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 6, 20:30 UTC
[Bugfix][ROCm] Strip block_size before attention backend validation (#36274)
Signed-off-by: jennyyyyzhen <yzhen@hmc.edu> Co-authored-by: Lu Fang <30275821+houseroad@users.noreply.github.com>
jennyyyyzhen committed
Mar 11, 2026 at 13:37 UTC
428bc718bd4a736c1bc129a23c51963c4f0b71b9
1 file changed
+2
vllm/platforms/rocm.py
+2
@@ -438,6 +438,8 @@ class RocmPlatform(Platform):
438
device_capability = cls.get_device_capability()
439
assert device_capability is not None
440
441
+ attn_selector_config = attn_selector_config._replace(block_size=None)
442
+
443
# First try checking just the selected backend, if there is one.
444
if selected_backend is not None:
445
try: