[ROCm][Refactor] Enable AWQMarlinConfig on ROCm to use choose_mp_linear_kernel (#36505)

Signed-off-by: Matthias Gehre <matthias.gehre@amd.com> Co-authored-by: Michael Goin <mgoin64@gmail.com>
2026-03-23 08:36:08 +01:00
parent d3fe857135
commit 410d300893
3 changed files with 143 additions and 102 deletions
--- a/vllm/model_executor/kernels/linear/mixed_precision/conch.py
+++ b/vllm/model_executor/kernels/linear/mixed_precision/conch.py
@@ -113,6 +113,8 @@ class ConchLinearKernel(MPLinearKernel):
        self._transform_param(layer, self.w_s_name, transform_w_s)
        if self.config.zero_points:
            self._transform_param(layer, self.w_zp_name, transform_w_zp)
+        elif self.w_zp_name is not None:
+            layer.register_parameter(self.w_zp_name, None)

    def apply_weights(
        self,