[Attention][AMD] Standardize kv layout to blocks first for AMD (#43660)

Signed-off-by: NickLucche <nlucches@redhat.com>
This commit is contained in:
Nicolò Lucchesi
2026-05-28 12:28:50 -05:00
committed by GitHub
parent 53a2088675
commit 5b115bb8a3
6 changed files with 57 additions and 44 deletions
+7 -3
View File
@@ -362,6 +362,7 @@ def flash_attn_triton_available() -> bool:
def _get_backend_priorities(
use_mla: bool,
use_sparse: bool,
use_kv_connector: bool = False,
) -> list[AttentionBackendEnum]:
from vllm._aiter_ops import is_aiter_found_and_supported, rocm_aiter_ops
@@ -380,9 +381,11 @@ def _get_backend_priorities(
AttentionBackendEnum.TRITON_MLA,
]
backends = [
AttentionBackendEnum.ROCM_ATTN,
]
backends = []
# ROCM_ATTN uses (2, num_blocks, ...) KV cache layout which is
# incompatible with KV connectors that require blocks-first layout.
if not use_kv_connector:
backends.append(AttentionBackendEnum.ROCM_ATTN)
if rocm_aiter_ops.is_mha_enabled():
backends.append(AttentionBackendEnum.ROCM_AITER_FA)
if is_aiter_found_and_supported():
@@ -461,6 +464,7 @@ class RocmPlatform(Platform):
backend_priorities = _get_backend_priorities(
attn_selector_config.use_mla,
attn_selector_config.use_sparse,
attn_selector_config.use_kv_connector,
)
for priority, backend in enumerate(backend_priorities):
try: