[Bugfix] Fall back to auto for unquantized MoE layers when --moe-backend is marlin (#60219)

Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>

tests/kernels/moe/test_unquantized_backend_selection.py

@@ -626,3 +626,17 @@ def test_select_explicit_triton_backend(is_lora_enabled):
 
     assert selected_backend == UnquantizedMoeBackend.TRITON
     assert experts_cls is not None
+
+
+@skipif_not_cuda_rocm
+@pytest.mark.parametrize("moe_backend", ["humming", "marlin"])
+def test_quantization_only_backend_falls_back_to_auto(moe_backend):
+    """A quantization-only --moe-backend also reaches the unquantized MoE
+    layers of a quantized model (e.g. a BF16 MTP draft), which must use the
+    auto-selected backend instead of failing."""
+    moe_config = make_dummy_moe_config()
+    moe_config.moe_backend = moe_backend
+
+    assert select_unquantized_moe_backend(
+        moe_config=moe_config
+    ) == select_unquantized_moe_backend(moe_config=make_dummy_moe_config())

vllm/model_executor/layers/fused_moe/oracle/unquantized.py

@@ -303,9 +303,10 @@ def select_unquantized_moe_backend(
         )
 
     runner_backend = moe_config.moe_backend
-    # 'humming' is quantization-only; an unquantized layer (e.g. excluded via
-    # modules_to_not_convert) falls through to auto instead of erroring.
-    if runner_backend not in ["auto", "humming"]:
+    # 'humming' and 'marlin' are quantization-only; an unquantized layer (e.g.
+    # excluded via modules_to_not_convert, or an unquantized MTP draft of a
+    # quantized model) falls through to auto instead of erroring.
+    if runner_backend not in ["auto", "humming", "marlin"]:
         requested_backend = map_unquantized_backend(runner_backend)
         if (
             activation_format == mk.FusedMoEActivationFormat.BatchedExperts