add kvint8

vllm/attention/ops/paged_attn.py

@@ -131,9 +131,10 @@ class PagedAttention:
         # TODO(woosuk): Tune this heuristic.
         # For context len > 8192, use V2 kernel to avoid shared memory shortage.
         
+        kvquant = False
         if (kv_cache_dtype == "int8"):
-            use_tc = False
-        if use_tc and head_size==128:
+            kvquant = True
+        if use_tc and head_size==128 and not kvquant:
             if envs.VLLM_USE_PA_PRINT_PARAM:
                 print("PA V1 SIZE:")
                 print(f"query.shape = {query.shape}, key_cache.shape = {key_cache.shape}, value_cache.shape = {value_cache.shape}")