diff --git a/vllm/models/deepseek_v4/attention.py b/vllm/models/deepseek_v4/attention.py index ff7b92c6b2..128debf70c 100644 --- a/vllm/models/deepseek_v4/attention.py +++ b/vllm/models/deepseek_v4/attention.py @@ -906,7 +906,10 @@ class DeepseekV4Indexer(nn.Module): attn_metadata = get_forward_context().attn_metadata if isinstance(attn_metadata, dict): indexer_metadata = cast(Any, attn_metadata[self.k_cache.prefix]) - if indexer_metadata.max_seq_len // self.compress_ratio <= self.topk_tokens: + if ( + indexer_metadata.max_seq_len // self.compress_ratio <= self.topk_tokens + and not torch.cuda.is_current_stream_capturing() + ): # candidates num smaller than topk, every candidate is selected # but we still need to build k cache compressor(compressed_kv_score, positions, rotary_emb)