[Bugfix] GPT-OSS instruction rendering (#44330)

Signed-off-by: Yifan Zong <yzong@redhat.com>
This commit is contained in:
yzong-rh
2026-06-05 13:52:32 -04:00
committed by GitHub
parent b593396c7a
commit 703fb17b13
9 changed files with 276 additions and 112 deletions
+15 -17
View File
@@ -18,8 +18,8 @@ from vllm.entrypoints.openai.engine.protocol import (
)
from vllm.entrypoints.openai.models.serving import OpenAIModelRegistry
from vllm.entrypoints.openai.parser.harmony_utils import (
get_developer_message,
get_system_message,
build_harmony_preamble,
extract_instructions_from_messages,
parse_chat_inputs_to_harmony_messages,
render_for_completion,
)
@@ -405,6 +405,9 @@ class OpenAIServingRender:
# for more info: see comment in `maybe_serialize_tool_calls`
_mt.maybe_serialize_tool_calls(request) # type: ignore[arg-type]
chat_messages = list(request.messages)
instructions, chat_messages = extract_instructions_from_messages(chat_messages)
# Add system message.
# NOTE: In Chat Completion API, browsing is enabled by default
# if the model supports it. TODO: Support browsing.
@@ -412,23 +415,18 @@ class OpenAIServingRender:
assert not self.supports_code_interpreter
if (reasoning_effort := request.reasoning_effort) == "none":
raise ValueError(f"Harmony does not support {reasoning_effort=}")
sys_msg = get_system_message(
reasoning_effort=reasoning_effort,
browser_description=None,
python_description=None,
with_custom_tools=should_include_tools,
)
messages.append(sys_msg)
# Add developer message.
if request.tools:
dev_msg = get_developer_message(
tools=request.tools if should_include_tools else None # type: ignore[arg-type]
tools = request.tools if should_include_tools else None
messages.extend(
build_harmony_preamble(
instructions=instructions,
tools=tools, # type: ignore[arg-type]
reasoning_effort=reasoning_effort,
with_custom_tools=should_include_tools,
)
messages.append(dev_msg)
)
# Add user message.
messages.extend(parse_chat_inputs_to_harmony_messages(request.messages))
# Add remaining conversation messages.
messages.extend(parse_chat_inputs_to_harmony_messages(chat_messages))
# Render prompt token ids.
prompt_token_ids = render_for_completion(messages)