biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

150
vllm/entrypoints/harmony_utils.py
View File

@@ -8,23 +8,44 @@ import json
from collections.abc import Iterable, Sequence
from typing import Literal, Optional, Union
from openai.types.responses import (ResponseFunctionToolCall,
ResponseOutputItem, ResponseOutputMessage,
ResponseOutputText, ResponseReasoningItem)
from openai.types.responses import (
ResponseFunctionToolCall,
ResponseOutputItem,
ResponseOutputMessage,
ResponseOutputText,
ResponseReasoningItem,
)
from openai.types.responses.response_function_web_search import (
ActionFind, ActionOpenPage, ActionSearch, ResponseFunctionWebSearch)
ActionFind,
ActionOpenPage,
ActionSearch,
ResponseFunctionWebSearch,
)
from openai.types.responses.response_reasoning_item import (
Content as ResponseReasoningTextContent)
Content as ResponseReasoningTextContent,
)
from openai.types.responses.tool import Tool
from openai_harmony import (Author, ChannelConfig, Conversation,
DeveloperContent, HarmonyEncodingName, Message,
ReasoningEffort, Role, StreamableParser,
SystemContent, TextContent, ToolDescription,
load_harmony_encoding)
from openai_harmony import (
Author,
ChannelConfig,
Conversation,
DeveloperContent,
HarmonyEncodingName,
Message,
ReasoningEffort,
Role,
StreamableParser,
SystemContent,
TextContent,
ToolDescription,
load_harmony_encoding,
)
from vllm import envs
from vllm.entrypoints.openai.protocol import (ChatCompletionToolsParam,
ResponseInputOutputItem)
from vllm.entrypoints.openai.protocol import (
ChatCompletionToolsParam,
ResponseInputOutputItem,
)
from vllm.utils import random_uuid
REASONING_EFFORT = {
@@ -53,8 +74,7 @@ def has_custom_tools(tool_types: list[str]) -> bool:
def get_encoding():
global _harmony_encoding
if _harmony_encoding is None:
_harmony_encoding = load_harmony_encoding(
HarmonyEncodingName.HARMONY_GPT_OSS)
_harmony_encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
return _harmony_encoding
@@ -71,15 +91,16 @@ def get_system_message(
sys_msg_content = SystemContent.new()
if model_identity is not None:
sys_msg_content = sys_msg_content.with_model_identity(model_identity)
if (instructions is not None
and envs.VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS):
if instructions is not None and envs.VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS:
current_identity = sys_msg_content.model_identity
new_identity = (f'{current_identity}\n{instructions}'
if current_identity else instructions)
new_identity = (
f"{current_identity}\n{instructions}" if current_identity else instructions
)
sys_msg_content = sys_msg_content.with_model_identity(new_identity)
if reasoning_effort is not None:
sys_msg_content = sys_msg_content.with_reasoning_effort(
REASONING_EFFORT[reasoning_effort])
REASONING_EFFORT[reasoning_effort]
)
if start_date is None:
# NOTE(woosuk): This brings non-determinism in vLLM. Be careful.
start_date = datetime.datetime.now().strftime("%Y-%m-%d")
@@ -94,7 +115,8 @@ def get_system_message(
channel_config = sys_msg_content.channel_config
invalid_channel = "commentary"
new_config = ChannelConfig.require_channels(
[c for c in channel_config.valid_channels if c != invalid_channel])
[c for c in channel_config.valid_channels if c != invalid_channel]
)
sys_msg_content = sys_msg_content.with_channel_config(new_config)
sys_msg = Message.from_role_and_content(Role.SYSTEM, sys_msg_content)
return sys_msg
@@ -119,14 +141,17 @@ def get_developer_message(
tools: Optional[list[Union[Tool, ChatCompletionToolsParam]]] = None,
) -> Message:
dev_msg_content = DeveloperContent.new()
if (instructions is not None
and not envs.VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS):
if instructions is not None and not envs.VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS:
dev_msg_content = dev_msg_content.with_instructions(instructions)
if tools is not None:
function_tools: list[Union[Tool, ChatCompletionToolsParam]] = []
for tool in tools:
if tool.type in ("web_search_preview", "code_interpreter",
"container", "mcp"):
if tool.type in (
"web_search_preview",
"code_interpreter",
"container",
"mcp",
):
