xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-01-25 18:34:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix AWQ MoE marlin check issue in marlin_utils.py for AMD backend (#30102)

Browse Source 
Signed-off-by: yuttian1 <yuttian@amd.com>
This commit is contained in:
yuttian1 2025-12-06 12:54:38 +08:00 committed by GitHub
parent 62079d8600
commit c4d62618ca
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/layers/quantization/utils/marlin_utils.py
View File

@ -179,6 +179,8 @@ def check_marlin_supports_shape(
def check_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
if current_platform.is_rocm():
return False
output_size_per_partition = (
getattr(layer, "output_size_per_partition", None) or layer.output_size
)
@ -195,6 +197,8 @@ def check_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
def check_moe_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
if current_platform.is_rocm():
return False
hidden_size = layer.hidden_size
intermediate_size_per_partition = layer.intermediate_size_per_partition
# apply_router_weight_on_input is not supported for moe marlin