set TRITON_HIP_USE_NEW_STREAM_PIPELINE

vllm/attention/ops/triton_decode_attention.py

@@ -28,6 +28,7 @@ Memory-efficient attention for decoding.
 It supports page size >= 1.
 """
 
+import os
 import logging
 
 import triton
@@ -36,6 +37,7 @@ import triton.language as tl
 from vllm.platforms import current_platform
 
 is_hip_ = current_platform.is_rocm()
+os.environ["TRITON_HIP_USE_NEW_STREAM_PIPELINE"] = f"0"
 
 logger = logging.getLogger(__name__)