Dholakia
GitHub
parent
5652c375b3
commit
f79365df6e
24 changed files with 1851 additions and 700 deletions
|
|
@ -35,6 +35,9 @@ class JsonFormatter(Formatter):
|
|||
"timestamp": self.formatTime(record),
|
||||
}
|
||||
|
||||
if record.exc_info:
|
||||
json_record["stacktrace"] = self.formatException(record.exc_info)
|
||||
|
||||
return json.dumps(json_record)
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -123,7 +123,7 @@ class AnthropicConfig:
|
|||
return headers
|
||||
|
||||
def _map_tool_choice(
|
||||
self, tool_choice: Optional[str], disable_parallel_tool_use: Optional[bool]
|
||||
self, tool_choice: Optional[str], parallel_tool_use: Optional[bool]
|
||||
) -> Optional[AnthropicMessagesToolChoice]:
|
||||
_tool_choice: Optional[AnthropicMessagesToolChoice] = None
|
||||
if tool_choice == "auto":
|
||||
|
|
@ -138,13 +138,15 @@ class AnthropicConfig:
|
|||
if _tool_name is not None:
|
||||
_tool_choice["name"] = _tool_name
|
||||
|
||||
if disable_parallel_tool_use is not None:
|
||||
if parallel_tool_use is not None:
|
||||
# Anthropic uses 'disable_parallel_tool_use' flag to determine if parallel tool use is allowed
|
||||
# this is the inverse of the openai flag.
|
||||
if _tool_choice is not None:
|
||||
_tool_choice["disable_parallel_tool_use"] = disable_parallel_tool_use
|
||||
_tool_choice["disable_parallel_tool_use"] = not parallel_tool_use
|
||||
else: # use anthropic defaults and make sure to send the disable_parallel_tool_use flag
|
||||
_tool_choice = AnthropicMessagesToolChoice(
|
||||
type="auto",
|
||||
disable_parallel_tool_use=disable_parallel_tool_use,
|
||||
disable_parallel_tool_use=not parallel_tool_use,
|
||||
)
|
||||
return _tool_choice
|
||||
|
||||
|
|
@ -255,9 +257,7 @@ class AnthropicConfig:
|
|||
_tool_choice: Optional[AnthropicMessagesToolChoice] = (
|
||||
self._map_tool_choice(
|
||||
tool_choice=non_default_params.get("tool_choice"),
|
||||
disable_parallel_tool_use=non_default_params.get(
|
||||
"parallel_tool_calls"
|
||||
),
|
||||
parallel_tool_use=non_default_params.get("parallel_tool_calls"),
|
||||
)
|
||||
)
|
||||
|
||||
|
|
|
|||
|
|
@ -2552,18 +2552,20 @@ def _bedrock_converse_messages_pt( # noqa: PLR0915
|
|||
BedrockContentBlock(image=assistants_part) # type: ignore
|
||||
)
|
||||
assistant_content.extend(assistants_parts)
|
||||
elif messages[msg_i].get(
|
||||
"tool_calls", []
|
||||
): # support assistant tool invoke convertion
|
||||
assistant_content.extend(
|
||||
_convert_to_bedrock_tool_call_invoke(messages[msg_i]["tool_calls"])
|
||||
)
|
||||
else:
|
||||
elif messages[msg_i].get("content", None) is not None and isinstance(
|
||||
messages[msg_i]["content"], str
|
||||
):
|
||||
assistant_text = (
|
||||
messages[msg_i].get("content") or ""
|
||||
) # either string or none
|
||||
if assistant_text:
|
||||
assistant_content.append(BedrockContentBlock(text=assistant_text))
|
||||
if messages[msg_i].get(
|
||||
"tool_calls", []
|
||||
): # support assistant tool invoke convertion [TODO]:
|
||||
assistant_content.extend(
|
||||
_convert_to_bedrock_tool_call_invoke(messages[msg_i]["tool_calls"])
|
||||
)
|
||||
|
||||
msg_i += 1
|
||||
|
||||
|
|
|
|||
|
|
@ -11,9 +11,9 @@ from litellm.types.llms.vertex_ai import CachedContentRequestBody, SystemInstruc
|
|||
from litellm.utils import is_cached_message
|
||||
|
||||
from ..common_utils import VertexAIError, get_supports_system_message
|
||||
from ..gemini.transformation import _transform_system_message
|
||||
from ..gemini.vertex_and_google_ai_studio_gemini import (
|
||||
from ..gemini.transformation import (
|
||||
_gemini_convert_messages_with_history,
|
||||
_transform_system_message,
|
||||
)
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -4,14 +4,34 @@ Transformation logic from OpenAI format to Gemini format.
|
|||
Why separate file? Make it easy to see how transformation works
|
||||
"""
|
||||
|
||||
from typing import List, Literal, Optional, Tuple, Union
|
||||
import os
|
||||
from typing import List, Literal, Optional, Tuple, Union, cast
|
||||
|
||||
import httpx
|
||||
from pydantic import BaseModel
|
||||
|
||||
import litellm
|
||||
from litellm._logging import verbose_logger
|
||||
from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
|
||||
from litellm.llms.prompt_templates.factory import response_schema_prompt
|
||||
from litellm.types.llms.openai import AllMessageValues
|
||||
from litellm.llms.prompt_templates.factory import (
|
||||
convert_to_anthropic_image_obj,
|
||||
convert_to_gemini_tool_call_invoke,
|
||||
convert_to_gemini_tool_call_result,
|
||||
response_schema_prompt,
|
||||
)
|
||||
from litellm.types.files import (
|
||||
get_file_mime_type_for_file_type,
|
||||
get_file_type_from_extension,
|
||||
is_gemini_1_5_accepted_file_type,
|
||||
is_video_file_type,
|
||||
)
|
||||
from litellm.types.llms.openai import (
|
||||
AllMessageValues,
|
||||
ChatCompletionAssistantMessage,
|
||||
ChatCompletionImageObject,
|
||||
ChatCompletionTextObject,
|
||||
)
|
||||
from litellm.types.llms.vertex_ai import *
|
||||
from litellm.types.llms.vertex_ai import (
|
||||
GenerationConfig,
|
||||
PartType,
|
||||
|
|
@ -21,9 +41,185 @@ from litellm.types.llms.vertex_ai import (
|
|||
ToolConfig,
|
||||
Tools,
|
||||
)
|
||||
from litellm.utils import CustomStreamWrapper, ModelResponse, Usage
|
||||
|
||||
from ..common_utils import get_supports_response_schema, get_supports_system_message
|
||||
from ..vertex_ai_non_gemini import _gemini_convert_messages_with_history
|
||||
from ..common_utils import (
|
||||
_check_text_in_content,
|
||||
get_supports_response_schema,
|
||||
get_supports_system_message,
|
||||
)
|
||||
|
||||
|
||||
def _process_gemini_image(image_url: str) -> PartType:
|
||||
try:
|
||||
# GCS URIs
|
||||
if "gs://" in image_url:
|
||||
# Figure out file type
|
||||
extension_with_dot = os.path.splitext(image_url)[-1] # Ex: ".png"
|
||||
extension = extension_with_dot[1:] # Ex: "png"
|
||||
|
||||
file_type = get_file_type_from_extension(extension)
|
||||
|
||||
# Validate the file type is supported by Gemini
|
||||
if not is_gemini_1_5_accepted_file_type(file_type):
|
||||
raise Exception(f"File type not supported by gemini - {file_type}")
|
||||
|
||||
mime_type = get_file_mime_type_for_file_type(file_type)
|
||||
file_data = FileDataType(mime_type=mime_type, file_uri=image_url)
|
||||
|
||||
return PartType(file_data=file_data)
|
||||
|
||||
# Direct links
|
||||
elif "https:/" in image_url or "base64" in image_url:
|
||||
image = convert_to_anthropic_image_obj(image_url)
|
||||
_blob = BlobType(data=image["data"], mime_type=image["media_type"])
|
||||
return PartType(inline_data=_blob)
|
||||
raise Exception("Invalid image received - {}".format(image_url))
|
||||
except Exception as e:
|
||||
raise e
|
||||
|
||||
|
||||
def _gemini_convert_messages_with_history( # noqa: PLR0915
|
||||
messages: List[AllMessageValues],
|
||||
) -> List[ContentType]:
|
||||
"""
|
||||
Converts given messages from OpenAI format to Gemini format
|
||||
|
||||
- Parts must be iterable
|
||||
- Roles must alternate b/w 'user' and 'model' (same as anthropic -> merge consecutive roles)
|
||||
- Please ensure that function response turn comes immediately after a function call turn
|
||||
"""
|
||||
user_message_types = {"user", "system"}
|
||||
contents: List[ContentType] = []
|
||||
|
||||
last_message_with_tool_calls = None
|
||||
|
||||
msg_i = 0
|
||||
tool_call_responses = []
|
||||
try:
|
||||
while msg_i < len(messages):
|
||||
user_content: List[PartType] = []
|
||||
init_msg_i = msg_i
|
||||
## MERGE CONSECUTIVE USER CONTENT ##
|
||||
while (
|
||||
msg_i < len(messages) and messages[msg_i]["role"] in user_message_types
|
||||
):
|
||||
_message_content = messages[msg_i].get("content")
|
||||
if _message_content is not None and isinstance(_message_content, list):
|
||||
_parts: List[PartType] = []
|
||||
for element in _message_content:
|
||||
if (
|
||||
element["type"] == "text"
|
||||
and "text" in element
|
||||
and len(element["text"]) > 0
|
||||
):
|
||||
element = cast(ChatCompletionTextObject, element)
|
||||
_part = PartType(text=element["text"])
|
||||
_parts.append(_part)
|
||||
elif element["type"] == "image_url":
|
||||
element = cast(ChatCompletionImageObject, element)
|
||||
img_element = element
|
||||
if isinstance(img_element["image_url"], dict):
|
||||
image_url = img_element["image_url"]["url"]
|
||||
else:
|
||||
image_url = img_element["image_url"]
|
||||
_part = _process_gemini_image(image_url=image_url)
|
||||
_parts.append(_part)
|
||||
user_content.extend(_parts)
|
||||
elif (
|
||||
_message_content is not None
|
||||
and isinstance(_message_content, str)
|
||||
and len(_message_content) > 0
|
||||
):
|
||||
_part = PartType(text=_message_content)
|
||||
user_content.append(_part)
|
||||
|
||||
msg_i += 1
|
||||
|
||||
if user_content:
|
||||
"""
|
||||
check that user_content has 'text' parameter.
