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

[BugFix] fix dsv3.1 service failed to start (#8207)

Browse Source 
### What this PR does / why we need it?

This PR fixes a service startup failure for DeepSeek-V3.1 models by
removing a strict type assertion for `MLAAttentionSpec` in
`NPUModelRunner.get_kv_cache_spec`. The assertion was failing due to
class identity mismatches caused by the runtime patching of
`MLAAttentionSpec` with `AscendMLAAttentionSpec`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Verified that the service starts correctly for DSV3.1 models.

Signed-off-by: mayumeng <m30059191@china.huawei.com>
Co-authored-by: mayumeng <m30059191@china.huawei.com>
This commit is contained in:
Yaphets24
2026-04-14 17:52:55 +08:00
committed by GitHub
parent d94b1dc2d0
commit 13c7392416
1 changed files with 0 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
vllm_ascend/worker/model_runner_v1.py
View File

@@ -3241,7 +3241,6 @@ class NPUModelRunner(GPUModelRunner):
cache_sparse_c8=self.use_sparse_c8_indexer, cache_sparse_c8=self.use_sparse_c8_indexer,
) )
elif spec := attn_module.get_kv_cache_spec(self.vllm_config): elif spec := attn_module.get_kv_cache_spec(self.vllm_config):
assert isinstance(spec, MLAAttentionSpec)
from vllm.v1.kv_cache_interface import MLAAttentionSpec as AscendMLAAttentionSpec from vllm.v1.kv_cache_interface import MLAAttentionSpec as AscendMLAAttentionSpec
if getattr(attn_module.impl, "fa_quant_layer", False): if getattr(attn_module.impl, "fa_quant_layer", False):
head_size = attn_module.head_size + attn_module.qk_rope_head_dim head_size = attn_module.head_size + attn_module.qk_rope_head_dim