[Model] Remove unused DeepSeek V4 top-k buffer helper (#53697)
Woosuk Kwon committed
Aug 26, 2026 at 11:59 UTC
0a5ad6f0d42c1ace54f27d7dd6a852be86e42148
1 file changed
-7
vllm/models/deepseek_v4/attention.py
-7
@@ -685,13 +685,6 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):
685
)
686
return q_fp8
687
688
- def _global_topk_output_buffers(
689
- self, topk_indices: torch.Tensor
690
- ) -> tuple[torch.Tensor, torch.Tensor] | None:
691
- if self.compress_ratio != 4 or self.eager_scratch_pool is None:
692
- return None
693
- return self.eager_scratch_pool.global_topk_outputs(topk_indices)
694
-
688
def bind_kv_cache(self, kv_cache: torch.Tensor) -> None:
689
# [B, H=1, N, C] -> [B, N, C]
690
self.kv_cache = kv_cache.squeeze(1)