xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-09 18:17:10 +08:00
Code Issues Packages Projects Releases Wiki Activity

refine

Browse Source 
Signed-off-by: inkcherry <mingzhi.liu@amd.com>
This commit is contained in:
inkcherry 2025-11-21 05:10:58 +00:00
parent 857d93cbfb
commit 96da87bfe0
1 changed files with 2 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/distributed/kv_transfer/kv_connector/v1/moriio_connector.py
View File

@ -1014,7 +1014,7 @@ class MoRIIOConnectorScheduler:
self._reqs_need_save: dict[ReqId, tuple[Request, list[int]]] = {} self._reqs_need_save: dict[ReqId, tuple[Request, list[int]]] = {}
# For chunked prefill, we perform layer-wise access within the final chunk. # For chunked prefill, we perform layer-wise access within the final chunk.
# TODO: Perform access at the end of each chunk. # TODO: Perform transfer at end chunk.
self._reqs_need_pending_save: dict[ReqId, tuple[Request, list[int]]] = {} self._reqs_need_pending_save: dict[ReqId, tuple[Request, list[int]]] = {}
if self.is_producer: if self.is_producer:
@ -1461,9 +1461,6 @@ class MoRIIOConnectorWorker:
self.model_config = vllm_config.model_config self.model_config = vllm_config.model_config
self.cache_config = vllm_config.cache_config self.cache_config = vllm_config.cache_config
# TODO(mgoin): remove this once we have hybrid memory allocator
# Optimization for models with local attention (Llama 4)
# List of block window sizes for each layer for local attention
self.block_window_per_layer: list[int | None] = [] self.block_window_per_layer: list[int | None] = []
self.use_mla = self.model_config.use_mla self.use_mla = self.model_config.use_mla
self.built_session = False self.built_session = False
@ -1775,8 +1772,6 @@ class MoRIIOConnectorWorker:
tp_size = int(meta.tp_size) tp_size = int(meta.tp_size)
remote_dp_size = int(meta.remote_dp_size) remote_dp_size = int(meta.remote_dp_size)
# TODO: handle failure state of future in the
# callback, we want to fail the request in this case.
def request_ready(_f: Future[Any], entry=(req_id, meta)): def request_ready(_f: Future[Any], entry=(req_id, meta)):
logger.info("MoRIIO handshake done for request %s", req_id) logger.info("MoRIIO handshake done for request %s", req_id)
self._ready_requests.put(entry) self._ready_requests.put(entry)
@ -1998,8 +1993,7 @@ class MoRIIOConnectorWorker:
meta.remote_engine_id = remote_engine_id meta.remote_engine_id = remote_engine_id
# TODO: mz get_remote_engine_id() for engine_id mapping. self.get_engine_name_with_dp(remote_engine_id, 0)
dp0_remote_engine_id = f"{remote_engine_id}_dp0"
if dp0_remote_engine_id not in self._remote_agents: if dp0_remote_engine_id not in self._remote_agents:
# Initiate handshake with remote engine to exchange metadata. # Initiate handshake with remote engine to exchange metadata.
with self._handshake_lock: with self._handshake_lock: