Skip to content

GitLab

"docs/vscode:/vscode.git/clone" did not exist on "c9e4c75d09c158bbe06765fead210abeca7f051d"
Unverified Commit 887c2b45 authored by Chang Su's avatar Chang Su Committed by GitHub
Browse files

[router][grpc] Add `serve_grpc` to `launch_server` and log id for HealthCheck (#11564)

parent 065ce815
Hide whitespace changes
Inline Side-by-side
Showing with 68 additions and 93 deletions
python/sglang/launch_server.py
View file @ 887c2b45
"""Launch the inference server."""
import asyncio
import os
import sys
from sglang.srt.entrypoints.http_server import launch_server
from sglang.srt.server_args import prepare_server_args
from sglang.srt.utils import kill_process_tree
......@@ -11,6 +11,13 @@ if __name__ == "__main__":
server_args = prepare_server_args(sys.argv[1:])
try:
launch_server(server_args)
if server_args.grpc_mode:
from sglang.srt.entrypoints.grpc_server import serve_grpc
asyncio.run(serve_grpc(server_args))
else:
from sglang.srt.entrypoints.http_server import launch_server
launch_server(server_args)
finally:
kill_process_tree(os.getpid(), include_parent=False)
python/sglang/srt/entrypoints/grpc_server.py
View file @ 887c2b45
......@@ -22,8 +22,8 @@ from grpc_reflection.v1alpha import reflection
import sglang
from sglang.srt.disaggregation.utils import FAKE_BOOTSTRAP_HOST, DisaggregationMode
from sglang.srt.entrypoints.grpc_request_manager import GrpcRequestManager
from sglang.srt.grpc import sglang_scheduler_pb2, sglang_scheduler_pb2_grpc
from sglang.srt.grpc.grpc_request_manager import GrpcRequestManager
from sglang.srt.managers.data_parallel_controller import (
run_data_parallel_controller_process,
)
......@@ -68,6 +68,8 @@ def _launch_scheduler_process_only(
# Configure global environment
configure_logger(server_args)
server_args.check_server_args()
# Fix CUDA multiprocessing issues - must be called before any CUDA operations
mp.set_start_method("spawn", force=True)
# Allocate ports for inter-process communications
if port_args is None:
......@@ -317,7 +319,8 @@ class SGLangSchedulerServicer(sglang_scheduler_pb2_grpc.SglangSchedulerServicer)
Check the health of the inference server by sending a special request to generate one token.
Similar to HTTP server's /health endpoint.
"""
logger.info("Receive health check request")
rid = f"HEALTH_CHECK_{time.time()}"
logger.info(f"Receive health check request: {rid}")
if self.request_manager.gracefully_exit:
logger.info(
......@@ -328,7 +331,6 @@ class SGLangSchedulerServicer(sglang_scheduler_pb2_grpc.SglangSchedulerServicer)
)
# Create a special health check request
rid = f"HEALTH_CHECK_{time.time()}"
sampling_params = SGLSamplingParams(max_new_tokens=1, temperature=0.0)
sampling_params.normalize(tokenizer=None)
......@@ -919,25 +921,3 @@ async def serve_grpc(
proc.join(timeout=1.0)
logger.info("All scheduler processes terminated")
def main():
"""Main entry point for standalone gRPC server."""
# Fix CUDA multiprocessing issues - must be called before any CUDA operations
mp.set_start_method("spawn", force=True)
parser = argparse.ArgumentParser(description="SGLang Standalone gRPC Server")
ServerArgs.add_cli_args(parser)
args = parser.parse_args()
server_args = ServerArgs.from_cli_args(args)
# Run server
asyncio.run(
serve_grpc(
server_args=server_args,
)
)
if __name__ == "__main__":
main()
python/sglang/srt/grpc/sglang_scheduler.proto
View file @ 887c2b45
......@@ -326,10 +326,7 @@ message EmbedError {
// Management Operations
// =====================
message HealthCheckRequest {
// Input for health test generation (must be tokenized)
TokenizedInput tokenized = 1;
}
message HealthCheckRequest {}
message HealthCheckResponse {
bool healthy = 1;
......
