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

[BugFix][DSv32] Fix DSA-CP PD role gating for deepseek v3.2 (v0.18.0) (#8291)

Browse Source 
### What this PR does / why we need it?
This PR backports the DSA-CP PD role gating fix to `releases/v0.18.0`.

The existing helper logic on the release branch does not handle the PD
mixed-role case correctly when deciding whether layer sharding or TP
`o_proj` handling should be enabled. Layer sharding should only run on
the P-side instance, while TP `o_proj` handling should stay enabled for
normal non-PD deployments and for the PD mixed-role (`kv_both`)
instance. This patch makes those conditions explicit and adds unit
coverage for the allowed and disallowed combinations, including the
DSA-CP-disabled path.

Such wrong condition lead to **vllm serve failures** in case: **FC1 +
PD-colocated KV pooling + no layer_sharding**, specifically causing:
1. insufficient Available KV cache memory
2. o_proj shape error in sfa_v1 attention module

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
E2E test with dsv32 + FC1 + FULL_DECODE_ONLY +
kv_transfer_config(kv_both) + no layer_sharding

---------

Signed-off-by: chenchuw886 <chenchuw@huawei.com>
Co-authored-by: chenchuw886 <chenchuw@huawei.com>
This commit is contained in:
Frank Chen
2026-04-17 10:05:40 +08:00
committed by GitHub
parent 1dd1de8153
commit f85144cc57
2 changed files with 53 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm_ascend/utils.py
View File

@@ -1252,11 +1252,10 @@ def enable_dsa_cp_with_layer_shard() -> bool:
from vllm.config import get_current_vllm_config
vllm_config = get_current_vllm_config()
# because the broadcast in layer sharding needs to be overlapped with a heavy compute stream to be
# effectively hidden, it is enabled only during the prefill stage.
is_prefill_instance = (
vllm_config.kv_transfer_config is not None and vllm_config.kv_transfer_config.kv_role == "kv_producer"
)
kv_transfer_config = vllm_config.kv_transfer_config
# Layer sharding broadcast only pays off when it can be hidden by the
# heavier prefill-stage compute, so enable it only on the P-side instance.
is_prefill_instance = kv_transfer_config is not None and kv_transfer_config.kv_role == "kv_producer"
return is_prefill_instance
@@ -1267,9 +1266,12 @@ def enable_dsa_cp_with_o_proj_tp() -> bool:
from vllm.config import get_current_vllm_config
vllm_config = get_current_vllm_config()
# if is PD mix stage, using original TP o_proj weight, and also need to
# full gather for o_proj weight for prefill stage.
return vllm_config.kv_transfer_config is None
kv_transfer_config = vllm_config.kv_transfer_config
# In PD-mixed mode, keep the original TP o_proj weight when:
# 1) KV pooling is disabled, or
# 2) KV pooling is enabled with kv_role == "kv_both".
return kv_transfer_config is None or kv_transfer_config.kv_role == "kv_both"
def check_gdn_layer(vllm_config) -> bool: