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

Revert [KV-Sharing] Support KV-Sharing feature in CLA models (#4138) (#5317)

Browse Source 
### What this PR does / why we need it?
Revert [KV-Sharing] Support KV-Sharing feature in CLA models (#4138) as
it causes deepseek v3.2 hang error


- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef

---------

Signed-off-by: MengqingCao <cmq0113@163.com>
This commit is contained in:
Mengqing Cao
2025-12-24 22:24:17 +08:00
committed by GitHub
parent fb3d6ca08c
commit e54630e01c
5 changed files with 18 additions and 96 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/worker/model_runner_v1.py
View File

@@ -1195,10 +1195,6 @@ class NPUModelRunner(GPUModelRunner):
def _build_attn_state(self, num_reqs, num_scheduled_tokens,
num_valid_tokens):
if self.shared_kv_cache_layers is not None:
# sharing kv across layers need to read the kvcache,
# directly return chunked prefill in this scenario
return AscendAttentionState.ChunkedPrefill
if np.array_equal(self.seq_lens.np[:num_reqs], num_scheduled_tokens):
attn_state = AscendAttentionState.PrefillNoCache
# We assume it is the decode stage, where prefill occurs but only one token is not hit in cache.