[Renderer] Warm up the renderer properly. (#50408)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
@@ -140,8 +140,6 @@ async def init_generate_state(
|
||||
if "generate" in supported_tasks
|
||||
else None
|
||||
)
|
||||
if state.openai_serving_chat is not None:
|
||||
state.openai_serving_chat.warmup()
|
||||
state.openai_serving_completion = (
|
||||
OpenAIServingCompletion(
|
||||
engine_client,
|
||||
|
||||
@@ -55,6 +55,7 @@ from vllm.v1.engine import PauseMode
|
||||
from vllm.v1.engine.llm_engine import LLMEngine
|
||||
from vllm.v1.sample.logits_processor import LogitsProcessor
|
||||
|
||||
from ..renderers import ChatParams
|
||||
from .offline_utils import _O, _R, OfflineInferenceMixin
|
||||
|
||||
if TYPE_CHECKING:
|
||||
@@ -352,6 +353,8 @@ class LLM(BeamSearchOfflineMixin, PoolingOfflineMixin, OfflineInferenceMixin):
|
||||
self.chat_template = load_chat_template(chat_template)
|
||||
self.input_processor = self.llm_engine.input_processor
|
||||
|
||||
self.renderer.warmup(ChatParams(chat_template=self.chat_template))
|
||||
|
||||
# The renderer thread pool is only consumed by the async renderer
|
||||
# path; the synchronous `LLM` entrypoint runs multimodal
|
||||
# preprocessing serially. Warn so the setting is not a silent
|
||||
|
||||
@@ -426,6 +426,7 @@ async def init_app_state(
|
||||
default_chat_template_kwargs=args.default_chat_template_kwargs,
|
||||
log_error_stack=args.log_error_stack,
|
||||
)
|
||||
state.online_renderer.warmup()
|
||||
|
||||
state.online_derenderer = OnlineDerenderer(
|
||||
model_config=engine_client.model_config,
|
||||
@@ -529,6 +530,7 @@ async def init_render_app_state(
|
||||
default_chat_template_kwargs=args.default_chat_template_kwargs,
|
||||
log_error_stack=args.log_error_stack,
|
||||
)
|
||||
state.online_renderer.warmup()
|
||||
|
||||
state.online_derenderer = OnlineDerenderer(
|
||||
model_config=vllm_config.model_config,
|
||||
|
||||
@@ -60,7 +60,6 @@ from vllm.logprobs import Logprob
|
||||
from vllm.outputs import RequestOutput
|
||||
from vllm.parser import ParserManager
|
||||
from vllm.parser.abstract_parser import Parser
|
||||
from vllm.renderers import ChatParams
|
||||
from vllm.renderers.online_renderer import OnlineRenderer
|
||||
from vllm.sampling_params import BeamSearchParams, SamplingParams
|
||||
from vllm.tokenizers import TokenizerLike
|
||||
@@ -178,15 +177,6 @@ class OpenAIServingChat(GenerateBaseServing):
|
||||
self.supports_code_interpreter = False
|
||||
self.python_tool = None
|
||||
|
||||
def warmup(self) -> None:
|
||||
self.renderer.warmup(
|
||||
ChatParams(
|
||||
chat_template=self.chat_template,
|
||||
chat_template_content_format=self.chat_template_content_format,
|
||||
chat_template_kwargs=self.default_chat_template_kwargs,
|
||||
)
|
||||
)
|
||||
|
||||
def _effective_chat_template_kwargs(
|
||||
self, request: ChatCompletionRequest
|
||||
) -> dict[str, Any]:
|
||||
|
||||
Reference in New Issue
Block a user