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 14, 05:30 UTC
[Misc] fix tests failure by using current_platform (#25825)
Signed-off-by: Juechen Liu <jueliu@meta.com>
Juechen Liu committed
Sep 28, 2025 at 21:18 UTC
a3ae45a38cb0ee920427e3f2db2a237afd557fd8
1 file changed
+1
-1
vllm/attention/ops/triton_reshape_and_cache_flash.py
+1
-1
@@ -137,7 +137,7 @@ def triton_reshape_and_cache_flash(
137
138
# heuristics instead of autotuning
139
TILE_SIZE = min(2048, triton.next_power_of_2(n))
140
- if torch.version.hip or torch.version.xpu:
140
+ if current_platform.is_rocm() or current_platform.is_xpu():
141
num_stages = 4
142
num_warps = 8
143
else: # cuda