[Bugfix][Models] Fix OpenPangu sleep mode with static sinks (#53696)
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Ronald committed
Sep 14, 2026 at 21:05 UTC
e0c04c7b4dddceb25eb9d23d87c3914bfc2ad3f9
2 files changed
+11
-9
vllm/model_executor/layers/attention/static_sink_attention.py
-2
@@ -152,8 +152,6 @@ class StaticSinkAttention(Attention, CustomOp):
152
attn_backend=attn_backend,
153
**kwargs,
154
)
155
- CustomOp.__init__(self)
156
-
155
self.sink_len = sink_len
156
self.sink_populated = False
157
self.sink_key = None
vllm/model_executor/models/openpangu.py
+11
-7
@@ -677,14 +677,18 @@ class OpenPanguSinkAttention(nn.Module):
677
},
678
)
679
else:
680
- self.param_sink_value = torch.zeros(
681
- (
682
- self.param_sink_number,
683
- self.num_kv_heads,
684
- self.v_channels,
680
+ self.register_buffer(
681
+ "param_sink_value",
682
+ torch.zeros(
683
+ (
684
+ self.param_sink_number,
685
+ self.num_kv_heads,
686
+ self.v_channels,
687
+ ),
688
+ device=current_platform.current_device(),
689
+ dtype=config.torch_dtype,
690
),
686
- device=current_platform.current_device(),
687
- dtype=config.torch_dtype,
691
+ persistent=False,
692
)
693
# To enable dummy run with out weight
694
self.post_weight_load()