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, 01:30 UTC
[Bug] Add TPU backend option (#32438)
Signed-off-by: Xiongfei Wei <isaacwxf23@gmail.com>
XiongfeiWei committed
Jan 15, 2026 at 21:17 UTC
73f635a75f8a7b6655d45a26e34ac7b9a3ad11fd
1 file changed
+4
vllm/model_executor/layers/fused_moe/oracle/unquantized.py
+4
@@ -32,6 +32,7 @@ class UnquantizedMoeBackend(Enum):
32
TRITON = "TRITON"
33
CPU = "CPU"
34
XPU = "XPU"
35
+ TPU = "TPU"
36
37
38
# NOTE(zyongye): Unsupported backend means backend
@@ -40,6 +41,7 @@ class UnquantizedMoeBackend(Enum):
41
UNSUPPORTED_BACKEND = [
42
UnquantizedMoeBackend.CPU,
43
UnquantizedMoeBackend.XPU,
44
+ UnquantizedMoeBackend.TPU,
45
]
46
47
@@ -91,6 +93,8 @@ def select_unquantized_moe_backend(
93
backend = UnquantizedMoeBackend.XPU
94
if current_platform.is_cpu():
95
backend = UnquantizedMoeBackend.CPU
96
+ if current_platform.is_tpu():
97
+ backend = UnquantizedMoeBackend.TPU
98
99
logger.info_once(_make_log_backend(backend), scope="local")
100
return backend