python/sglang/srt/grpc/sglang_scheduler_pb2.py
View file @ 887c2b45
......@@ -29,7 +29,7 @@ from google.protobuf import timestamp_pb2 as google_dot_protobuf_dot_timestamp__
from google.protobuf import struct_pb2 as google_dot_protobuf_dot_struct__pb2
DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n\x16sglang_scheduler.proto\x12\x15sglang.grpc.scheduler\x1a\x1fgoogle/protobuf/timestamp.proto\x1a\x1cgoogle/protobuf/struct.proto\"\xd0\x05\n\x0eSamplingParams\x12\x13\n\x0btemperature\x18\x01 \x01(\x02\x12\r\n\x05top_p\x18\x02 \x01(\x02\x12\r\n\x05top_k\x18\x03 \x01(\x05\x12\r\n\x05min_p\x18\x04 \x01(\x02\x12\x19\n\x11\x66requency_penalty\x18\x05 \x01(\x02\x12\x18\n\x10presence_penalty\x18\x06 \x01(\x02\x12\x1a\n\x12repetition_penalty\x18\x07 \x01(\x02\x12\x1b\n\x0emax_new_tokens\x18\x08 \x01(\x05H\x01\x88\x01\x01\x12\x0c\n\x04stop\x18\t \x03(\t\x12\x16\n\x0estop_token_ids\x18\n \x03(\r\x12\x1b\n\x13skip_special_tokens\x18\x0b \x01(\x08\x12%\n\x1dspaces_between_special_tokens\x18\x0c \x01(\x08\x12\x0f\n\x05regex\x18\r \x01(\tH\x00\x12\x15\n\x0bjson_schema\x18\x0e \x01(\tH\x00\x12\x16\n\x0c\x65\x62nf_grammar\x18\x0f \x01(\tH\x00\x12\x18\n\x0estructural_tag\x18\x10 \x01(\tH\x00\x12\t\n\x01n\x18\x11 \x01(\x05\x12\x16\n\x0emin_new_tokens\x18\x12 \x01(\x05\x12\x12\n\nignore_eos\x18\x13 \x01(\x08\x12\x14\n\x0cno_stop_trim\x18\x14 \x01(\x08\x12\x1c\n\x0fstream_interval\x18\x15 \x01(\x05H\x02\x88\x01\x01\x12H\n\nlogit_bias\x18\x16 \x03(\x0b\x32\x34.sglang.grpc.scheduler.SamplingParams.LogitBiasEntry\x12.\n\rcustom_params\x18\x17 \x01(\x0b\x32\x17.google.protobuf.Struct\x1a\x30\n\x0eLogitBiasEntry\x12\x0b\n\x03key\x18\x01 \x01(\t\x12\r\n\x05value\x18\x02 \x01(\x02:\x02\x38\x01\x42\x0c\n\nconstraintB\x11\n\x0f_max_new_tokensB\x12\n\x10_stream_interval\"]\n\x13\x44isaggregatedParams\x12\x16\n\x0e\x62ootstrap_host\x18\x01 \x01(\t\x12\x16\n\x0e\x62ootstrap_port\x18\x02 \x01(\x05\x12\x16\n\x0e\x62ootstrap_room\x18\x03 \x01(\x05\"\xe2\x04\n\x0fGenerateRequest\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x38\n\ttokenized\x18\x02 \x01(\x0b\x32%.sglang.grpc.scheduler.TokenizedInput\x12:\n\tmm_inputs\x18\x03 \x01(\x0b\x32\'.sglang.grpc.scheduler.MultimodalInputs\x12>\n\x0fsampling_params\x18\x04 \x01(\x0b\x32%.sglang.grpc.scheduler.SamplingParams\x12\x16\n\x0ereturn_logprob\x18\x05 \x01(\x08\x12\x19\n\x11logprob_start_len\x18\x06 \x01(\x05\x12\x18\n\x10top_logprobs_num\x18\x07 \x01(\x05\x12\x19\n\x11token_ids_logprob\x18\x08 \x03(\r\x12\x1c\n\x14return_hidden_states\x18\t \x01(\x08\x12H\n\x14\x64isaggregated_params\x18\n \x01(\x0b\x32*.sglang.grpc.scheduler.DisaggregatedParams\x12\x1e\n\x16\x63ustom_logit_processor\x18\x0b \x01(\t\x12-\n\ttimestamp\x18\x0c \x01(\x0b\x32\x1a.google.protobuf.Timestamp\x12\x13\n\x0blog_metrics\x18\r \x01(\x08\x12\x14\n\x0cinput_embeds\x18\x0e \x03(\x02\x12\x0f\n\x07lora_id\x18\x0f \x01(\t\x12\x1a\n\x12\x64\x61ta_parallel_rank\x18\x10 \x01(\x05\x12\x0e\n\x06stream\x18\x11 \x01(\x08\":\n\x0eTokenizedInput\x12\x15\n\roriginal_text\x18\x01 \x01(\t\x12\x11\n\tinput_ids\x18\x02 \x03(\r\"\xd3\x01\n\x10MultimodalInputs\x12\x12\n\nimage_urls\x18\x01 \x03(\t\x12\x12\n\nvideo_urls\x18\x02 \x03(\t\x12\x12\n\naudio_urls\x18\x03 \x03(\t\x12\x33\n\x12processed_features\x18\x04 \x01(\x0b\x32\x17.google.protobuf.Struct\x12\x12\n\nimage_data\x18\x05 \x03(\x0c\x12\x12\n\nvideo_data\x18\x06 \x03(\x0c\x12\x12\n\naudio_data\x18\x07 \x03(\x0c\x12\x12\n\nmodalities\x18\x08 \x03(\t\"\xe3\x01\n\x10GenerateResponse\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12;\n\x05\x63hunk\x18\x02 \x01(\x0b\x32*.sglang.grpc.scheduler.GenerateStreamChunkH\x00\x12;\n\x08\x63omplete\x18\x03 \x01(\x0b\x32\'.sglang.grpc.scheduler.GenerateCompleteH\x00\x12\x35\n\x05\x65rror\x18\x04 \x01(\x0b\x32$.sglang.grpc.scheduler.GenerateErrorH\x00\x42\n\n\x08response\"\x95\x02\n\x13GenerateStreamChunk\x12\x11\n\ttoken_ids\x18\x01 \x03(\r\x12\x15\n\rprompt_tokens\x18\x02 \x01(\x05\x12\x19\n\x11\x63ompletion_tokens\x18\x03 \x01(\x05\x12\x15\n\rcached_tokens\x18\x04 \x01(\x05\x12>\n\x0foutput_logprobs\x18\x05 \x01(\x0b\x32%.sglang.grpc.scheduler.OutputLogProbs\x12\x15\n\rhidden_states\x18\x06 \x03(\x02\x12<\n\x0einput_logprobs\x18\x07 \x01(\x0b\x32$.sglang.grpc.scheduler.InputLogProbs\x12\r\n\x05index\x18\x08 \x01(\r\"\x9b\x03\n\x10GenerateComplete\x12\x12\n\noutput_ids\x18\x01 \x03(\r\x12\x15\n\rfinish_reason\x18\x02 \x01(\t\x12\x15\n\rprompt_tokens\x18\x03 \x01(\x05\x12\x19\n\x11\x63ompletion_tokens\x18\x04 \x01(\x05\x12\x15\n\rcached_tokens\x18\x05 \x01(\x05\x12>\n\x0foutput_logprobs\x18\x06 \x01(\x0b\x32%.sglang.grpc.scheduler.OutputLogProbs\x12>\n\x11\x61ll_hidden_states\x18\x07 \x03(\x0b\x32#.sglang.grpc.scheduler.HiddenStates\x12\x1a\n\x10matched_token_id\x18\x08 \x01(\rH\x00\x12\x1a\n\x10matched_stop_str\x18\t \x01(\tH\x00\x12<\n\x0einput_logprobs\x18\n \x01(\x0b\x32$.sglang.grpc.scheduler.InputLogProbs\x12\r\n\x05index\x18\x0b \x01(\rB\x0e\n\x0cmatched_stop\"K\n\rGenerateError\x12\x0f\n\x07message\x18\x01 \x01(\t\x12\x18\n\x10http_status_code\x18\x02 \x01(\t\x12\x0f\n\x07\x64\x65tails\x18\x03 \x01(\t\"u\n\x0eOutputLogProbs\x12\x16\n\x0etoken_logprobs\x18\x01 \x03(\x02\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\x12\x38\n\x0ctop_logprobs\x18\x03 \x03(\x0b\x32\".sglang.grpc.scheduler.TopLogProbs\"\x9e\x01\n\rInputLogProbs\x12@\n\x0etoken_logprobs\x18\x01 \x03(\x0b\x32(.sglang.grpc.scheduler.InputTokenLogProb\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\x12\x38\n\x0ctop_logprobs\x18\x03 \x03(\x0b\x32\".sglang.grpc.scheduler.TopLogProbs\"1\n\x11InputTokenLogProb\x12\x12\n\x05value\x18\x01 \x01(\x02H\x00\x88\x01\x01\x42\x08\n\x06_value\"0\n\x0bTopLogProbs\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\"?\n\x0cHiddenStates\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\r\n\x05layer\x18\x02 \x01(\x05\x12\x10\n\x08position\x18\x03 \x01(\x05\"\xca\x02\n\x0c\x45mbedRequest\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x38\n\ttokenized\x18\x02 \x01(\x0b\x32%.sglang.grpc.scheduler.TokenizedInput\x12:\n\tmm_inputs\x18\x04 \x01(\x0b\x32\'.sglang.grpc.scheduler.MultimodalInputs\x12>\n\x0fsampling_params\x18\x05 \x01(\x0b\x32%.sglang.grpc.scheduler.SamplingParams\x12\x13\n\x0blog_metrics\x18\x06 \x01(\x08\x12\x16\n\x0etoken_type_ids\x18\x07 \x03(\x05\x12\x1a\n\x12\x64\x61ta_parallel_rank\x18\x08 \x01(\x05\x12\x18\n\x10is_cross_encoder\x18\t \x01(\x08\x12\r\n\x05texts\x18\n \x03(\t\"\x9d\x01\n\rEmbedResponse\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x38\n\x08\x63omplete\x18\x02 \x01(\x0b\x32$.sglang.grpc.scheduler.EmbedCompleteH\x00\x12\x32\n\x05\x65rror\x18\x03 \x01(\x0b\x32!.sglang.grpc.scheduler.EmbedErrorH\x00\x42\n\n\x08response\"\xa3\x01\n\rEmbedComplete\x12\x11\n\tembedding\x18\x01 \x03(\x02\x12\x15\n\rprompt_tokens\x18\x02 \x01(\x05\x12\x15\n\rcached_tokens\x18\x03 \x01(\x05\x12\x15\n\rembedding_dim\x18\x04 \x01(\x05\x12:\n\x10\x62\x61tch_embeddings\x18\x05 \x03(\x0b\x32 .sglang.grpc.scheduler.Embedding\"*\n\tEmbedding\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\r\n\x05index\x18\x02 \x01(\x05\"<\n\nEmbedError\x12\x0f\n\x07message\x18\x01 \x01(\t\x12\x0c\n\x04\x63ode\x18\x02 \x01(\t\x12\x0f\n\x07\x64\x65tails\x18\x03 \x01(\t\"N\n\x12HealthCheckRequest\x12\x38\n\ttokenized\x18\x01 \x01(\x0b\x32%.sglang.grpc.scheduler.TokenizedInput\"7\n\x13HealthCheckResponse\x12\x0f\n\x07healthy\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"2\n\x0c\x41\x62ortRequest\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x0e\n\x06reason\x18\x02 \x01(\t\"1\n\rAbortResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"I\n\x0fLoadLoRARequest\x12\x12\n\nadapter_id\x18\x01 \x01(\t\x12\x14\n\x0c\x61\x64\x61pter_path\x18\x02 \x01(\t\x12\x0c\n\x04rank\x18\x03 \x01(\x05\"H\n\x10LoadLoRAResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x12\n\nadapter_id\x18\x02 \x01(\t\x12\x0f\n\x07message\x18\x03 \x01(\t\"\'\n\x11UnloadLoRARequest\x12\x12\n\nadapter_id\x18\x01 \x01(\t\"6\n\x12UnloadLoRAResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"w\n\x14UpdateWeightsRequest\x12\x13\n\tdisk_path\x18\x01 \x01(\tH\x00\x12\x15\n\x0btensor_data\x18\x02 \x01(\x0cH\x00\x12\x14\n\nremote_url\x18\x03 \x01(\tH\x00\x12\x13\n\x0bweight_name\x18\x04 \x01(\tB\x08\n\x06source\"9\n\x15UpdateWeightsResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"-\n\x17GetInternalStateRequest\x12\x12\n\nstate_keys\x18\x01 \x03(\t\"B\n\x18GetInternalStateResponse\x12&\n\x05state\x18\x01 \x01(\x0b\x32\x17.google.protobuf.Struct\"A\n\x17SetInternalStateRequest\x12&\n\x05state\x18\x01 \x01(\x0b\x32\x17.google.protobuf.Struct\"<\n\x18SetInternalStateResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"\x15\n\x13GetModelInfoRequest\"\xea\x02\n\x14GetModelInfoResponse\x12\x12\n\nmodel_path\x18\x01 \x01(\t\x12\x16\n\x0etokenizer_path\x18\x02 \x01(\t\x12\x15\n\ris_generation\x18\x03 \x01(\x08\x12!\n\x19preferred_sampling_params\x18\x04 \x01(\t\x12\x16\n\x0eweight_version\x18\x05 \x01(\t\x12\x19\n\x11served_model_name\x18\x06 \x01(\t\x12\x1a\n\x12max_context_length\x18\x07 \x01(\x05\x12\x12\n\nvocab_size\x18\x08 \x01(\x05\x12\x17\n\x0fsupports_vision\x18\t \x01(\x08\x12\x12\n\nmodel_type\x18\n \x01(\t\x12\x15\n\reos_token_ids\x18\x0b \x03(\x05\x12\x14\n\x0cpad_token_id\x18\x0c \x01(\x05\x12\x14\n\x0c\x62os_token_id\x18\r \x01(\x05\x12\x19\n\x11max_req_input_len\x18\x0e \x01(\x05\"\x16\n\x14GetServerInfoRequest\"\xb7\x02\n\x15GetServerInfoResponse\x12,\n\x0bserver_args\x18\x01 \x01(\x0b\x32\x17.google.protobuf.Struct\x12/\n\x0escheduler_info\x18\x02 \x01(\x0b\x32\x17.google.protobuf.Struct\x12\x17\n\x0f\x61\x63tive_requests\x18\x03 \x01(\x05\x12\x11\n\tis_paused\x18\x04 \x01(\x08\x12\x1e\n\x16last_receive_timestamp\x18\x05 \x01(\x01\x12\x16\n\x0euptime_seconds\x18\x06 \x01(\x01\x12\x16\n\x0esglang_version\x18\x07 \x01(\t\x12\x13\n\x0bserver_type\x18\x08 \x01(\t\x12.\n\nstart_time\x18\t \x01(\x0b\x32\x1a.google.protobuf.Timestamp2\xd3\x04\n\x0fSglangScheduler\x12]\n\x08Generate\x12&.sglang.grpc.scheduler.GenerateRequest\x1a\'.sglang.grpc.scheduler.GenerateResponse0\x01\x12R\n\x05\x45mbed\x12#.sglang.grpc.scheduler.EmbedRequest\x1a$.sglang.grpc.scheduler.EmbedResponse\x12\x64\n\x0bHealthCheck\x12).sglang.grpc.scheduler.HealthCheckRequest\x1a*.sglang.grpc.scheduler.HealthCheckResponse\x12R\n\x05\x41\x62ort\x12#.sglang.grpc.scheduler.AbortRequest\x1a$.sglang.grpc.scheduler.AbortResponse\x12g\n\x0cGetModelInfo\x12*.sglang.grpc.scheduler.GetModelInfoRequest\x1a+.sglang.grpc.scheduler.GetModelInfoResponse\x12j\n\rGetServerInfo\x12+.sglang.grpc.scheduler.GetServerInfoRequest\x1a,.sglang.grpc.scheduler.GetServerInfoResponseb\x06proto3')
DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n\x16sglang_scheduler.proto\x12\x15sglang.grpc.scheduler\x1a\x1fgoogle/protobuf/timestamp.proto\x1a\x1cgoogle/protobuf/struct.proto\"\xd0\x05\n\x0eSamplingParams\x12\x13\n\x0btemperature\x18\x01 \x01(\x02\x12\r\n\x05top_p\x18\x02 \x01(\x02\x12\r\n\x05top_k\x18\x03 \x01(\x05\x12\r\n\x05min_p\x18\x04 \x01(\x02\x12\x19\n\x11\x66requency_penalty\x18\x05 \x01(\x02\x12\x18\n\x10presence_penalty\x18\x06 \x01(\x02\x12\x1a\n\x12repetition_penalty\x18\x07 \x01(\x02\x12\x1b\n\x0emax_new_tokens\x18\x08 \x01(\x05H\x01\x88\x01\x01\x12\x0c\n\x04stop\x18\t \x03(\t\x12\x16\n\x0estop_token_ids\x18\n \x03(\r\x12\x1b\n\x13skip_special_tokens\x18\x0b \x01(\x08\x12%\n\x1dspaces_between_special_tokens\x18\x0c \x01(\x08\x12\x0f\n\x05regex\x18\r \x01(\tH\x00\x12\x15\n\x0bjson_schema\x18\x0e \x01(\tH\x00\x12\x16\n\x0c\x65\x62nf_grammar\x18\x0f \x01(\tH\x00\x12\x18\n\x0estructural_tag\x18\x10 \x01(\tH\x00\x12\t\n\x01n\x18\x11 \x01(\x05\x12\x16\n\x0emin_new_tokens\x18\x12 \x01(\x05\x12\x12\n\nignore_eos\x18\x13 \x01(\x08\x12\x14\n\x0cno_stop_trim\x18\x14 \x01(\x08\x12\x1c\n\x0fstream_interval\x18\x15 \x01(\x05H\x02\x88\x01\x01\x12H\n\nlogit_bias\x18\x16 \x03(\x0b\x32\x34.sglang.grpc.scheduler.SamplingParams.LogitBiasEntry\x12.