|
||||
- Known Vertex Error: Unable to submit request because it must have a text parameter.
|
||||
- Relevant Issue: https://github.com/BerriAI/litellm/issues/5515
|
||||
"""
|
||||
has_text_in_content = _check_text_in_content(user_content)
|
||||
if has_text_in_content is False:
|
||||
verbose_logger.warning(
|
||||
"No text in user content. Adding a blank text to user content, to ensure Gemini doesn't fail the request. Relevant Issue - https://github.com/BerriAI/litellm/issues/5515"
|
||||
)
|
||||
user_content.append(
|
||||
PartType(text=" ")
|
||||
) # add a blank text, to ensure Gemini doesn't fail the request.
|
||||
contents.append(ContentType(role="user", parts=user_content))
|
||||
assistant_content = []
|
||||
## MERGE CONSECUTIVE ASSISTANT CONTENT ##
|
||||
while msg_i < len(messages) and messages[msg_i]["role"] == "assistant":
|
||||
if isinstance(messages[msg_i], BaseModel):
|
||||
msg_dict: Union[ChatCompletionAssistantMessage, dict] = messages[msg_i].model_dump() # type: ignore
|
||||
else:
|
||||
msg_dict = messages[msg_i] # type: ignore
|
||||
assistant_msg = ChatCompletionAssistantMessage(**msg_dict) # type: ignore
|
||||
_message_content = assistant_msg.get("content", None)
|
||||
if _message_content is not None and isinstance(_message_content, list):
|
||||
_parts = []
|
||||
for element in _message_content:
|
||||
if isinstance(element, dict):
|
||||
if element["type"] == "text":
|
||||
_part = PartType(text=element["text"])
|
||||
_parts.append(_part)
|
||||
assistant_content.extend(_parts)
|
||||
elif (
|
||||
_message_content is not None
|
||||
and isinstance(_message_content, str)
|
||||
and _message_content
|
||||
):
|
||||
assistant_text = _message_content # either string or none
|
||||
assistant_content.append(PartType(text=assistant_text)) # type: ignore
|
||||
|
||||
## HANDLE ASSISTANT FUNCTION CALL
|
||||
if (
|
||||
assistant_msg.get("tool_calls", []) is not None
|
||||
or assistant_msg.get("function_call") is not None
|
||||
): # support assistant tool invoke conversion
|
||||
assistant_content.extend(
|
||||
convert_to_gemini_tool_call_invoke(assistant_msg)
|
||||
)
|
||||
last_message_with_tool_calls = assistant_msg
|
||||
|
||||
msg_i += 1
|
||||
|
||||
if assistant_content:
|
||||
contents.append(ContentType(role="model", parts=assistant_content))
|
||||
|
||||
## APPEND TOOL CALL MESSAGES ##
|
||||
tool_call_message_roles = ["tool", "function"]
|
||||
if (
|
||||
msg_i < len(messages)
|
||||
and messages[msg_i]["role"] in tool_call_message_roles
|
||||
):
|
||||
_part = convert_to_gemini_tool_call_result(
|
||||
messages[msg_i], last_message_with_tool_calls # type: ignore
|
||||
)
|
||||
msg_i += 1
|
||||
tool_call_responses.append(_part)
|
||||
if msg_i < len(messages) and (
|
||||
messages[msg_i]["role"] not in tool_call_message_roles
|
||||
):
|
||||
if len(tool_call_responses) > 0:
|
||||
contents.append(ContentType(parts=tool_call_responses))
|
||||
tool_call_responses = []
|
||||
|
||||
if msg_i == init_msg_i: # prevent infinite loops
|
||||
raise Exception(
|
||||
"Invalid Message passed in - {}. File an issue https://github.com/BerriAI/litellm/issues".format(
|
||||
messages[msg_i]
|
||||
)
|
||||
)
|
||||
if len(tool_call_responses) > 0:
|
||||
contents.append(ContentType(parts=tool_call_responses))
|
||||
return contents
|
||||
except Exception as e:
|
||||
raise e
|
||||
|
||||
|
||||
def _transform_request_body(
|
||||
|
|
|
|||
|
|
@ -35,13 +35,6 @@ from litellm.llms.custom_httpx.http_handler import (
|
|||
HTTPHandler,
|
||||
get_async_httpx_client,
|
||||
)
|
||||
from litellm.llms.prompt_templates.factory import (
|
||||
convert_url_to_base64,
|
||||
response_schema_prompt,
|
||||
)
|
||||
from litellm.llms.vertex_ai_and_google_ai_studio.vertex_ai_non_gemini import (
|
||||
_gemini_convert_messages_with_history,
|
||||
)
|
||||
from litellm.types.llms.openai import (
|
||||
ChatCompletionResponseMessage,
|
||||
ChatCompletionToolCallChunk,
|
||||
|
|
@ -57,6 +50,7 @@ from litellm.types.llms.vertex_ai import (
|
|||
GenerateContentResponseBody,
|
||||
GenerationConfig,
|
||||
HttpxPartType,
|
||||
LogprobsResult,
|
||||
PartType,
|
||||
RequestBody,
|
||||
SafetSettingsConfig,
|
||||
|
|
@ -64,7 +58,12 @@ from litellm.types.llms.vertex_ai import (
|
|||
ToolConfig,
|
||||
Tools,
|
||||
)
|
||||
from litellm.types.utils import GenericStreamingChunk
|
||||
from litellm.types.utils import (
|
||||
ChatCompletionTokenLogprob,
|
||||
ChoiceLogprobs,
|
||||
GenericStreamingChunk,
|
||||
TopLogprob,
|
||||
)
|
||||
from litellm.utils import CustomStreamWrapper, ModelResponse, Usage
|
||||
|
||||
from ....utils import _remove_additional_properties, _remove_strict_from_schema
|
||||
|
|
@ -365,6 +364,7 @@ class VertexGeminiConfig:
|
|||
"presence_penalty",
|
||||
"extra_headers",
|
||||
"seed",
|
||||
"logprobs",
|
||||
]
|
||||
|
||||
def map_tool_choice_values(
|
||||
|
|
@ -454,6 +454,16 @@ class VertexGeminiConfig:
|
|||
_tools["code_execution"] = code_execution
|
||||
return [_tools]
|
||||
|
||||
def _map_response_schema(self, value: dict) -> dict:
|
||||
old_schema = deepcopy(value)
|
||||
if isinstance(old_schema, list):
|
||||
for item in old_schema:
|
||||
if isinstance(item, dict):
|
||||
item = _build_vertex_schema(parameters=item)
|
||||
elif isinstance(old_schema, dict):
|
||||
old_schema = _build_vertex_schema(parameters=old_schema)
|
||||
return old_schema
|
||||
|
||||
def map_openai_params(
|
||||
self,
|
||||
model: str,
|
||||
|
|
@ -461,6 +471,7 @@ class VertexGeminiConfig:
|
|||
optional_params: dict,
|
||||
drop_params: bool,
|
||||
):
|
||||
|
||||
for param, value in non_default_params.items():
|
||||
if param == "temperature":
|
||||
optional_params["temperature"] = value
|
||||
|
|
@ -499,19 +510,15 @@ class VertexGeminiConfig:
|
|||
if "response_schema" in optional_params and isinstance(
|
||||
optional_params["response_schema"], dict
|
||||
):
|
||||
old_schema = deepcopy(optional_params["response_schema"])
|
||||
|
||||
if isinstance(old_schema, list):
|
||||
for item in old_schema:
|
||||
if isinstance(item, dict):
|
||||
item = _build_vertex_schema(parameters=item)
|
||||
elif isinstance(old_schema, dict):
|
||||
old_schema = _build_vertex_schema(parameters=old_schema)
|
||||
optional_params["response_schema"] = old_schema
|
||||
optional_params["response_schema"] = self._map_response_schema(
|
||||
value=optional_params["response_schema"]
|
||||
)
|
||||
if param == "frequency_penalty":
|
||||
optional_params["frequency_penalty"] = value
|
||||
if param == "presence_penalty":
