Skip to content

Commit 24d2a20

Browse files
committed
fix lint
Signed-off-by: MengqingCao <[email protected]>
1 parent 81bb71b commit 24d2a20

File tree

1 file changed

+4
-2
lines changed

1 file changed

+4
-2
lines changed

vllm_ascend/worker/model_runner_v1.py

Lines changed: 4 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -2823,7 +2823,7 @@ def _allocate_kv_cache_tensors(
28232823
v_tensor = self._align_memory(
28242824
v_tensor, alignment)[:v_tensor_size]
28252825
#### k cache: for deepseek sparse attention
2826-
if dsa_k_cache_factor is not None:
2826+
if dsa_k_cache_factor is not None and dsa_k_cache_size is not None:
28272827
k_cache_tensor = torch.zeros(dsa_k_cache_size +
28282828
alignment,
28292829
dtype=torch.int8,
@@ -2848,6 +2848,8 @@ def _allocate_kv_cache_tensors(
28482848
assert layer_names == set(kv_cache_raw_tensors.keys(
28492849
)), "Some layers are not correctly initialized"
28502850

2851+
return kv_cache_raw_tensors
2852+
28512853
def _reshape_kv_cache_tensors(
28522854
self,
28532855
kv_cache_config: KVCacheConfig,
@@ -2926,7 +2928,7 @@ def _reshape_kv_cache_tensors(
29262928
k_cache = self._convert_torch_format(k_cache)
29272929
v_cache = raw_v_tensor.view(dtype).view(kv_cache_shape[1:])
29282930
v_cache = self._convert_torch_format(v_cache)
2929-
if self.use_sparse:
2931+
if self.use_sparse and raw_dsa_k_cache is not None:
29302932
dsa_k_cache_shape = (num_blocks, block_size, 1, 128)
29312933
dsa_k_cache = raw_dsa_k_cache.view(dtype).view(
29322934
dsa_k_cache_shape)

0 commit comments

Comments
 (0)