[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:
co-authored by
Rohan138
parent
bf292b5f6b
commit
d980a3cc6e
@@ -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,
|
||||
|
||||
Reference in New Issue
Block a user