[ROCm] Add XGMI backend for MoRI Connector (#41753)

Signed-off-by: simondanielsson <simon.danielsson99@hotmail.com>
This commit is contained in:
Simon Danielsson
2026-05-22 03:06:40 +00:00
committed by GitHub
parent 2998a047aa
commit 86ccef7d44
3 changed files with 34 additions and 12 deletions
@@ -189,6 +189,7 @@ class MoRIIOConfig:
dp_rank: int dp_rank: int
dp_size: int dp_size: int
tp_size: int tp_size: int
backend: str = "rdma"
@classmethod @classmethod
def from_vllm_config(cls, vllm_config: VllmConfig) -> "MoRIIOConfig": def from_vllm_config(cls, vllm_config: VllmConfig) -> "MoRIIOConfig":
@@ -213,6 +214,12 @@ class MoRIIOConfig:
dp_size = vllm_config.parallel_config.data_parallel_size dp_size = vllm_config.parallel_config.data_parallel_size
tp_size = get_tensor_model_parallel_world_size() tp_size = get_tensor_model_parallel_world_size()
port_offset = get_port_offset(dp_rank, tp_rank) port_offset = get_port_offset(dp_rank, tp_rank)
backend = str(extra_config.get("backend", "rdma")).lower()
if backend not in ("rdma", "xgmi"):
raise ValueError(
f"Invalid MoRIIO backend {backend!r} in kv_connector_extra_config; "
"must be one of 'rdma' or 'xgmi'."
)
return cls( return cls(
local_ip=get_ip(), local_ip=get_ip(),
@@ -227,6 +234,7 @@ class MoRIIOConfig:
dp_rank=dp_rank, dp_rank=dp_rank,
dp_size=dp_size, dp_size=dp_size,
tp_size=tp_size, tp_size=tp_size,
backend=backend,
) )
@@ -695,7 +695,12 @@ class MoRIIOConnectorWorker:
# Agent. # Agent.
self.moriio_wrapper = MoRIIOWrapper(tp_rank=self.tp_rank, dp_rank=self.dp_rank) self.moriio_wrapper = MoRIIOWrapper(tp_rank=self.tp_rank, dp_rank=self.dp_rank)
self.moriio_wrapper.set_moriio_engine(self.moriio_engine) self.moriio_wrapper.set_moriio_engine(self.moriio_engine)
self.moriio_wrapper.set_backend_type(BackendType.RDMA) backend = (
BackendType.XGMI
if self.moriio_config.backend == "xgmi"
else BackendType.RDMA
)
self.moriio_wrapper.set_backend_type(backend)
self.moriio_wrapper.notify_port = self.moriio_config.notify_port self.moriio_wrapper.notify_port = self.moriio_config.notify_port
self.local_kv_cache_metadata: list[bytes] = [] self.local_kv_cache_metadata: list[bytes] = []
self.local_kv_cache_size: list[int] = [] self.local_kv_cache_size: list[int] = []
@@ -44,11 +44,13 @@ if TYPE_CHECKING:
logger = init_logger(__name__) logger = init_logger(__name__)
try: try:
from mori.io import ( from mori.io import (
BackendType,
EngineDesc, EngineDesc,
IOEngine, IOEngine,
MemoryDesc, MemoryDesc,
PollCqMode, PollCqMode,
RdmaBackendConfig, RdmaBackendConfig,
XgmiBackendConfig,
) )
logger.info("MoRIIO is available") logger.info("MoRIIO is available")
@@ -376,17 +378,24 @@ class MoRIIOWrapper:
def set_backend_type(self, backend_type): def set_backend_type(self, backend_type):
assert self.moriio_engine is not None, "MoRIIO engine must be set first" assert self.moriio_engine is not None, "MoRIIO engine must be set first"
qp_per_transfer = envs.VLLM_MORIIO_QP_PER_TRANSFER if backend_type == BackendType.XGMI:
post_batch_size = envs.VLLM_MORIIO_POST_BATCH_SIZE logger.info("Using MoRIIO backend: XGMI")
num_worker_threads = envs.VLLM_MORIIO_NUM_WORKERS self.moriio_engine.create_backend(backend_type, XgmiBackendConfig())
poll_mode = PollCqMode.POLLING else:
rdma_cfg = RdmaBackendConfig( logger.info(
qp_per_transfer, "Using MoRIIO backend: RDMA "
post_batch_size, "(qp_per_transfer=%d, post_batch_size=%d, num_workers=%d)",
num_worker_threads, envs.VLLM_MORIIO_QP_PER_TRANSFER,
poll_mode, envs.VLLM_MORIIO_POST_BATCH_SIZE,
) envs.VLLM_MORIIO_NUM_WORKERS,
self.moriio_engine.create_backend(backend_type, rdma_cfg) )
rdma_cfg = RdmaBackendConfig(
envs.VLLM_MORIIO_QP_PER_TRANSFER,
envs.VLLM_MORIIO_POST_BATCH_SIZE,
envs.VLLM_MORIIO_NUM_WORKERS,
PollCqMode.POLLING,
)
self.moriio_engine.create_backend(backend_type, rdma_cfg)
def get_agent_metadata(self): def get_agent_metadata(self):
assert self.moriio_engine is not None, "MoRIIO engine must be set first" assert self.moriio_engine is not None, "MoRIIO engine must be set first"