update prefix_prefill params

vllm/attention/ops/prefix_prefill.py

@@ -686,7 +686,7 @@ if triton.__version__ >= "2.1.0":
                               sliding_window=None):
 
         cap = current_platform.get_device_capability()
-        BLOCK = 128 if cap[0] >= 8 else 64
+        BLOCK = 128 if cap[0] >= 8 else 32
 
         # need to reduce num. blocks when using fp32
         # due to increased use of GPU shared memory
@@ -709,7 +709,7 @@ if triton.__version__ >= "2.1.0":
         if sliding_window is None or sliding_window <= 0:
             sliding_window = 0
 
-        num_warps = 8 if Lk <= 64 else 8
+        num_warps = 8 if Lk <= 64 else 4
         if alibi_slopes is not None:
             _fwd_kernel_alibi[grid](
                 q,