|
||||
optional_params["presence_penalty"] = value
|
||||
if param == "logprobs":
|
||||
optional_params["responseLogprobs"] = value
|
||||
if (param == "tools" or param == "functions") and isinstance(value, list):
|
||||
optional_params["tools"] = self._map_function(value=value)
|
||||
optional_params["litellm_param_is_function_call"] = (
|
||||
|
|
@ -527,6 +534,7 @@ class VertexGeminiConfig:
|
|||
optional_params["tool_choice"] = _tool_choice_value
|
||||
if param == "seed":
|
||||
optional_params["seed"] = value
|
||||
|
||||
return optional_params
|
||||
|
||||
def get_mapped_special_auth_params(self) -> dict:
|
||||
|
|
@ -584,12 +592,325 @@ class VertexGeminiConfig:
|
|||
)
|
||||
return exception_string
|
||||
|
||||
def get_assistant_content_message(self, parts: List[HttpxPartType]) -> str:
|
||||
content_str = ""
|
||||
def get_assistant_content_message(
|
||||
self, parts: List[HttpxPartType]
|
||||
) -> Optional[str]:
|
||||
_content_str = ""
|
||||
for part in parts:
|
||||
if "text" in part:
|
||||
content_str += part["text"]
|
||||
return content_str
|
||||
_content_str += part["text"]
|
||||
if _content_str:
|
||||
return _content_str
|
||||
return None
|
||||
|
||||
def _transform_parts(
|
||||
self,
|
||||
parts: List[HttpxPartType],
|
||||
index: int,
|
||||
is_function_call: Optional[bool],
|
||||
) -> Tuple[
|
||||
Optional[ChatCompletionToolCallFunctionChunk],
|
||||
Optional[List[ChatCompletionToolCallChunk]],
|
||||
]:
|
||||
function: Optional[ChatCompletionToolCallFunctionChunk] = None
|
||||
_tools: List[ChatCompletionToolCallChunk] = []
|
||||
for part in parts:
|
||||
if "functionCall" in part:
|
||||
_function_chunk = ChatCompletionToolCallFunctionChunk(
|
||||
name=part["functionCall"]["name"],
|
||||
arguments=json.dumps(part["functionCall"]["args"]),
|
||||
)
|
||||
if is_function_call is True:
|
||||
function = _function_chunk
|
||||
else:
|
||||
_tool_response_chunk = ChatCompletionToolCallChunk(
|
||||
id=f"call_{str(uuid.uuid4())}",
|
||||
type="function",
|
||||
function=_function_chunk,
|
||||
index=index,
|
||||
)
|
||||
_tools.append(_tool_response_chunk)
|
||||
if len(_tools) == 0:
|
||||
tools: Optional[List[ChatCompletionToolCallChunk]] = None
|
||||
else:
|
||||
tools = _tools
|
||||
return function, tools
|
||||
|
||||
def _transform_logprobs(
|
||||
self, logprobs_result: Optional[LogprobsResult]
|
||||
) -> Optional[ChoiceLogprobs]:
|
||||
if logprobs_result is None:
|
||||
return None
|
||||
if "chosenCandidates" not in logprobs_result:
|
||||
return None
|
||||
logprobs_list: List[ChatCompletionTokenLogprob] = []
|
||||
for index, candidate in enumerate(logprobs_result["chosenCandidates"]):
|
||||
top_logprobs: List[TopLogprob] = []
|
||||
if "topCandidates" in logprobs_result and index < len(
|
||||
logprobs_result["topCandidates"]
|
||||
):
|
||||
top_candidates_for_index = logprobs_result["topCandidates"][index][
|
||||
"candidates"
|
||||
]
|
||||
|
||||
for options in top_candidates_for_index:
|
||||
top_logprobs.append(
|
||||
TopLogprob(
|
||||
token=options["token"], logprob=options["logProbability"]
|
||||
)
|
||||
)
|
||||
logprobs_list.append(
|
||||
ChatCompletionTokenLogprob(
|
||||
token=candidate["token"],
|
||||
logprob=candidate["logProbability"],
|
||||
top_logprobs=top_logprobs,
|
||||
)
|
||||
)
|
||||
return ChoiceLogprobs(content=logprobs_list)
|
||||
|
||||
def _handle_blocked_response(
|
||||
self,
|
||||
model_response: ModelResponse,
|
||||
completion_response: GenerateContentResponseBody,
|
||||
) -> ModelResponse:
|
||||
# If set, the prompt was blocked and no candidates are returned. Rephrase your prompt
|
||||
model_response.choices[0].finish_reason = "content_filter"
|
||||
|
||||
chat_completion_message: ChatCompletionResponseMessage = {
|
||||
"role": "assistant",
|
||||
"content": None,
|
||||
}
|
||||
|
||||
choice = litellm.Choices(
|
||||
finish_reason="content_filter",
|
||||
index=0,
|
||||
message=chat_completion_message, # type: ignore
|
||||
logprobs=None,
|
||||
enhancements=None,
|
||||
)
|
||||
|
||||
model_response.choices = [choice]
|
||||
|
||||
## GET USAGE ##
|
||||
usage = litellm.Usage(
|
||||
prompt_tokens=completion_response["usageMetadata"].get(
|
||||
"promptTokenCount", 0
|
||||
),
|
||||
completion_tokens=completion_response["usageMetadata"].get(
|
||||
"candidatesTokenCount", 0
|
||||
),
|
||||
total_tokens=completion_response["usageMetadata"].get("totalTokenCount", 0),
|
||||
)
|
||||
|
||||
setattr(model_response, "usage", usage)
|
||||
|
||||
return model_response
|
||||
|
||||
def _handle_content_policy_violation(
|
||||
self,
|
||||
model_response: ModelResponse,
|
||||
completion_response: GenerateContentResponseBody,
|
||||
) -> ModelResponse:
|
||||
## CONTENT POLICY VIOLATION ERROR
|
||||
model_response.choices[0].finish_reason = "content_filter"
|
||||
|
||||
_chat_completion_message = {
|
||||
"role": "assistant",
|
||||
"content": None,
|
||||
}
|
||||
|
||||
choice = litellm.Choices(
|
||||
finish_reason="content_filter",
|
||||
index=0,
|
||||
message=_chat_completion_message,
|
||||
logprobs=None,
|
||||
enhancements=None,
|
||||
)
|
||||
|
||||
model_response.choices = [choice]
|
||||
|
||||
## GET USAGE ##
|
||||
usage = litellm.Usage(
|
||||
prompt_tokens=completion_response["usageMetadata"].get(
|
||||
"promptTokenCount", 0
|
||||
),
|
||||
completion_tokens=completion_response["usageMetadata"].get(
|
||||
"candidatesTokenCount", 0
|
||||
),
|
||||
total_tokens=completion_response["usageMetadata"].get("totalTokenCount", 0),
|
||||
)
|
||||
|
||||
setattr(model_response, "usage", usage)
|
||||
|
||||
return model_response
|
||||
|
||||
def _transform_response(
|
||||
self,
|
||||
model: str,
|
||||
response: httpx.Response,
|
||||
model_response: ModelResponse,
|
||||
logging_obj: litellm.litellm_core_utils.litellm_logging.Logging,
|
||||
optional_params: dict,
|
||||
litellm_params: dict,
|
||||
api_key: str,
|
||||
data: Union[dict, str, RequestBody],
|
||||
messages: List,
|
||||
print_verbose,
|
||||
encoding,
|
||||
) -> ModelResponse:
|
||||
|
||||
## LOGGING
|
||||
logging_obj.post_call(
|
||||
input=messages,
|
||||
api_key="",
|
||||
original_response=response.text,
|
||||
additional_args={"complete_input_dict": data},
|
||||
)
|
||||
|
||||
## RESPONSE OBJECT
|
||||
try:
|
||||
completion_response = GenerateContentResponseBody(**response.json()) # type: ignore
|
||||
except Exception as e:
|
||||
raise VertexAIError(
|
||||
message="Received={}, Error converting to valid response block={}. File an issue if litellm error - https://github.com/BerriAI/litellm/issues".format(
