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 11, 22:30 UTC
Fix AWQ MoE marlin check issue in marlin_utils.py for AMD backend (#30102)
Signed-off-by: yuttian1 <yuttian@amd.com>
yuttian1 committed
Dec 6, 2025 at 12:54 UTC
c4d62618ca7cf8507c1e357fa1180fb162c670fa
1 file changed
+4
vllm/model_executor/layers/quantization/utils/marlin_utils.py
+4
@@ -179,6 +179,8 @@ def check_marlin_supports_shape(
179
180
181
def check_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
182
+ if current_platform.is_rocm():
183
+ return False
184
output_size_per_partition = (
185
getattr(layer, "output_size_per_partition", None) or layer.output_size
186
)
@@ -195,6 +197,8 @@ def check_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
197
198
199
def check_moe_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
200
+ if current_platform.is_rocm():
201
+ return False
202
hidden_size = layer.hidden_size
203
intermediate_size_per_partition = layer.intermediate_size_per_partition
204
# apply_router_weight_on_input is not supported for moe marlin