diff --git a/python/sglang/srt/layers/attention/deepseek_v4_backend.py b/python/sglang/srt/layers/attention/deepseek_v4_backend.py index 9ef4e245be..1891f46fd9 100644 --- a/python/sglang/srt/layers/attention/deepseek_v4_backend.py +++ b/python/sglang/srt/layers/attention/deepseek_v4_backend.py @@ -2951,7 +2951,7 @@ class DeepseekV4AttnBackend( self, forward_batch: ForwardBatch ): max_seq_len = forward_batch.max_seq_len_override or self.MAX_SEQ_LEN_FOR_CAPTURE - if self.token_to_kv_pool.request_window is None and self.low_ratios == (1, 2): + if self.low_ratios == (1, 2): graph_context_limit = _prefill_graph_max_seq_len() if graph_context_limit is not None: max_seq_len = min(max_seq_len, graph_context_limit) @@ -2965,7 +2965,7 @@ class DeepseekV4AttnBackend( self.enable_decoder_swa_bounded_replay and forward_batch.forward_mode.is_extend_without_speculative() ) - if self.token_to_kv_pool.request_window is None and self.low_ratios == (1, 2): + if self.low_ratios == (1, 2): self._share_bcg_low_ratio_index_buffers(self.forward_metadata.core_metadata) self.forward_metadata.prefill_graph_tail = with_tail self.tail_forward_metadata = ( @@ -3041,7 +3041,7 @@ class DeepseekV4AttnBackend( max_seq_len = ( metadata_batch.max_seq_len_override or self.MAX_SEQ_LEN_FOR_CAPTURE ) - if self.token_to_kv_pool.request_window is None and self.low_ratios == (1, 2): + if self.low_ratios == (1, 2): graph_context_limit = _prefill_graph_max_seq_len() if graph_context_limit is not None: max_seq_len = min(max_seq_len, graph_context_limit)