|
||||
response.text, str(e)
|
||||
),
|
||||
status_code=422,
|
||||
)
|
||||
|
||||
## GET MODEL ##
|
||||
model_response.model = model
|
||||
|
||||
## CHECK IF RESPONSE FLAGGED
|
||||
if (
|
||||
"promptFeedback" in completion_response
|
||||
and "blockReason" in completion_response["promptFeedback"]
|
||||
):
|
||||
return self._handle_blocked_response(
|
||||
model_response=model_response,
|
||||
completion_response=completion_response,
|
||||
)
|
||||
|
||||
_candidates = completion_response.get("candidates")
|
||||
if _candidates and len(_candidates) > 0:
|
||||
content_policy_violations = (
|
||||
VertexGeminiConfig().get_flagged_finish_reasons()
|
||||
)
|
||||
if (
|
||||
"finishReason" in _candidates[0]
|
||||
and _candidates[0]["finishReason"] in content_policy_violations.keys()
|
||||
):
|
||||
return self._handle_content_policy_violation(
|
||||
model_response=model_response,
|
||||
completion_response=completion_response,
|
||||
)
|
||||
|
||||
model_response.choices = [] # type: ignore
|
||||
|
||||
try:
|
||||
## CHECK IF GROUNDING METADATA IN REQUEST
|
||||
grounding_metadata: List[dict] = []
|
||||
safety_ratings: List = []
|
||||
citation_metadata: List = []
|
||||
## GET TEXT ##
|
||||
chat_completion_message: ChatCompletionResponseMessage = {
|
||||
"role": "assistant"
|
||||
}
|
||||
chat_completion_logprobs: Optional[ChoiceLogprobs] = None
|
||||
tools: Optional[List[ChatCompletionToolCallChunk]] = []
|
||||
functions: Optional[ChatCompletionToolCallFunctionChunk] = None
|
||||
if _candidates:
|
||||
for idx, candidate in enumerate(_candidates):
|
||||
if "content" not in candidate:
|
||||
continue
|
||||
|
||||
if "groundingMetadata" in candidate:
|
||||
grounding_metadata.append(candidate["groundingMetadata"]) # type: ignore
|
||||
|
||||
if "safetyRatings" in candidate:
|
||||
safety_ratings.append(candidate["safetyRatings"])
|
||||
|
||||
if "citationMetadata" in candidate:
|
||||
citation_metadata.append(candidate["citationMetadata"])
|
||||
if "parts" in candidate["content"]:
|
||||
chat_completion_message[
|
||||
"content"
|
||||
] = VertexGeminiConfig().get_assistant_content_message(
|
||||
parts=candidate["content"]["parts"]
|
||||
)
|
||||
|
||||
functions, tools = self._transform_parts(
|
||||
parts=candidate["content"]["parts"],
|
||||
index=candidate.get("index", idx),
|
||||
is_function_call=litellm_params.get(
|
||||
"litellm_param_is_function_call"
|
||||
),
|
||||
)
|
||||
|
||||
if "logprobsResult" in candidate:
|
||||
chat_completion_logprobs = self._transform_logprobs(
|
||||
logprobs_result=candidate["logprobsResult"]
|
||||
)
|
||||
|
||||
if tools:
|
||||
chat_completion_message["tool_calls"] = tools
|
||||
|
||||
if functions is not None:
|
||||
chat_completion_message["function_call"] = functions
|
||||
choice = litellm.Choices(
|
||||
finish_reason=candidate.get("finishReason", "stop"),
|
||||
index=candidate.get("index", idx),
|
||||
message=chat_completion_message, # type: ignore
|
||||
logprobs=chat_completion_logprobs,
|
||||
enhancements=None,
|
||||
)
|
||||
|
||||
model_response.choices.append(choice)
|
||||
|
||||
## GET USAGE ##
|
||||
usage = litellm.Usage(
|
||||
prompt_tokens=completion_response["usageMetadata"].get(
|
||||
"promptTokenCount", 0
|
||||
),
|
||||
completion_tokens=completion_response["usageMetadata"].get(
|
||||
"candidatesTokenCount", 0
|
||||
),
|
||||
total_tokens=completion_response["usageMetadata"].get(
|
||||
"totalTokenCount", 0
|
||||
),
|
||||
)
|
||||
|
||||
setattr(model_response, "usage", usage)
|
||||
|
||||
## ADD GROUNDING METADATA ##
|
||||
setattr(model_response, "vertex_ai_grounding_metadata", grounding_metadata)
|
||||
model_response._hidden_params[
|
||||
"vertex_ai_grounding_metadata"
|
||||
] = ( # older approach - maintaining to prevent regressions
|
||||
grounding_metadata
|
||||
)
|
||||
|
||||
## ADD SAFETY RATINGS ##
|
||||
setattr(model_response, "vertex_ai_safety_results", safety_ratings)
|
||||
model_response._hidden_params["vertex_ai_safety_results"] = (
|
||||
safety_ratings # older approach - maintaining to prevent regressions
|
||||
)
|
||||
|
||||
## ADD CITATION METADATA ##
|
||||
setattr(model_response, "vertex_ai_citation_metadata", citation_metadata)
|
||||
model_response._hidden_params["vertex_ai_citation_metadata"] = (
|
||||
citation_metadata # older approach - maintaining to prevent regressions
|
||||
)
|
||||
|
||||
except Exception as e:
|
||||
raise VertexAIError(
|
||||
message="Received={}, Error converting to valid response block={}. File an issue if litellm error - https://github.com/BerriAI/litellm/issues".format(
|
||||
completion_response, str(e)
|
||||
),
|
||||
status_code=422,
|
||||
)
|
||||
|
||||
return model_response
|
||||
|
||||
|
||||
class GoogleAIStudioGeminiConfig(
|
||||
|
|
@ -675,6 +996,7 @@ class GoogleAIStudioGeminiConfig(
|
|||
"response_format",
|
||||
"n",
|
||||
"stop",
|
||||
"logprobs",
|
||||
]
|
||||
|
||||
def map_openai_params(
|
||||
|
|
@ -771,243 +1093,6 @@ class VertexLLM(VertexBase):
|
|||
def __init__(self) -> None:
|
||||
super().__init__()
|
||||
|
||||
def _process_response( # noqa: PLR0915
|
||||
self,
|
||||
model: str,
|
||||
response: httpx.Response,
|
||||
model_response: ModelResponse,
|
||||
logging_obj: litellm.litellm_core_utils.litellm_logging.Logging,
|
||||
optional_params: dict,
|
||||
litellm_params: dict,
|
||||
api_key: str,
|
||||
data: Union[dict, str, RequestBody],
|
||||
messages: List,
|
||||
print_verbose,
|
||||
encoding,
|
||||
) -> ModelResponse:
|
||||
|
||||
## LOGGING
|
||||
logging_obj.post_call(
|
||||
input=messages,
|
||||
api_key="",
|
||||
original_response=response.text,
|
||||
additional_args={"complete_input_dict": data},
|
||||
)
|
||||
|
||||
print_verbose(f"raw model_response: {response.text}")
|
||||
## RESPONSE OBJECT
|
||||
try:
|
||||
completion_response = GenerateContentResponseBody(**response.json()) # type: ignore
|
||||
except Exception as e:
|
||||
raise VertexAIError(
|
||||
message="Received={}, Error converting to valid response block={}. File an issue if litellm error - https://github.com/BerriAI/litellm/issues".format(
|
||||
response.text, str(e)
|
||||
),
|
||||
status_code=422,
|
||||
)
|
||||
|
||||
## GET MODEL ##
|
||||
model_response.model = model
|
||||
|
||||
## CHECK IF RESPONSE FLAGGED
|
||||
if "promptFeedback" in completion_response:
|
||||
if "blockReason" in completion_response["promptFeedback"]:
|
||||
# If set, the prompt was blocked and no candidates are returned. Rephrase your prompt
|
