diff --git a/vllm/models/kimi_k3/nvidia/dspark_mla.py b/vllm/models/kimi_k3/nvidia/dspark_mla.py index 0baff884cc..37207d7bc9 100644 --- a/vllm/models/kimi_k3/nvidia/dspark_mla.py +++ b/vllm/models/kimi_k3/nvidia/dspark_mla.py @@ -376,7 +376,12 @@ class K3DSparkModel(nn.Module): self, cache_layers: list[MultiHeadLatentAttention], ) -> bool: - if not hasattr(self, "_layers_share_kv_block_layout"): + key = tuple(cl.kv_cache.data_ptr() for cl in cache_layers) + if getattr(self, "_kv_cache_ptrs_key", None) != key: + assert not torch.cuda.is_current_stream_capturing() + self._kv_cache_ptrs_key = key + if hasattr(self, "_context_cache_ptrs"): + del self._context_cache_ptrs ref_cache = cache_layers[0].kv_cache self._layers_share_kv_block_layout = all( cl.kv_cache.size(1) == ref_cache.size(1)