diff --git a/vllm/v1/attention/backends/gdn_attn.py b/vllm/v1/attention/backends/gdn_attn.py index 0b32dfcf94..7332d3e365 100644 --- a/vllm/v1/attention/backends/gdn_attn.py +++ b/vllm/v1/attention/backends/gdn_attn.py @@ -177,7 +177,6 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata] query_start_loc = m.query_start_loc query_start_loc_cpu = m.query_start_loc_cpu - context_lens_tensor = m.compute_num_computed_tokens() nums_dict, batch_ptr, token_chunk_offset_ptr = None, None, None block_table_tensor = mamba_get_block_table_tensor( m.block_table_tensor, @@ -389,6 +388,7 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata] ) if num_prefills > 0: + context_lens_tensor = m.compute_num_computed_tokens() has_initial_state = context_lens_tensor > 0 if spec_sequence_masks_cpu is not None: has_initial_state = has_initial_state[~spec_sequence_masks_cpu]