||||
model_response.choices[0].finish_reason = "content_filter"
|
||||
|
||||
chat_completion_message: ChatCompletionResponseMessage = {
|
||||
"role": "assistant",
|
||||
"content": None,
|
||||
}
|
||||
|
||||
choice = litellm.Choices(
|
||||
finish_reason="content_filter",
|
||||
index=0,
|
||||
message=chat_completion_message, # type: ignore
|
||||
logprobs=None,
|
||||
enhancements=None,
|
||||
)
|
||||
|
||||
model_response.choices = [choice]
|
||||
|
||||
## GET USAGE ##
|
||||
usage = litellm.Usage(
|
||||
prompt_tokens=completion_response["usageMetadata"].get(
|
||||
"promptTokenCount", 0
|
||||
),
|
||||
completion_tokens=completion_response["usageMetadata"].get(
|
||||
"candidatesTokenCount", 0
|
||||
),
|
||||
total_tokens=completion_response["usageMetadata"].get(
|
||||
"totalTokenCount", 0
|
||||
),
|
||||
)
|
||||
|
||||
setattr(model_response, "usage", usage)
|
||||
|
||||
return model_response
|
||||
|
||||
_candidates = completion_response.get("candidates")
|
||||
if _candidates and len(_candidates) > 0:
|
||||
content_policy_violations = (
|
||||
VertexGeminiConfig().get_flagged_finish_reasons()
|
||||
)
|
||||
if (
|
||||
"finishReason" in _candidates[0]
|
||||
and _candidates[0]["finishReason"] in content_policy_violations.keys()
|
||||
):
|
||||
## CONTENT POLICY VIOLATION ERROR
|
||||
model_response.choices[0].finish_reason = "content_filter"
|
||||
|
||||
_chat_completion_message = {
|
||||
"role": "assistant",
|
||||
"content": None,
|
||||
}
|
||||
|
||||
choice = litellm.Choices(
|
||||
finish_reason="content_filter",
|
||||
index=0,
|
||||
message=_chat_completion_message,
|
||||
logprobs=None,
|
||||
enhancements=None,
|
||||
)
|
||||
|
||||
model_response.choices = [choice]
|
||||
|
||||
## GET USAGE ##
|
||||
usage = litellm.Usage(
|
||||
prompt_tokens=completion_response["usageMetadata"].get(
|
||||
"promptTokenCount", 0
|
||||
),
|
||||
completion_tokens=completion_response["usageMetadata"].get(
|
||||
"candidatesTokenCount", 0
|
||||
),
|
||||
total_tokens=completion_response["usageMetadata"].get(
|
||||
"totalTokenCount", 0
|
||||
),
|
||||
)
|
||||
|
||||
setattr(model_response, "usage", usage)
|
||||
|
||||
return model_response
|
||||
|
||||
model_response.choices = [] # type: ignore
|
||||
|
||||
try:
|
||||
## CHECK IF GROUNDING METADATA IN REQUEST
|
||||
grounding_metadata: List[dict] = []
|
||||
safety_ratings: List = []
|
||||
citation_metadata: List = []
|
||||
## GET TEXT ##
|
||||
chat_completion_message = {"role": "assistant"}
|
||||
content_str: str = ""
|
||||
tools: List[ChatCompletionToolCallChunk] = []
|
||||
functions: Optional[ChatCompletionToolCallFunctionChunk] = None
|
||||
if _candidates:
|
||||
for idx, candidate in enumerate(_candidates):
|
||||
if "content" not in candidate:
|
||||
continue
|
||||
|
||||
if "groundingMetadata" in candidate:
|
||||
grounding_metadata.append(candidate["groundingMetadata"]) # type: ignore
|
||||
|
||||
if "safetyRatings" in candidate:
|
||||
safety_ratings.append(candidate["safetyRatings"])
|
||||
|
||||
if "citationMetadata" in candidate:
|
||||
citation_metadata.append(candidate["citationMetadata"])
|
||||
if "parts" in candidate["content"]:
|
||||
content_str = (
|
||||
VertexGeminiConfig().get_assistant_content_message(
|
||||
parts=candidate["content"]["parts"]
|
||||
)
|
||||
)
|
||||
|
||||
if (
|
||||
"parts" in candidate["content"]
|
||||
and "functionCall" in candidate["content"]["parts"][0]
|
||||
):
|
||||
_function_chunk = ChatCompletionToolCallFunctionChunk(
|
||||
name=candidate["content"]["parts"][0]["functionCall"][
|
||||
"name"
|
||||
],
|
||||
arguments=json.dumps(
|
||||
candidate["content"]["parts"][0]["functionCall"]["args"]
|
||||
),
|
||||
)
|
||||
if litellm_params.get("litellm_param_is_function_call") is True:
|
||||
functions = _function_chunk
|
||||
else:
|
||||
_tool_response_chunk = ChatCompletionToolCallChunk(
|
||||
id=f"call_{str(uuid.uuid4())}",
|
||||
type="function",
|
||||
function=_function_chunk,
|
||||
index=candidate.get("index", idx),
|
||||
)
|
||||
tools.append(_tool_response_chunk)
|
||||
chat_completion_message["content"] = (
|
||||
content_str if len(content_str) > 0 else None
|
||||
)
|
||||
if len(tools) > 0:
|
||||
chat_completion_message["tool_calls"] = tools
|
||||
|
||||
if functions is not None:
|
||||
chat_completion_message["function_call"] = functions
|
||||
choice = litellm.Choices(
|
||||
finish_reason=candidate.get("finishReason", "stop"),
|
||||
index=candidate.get("index", idx),
|
||||
message=chat_completion_message, # type: ignore
|
||||
logprobs=None,
|
||||
enhancements=None,
|
||||
)
|
||||
|
||||
model_response.choices.append(choice)
|
||||
|
||||
## GET USAGE ##
|
||||
usage = litellm.Usage(
|
||||
prompt_tokens=completion_response["usageMetadata"].get(
|
||||
"promptTokenCount", 0
|
||||
),
|
||||
completion_tokens=completion_response["usageMetadata"].get(
|
||||
"candidatesTokenCount", 0
|
||||
),
|
||||
total_tokens=completion_response["usageMetadata"].get(
|
||||
"totalTokenCount", 0
|
||||
),
|
||||
)
|
||||
|
||||
setattr(model_response, "usage", usage)
|
||||
|
||||
## ADD GROUNDING METADATA ##
|
||||
setattr(model_response, "vertex_ai_grounding_metadata", grounding_metadata)
|
||||
model_response._hidden_params[
|
||||
"vertex_ai_grounding_metadata"
|
||||
] = ( # older approach - maintaining to prevent regressions
|
||||
grounding_metadata
|
||||
)
|
||||
|
||||
## ADD SAFETY RATINGS ##
|
||||
setattr(model_response, "vertex_ai_safety_results", safety_ratings)
|
||||
model_response._hidden_params["vertex_ai_safety_results"] = (
|
||||
safety_ratings # older approach - maintaining to prevent regressions
|
||||
)
|
||||
|
||||
## ADD CITATION METADATA ##
|
||||
setattr(model_response, "vertex_ai_citation_metadata", citation_metadata)
|
||||
model_response._hidden_params["vertex_ai_citation_metadata"] = (
|
||||
citation_metadata # older approach - maintaining to prevent regressions
|
||||
)
|
||||
|
||||
except Exception as e:
|
||||
raise VertexAIError(
|
||||
message="Received={}, Error converting to valid response block={}. File an issue if litellm error - https://github.com/BerriAI/litellm/issues".format(
|
||||
completion_response, str(e)
|
||||
),
|
||||
status_code=422,
|
||||
)
|
||||
|
||||
return model_response
|
||||
|
||||
async def async_streaming(
|
||||
self,
|
||||
model: str,
|
||||
|
|
@ -1171,7 +1256,7 @@ class VertexLLM(VertexBase):
|
|||
except httpx.TimeoutException:
|
||||
raise VertexAIError(status_code=408, message="Timeout error occurred.")
|
||||
|
||||
return self._process_response(
|
||||
return VertexGeminiConfig()._transform_response(
|
||||
model=model,
|
||||
response=response,
|
||||
model_response=model_response,
|
||||
|
|
@ -1359,7 +1444,7 @@ class VertexLLM(VertexBase):
|
|||
except httpx.TimeoutException:
|
||||
raise VertexAIError(status_code=408, message="Timeout error occurred.")
|
||||
|
||||
return self._process_response(
|
||||
return VertexGeminiConfig()._transform_response(
|
||||
model=model,
|
||||
response=response,
|
||||
model_response=model_response,
|
||||
|
|
|
|||
|
|
@ -85,185 +85,6 @@ class TextStreamer:
|
|||
raise StopAsyncIteration # once we run out of data to stream, we raise this error
|
||||
|
||||
|
||||
def _get_image_bytes_from_url(image_url: str) -> bytes:
|
||||
try:
|
||||
response = requests.get(image_url)
|
||||
response.raise_for_status() # Raise an error for bad responses (4xx and 5xx)
|
||||
image_bytes = response.content
|
||||
return image_bytes
|
||||
except requests.exceptions.RequestException as e:
|
||||
raise Exception(f"An exception occurs with this image - {str(e)}")
|
||||
|
||||
|
||||
def _convert_gemini_role(role: str) -> Literal["user", "model"]:
|
||||
if role == "user":
|
||||
return "user"
|
||||
else:
|
||||
return "model"
|
||||
|
||||
|
||||
def _process_gemini_image(image_url: str) -> PartType:
|
||||
try:
|
||||
# GCS URIs
|
||||
if "gs://" in image_url:
|
||||
# Figure out file type
|
||||
extension_with_dot = os.path.splitext(image_url)[-1] # Ex: ".png"
|
||||
extension = extension_with_dot[1:] # Ex: "png"
|
||||
|
||||
file_type = get_file_type_from_extension(extension)
|
||||
|
||||
# Validate the file type is supported by Gemini
|
||||
if not is_gemini_1_5_accepted_file_type(file_type):
|
||||
raise Exception(f"File type not supported by gemini - {file_type}")
|
||||
|
||||
mime_type = get_file_mime_type_for_file_type(file_type)
|
||||
file_data = FileDataType(mime_type=mime_type, file_uri=image_url)
|
||||
|
||||
return PartType(file_data=file_data)
|
||||
|
||||
# Direct links
|
||||
elif "https:/" in image_url or "base64" in image_url:
|
||||
image = convert_to_anthropic_image_obj(image_url)
|
||||
_blob = BlobType(data=image["data"], mime_type=image["media_type"])
|
||||
return PartType(inline_data=_blob)
|
||||
raise Exception("Invalid image received - {}".format(image_url))
|
||||
except Exception as e:
|
||||
raise e
|
||||
|
||||
|
||||
def _gemini_convert_messages_with_history( # noqa: PLR0915
|
||||
messages: List[AllMessageValues],
|
||||
) -> List[ContentType]:
|
||||
"""
|
||||
Converts given messages from OpenAI format to Gemini format
|
||||
|
||||
- Parts must be iterable
|
||||
- Roles must alternate b/w 'user' and 'model' (same as anthropic -> merge consecutive roles)
|
||||
- Please ensure that function response turn comes immediately after a function call turn
|
||||
"""
|
||||
user_message_types = {"user", "system"}
|
||||
contents: List[ContentType] = []
|
||||
|
||||
last_message_with_tool_calls = None
|
||||
|
||||
msg_i = 0
|
||||
try:
|
||||
while msg_i < len(messages):
|
||||
user_content: List[PartType] = []
|
||||
init_msg_i = msg_i
|
||||
## MERGE CONSECUTIVE USER CONTENT ##
|
||||
while (
|
||||
msg_i < len(messages) and messages[msg_i]["role"] in user_message_types
|
||||
):
|
||||
_message_content = messages[msg_i].get("content")
|
||||
if _message_content is not None and isinstance(_message_content, list):
|
||||
_parts: List[PartType] = []
|
||||
for element in _message_content:
|
||||
if (
|
||||
element["type"] == "text"
|
||||
and "text" in element
|
||||
and len(element["text"]) > 0
|
||||
):
|
||||
element = cast(ChatCompletionTextObject, element)
|
||||
_part = PartType(text=element["text"])
|
||||
_parts.append(_part)
|
||||
elif element["type"] == "image_url":
|
||||
element = cast(ChatCompletionImageObject, element)
|
||||
img_element = element
|
||||
if isinstance(img_element["image_url"], dict):
|
||||
image_url = img_element["image_url"]["url"]
|
||||
else:
|
||||
image_url = img_element["image_url"]
|
||||
_part = _process_gemini_image(image_url=image_url)
|
||||
_parts.append(_part)
|
||||
user_content.extend(_parts)
|
||||
elif (
|
||||
_message_content is not None
|
||||
and isinstance(_message_content, str)
|
||||
and len(_message_content) > 0
|
||||
):
|
||||
_part = PartType(text=_message_content)
|
||||
user_content.append(_part)
|
||||
|
||||
msg_i += 1
|
||||
|
||||
if user_content:
|
||||
"""
|
||||
check that user_content has 'text' parameter.
|
||||
- Known Vertex Error: Unable to submit request because it must have a text parameter.
|
||||
- Relevant Issue: https://github.com/BerriAI/litellm/issues/5515
|
||||
"""
|
||||
has_text_in_content = _check_text_in_content(user_content)
|
||||
if has_text_in_content is False:
|
||||
verbose_logger.warning(
|
||||
"No text in user content. Adding a blank text to user content, to ensure Gemini doesn't fail the request. Relevant Issue - https://github.com/BerriAI/litellm/issues/5515"
|
||||
)
|
||||
user_content.append(
|
||||
PartType(text=" ")
|
||||
) # add a blank text, to ensure Gemini doesn't fail the request.
