[Core][Kernel] Share persistent workspaces for Marlin and Humming (#57421)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: ErinYin <liang.yin@daocloud.io>
This commit is contained in:
Misha Goin
2026-09-20 17:28:33 +08:00
committed by GitHub
co-authored by ErinYin
parent 0e110f696d
commit e5fce7b56b
21 changed files with 598 additions and 211 deletions
@@ -48,7 +48,6 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils import (
check_marlin_supports_layer,
check_moe_marlin_supports_layer,
get_marlin_input_dtype,
marlin_make_workspace_new,
verify_marlin_supported,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
@@ -642,9 +641,6 @@ class AutoAWQMoEMethod(FusedMoEMethodBase):
layer.register_parameter("w2_qzeros", w2_qzeros)
set_weight_attrs(w2_qzeros, extra_weight_attrs)
device = layer.w13_qweight.device
layer.workspace = marlin_make_workspace_new(device, 4)
def process_weights_after_loading(self, layer: RoutedExperts) -> None:
converted = convert_to_wna16_moe_kernel_format(
backend=self.wna16_moe_backend,