删除v2 8192限制

vllm/attention/ops/paged_attn.py

@@ -227,7 +227,7 @@ class PagedAttention:
                 print(f"num_kv_heads = {num_kv_heads}, scale = {scale:.3f}, block_tables.shape = {block_tables.shape}, seq_lens.shape = {seq_lens.shape}, block_size = {block_size}, max_seq_len = {max_seq_len}")
 
             if envs.VLLM_USE_OPT_OP:
-                if envs.VLLM_USE_TC_PAGED_ATTN and max_seq_len < 8192:
+                if envs.VLLM_USE_TC_PAGED_ATTN:
                     ops.paged_attention_v2_opt_tc(
                         output,
                         exp_sums,