\n\rcustom_params\x18\x17 \x01(\x0b\x32\x17.google.protobuf.Struct\x1a\x30\n\x0eLogitBiasEntry\x12\x0b\n\x03key\x18\x01 \x01(\t\x12\r\n\x05value\x18\x02 \x01(\x02:\x02\x38\x01\x42\x0c\n\nconstraintB\x11\n\x0f_max_new_tokensB\x12\n\x10_stream_interval\"]\n\x13\x44isaggregatedParams\x12\x16\n\x0e\x62ootstrap_host\x18\x01 \x01(\t\x12\x16\n\x0e\x62ootstrap_port\x18\x02 \x01(\x05\x12\x16\n\x0e\x62ootstrap_room\x18\x03 \x01(\x05\"\xe2\x04\n\x0fGenerateRequest\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x38\n\ttokenized\x18\x02 \x01(\x0b\x32%.sglang.grpc.scheduler.TokenizedInput\x12:\n\tmm_inputs\x18\x03 \x01(\x0b\x32\'.sglang.grpc.scheduler.MultimodalInputs\x12>\n\x0fsampling_params\x18\x04 \x01(\x0b\x32%.sglang.grpc.scheduler.SamplingParams\x12\x16\n\x0ereturn_logprob\x18\x05 \x01(\x08\x12\x19\n\x11logprob_start_len\x18\x06 \x01(\x05\x12\x18\n\x10top_logprobs_num\x18\x07 \x01(\x05\x12\x19\n\x11token_ids_logprob\x18\x08 \x03(\r\x12\x1c\n\x14return_hidden_states\x18\t \x01(\x08\x12H\n\x14\x64isaggregated_params\x18\n \x01(\x0b\x32*.sglang.grpc.scheduler.DisaggregatedParams\x12\x1e\n\x16\x63ustom_logit_processor\x18\x0b \x01(\t\x12-\n\ttimestamp\x18\x0c \x01(\x0b\x32\x1a.google.protobuf.Timestamp\x12\x13\n\x0blog_metrics\x18\r \x01(\x08\x12\x14\n\x0cinput_embeds\x18\x0e \x03(\x02\x12\x0f\n\x07lora_id\x18\x0f \x01(\t\x12\x1a\n\x12\x64\x61ta_parallel_rank\x18\x10 \x01(\x05\x12\x0e\n\x06stream\x18\x11 \x01(\x08\":\n\x0eTokenizedInput\x12\x15\n\roriginal_text\x18\x01 \x01(\t\x12\x11\n\tinput_ids\x18\x02 \x03(\r\"\xd3\x01\n\x10MultimodalInputs\x12\x12\n\nimage_urls\x18\x01 \x03(\t\x12\x12\n\nvideo_urls\x18\x02 \x03(\t\x12\x12\n\naudio_urls\x18\x03 \x03(\t\x12\x33\n\x12processed_features\x18\x04 \x01(\x0b\x32\x17.google.protobuf.Struct\x12\x12\n\nimage_data\x18\x05 \x03(\x0c\x12\x12\n\nvideo_data\x18\x06 \x03(\x0c\x12\x12\n\naudio_data\x18\x07 \x03(\x0c\x12\x12\n\nmodalities\x18\x08 \x03(\t\"\xe3\x01\n\x10GenerateResponse\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12;\n\x05\x63hunk\x18\x02 \x01(\x0b\x32*.sglang.grpc.scheduler.GenerateStreamChunkH\x00\x12;\n\x08\x63omplete\x18\x03 \x01(\x0b\x32\'.sglang.grpc.scheduler.GenerateCompleteH\x00\x12\x35\n\x05\x65rror\x18\x04 \x01(\x0b\x32$.sglang.grpc.scheduler.GenerateErrorH\x00\x42\n\n\x08response\"\x95\x02\n\x13GenerateStreamChunk\x12\x11\n\ttoken_ids\x18\x01 \x03(\r\x12\x15\n\rprompt_tokens\x18\x02 \x01(\x05\x12\x19\n\x11\x63ompletion_tokens\x18\x03 \x01(\x05\x12\x15\n\rcached_tokens\x18\x04 \x01(\x05\x12>\n\x0foutput_logprobs\x18\x05 \x01(\x0b\x32%.sglang.grpc.scheduler.OutputLogProbs\x12\x15\n\rhidden_states\x18\x06 \x03(\x02\x12<\n\x0einput_logprobs\x18\x07 \x01(\x0b\x32$.sglang.grpc.scheduler.InputLogProbs\x12\r\n\x05index\x18\x08 \x01(\r\"\x9b\x03\n\x10GenerateComplete\x12\x12\n\noutput_ids\x18\x01 \x03(\r\x12\x15\n\rfinish_reason\x18\x02 \x01(\t\x12\x15\n\rprompt_tokens\x18\x03 \x01(\x05\x12\x19\n\x11\x63ompletion_tokens\x18\x04 \x01(\x05\x12\x15\n\rcached_tokens\x18\x05 \x01(\x05\x12>\n\x0foutput_logprobs\x18\x06 \x01(\x0b\x32%.sglang.grpc.scheduler.OutputLogProbs\x12>\n\x11\x61ll_hidden_states\x18\x07 \x03(\x0b\x32#.sglang.grpc.scheduler.HiddenStates\x12\x1a\n\x10matched_token_id\x18\x08 \x01(\rH\x00\x12\x1a\n\x10matched_stop_str\x18\t \x01(\tH\x00\x12<\n\x0einput_logprobs\x18\n \x01(\x0b\x32$.sglang.grpc.scheduler.InputLogProbs\x12\r\n\x05index\x18\x0b \x01(\rB\x0e\n\x0cmatched_stop\"K\n\rGenerateError\x12\x0f\n\x07message\x18\x01 \x01(\t\x12\x18\n\x10http_status_code\x18\x02 \x01(\t\x12\x0f\n\x07\x64\x65tails\x18\x03 \x01(\t\"u\n\x0eOutputLogProbs\x12\x16\n\x0etoken_logprobs\x18\x01 \x03(\x02\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\x12\x38\n\x0ctop_logprobs\x18\x03 \x03(\x0b\x32\".sglang.grpc.scheduler.TopLogProbs\"\x9e\x01\n\rInputLogProbs\x12@\n\x0etoken_logprobs\x18\x01 \x03(\x0b\x32(.sglang.grpc.scheduler.InputTokenLogProb\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\x12\x38\n\x0ctop_logprobs\x18\x03 \x03(\x0b\x32\".sglang.grpc.scheduler.TopLogProbs\"1\n\x11InputTokenLogProb\x12\x12\n\x05value\x18\x01 \x01(\x02H\x00\x88\x01\x01\x42\x08\n\x06_value\"0\n\x0bTopLogProbs\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\"?