Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

support minimax_m2 ignore layer: block_sparse_moe.gate (#1508)

Xin Hea32224d36d04
diff --git a/auto_round/special_model_handler.py b/auto_round/special_model_handler.py
index f051b9673..a4573a3e8 100644
--- a/auto_round/special_model_handler.py
+++ b/auto_round/special_model_handler.py
@@ -256,6 +256,15 @@ register_ignore_layers(
)
+# minimax_m2
+register_ignore_layers(
+ matchers=[
+ ModelTypeMatcher(r"minimax_m2", mode="full"),
+ ],
+ ignore_layers=["block_sparse_moe.gate"],
+)
+
+
def get_predefined_ignore_layers(model: torch.nn.Module) -> list[str]:
layers = []
for rule in _PRE_DEFINED_IGNORE_LAYERS: