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

[bugfix][main]KV Pool for KV Transfer in PD Disaggregation Scenarios (#5398)

Browse Source 
### What this PR does / why we need it?
1.KV Pool for KV Transfer in PD Disaggregation Scenarios Error
Resolution
2.Update KV Pool Documentation

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: release/v0.13.0
- vLLM main:
254f6b9867

---------

Signed-off-by: fems14 <1804143737@qq.com>
This commit is contained in:
fems14
2025-12-27 09:53:57 +08:00
committed by GitHub
parent ce52e17bf3
commit 2ef4d1979e
3 changed files with 40 additions and 41 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
if vllm_config.parallel_config.rank == 0 and self.kv_role != "kv_consumer":
if vllm_config.parallel_config.rank == 0:
self.lookup_server = LookupKeyServer(self.connector_worker,
vllm_config,
self.use_layerwise)

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

@@ -26,8 +26,7 @@ class KVPoolScheduler:
"consumer_is_to_load", False)
self.load_async = vllm_config.kv_transfer_config.kv_connector_extra_config.get(
"load_async", False)
self.client = LookupKeyClient(
vllm_config) if self.kv_role != "kv_consumer" else None
self.client = LookupKeyClient(vllm_config)
# request_id -> (vllm cached tokes, kvpool cached tokens)
self.load_specs: dict[str, LoadSpec] = {}
self.pcp_size = getattr(vllm_config.parallel_config,
@@ -75,8 +74,8 @@ class KVPoolScheduler:
else:
token_len = len(request.prompt_token_ids)
num_external_hit_tokens = self.client.lookup( # type: ignore[union-attr]
token_len, request.block_hashes)
num_external_hit_tokens = self.client.lookup(token_len,
request.block_hashes)
if num_external_hit_tokens == request.num_tokens:
num_external_hit_tokens -= 1