\n\x0cHiddenStates\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\r\n\x05layer\x18\x02 \x01(\x05\x12\x10\n\x08position\x18\x03 \x01(\x05\"\xca\x02\n\x0c\x45mbedRequest\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x38\n\ttokenized\x18\x02 \x01(\x0b\x32%.sglang.grpc.scheduler.TokenizedInput\x12:\n\tmm_inputs\x18\x04 \x01(\x0b\x32\'.sglang.grpc.scheduler.MultimodalInputs\x12>\n\x0fsampling_params\x18\x05 \x01(\x0b\x32%.sglang.grpc.scheduler.SamplingParams\x12\x13\n\x0blog_metrics\x18\x06 \x01(\x08\x12\x16\n\x0etoken_type_ids\x18\x07 \x03(\x05\x12\x1a\n\x12\x64\x61ta_parallel_rank\x18\x08 \x01(\x05\x12\x18\n\x10is_cross_encoder\x18\t \x01(\x08\x12\r\n\x05texts\x18\n \x03(\t\"\x9d\x01\n\rEmbedResponse\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x38\n\x08\x63omplete\x18\x02 \x01(\x0b\x32$.sglang.grpc.scheduler.EmbedCompleteH\x00\x12\x32\n\x05\x65rror\x18\x03 \x01(\x0b\x32!.sglang.grpc.scheduler.EmbedErrorH\x00\x42\n\n\x08response\"\xa3\x01\n\rEmbedComplete\x12\x11\n\tembedding\x18\x01 \x03(\x02\x12\x15\n\rprompt_tokens\x18\x02 \x01(\x05\x12\x15\n\rcached_tokens\x18\x03 \x01(\x05\x12\x15\n\rembedding_dim\x18\x04 \x01(\x05\x12:\n\x10\x62\x61tch_embeddings\x18\x05 \x03(\x0b\x32 .sglang.grpc.scheduler.Embedding\"*\n\tEmbedding\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\r\n\x05index\x18\x02 \x01(\x05\"<\n\nEmbedError\x12\x0f\n\x07message\x18\x01 \x01(\t\x12\x0c\n\x04\x63ode\x18\x02 \x01(\t\x12\x0f\n\x07\x64\x65tails\x18\x03 \x01(\t\"\x14\n\x12HealthCheckRequest\"7\n\x13HealthCheckResponse\x12\x0f\n\x07healthy\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"2\n\x0c\x41\x62ortRequest\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x0e\n\x06reason\x18\x02 \x01(\t\"1\n\rAbortResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"I\n\x0fLoadLoRARequest\x12\x12\n\nadapter_id\x18\x01 \x01(\t\x12\x14\n\x0c\x61\x64\x61pter_path\x18\x02 \x01(\t\x12\x0c\n\x04rank\x18\x03 \x01(\x05\"H\n\x10LoadLoRAResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x12\n\nadapter_id\x18\x02 \x01(\t\x12\x0f\n\x07message\x18\x03 \x01(\t\"\'\n\x11UnloadLoRARequest\x12\x12\n\nadapter_id\x18\x01 \x01(\t\"6\n\x12UnloadLoRAResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"w\n\x14UpdateWeightsRequest\x12\x13\n\tdisk_path\x18\x01 \x01(\tH\x00\x12\x15\n\x0btensor_data\x18\x02 \x01(\x0cH\x00\x12\x14\n\nremote_url\x18\x03 \x01(\tH\x00\x12\x13\n\x0bweight_name\x18\x04 \x01(\tB\x08\n\x06source\"9\n\x15UpdateWeightsResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"-\n\x17GetInternalStateRequest\x12\x12\n\nstate_keys\x18\x01 \x03(\t\"B\n\x18GetInternalStateResponse\x12&\n\x05state\x18\x01 \x01(\x0b\x32\x17.google.protobuf.Struct\"A\n\x17SetInternalStateRequest\x12&\n\x05state\x18\x01 \x01(\x0b\x32\x17.google.protobuf.Struct\"<\n\x18SetInternalStateResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"\x15\n\x13GetModelInfoRequest\"\xea\x02\n\x14GetModelInfoResponse\x12\x12\n\nmodel_path\x18\x01 \x01(\t\x12\x16\n\x0etokenizer_path\x18\x02 \x01(\t\x12\x15\n\ris_generation\x18\x03 \x01(\x08\x12!\n\x19preferred_sampling_params\x18\x04 \x01(\t\x12\x16\n\x0eweight_version\x18\x05 \x01(\t\x12\x19\n\x11served_model_name\x18\x06 \x01(\t\x12\x1a\n\x12max_context_length\x18\x07 \x01(\x05\x12\x12\n\nvocab_size\x18\x08 \x01(\x05\x12\x17\n\x0fsupports_vision\x18\t \x01(\x08\x12\x12\n\nmodel_type\x18\n \x01(\t\x12\x15\n\reos_token_ids\x18\x0b \x03(\x05\x12\x14\n\x0cpad_token_id\x18\x0c \x01(\x05\x12\x14\n\x0c\x62os_token_id\x18\r \x01(\x05\x12\x19\n\x11max_req_input_len\x18\x0e \x01(\x05\"\x16\n\x14GetServerInfoRequest\"\xb7\x02\n\x15GetServerInfoResponse\x12,\n\x0bserver_args\x18\x01 \x01(\x0b\x32\x17.google.protobuf.Struct\x12/\n\x0escheduler_info\x18\x02 \x01(\x0b\x32\x17.google.protobuf.Struct\x12\x17\n\x0f\x61\x63tive_requests\x18\x03 \x01(\x05\x12\x11\n\tis_paused\x18\x04 \x01(\x08\x12\x1e\n\x16last_receive_timestamp\x18\x05 \x01(\x01\x12\x16\n\x0euptime_seconds\x18\x06 \x01(\x01\x12\x16\n\x0esglang_version\x18\x07 \x01(\t\x12\x13\n\x0bserver_type\x18\x08 \x01(\t\x12.