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 15, 08:30 UTC
[rocm] Fix wrong attention log (#18764)
Signed-off-by: Felix Marty <felmarty@amd.com>
fxmarty-amd committed
May 28, 2025 at 04:45 UTC
794ae1f551e4152a258f88c0c914ce5ea130bd05
1 file changed
+3
-2
vllm/platforms/rocm.py
+3
-2
@@ -194,8 +194,9 @@ class RocmPlatform(Platform):
194
f" The selected backend, {selected_backend.name},"
195
f"is not MLA type while requested for MLA backend.")
196
197
- selected_backend = (_Backend.ROCM_FLASH if selected_backend
198
- == _Backend.FLASH_ATTN else selected_backend)
197
+ if selected_backend is None or selected_backend == _Backend.FLASH_ATTN:
198
+ selected_backend = _Backend.ROCM_FLASH
199
+
200
if envs.VLLM_USE_V1:
201
logger.info("Using Triton Attention backend on V1 engine.")
202
return ("vllm.v1.attention.backends."