[PD Disaggregation] support different tp_size for prefill and decode (#5296)

* up

* up

* up

* fix
This commit is contained in:
Juncai
2025-12-01 17:50:20 +08:00
committed by GitHub
parent 54119cf07e
commit 0925d44f18
13 changed files with 584 additions and 36 deletions

View File

@@ -1888,6 +1888,7 @@ class FDConfig:
logger.info(f"disaggregate_info: {self.disaggregate_info}")
if self.router_config:
# the information for registering this server to router
self.register_info = {
"role": self.scheduler_config.splitwise_role,
"host_ip": self.host_ip,
@@ -1897,6 +1898,7 @@ class FDConfig:
"engine_worker_queue_port": engine_worker_queue_port,
"device_ids": self.local_device_ids,
"transfer_protocol": self.cache_config.cache_transfer_protocol.split(","),
"tp_size": self.parallel_config.tensor_parallel_size,
}
logger.info(f"register_info: {self.register_info}")