From e0c04c7b4dddceb25eb9d23d87c3914bfc2ad3f9 Mon Sep 17 00:00:00 2001 From: Ronald Date: Mon, 14 Sep 2026 21:05:47 +0800 Subject: [PATCH] [Bugfix][Models] Fix OpenPangu sleep mode with static sinks (#53696) Signed-off-by: Ronald1995 --- .../layers/attention/static_sink_attention.py | 2 -- vllm/model_executor/models/openpangu.py | 18 +++++++++++------- 2 files changed, 11 insertions(+), 9 deletions(-) diff --git a/vllm/model_executor/layers/attention/static_sink_attention.py b/vllm/model_executor/layers/attention/static_sink_attention.py index d27ccc728a..d321d675a5 100644 --- a/vllm/model_executor/layers/attention/static_sink_attention.py +++ b/vllm/model_executor/layers/attention/static_sink_attention.py @@ -152,8 +152,6 @@ class StaticSinkAttention(Attention, CustomOp): attn_backend=attn_backend, **kwargs, ) - CustomOp.__init__(self) - self.sink_len = sink_len self.sink_populated = False self.sink_key = None diff --git a/vllm/model_executor/models/openpangu.py b/vllm/model_executor/models/openpangu.py index 60b5b301d5..b8937f2e09 100644 --- a/vllm/model_executor/models/openpangu.py +++ b/vllm/model_executor/models/openpangu.py @@ -677,14 +677,18 @@ class OpenPanguSinkAttention(nn.Module): }, ) else: - self.param_sink_value = torch.zeros( - ( - self.param_sink_number, - self.num_kv_heads, - self.v_channels, + self.register_buffer( + "param_sink_value", + torch.zeros( + ( + self.param_sink_number, + self.num_kv_heads, + self.v_channels, + ), + device=current_platform.current_device(), + dtype=config.torch_dtype, ), - device=current_platform.current_device(), - dtype=config.torch_dtype, + persistent=False, ) # To enable dummy run with out weight self.post_weight_load()