\n\nstart_time\x18\t \x01(\x0b\x32\x1a.google.protobuf.Timestamp2\xd3\x04\n\x0fSglangScheduler\x12]\n\x08Generate\x12&.sglang.grpc.scheduler.GenerateRequest\x1a\'.sglang.grpc.scheduler.GenerateResponse0\x01\x12R\n\x05\x45mbed\x12#.sglang.grpc.scheduler.EmbedRequest\x1a$.sglang.grpc.scheduler.EmbedResponse\x12\x64\n\x0bHealthCheck\x12).sglang.grpc.scheduler.HealthCheckRequest\x1a*.sglang.grpc.scheduler.HealthCheckResponse\x12R\n\x05\x41\x62ort\x12#.sglang.grpc.scheduler.AbortRequest\x1a$.sglang.grpc.scheduler.AbortResponse\x12g\n\x0cGetModelInfo\x12*.sglang.grpc.scheduler.GetModelInfoRequest\x1a+.sglang.grpc.scheduler.GetModelInfoResponse\x12j\n\rGetServerInfo\x12+.sglang.grpc.scheduler.GetServerInfoRequest\x1a,.sglang.grpc.scheduler.GetServerInfoResponseb\x06proto3')
_globals = globals()
_builder.BuildMessageAndEnumDescriptors(DESCRIPTOR, _globals)
......@@ -79,41 +79,41 @@ if not _descriptor._USE_C_DESCRIPTORS:
_globals['_EMBEDERROR']._serialized_start=3967
_globals['_EMBEDERROR']._serialized_end=4027
_globals['_HEALTHCHECKREQUEST']._serialized_start=4029
_globals['_HEALTHCHECKREQUEST']._serialized_end=4107
_globals['_HEALTHCHECKRESPONSE']._serialized_start=4109
_globals['_HEALTHCHECKRESPONSE']._serialized_end=4164
_globals['_ABORTREQUEST']._serialized_start=4166
_globals['_ABORTREQUEST']._serialized_end=4216
_globals['_ABORTRESPONSE']._serialized_start=4218
_globals['_ABORTRESPONSE']._serialized_end=4267
_globals['_LOADLORAREQUEST']._serialized_start=4269
_globals['_LOADLORAREQUEST']._serialized_end=4342
_globals['_LOADLORARESPONSE']._serialized_start=4344
_globals['_LOADLORARESPONSE']._serialized_end=4416
_globals['_UNLOADLORAREQUEST']._serialized_start=4418
_globals['_UNLOADLORAREQUEST']._serialized_end=4457
_globals['_UNLOADLORARESPONSE']._serialized_start=4459
_globals['_UNLOADLORARESPONSE']._serialized_end=4513
_globals['_UPDATEWEIGHTSREQUEST']._serialized_start=4515
_globals['_UPDATEWEIGHTSREQUEST']._serialized_end=4634
_globals['_UPDATEWEIGHTSRESPONSE']._serialized_start=4636
_globals['_UPDATEWEIGHTSRESPONSE']._serialized_end=4693
_globals['_GETINTERNALSTATEREQUEST']._serialized_start=4695
_globals['_GETINTERNALSTATEREQUEST']._serialized_end=4740
_globals['_GETINTERNALSTATERESPONSE']._serialized_start=4742
_globals['_GETINTERNALSTATERESPONSE']._serialized_end=4808
_globals['_SETINTERNALSTATEREQUEST']._serialized_start=4810
_globals['_SETINTERNALSTATEREQUEST']._serialized_end=4875
_globals['_SETINTERNALSTATERESPONSE']._serialized_start=4877
_globals['_SETINTERNALSTATERESPONSE']._serialized_end=4937
_globals['_GETMODELINFOREQUEST']._serialized_start=4939
_globals['_GETMODELINFOREQUEST']._serialized_end=4960
_globals['_GETMODELINFORESPONSE']._serialized_start=4963
_globals['_GETMODELINFORESPONSE']._serialized_end=5325
_globals['_GETSERVERINFOREQUEST']._serialized_start=5327
_globals['_GETSERVERINFOREQUEST']._serialized_end=5349
_globals['_GETSERVERINFORESPONSE']._serialized_start=5352
_globals['_GETSERVERINFORESPONSE']._serialized_end=5663
_globals['_SGLANGSCHEDULER']._serialized_start=5666
_globals['_SGLANGSCHEDULER']._serialized_end=6261
_globals['_HEALTHCHECKREQUEST']._serialized_end=4049
_globals['_HEALTHCHECKRESPONSE']._serialized_start=4051
_globals['_HEALTHCHECKRESPONSE']._serialized_end=4106
_globals['_ABORTREQUEST']._serialized_start=4108
_globals['_ABORTREQUEST']._serialized_end=4158
_globals['_ABORTRESPONSE']._serialized_start=4160
_globals['_ABORTRESPONSE']._serialized_end=4209
_globals['_LOADLORAREQUEST']._serialized_start=4211
_globals['_LOADLORAREQUEST']._serialized_end=4284
_globals['_LOADLORARESPONSE']._serialized_start=4286
_globals['_LOADLORARESPONSE']._serialized_end=4358
_globals['_UNLOADLORAREQUEST']._serialized_start=4360
_globals['_UNLOADLORAREQUEST']._serialized_end=4399
_globals['_UNLOADLORARESPONSE']._serialized_start=4401
_globals['_UNLOADLORARESPONSE']._serialized_end=4455
_globals['_UPDATEWEIGHTSREQUEST']._serialized_start=4457
_globals['_UPDATEWEIGHTSREQUEST']._serialized_end=4576
_globals['_UPDATEWEIGHTSRESPONSE']._serialized_start=4578
_globals['_UPDATEWEIGHTSRESPONSE']._serialized_end=4635
_globals['_GETINTERNALSTATEREQUEST']._serialized_start=4637
_globals['_GETINTERNALSTATEREQUEST']._serialized_end=4682