|
||||
contents.append(ContentType(role="user", parts=user_content))
|
||||
assistant_content = []
|
||||
## MERGE CONSECUTIVE ASSISTANT CONTENT ##
|
||||
while msg_i < len(messages) and messages[msg_i]["role"] == "assistant":
|
||||
if isinstance(messages[msg_i], BaseModel):
|
||||
msg_dict: Union[ChatCompletionAssistantMessage, dict] = messages[msg_i].model_dump() # type: ignore
|
||||
else:
|
||||
msg_dict = messages[msg_i] # type: ignore
|
||||
assistant_msg = ChatCompletionAssistantMessage(**msg_dict) # type: ignore
|
||||
_message_content = assistant_msg.get("content", None)
|
||||
if _message_content is not None and isinstance(_message_content, list):
|
||||
_parts = []
|
||||
for element in _message_content:
|
||||
if isinstance(element, dict):
|
||||
if element["type"] == "text":
|
||||
_part = PartType(text=element["text"])
|
||||
_parts.append(_part)
|
||||
assistant_content.extend(_parts)
|
||||
elif (
|
||||
_message_content is not None
|
||||
and isinstance(_message_content, str)
|
||||
and _message_content
|
||||
):
|
||||
assistant_text = _message_content # either string or none
|
||||
assistant_content.append(PartType(text=assistant_text)) # type: ignore
|
||||
|
||||
## HANDLE ASSISTANT FUNCTION CALL
|
||||
if (
|
||||
assistant_msg.get("tool_calls", []) is not None
|
||||
or assistant_msg.get("function_call") is not None
|
||||
): # support assistant tool invoke conversion
|
||||
assistant_content.extend(
|
||||
convert_to_gemini_tool_call_invoke(assistant_msg)
|
||||
)
|
||||
last_message_with_tool_calls = assistant_msg
|
||||
|
||||
msg_i += 1
|
||||
|
||||
if assistant_content:
|
||||
contents.append(ContentType(role="model", parts=assistant_content))
|
||||
|
||||
## APPEND TOOL CALL MESSAGES ##
|
||||
if msg_i < len(messages) and (
|
||||
messages[msg_i]["role"] == "tool"
|
||||
or messages[msg_i]["role"] == "function"
|
||||
):
|
||||
_part = convert_to_gemini_tool_call_result(
|
||||
messages[msg_i], last_message_with_tool_calls # type: ignore
|
||||
)
|
||||
contents.append(ContentType(parts=[_part])) # type: ignore
|
||||
msg_i += 1
|
||||
|
||||
if msg_i == init_msg_i: # prevent infinite loops
|
||||
raise Exception(
|
||||
"Invalid Message passed in - {}. File an issue https://github.com/BerriAI/litellm/issues".format(
|
||||
messages[msg_i]
|
||||
)
|
||||
)
|
||||
return contents
|
||||
except Exception as e:
|
||||
raise e
|
||||
|
||||
|
||||
def _get_client_cache_key(
|
||||
model: str, vertex_project: Optional[str], vertex_location: Optional[str]
|
||||
):
|
||||
|
|
@ -487,91 +308,7 @@ def completion( # noqa: PLR0915
|
|||
return async_completion(**data)
|
||||
|
||||
completion_response = None
|
||||
if mode == "vision":
|
||||
print_verbose("\nMaking VertexAI Gemini Pro / Pro Vision Call")
|
||||
print_verbose(f"\nProcessing input messages = {messages}")
|
||||
tools = optional_params.pop("tools", None)
|
||||
content = _gemini_convert_messages_with_history(messages=messages)
|
||||
stream = optional_params.pop("stream", False)
|
||||
if stream is True:
|
||||
request_str += f"response = llm_model.generate_content({content}, generation_config=GenerationConfig(**{optional_params}), safety_settings={safety_settings}, stream={stream})\n"
|
||||
logging_obj.pre_call(
|
||||
input=prompt,
|
||||
api_key=None,
|
||||
additional_args={
|
||||
"complete_input_dict": optional_params,
|
||||
"request_str": request_str,
|
||||
},
|
||||
)
|
||||
|
||||
_model_response = llm_model.generate_content(
|
||||
contents=content,
|
||||
generation_config=optional_params,
|
||||
safety_settings=safety_settings,
|
||||
stream=True,
|
||||
tools=tools,
|
||||
)
|
||||
|
||||
return _model_response
|
||||
|
||||
request_str += f"response = llm_model.generate_content({content})\n"
|
||||
## LOGGING
|
||||
logging_obj.pre_call(
|
||||
input=prompt,
|
||||
api_key=None,
|
||||
additional_args={
|
||||
"complete_input_dict": optional_params,
|
||||
"request_str": request_str,
|
||||
},
|
||||
)
|
||||
|
||||
## LLM Call
|
||||
response = llm_model.generate_content(
|
||||
contents=content,
|
||||
generation_config=optional_params,
|
||||
safety_settings=safety_settings,
|
||||
tools=tools,
|
||||
)
|
||||
|
||||
if tools is not None and bool(
|
||||
getattr(response.candidates[0].content.parts[0], "function_call", None)
|
||||
):
|
||||
function_call = response.candidates[0].content.parts[0].function_call
|
||||
args_dict = {}
|
||||
|
||||
# Check if it's a RepeatedComposite instance
|
||||
for key, val in function_call.args.items():
|
||||
if isinstance(
|
||||
val, proto.marshal.collections.repeated.RepeatedComposite # type: ignore
|
||||
):
|
||||
# If so, convert to list
|
||||
args_dict[key] = [v for v in val]
|
||||
else:
|
||||
args_dict[key] = val
|
||||
|
||||
try:
|
||||
args_str = json.dumps(args_dict)
|
||||
except Exception as e:
|
||||
raise VertexAIError(status_code=422, message=str(e))
|
||||
message = litellm.Message(
|
||||
content=None,
|
||||
tool_calls=[
|
||||
{
|
||||
"id": f"call_{str(uuid.uuid4())}",
|
||||
"function": {
|
||||
"arguments": args_str,
|
||||
"name": function_call.name,
|
||||
},
|
||||
"type": "function",
|
||||
}
|
||||
],
|
||||
)
|
||||
completion_response = message
|
||||
else:
|
||||
completion_response = response.text
|
||||
response_obj = response._raw_response
|
||||
optional_params["tools"] = tools
|
||||
elif mode == "chat":
|
||||
if mode == "chat":
|
||||
chat = llm_model.start_chat()
|
||||
request_str += "chat = llm_model.start_chat()\n"
|
||||
|
||||
|
|
@ -796,82 +533,7 @@ async def async_completion( # noqa: PLR0915
|
|||
|
||||
response_obj = None
|
||||
completion_response = None
|
||||
if mode == "vision":
|
||||
print_verbose("\nMaking VertexAI Gemini Pro/Vision Call")
|
||||
print_verbose(f"\nProcessing input messages = {messages}")
|
||||
tools = optional_params.pop("tools", None)
|
||||
optional_params.pop("stream", False)
|
||||
|
||||
content = _gemini_convert_messages_with_history(messages=messages)
|
||||
|
||||
request_str += f"response = llm_model.generate_content({content})\n"
|
||||
## LOGGING
|
||||
logging_obj.pre_call(
|
||||
input=prompt,
|
||||
api_key=None,
|
||||
additional_args={
|
||||
"complete_input_dict": optional_params,
|
||||
"request_str": request_str,
|
||||
},
|
||||
)
|
||||
|
||||
## LLM Call
|
||||
# print(f"final content: {content}")
|
||||
response = await llm_model._generate_content_async(
|
||||
contents=content,
|
||||
generation_config=optional_params,
|
||||
safety_settings=safety_settings,
|
||||
tools=tools,
|
||||
)
|
||||
|
||||
_cache_key = _get_client_cache_key(
|
||||
model=model,
|
||||
vertex_project=vertex_project,
|
||||
vertex_location=vertex_location,
|
||||
)
|
||||
_set_client_in_cache(
|
||||
client_cache_key=_cache_key, vertex_llm_model=llm_model
|
||||
)
|
||||
|
||||
if tools is not None and bool(
|
||||
getattr(response.candidates[0].content.parts[0], "function_call", None)
|
||||
):
|
||||
function_call = response.candidates[0].content.parts[0].function_call
|
||||
args_dict = {}
|
||||
|
||||
# Check if it's a RepeatedComposite instance
|
||||
for key, val in function_call.args.items():
|
||||
if isinstance(
|
||||
val, proto.marshal.collections.repeated.RepeatedComposite # type: ignore
|
||||
):
|
||||
# If so, convert to list
|
||||
args_dict[key] = [v for v in val]
|
||||
else:
|
||||
args_dict[key] = val
|
||||
|
||||
try:
|
||||
args_str = json.dumps(args_dict)
|
||||
except Exception as e:
|
||||
raise VertexAIError(status_code=422, message=str(e))
|
||||
message = litellm.Message(
|
||||
content=None,
|
||||
tool_calls=[
|
||||
{
|
||||
"id": f"call_{str(uuid.uuid4())}",
|
||||
"function": {
|
||||
"arguments": args_str,
|
||||
"name": function_call.name,
|
||||
},
|
||||
"type": "function",
|
||||
}
|
||||
],
|
||||
)
|
||||
completion_response = message
|
||||
else:
|
||||
completion_response = response.text
|
||||
response_obj = response._raw_response
|
||||
optional_params["tools"] = tools
|
||||
elif mode == "chat":
|
||||
if mode == "chat":
