Unverified Commit b51255f3 authored by Andreas Karatzas's avatar Andreas Karatzas Committed by GitHub
Browse files

[ROCm] Fix broken import in platform attention backend dispatching (#30432)


Signed-off-by: default avatarAndreas Karatzas <akaratza@amd.com>
parent b4054c8a
......@@ -403,6 +403,20 @@ class RocmPlatform(Platform):
compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
if cache_config and cache_config.block_size is None:
if (
envs.VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION and envs.VLLM_ROCM_USE_AITER
# NOTE: This block has been deprecated
# or get_env_variable_attn_backend()
# == AttentionBackendEnum.ROCM_AITER_UNIFIED_ATTN
# TODO: monitor https://github.com/vllm-project/vllm/pull/30396
# to see how we can transition to the new way of selecting
# attention backends
):
cache_config.block_size = 64
logger.warning(
"[ROCM_AITER_UNIFIED_ATTN]: Setting kv cache block size to 64."
)
else:
cache_config.block_size = 16
if parallel_config.worker_cls == "auto":
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment