[Kernel] Extend Marlin thread-tile padding to MoE (WNA16 + FP8/MXFP8) (#45703)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Michael Goin
2026-06-23 11:45:10 -07:00
committed by GitHub
co-authored by Claude
parent d8e422ccda
commit 37a682d392
7 changed files with 573 additions and 29 deletions
@@ -243,7 +243,9 @@ class AutoGPTQConfig(QuantizationConfig):
if isinstance(layer, RoutedExperts):
from vllm.model_executor.layers.quantization.moe_wna16 import MoeWNA16Config
if not check_moe_marlin_supports_layer(layer, self.group_size):
if not check_moe_marlin_supports_layer(
layer, self.group_size, allow_tile_padding=not self.desc_act
):
logger.warning_once(
f"Layer '{prefix}' is not supported by GPTQMoeMarlin. "
"Falling back to Moe WNA16 kernels."