[PD] Add transfer backend abstraction (#5328)

This commit is contained in:
Byron Hsu
2025-04-13 10:39:39 -07:00
committed by GitHub
parent f765579046
commit a9499885e9
12 changed files with 236 additions and 41 deletions

View File

@@ -195,6 +195,7 @@ class ServerArgs:
# For PD disaggregation: can be "null" (not disaggregated), "prefill" (prefill-only), or "decode" (decode-only)
disaggregation_mode: str = "null"
disaggregation_bootstrap_port: int = 8998
disaggregation_transfer_backend: str = "mooncake"
# multimodal
disable_fast_image_processor: bool = False
@@ -1173,6 +1174,12 @@ class ServerArgs:
default=ServerArgs.disaggregation_bootstrap_port,
help="Bootstrap server port on the prefill server. Default is 8998.",
)
parser.add_argument(
"--disaggregation-transfer-backend",
type=str,
default=ServerArgs.disaggregation_transfer_backend,
help="The backend for disaggregation transfer. Default is mooncake.",
)
# Multimodal
parser.add_argument(