_globals['_GETINTERNALSTATERESPONSE']._serialized_start=4684
_globals['_GETINTERNALSTATERESPONSE']._serialized_end=4750
_globals['_SETINTERNALSTATEREQUEST']._serialized_start=4752
_globals['_SETINTERNALSTATEREQUEST']._serialized_end=4817
_globals['_SETINTERNALSTATERESPONSE']._serialized_start=4819
_globals['_SETINTERNALSTATERESPONSE']._serialized_end=4879
_globals['_GETMODELINFOREQUEST']._serialized_start=4881
_globals['_GETMODELINFOREQUEST']._serialized_end=4902
_globals['_GETMODELINFORESPONSE']._serialized_start=4905
_globals['_GETMODELINFORESPONSE']._serialized_end=5267
_globals['_GETSERVERINFOREQUEST']._serialized_start=5269
_globals['_GETSERVERINFOREQUEST']._serialized_end=5291
_globals['_GETSERVERINFORESPONSE']._serialized_start=5294
_globals['_GETSERVERINFORESPONSE']._serialized_end=5605
_globals['_SGLANGSCHEDULER']._serialized_start=5608
_globals['_SGLANGSCHEDULER']._serialized_end=6203
# @@protoc_insertion_point(module_scope)
python/sglang/srt/grpc/sglang_scheduler_pb2.pyi
View file @ 887c2b45
......@@ -320,10 +320,8 @@ class EmbedError(_message.Message):
def __init__(self, message: _Optional[str] = ..., code: _Optional[str] = ..., details: _Optional[str] = ...) -> None: ...
class HealthCheckRequest(_message.Message):
__slots__ = ("tokenized",)
TOKENIZED_FIELD_NUMBER: _ClassVar[int]
tokenized: TokenizedInput
def __init__(self, tokenized: _Optional[_Union[TokenizedInput, _Mapping]] = ...) -> None: ...
__slots__ = ()
def __init__(self) -> None: ...
class HealthCheckResponse(_message.Message):
__slots__ = ("healthy", "message")
......
python/sglang/srt/server_args.py
View file @ 887c2b45
......@@ -194,6 +194,7 @@ class ServerArgs:
# HTTP server
host: str = "127.0.0.1"
port: int = 30000
grpc_mode: bool = False
skip_server_warmup: bool = False
warmups: Optional[str] = None
nccl_port: Optional[int] = None
......@@ -1516,6 +1517,11 @@ class ServerArgs:
default=ServerArgs.port,
help="The port of the HTTP server.",
)
parser.add_argument(
"--grpc-mode",
action="store_true",
help="If set, use gRPC server instead of HTTP server.",
)
parser.add_argument(
"--skip-server-warmup",
action="store_true",
......
sgl-router/src/grpc_client/sglang_scheduler.rs
View file @ 887c2b45
......@@ -169,8 +169,8 @@ impl SglangSchedulerClient {
&self,
) -> Result<proto::HealthCheckResponse, Box<dyn std::error::Error + Send + Sync>> {
debug!("Sending health check request");
// Server ignores the request body and creates its own health check internally
let request = Request::new(proto::HealthCheckRequest { tokenized: None });
// HealthCheckRequest is now empty - server generates its own health check internally
let request = Request::new(proto::HealthCheckRequest {});
let mut client = self.client.clone();
let response = client.health_check(request).await?;
......@@ -510,13 +510,8 @@ mod tests {
#[test]
fn test_proto_types_compilation() {
let health_req = proto::HealthCheckRequest {
tokenized: Some(proto::TokenizedInput {
original_text: "test".to_string(),
input_ids: vec![1296],
}),
};
assert!(health_req.tokenized.is_some());
let _health_req = proto::HealthCheckRequest {};
// HealthCheckRequest is now empty - no fields to test
}
#[test]
......@@ -558,13 +553,8 @@ mod tests {
#[test]
fn test_health_check_request() {
let health_req = proto::HealthCheckRequest {
tokenized: Some(proto::TokenizedInput {
original_text: "test".to_string(),
input_ids: vec![1296], // Mock token ID for "test"
}),
};
assert!(health_req.tokenized.is_some());
let _health_req = proto::HealthCheckRequest {};
// HealthCheckRequest is now empty - server generates its own test internally
}
#[test]
......
sgl-router/src/proto/sglang_scheduler.proto
View file @ 887c2b45
......@@ -326,10 +326,7 @@ message EmbedError {
// Management Operations
// =====================
message HealthCheckRequest {
// Input for health test generation (must be tokenized)
TokenizedInput tokenized = 1;
}
message HealthCheckRequest {}
message HealthCheckResponse {
bool healthy = 1;
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment