[Bugfix][Frontend] Apply model-default reasoning parser in GPU-less render server (#54835)
Signed-off-by: Shimi Bandiel <shimib@google.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
co-authored by
Claude
Cyrus Leung
parent
68088ed392
commit
09e0ce3d82
@@ -50,6 +50,16 @@ async def init_render_app_state(
|
||||
default_chat_template_kwargs = resolve_default_chat_template_kwargs(args)
|
||||
state.tool_server = await init_tool_server(args)
|
||||
|
||||
# The config-resolved reasoning parser carries both the CLI flag (merged
|
||||
# in the entrypoint via `create_structured_outputs_config`) and any
|
||||
# model-specific default applied by `verify_and_update_config`
|
||||
# (e.g. "openai_gptoss" for gpt_oss), matching the main API server.
|
||||
# Keep `args.reasoning_parser` as the first source so callers that build
|
||||
# a VllmConfig without that merge still honor an explicit flag.
|
||||
reasoning_parser = (
|
||||
args.reasoning_parser or vllm_config.structured_outputs_config.reasoning_parser
|
||||
)
|
||||
|
||||
state.online_renderer = OnlineRenderer(
|
||||
model_config=vllm_config.model_config,
|
||||
renderer=renderer,
|
||||
@@ -62,7 +72,7 @@ async def init_render_app_state(
|
||||
exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none,
|
||||
tool_parser=args.tool_call_parser,
|
||||
tool_strict_level=args.tool_strict_level,
|
||||
reasoning_parser=args.reasoning_parser,
|
||||
reasoning_parser=reasoning_parser,
|
||||
default_chat_template_kwargs=default_chat_template_kwargs,
|
||||
log_error_stack=args.log_error_stack,
|
||||
)
|
||||
@@ -79,7 +89,7 @@ async def init_render_app_state(
|
||||
exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none,
|
||||
tool_parser=args.tool_call_parser,
|
||||
tool_strict_level=args.tool_strict_level,
|
||||
reasoning_parser=args.reasoning_parser,
|
||||
reasoning_parser=reasoning_parser,
|
||||
default_chat_template_kwargs=default_chat_template_kwargs,
|
||||
log_error_stack=args.log_error_stack,
|
||||
)
|
||||
|
||||
@@ -87,9 +87,14 @@ async def run_launch_fastapi(args: argparse.Namespace) -> None:
|
||||
# cache space warning from CpuPlatform.check_and_update_config.
|
||||
envs.VLLM_CPU_KVCACHE_SPACE = 0
|
||||
|
||||
# Merge frontend parser flags (e.g. --reasoning-parser) into the
|
||||
# structured outputs config, mirroring `create_engine_config`, which this
|
||||
# GPU-less entrypoint bypasses. Model-specific defaults are then applied
|
||||
# by `verify_and_update_config` during VllmConfig construction.
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
logging_config=engine_args.create_logging_config(),
|
||||
structured_outputs_config=engine_args.create_structured_outputs_config(),
|
||||
)
|
||||
shutdown_task = await build_and_serve_renderer(
|
||||
vllm_config, listen_address, sock, args
|
||||
|
||||
Reference in New Issue
Block a user