vllm - 💡(How to fix) Fix [Quantization] Humming MoE import uses a non-existent RoutedExperts submodule [1 pull requests]

What happened

vllm/model_executor/layers/quantization/utils/humming_utils.py currently imports:

from vllm.model_executor.layers.fused_moe.routed_experts import RoutedExperts

That submodule does not exist on main.

RoutedExperts is exported from vllm.model_executor.layers.fused_moe itself, so the Humming helper is importing from a stale path and raises ModuleNotFoundError as soon as the Humming MoE helper gets imported.

Why it matters

This is on the Humming / MXFP4 MoE path, so the backend crashes when code reaches get_humming_moe_quant_config or prepare_humming_moe_layer.

Repro

With Humming installed:

import vllm.model_executor.layers.quantization.utils.humming_utils

The import fails because vllm.model_executor.layers.fused_moe.routed_experts is not present in the tree.

History

This looks like a regression from #40735 (4d591db47), which switched humming_utils.py over to RoutedExperts but pointed it at fused_moe.routed_experts instead of the alias exported from fused_moe.__init__.

Expected fix

Import from:

from vllm.model_executor.layers.fused_moe import RoutedExperts

instead of the missing fused_moe.routed_experts submodule.