Prefix Caching- fix t4 triton error (#2517)

vllm/model_executor/layers/triton_kernel/prefix_prefill.py

@@ -618,7 +618,9 @@ if triton.__version__ >= "2.1.0":
                               b_ctx_len,
                               max_input_len,
                               alibi_slopes=None):
-        BLOCK = 128
+
+        cap = torch.cuda.get_device_capability()
+        BLOCK = 128 if cap[0] >= 8 else 64
         # shape constraints
         Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
         assert Lq == Lk and Lk == Lv