when v1 is set to block_size 16, switch to triton implementation

vllm/platforms/rocm.py

@@ -299,6 +299,7 @@ class RocmPlatform(Platform):
                 return FLASH_ATTN_V1
 
             else:
+                os.environ['VLLM_USE_FLASH_ATTN_PA'] = '0'
                 logger.info_once("Using Triton backend on V1 engine.")
                 return TRITON_ATTN_VLLM_V1