[Feature] Support batched tokens for EP (#3415)

* Support batched tokens for EP

* Support batched tokens for EP

* Support batched tokens for EP

* Support batched tokens for EP

* Support batched tokens for EP and fix bug

* Support batched tokens for EP and fix bug

* Support batched tokens for EP and fix bug

* Support batched tokens for EP and fix bug
This commit is contained in:
chenjian
2025-08-18 11:43:36 +08:00
committed by GitHub
parent 3f86ae0007
commit aba94169dc
9 changed files with 235 additions and 97 deletions

View File

@@ -88,6 +88,8 @@ environment_variables: dict[str, Callable[[], Any]] = {
"FD_ZMQ_SEND_RESPONSE_SERVER_PORT": lambda: os.getenv("FD_ZMQ_SEND_RESPONSE_SERVER_PORT", "8201"),
# LLMEngine recieve control command port, used when FD_ENABLE_INTERNAL_ADAPTER=1
"FD_ZMQ_CONTROL_CMD_SERVER_PORTS": lambda: os.getenv("FD_ZMQ_CONTROL_CMD_SERVER_PORTS", "8202"),
# Batched token timeout in EP
"FD_EP_BATCHED_TOKEN_TIMEOUT": lambda: float(os.getenv("FD_EP_BATCHED_TOKEN_TIMEOUT", "0.1")),
# Whether to use PLUGINS.
"FD_PLUGINS": lambda: None if "FD_PLUGINS" not in os.environ else os.environ["FD_PLUGINS"].split(","),
# Whether to enable cache task in decode node