[V1][Bugfix]: vllm v1 verison metric num_gpu_blocks is None (#15755)

Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
This commit is contained in:
rongfu.leng
2025-04-30 18:20:39 +08:00
committed by GitHub
parent 1534d389af
commit d803786731
4 changed files with 37 additions and 11 deletions
+2 -1
View File
@@ -120,7 +120,8 @@ class AsyncLLM(EngineClient):
executor_class=executor_class,
log_stats=self.log_stats,
)
for stat_logger in self.stat_loggers[0]:
stat_logger.log_engine_initialized()
self.output_handler: Optional[asyncio.Task] = None
try:
# Start output handler eagerly if we are in the asyncio eventloop.
+8 -1
View File
@@ -1,4 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
import json
import os
import queue
import signal
@@ -116,6 +117,7 @@ class EngineCore:
logger.info("Batch queue is enabled with size %d",
self.batch_queue_size)
self.batch_queue = queue.Queue(self.batch_queue_size)
self.vllm_config = vllm_config
def _initialize_kv_caches(
self, vllm_config: VllmConfig) -> tuple[int, int, KVCacheConfig]:
@@ -507,7 +509,12 @@ class EngineCoreProc(EngineCore):
bind=False) as socket:
# Send ready message to front-end once input socket is connected.
socket.send(b'READY')
message_dict = {
'type': 'READY',
'num_gpu_blocks': self.vllm_config.cache_config.num_gpu_blocks,
}
message = json.dumps(message_dict).encode('utf-8')
socket.send(message)
while True:
# (RequestType, RequestData)
+10 -3
View File
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
import asyncio
import contextlib
import json
import queue
import uuid
import weakref
@@ -362,6 +363,7 @@ class MPClient(EngineCoreClient):
executor_class: type[Executor],
log_stats: bool,
):
self.vllm_config = vllm_config
# Serialization setup.
self.encoder = MsgpackEncoder()
self.decoder = MsgpackDecoder(EngineCoreOutputs)
@@ -430,14 +432,19 @@ class MPClient(EngineCoreClient):
raise RuntimeError("Engine core initialization failed. "
"See root cause above.")
eng_id_bytes, msg = sync_input_socket.recv_multipart()
eng_id_bytes, data = sync_input_socket.recv_multipart()
eng_id = int.from_bytes(eng_id_bytes, byteorder="little")
if eng_id not in identities:
raise RuntimeError(f"Unexpected or duplicate engine: {eng_id}")
if msg != b'READY':
raise RuntimeError(f"Engine {eng_id} failed: {msg.decode()}")
message_dict = json.loads(data.decode('utf-8'))
if message_dict['type'] != 'READY':
raise RuntimeError(f"Engine {eng_id} failed: {data.decode()}")
logger.info("Core engine process %d ready.", eng_id)
identities.discard(eng_id)
# Setup KV cache config with initialization state from
# engine core process.
self.vllm_config.cache_config.num_gpu_blocks = message_dict[
'num_gpu_blocks']
def _init_core_engines(
self,