# These are built-in tools that are added to the system message.
# Adding in MCP for now until we support MCP tools executed
# server side
@@ -141,7 +166,8 @@ def get_developer_message(
create_tool_definition(tool) for tool in function_tools
]
dev_msg_content = dev_msg_content.with_function_tools(
function_tool_descriptions)
function_tool_descriptions
)
dev_msg = Message.from_role_and_content(Role.DEVELOPER, dev_msg_content)
return dev_msg
@@ -152,7 +178,7 @@ def get_user_message(content: str) -> Message:
def parse_response_input(
response_msg: ResponseInputOutputItem,
prev_responses: list[Union[ResponseOutputItem, ResponseReasoningItem]]
prev_responses: list[Union[ResponseOutputItem, ResponseReasoningItem]],
) -> Message:
if not isinstance(response_msg, dict):
response_msg = response_msg.model_dump()
@@ -170,9 +196,7 @@ def parse_response_input(
if isinstance(content, str):
msg = Message.from_role_and_content(role, text_prefix + content)
else:
contents = [
TextContent(text=text_prefix + c["text"]) for c in content
]
contents = [TextContent(text=text_prefix + c["text"]) for c in content]
msg = Message.from_role_and_contents(role, contents)
if role == "assistant":
msg = msg.with_channel("final")
@@ -180,22 +204,24 @@ def parse_response_input(
call_id = response_msg["call_id"]
call_response: Optional[ResponseFunctionToolCall] = None
for prev_response in reversed(prev_responses):
if isinstance(prev_response, ResponseFunctionToolCall
) and prev_response.call_id == call_id:
if (
isinstance(prev_response, ResponseFunctionToolCall)
and prev_response.call_id == call_id
):
call_response = prev_response
break
if call_response is None:
raise ValueError(f"No call message found for {call_id}")
msg = Message.from_author_and_content(
Author.new(Role.TOOL, f"functions.{call_response.name}"),
response_msg["output"])
response_msg["output"],
)
elif response_msg["type"] == "reasoning":
content = response_msg["content"]
assert len(content) == 1
msg = Message.from_role_and_content(Role.ASSISTANT, content[0]["text"])
elif response_msg["type"] == "function_call":
msg = Message.from_role_and_content(Role.ASSISTANT,
response_msg["arguments"])
msg = Message.from_role_and_content(Role.ASSISTANT, response_msg["arguments"])
msg = msg.with_channel("commentary")
msg = msg.with_recipient(f"functions.{response_msg['name']}")
msg = msg.with_content_type("json")
@@ -231,8 +257,8 @@ def parse_chat_input(chat_msg) -> list[Message]:
name = chat_msg.get("name", "")
content = chat_msg.get("content", "") or ""
msg = Message.from_author_and_content(
Author.new(Role.TOOL, f"functions.{name}"),
content).with_channel("commentary")
Author.new(Role.TOOL, f"functions.{name}"), content
).with_channel("commentary")
return [msg]
# Default: user/assistant/system messages with content
@@ -249,7 +275,8 @@ def parse_chat_input(chat_msg) -> list[Message]:
def render_for_completion(messages: list[Message]) -> list[int]:
conversation = Conversation.from_messages(messages)
token_ids = get_encoding().render_conversation_for_completion(
conversation, Role.ASSISTANT)
conversation, Role.ASSISTANT
)
return token_ids
@@ -273,14 +300,18 @@ def parse_output_message(message: Message) -> list[ResponseOutputItem]:
# TODO: translate to url properly!
if recipient == "browser.search":
action = ActionSearch(
query=f"cursor:{browser_call.get('query', '')}", type="search")
query=f"cursor:{browser_call.get('query', '')}", type="search"
)
elif recipient == "browser.open":
action = ActionOpenPage(
url=f"cursor:{browser_call.get('url', '')}", type="open_page")
url=f"cursor:{browser_call.get('url', '')}", type="open_page"
)
elif recipient == "browser.find":
action = ActionFind(pattern=browser_call["pattern"],
url=f"cursor:{browser_call.get('url', '')}",
type="find")
action = ActionFind(
pattern=browser_call["pattern"],
url=f"cursor:{browser_call.get('url', '')}",
type="find",
)
else:
raise ValueError(f"Unknown browser action: {recipient}")
web_search_item = ResponseFunctionWebSearch(
@@ -297,8 +328,9 @@ def parse_output_message(message: Message) -> list[ResponseOutputItem]:
summary=[],
type="reasoning",
content=[
ResponseReasoningTextContent(text=content.text,
type="reasoning_text")
ResponseReasoningTextContent(
text=content.text, type="reasoning_text"
)
],
status=None,
)
@@ -316,17 +348,20 @@ def parse_output_message(message: Message) -> list[ResponseOutputItem]:
id=f"fc_{random_id}",
)
output_items.append(response_item)
elif recipient is not None and (recipient.startswith("python")
or recipient.startswith("browser")
or recipient.startswith("container")):
elif recipient is not None and (
recipient.startswith("python")
or recipient.startswith("browser")
or recipient.startswith("container")
):
for content in message.content:
reasoning_item = ResponseReasoningItem(
id=f"rs_{random_uuid()}",
summary=[],
type="reasoning",
content=[
ResponseReasoningTextContent(text=content.text,
type="reasoning_text")
ResponseReasoningTextContent(
text=content.text, type="reasoning_text"
)
],
status=None,
)
@@ -356,15 +391,13 @@ def parse_output_message(message: Message) -> list[ResponseOutputItem]:
return output_items
def parse_remaining_state(
parser: StreamableParser) -> list[ResponseOutputItem]:
def parse_remaining_state(parser: StreamableParser) -> list[ResponseOutputItem]:
if not parser.current_content:
return []
if parser.current_role != Role.ASSISTANT:
return []
current_recipient = parser.current_recipient
if (current_recipient is not None
and current_recipient.startswith("browser.")):
if current_recipient is not None and current_recipient.startswith("browser."):
return []
if parser.current_channel == "analysis":
@@ -373,8 +406,9 @@ def parse_remaining_state(
summary=[],
type="reasoning",
content=[
ResponseReasoningTextContent(text=parser.current_content,
type="reasoning_text")
ResponseReasoningTextContent(
text=parser.current_content, type="reasoning_text"
)
],
status=None,
)
@@ -415,7 +449,8 @@ def parse_output_into_messages(token_ids: Iterable[int]) -> StreamableParser:
def parse_chat_output(
token_ids: Sequence[int]) -> tuple[Optional[str], Optional[str], bool]:
token_ids: Sequence[int],
) -> tuple[Optional[str], Optional[str], bool]:
parser = parse_output_into_messages(token_ids)
output_msgs = parser.messages
is_tool_call = False # TODO: update this when tool call is supported
@@ -430,7 +465,6 @@ def parse_chat_output(
else:
reasoning_msg = output_msgs[:-1]
final_msg = output_msgs[-1]
reasoning_content = "\n".join(
[msg.content[0].text for msg in reasoning_msg])
reasoning_content = "\n".join([msg.content[0].text for msg in reasoning_msg])
final_content = final_msg.content[0].text
return reasoning_content, final_content, is_tool_call