[ROCm] Fix AITER_UNIFIED_ATTN Dispatching After AITER Bump (#46780)

Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Co-authored-by: Rohan138 <rohanpotdar138@gmail.com>
This commit is contained in:
Micah Williamson
2026-06-26 02:09:56 -07:00
committed by GitHub
co-authored by Rohan138
parent bf292b5f6b
commit d980a3cc6e
6 changed files with 30 additions and 3 deletions
+5
View File
@@ -306,6 +306,11 @@ def test_attention_quant_pattern(
torch.manual_seed(42)
backend_cls = backend.get_class()
# TODO: drop once AITER reenables fp16 unified attention.
if dtype not in backend_cls.supported_dtypes:
pytest.skip(f"{backend.name} does not support dtype {dtype}")
block_size = backend_cls.get_preferred_block_size(16)
model_config = ModelConfig(
@@ -30,7 +30,8 @@ NUM_Q_HEADS = 8
NUM_KV_HEADS = 8
HEAD_SIZES = [128, 256]
BLOCK_SIZES = [16, 64]
DTYPES = [torch.bfloat16, torch.float16]
# TODO: re-add torch.float16 once AITER reenables fp16 unified attention.
DTYPES = [torch.bfloat16]
FP8_DTYPE = current_platform.fp8_dtype()
# (query_len, kv_len) per sequence
@@ -136,9 +136,17 @@ def test_standard_attention_backend_selection(
# Get the backend class path
from vllm.platforms.rocm import RocmPlatform
# The AITER unified attention kernel only supports BF16/FP8 KV caches
# (its 3D kernel asserts on fp16), so it must be selected with bf16.
dtype = (
torch.bfloat16
if selected_backend == "ROCM_AITER_UNIFIED_ATTN"
else torch.float16
)
attn_selector_config = AttentionSelectorConfig(
head_size=128,
dtype=torch.float16,
dtype=dtype,
kv_cache_dtype="auto",
block_size=16,
use_mla=False,