|
||||
# chat-bison etc.
|
||||
chat = llm_model.start_chat()
|
||||
## LOGGING
|
||||
|
|
@ -1032,32 +694,7 @@ async def async_streaming( # noqa: PLR0915
|
|||
Add support for async streaming calls for gemini-pro
|
||||
"""
|
||||
response: Any = None
|
||||
if mode == "vision":
|
||||
stream = optional_params.pop("stream")
|
||||
tools = optional_params.pop("tools", None)
|
||||
print_verbose("\nMaking VertexAI Gemini Pro Vision Call")
|
||||
print_verbose(f"\nProcessing input messages = {messages}")
|
||||
|
||||
content = _gemini_convert_messages_with_history(messages=messages)
|
||||
|
||||
request_str += f"response = llm_model.generate_content({content}, generation_config=GenerationConfig(**{optional_params}), stream={stream})\n"
|
||||
logging_obj.pre_call(
|
||||
input=prompt,
|
||||
api_key=None,
|
||||
additional_args={
|
||||
"complete_input_dict": optional_params,
|
||||
"request_str": request_str,
|
||||
},
|
||||
)
|
||||
|
||||
response = await llm_model._generate_content_streaming_async(
|
||||
contents=content,
|
||||
generation_config=optional_params,
|
||||
safety_settings=safety_settings,
|
||||
tools=tools,
|
||||
)
|
||||
|
||||
elif mode == "chat":
|
||||
if mode == "chat":
|
||||
chat = llm_model.start_chat()
|
||||
optional_params.pop(
|
||||
"stream", None
|
||||
|
|
|
|||
|
|
@ -15,7 +15,6 @@ import json
|
|||
import os
|
||||
import random
|
||||
import sys
|
||||
import threading
|
||||
import time
|
||||
import traceback
|
||||
import uuid
|
||||
|
|
|
|||
|
|
@ -10,10 +10,9 @@ model_list:
|
|||
output_cost_per_token: 0.000015 # 15$/M
|
||||
api_base: "https://exampleopenaiendpoint-production.up.railway.app"
|
||||
api_key: my-fake-key
|
||||
- model_name: my-custom-model
|
||||
- model_name: gemini-1.5-flash-002
|
||||
litellm_params:
|
||||
model: my-custom-llm/my-custom-model
|
||||
api_key: my-fake-key
|
||||
model: gemini/gemini-1.5-flash-002
|
||||
|
||||
litellm_settings:
|
||||
fallbacks: [{ "claude-3-5-sonnet-20240620": ["claude-3-5-sonnet-aihubmix"] }]
|
||||
|
|
|
|||
|
|
@ -97,9 +97,9 @@ class PassThroughEndpointLogging:
|
|||
if "generateContent" in url_route:
|
||||
model = self.extract_model_from_url(url_route)
|
||||
|
||||
instance_of_vertex_llm = VertexLLM()
|
||||
instance_of_vertex_llm = litellm.VertexGeminiConfig()
|
||||
litellm_model_response: litellm.ModelResponse = (
|
||||
instance_of_vertex_llm._process_response(
|
||||
instance_of_vertex_llm._transform_response(
|
||||
model=model,
|
||||
messages=[
|
||||
{"role": "user", "content": "no-message-pass-through-endpoint"}
|
||||
|
|
|
|||
|
|
@ -1,33 +0,0 @@
|
|||
import os
|
||||
import sys
|
||||
import time
|
||||
|
||||
import pytest
|
||||
from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
|
||||
from langtrace_python_sdk import langtrace
|
||||
|
||||
import litellm
|
||||
|
||||
sys.path.insert(0, os.path.abspath("../.."))
|
||||
|
||||
|
||||
@pytest.fixture()
|
||||
def exporter():
|
||||
exporter = InMemorySpanExporter()
|
||||
langtrace.init(batch=False, custom_remote_exporter=exporter)
|
||||
litellm.success_callback = ["langtrace"]
|
||||
litellm.set_verbose = True
|
||||
|
||||
return exporter
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model", ["claude-2.1", "gpt-3.5-turbo"])
|
||||
def test_langtrace_logging(exporter, model):
|
||||
litellm.completion(
|
||||
model=model,
|
||||
messages=[{"role": "user", "content": "This is a test"}],
|
||||
max_tokens=1000,
|
||||
temperature=0.7,
|
||||
timeout=5,
|
||||
mock_response="hi",
|
||||
)
|
||||
|
|
@ -500,9 +500,9 @@ ChatCompletionAssistantContentValue = (
|
|||
|
||||
class ChatCompletionResponseMessage(TypedDict, total=False):
|
||||
content: Optional[ChatCompletionAssistantContentValue]
|
||||
tool_calls: List[ChatCompletionToolCallChunk]
|
||||
tool_calls: Optional[List[ChatCompletionToolCallChunk]]
|
||||
role: Literal["assistant"]
|
||||
function_call: ChatCompletionToolCallFunctionChunk
|
||||
function_call: Optional[ChatCompletionToolCallFunctionChunk]
|
||||
|
||||
|
||||
class ChatCompletionUsageBlock(TypedDict):
|
||||
|
|
|
|||
|
|
@ -167,6 +167,8 @@ class GenerationConfig(TypedDict, total=False):
|
|||
response_mime_type: Literal["text/plain", "application/json"]
|
||||
response_schema: dict
|
||||
seed: int
|
||||
responseLogprobs: bool
|
||||
logprobs: int
|
||||
|
||||
|
||||
class Tools(TypedDict, total=False):
|
||||
|
|
@ -270,6 +272,21 @@ class GroundingMetadata(TypedDict, total=False):
|
|||
groundingAttributions: List[dict]
|
||||
|
||||
|
||||
class LogprobsCandidate(TypedDict):
|
||||
token: str
|
||||
tokenId: int
|
||||
logProbability: float
|
||||
|
||||
|
||||
class LogprobsTopCandidate(TypedDict):
|
||||
candidates: List[LogprobsCandidate]
|
||||
|
||||
|
||||
class LogprobsResult(TypedDict, total=False):
|
||||
topCandidates: List[LogprobsTopCandidate]
|
||||
chosenCandidates: List[LogprobsCandidate]
|
||||
|
||||
|
||||
class Candidates(TypedDict, total=False):
|
||||
index: int
|
||||
content: HttpxContentType
|
||||
|
|
@ -288,6 +305,7 @@ class Candidates(TypedDict, total=False):
|
|||
citationMetadata: CitationMetadata
|
||||
groundingMetadata: GroundingMetadata
|
||||
finishMessage: str
|
||||
logprobsResult: LogprobsResult
|
||||
|
||||
|
||||
class PromptFeedback(TypedDict):
|
||||
|
|
|
|||
|
|
@ -7708,10 +7708,17 @@ class CustomStreamWrapper:
|
|||
continue
|
||||
## LOGGING
|
||||
## LOGGING
|
||||
threading.Thread(
|
||||
target=self.logging_obj.success_handler,
|
||||
args=(processed_chunk, None, None, cache_hit),
|
||||
).start() # log response
|
||||
executor.submit(
|
||||
self.logging_obj.success_handler,
|
||||
result=processed_chunk,
|
||||
start_time=None,
|
||||
end_time=None,
|
||||
cache_hit=cache_hit,
|
||||
)
|
||||
# threading.Thread(
|
||||
# target=self.logging_obj.success_handler,
|
||||
# args=(processed_chunk, None, None, cache_hit),
|
||||
# ).start() # log response
|
||||
asyncio.create_task(
|
||||
self.logging_obj.async_success_handler(
|
||||
processed_chunk, cache_hit=cache_hit
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue