EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[main][Bugfix] Remove the ZMQ communication setup on the D node (#4926)

Browse Source 
In the PD separation scenario, the D node does not need to perform get
operations, and therefore does not need to create ZeroMQ (ZMQ)
communication.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
This commit is contained in:
Slightwind
2025-12-12 14:37:26 +08:00
committed by GitHub
parent d54db76dd2
commit b8a317caac
2 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm_ascend/distributed/kvpool/ascend_store_connector.py
View File

@@ -55,7 +55,7 @@ class AscendStoreConnector(KVConnectorBase_V1):
) )
assert self.connector_worker is not None assert self.connector_worker is not None
if vllm_config.parallel_config.rank == 0: if vllm_config.parallel_config.rank == 0 and self.kv_role != "kv_consumer":
self.lookup_server = LookupKeyServer(self.connector_worker, self.lookup_server = LookupKeyServer(self.connector_worker,
vllm_config, vllm_config,
self.use_layerwise) self.use_layerwise)

7
vllm_ascend/distributed/kvpool/pool_scheduler.py
View File

@@ -20,13 +20,14 @@ from vllm_ascend.distributed.kvpool.config_data import (
class KVPoolScheduler: class KVPoolScheduler:
def __init__(self, vllm_config: "VllmConfig", use_layerwise): def __init__(self, vllm_config: "VllmConfig", use_layerwise):
self.client = LookupKeyClient(vllm_config)
self.use_layerwise = use_layerwise self.use_layerwise = use_layerwise
self.kv_role = vllm_config.kv_transfer_config.kv_role self.kv_role = vllm_config.kv_transfer_config.kv_role
self.consumer_is_to_load = vllm_config.kv_transfer_config.kv_connector_extra_config.get( self.consumer_is_to_load = vllm_config.kv_transfer_config.kv_connector_extra_config.get(
"consumer_is_to_load", False) "consumer_is_to_load", False)
self.load_async = vllm_config.kv_transfer_config.kv_connector_extra_config.get( self.load_async = vllm_config.kv_transfer_config.kv_connector_extra_config.get(
"load_async", False) "load_async", False)
self.client = LookupKeyClient(
vllm_config) if self.kv_role != "kv_consumer" else None
# request_id -> (vllm cached tokes, kvpool cached tokens) # request_id -> (vllm cached tokes, kvpool cached tokens)
self.load_specs: dict[str, LoadSpec] = {} self.load_specs: dict[str, LoadSpec] = {}
self.pcp_size = getattr(vllm_config.parallel_config, self.pcp_size = getattr(vllm_config.parallel_config,
@@ -74,8 +75,8 @@ class KVPoolScheduler:
else: else:
token_len = len(request.prompt_token_ids) token_len = len(request.prompt_token_ids)
num_external_hit_tokens = self.client.lookup(token_len, num_external_hit_tokens = self.client.lookup( # type: ignore[union-attr]
request.block_hashes) token_len, request.block_hashes)
if num_external_hit_tokens == request.num_tokens: if num_external_hit_tokens == request.num_tokens:
num_external_hit_tokens -= 1 num_external_hit_tokens -= 1