xx

Author: wangxiaoxin-sherie

vllm_ascend/attention/attention_v1.py

@@ -327,7 +327,7 @@ def build(
         num_computed_tokens_cpu = (seq_lens - query_lens)
 
         if attn_state == AscendAttentionState.DecodeOnly and \
-            common_attn_metadata.num_input_tokens > num_actual_tokens:
+                common_attn_metadata.num_input_tokens > num_actual_tokens:
             padded_num_tokens = common_attn_metadata.num_input_tokens - num_actual_tokens
             seq_lens = torch.cat([
                 seq_lens,