# SPDX-FileCopyrightText: 2024-2025 ModelCloud.ai
# SPDX-FileCopyrightText: 2024-2025 qubitium@modelcloud.ai
# SPDX-License-Identifier: Apache-2.0
# Contact: qubitium@modelcloud.ai, x.com/qubitium

from ..base import BaseQModel
from ..moe_lifecycle import GateUpDownMoELifecycleHooks


class BailingMoeQModel(BaseQModel):
    # allow dynamic expert index for layer_modules so we don't need to write out 64 layers here
    # config.num_experts contains the actual expert count used for index
    dynamic_expert_index = "num_experts"
    layer_modules_strict = False

    pre_lm_head_norm_module = "model.norm"

    # MoE lifecycle hooks for gate_proj/up_proj/down_proj pattern
    moe_lifecycle_hooks = GateUpDownMoELifecycleHooks()

    module_tree = [
        "model",
        "layers",
        "#",
        {
            "input_layernorm": ("input_layernorm:!",),
            "attention": ("query_key_value"),
            "post_attention_layernorm": ("post_attention_layernorm:!",),
            "mlp:moe": {
                "gate": ("gate:!",), # <-- 0.5MB per layer. Not worth quantizing
                "shared_experts": ("gate_proj", "up_proj", "down_proj"),
                "experts": {
                    "#": ("gate_proj:0", "up_proj:0", "down_proj:1"),
                },
            },
        }
    ]