[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
View File
@@ -240,6 +240,10 @@ class AttentionBackend(ABC):
def supports_batch_invariance(cls) -> bool:
return False
@classmethod
def supports_kv_connector(cls) -> bool:
return True
@classmethod
def supports_attn_type(cls, attn_type: str) -> bool:
"""Check if backend supports a given attention type.
@@ -283,6 +287,7 @@ class AttentionBackend(ABC):
attn_type: str,
use_non_causal: bool = False,
use_batch_invariant: bool = False,
use_kv_connector: bool = False,
) -> list[str]:
invalid_reasons = []
if not cls.supports_head_size(head_size):
@@ -319,6 +324,8 @@ class AttentionBackend(ABC):
invalid_reasons.append("non-causal attention not supported")
if use_batch_invariant and not cls.supports_batch_invariance():
invalid_reasons.append("batch invariance not supported")
if use_kv_connector and not cls.supports_kv_connector():
invalid_reasons.append("KV connector not supported")
combination_reason = cls.supports_combination(
head_size,
dtype,