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

[BugFix][PCP] Fix presion bugs for pcp/dcp in PD disaggregate (#6876)

Browse Source 
### What this PR does / why we need it?
Fix a bug for PD disaggregate of PCP/DCP, some conditions only consider
MLA while ignoring DSA.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.15.0
- vLLM main:
15d76f74e2
- vLLM Ascend main: 81fb7d5779

Signed-off-by: tongyuzhou <tongyuzhou1@huawei.com>
Co-authored-by: tongyuzhou <tongyuzhou1@huawei.com>
This commit is contained in:
Yuzhou Tong
2026-03-02 16:11:00 +08:00
committed by GitHub
parent ddc78dbade
commit 9180dd6c51
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_connector.py
View File

@@ -1323,13 +1323,13 @@ class MooncakeConnectorWorker:
def context_parallel_parameters_check(): def context_parallel_parameters_check():
assert (meta.remote_pcp_size * meta.remote_dcp_size) % (self.pcp_size * self.dcp_size) == 0 assert (meta.remote_pcp_size * meta.remote_dcp_size) % (self.pcp_size * self.dcp_size) == 0
if not self.use_mla: if not (self.use_mla or self.use_sparse):
p_node_heads_per_rank = math.ceil(self.num_key_value_heads / prefill_tp_size) p_node_heads_per_rank = math.ceil(self.num_key_value_heads / prefill_tp_size)
d_node_heads_per_rank = math.ceil(self.num_key_value_heads / self.tp_size) d_node_heads_per_rank = math.ceil(self.num_key_value_heads / self.tp_size)
assert d_node_heads_per_rank % p_node_heads_per_rank == 0 assert d_node_heads_per_rank % p_node_heads_per_rank == 0
def get_kv_head_groups(tp_size): def get_kv_head_groups(tp_size):
if self.use_mla: if self.use_mla or self.use_sparse:
kv_head_groups = [] kv_head_groups = []
kv_head_ids = [0] kv_head_ids = [0]
kv_head_groups.append(tuple(kv_head_ids)) kv_head_groups.append(tuple(kv_head_ids))