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, 13:15 UTC
[ROCm] Validate block_size for explicitly selected attention backends (#36846)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Mar 17, 2026 at 17:04 UTC
3ed7b1e6e0d42a704626a622a79c169bdf51ee84
1 file changed
-2
vllm/platforms/rocm.py
-2
@@ -438,8 +438,6 @@ 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
-
441
# First try checking just the selected backend, if there is one.
442
if selected_backend is not None:
443
try: