[ROCm] Attention selector reordering (#36702)

Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Co-authored-by: Micah Williamson <micah.williamson@amd.com>
This commit is contained in:
Gregory Shtrasberg
2026-03-25 17:42:56 +08:00
committed by GitHub
co-authored by Micah Williamson
parent 09c3dc9186
commit 189ddefbfd
8 changed files with 29 additions and 56 deletions
-5
View File
@@ -117,7 +117,6 @@ if TYPE_CHECKING:
VLLM_ROCM_USE_SKINNY_GEMM: bool = True
VLLM_ROCM_FP8_PADDING: bool = True
VLLM_ROCM_MOE_PADDING: bool = True
VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT: bool = False
VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
@@ -1001,10 +1000,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_ROCM_FP8_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
# Pad the weights for the moe kernel
"VLLM_ROCM_MOE_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),
# custom paged attention kernel for MI3* cards
"VLLM_ROCM_CUSTOM_PAGED_ATTN": lambda: (
os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in ("true", "1")
),
# Whether to use the shuffled kv cache layout
"VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT": lambda: (
os.getenv("VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT", "False").lower() in ("true", "1")