A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 12:45 UTC
[ROCm][CI] v1 cpu offloading attention backend fix (#31833)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Jan 8, 2026 at 00:37 UTC
5f2a473ff3245c0db065c62568a515b72007ac57
1 file changed
+4
-2
tests/v1/kv_offload/test_cpu_offloading.py
+4
-2
@@ -15,10 +15,12 @@ from vllm.distributed.kv_events import BlockStored, KVEventBatch
15
from vllm.platforms import current_platform
16
17
CPU_BLOCK_SIZES = [48]
18
-ATTN_BACKENDS = ["FLASH_ATTN", "TRITON_ATTN"]
18
+ATTN_BACKENDS = []
19
20
if current_platform.is_cuda():
21
- ATTN_BACKENDS.append("FLASHINFER")
21
+ ATTN_BACKENDS = ["FLASH_ATTN", "FLASHINFER", "TRITON_ATTN"]
22
+elif current_platform.is_rocm():
23
+ ATTN_BACKENDS = ["TRITON_ATTN"]
24
25
26
class MockSubscriber: