"""Build a TITAN model from an MLX checkpoint: the host architecture (attention, norms, router) is
mlx-lm's own module tree; every layer's `switch_mlp` is replaced by a TitanSwitchGLU over a slot table.
Expert weights are never materialised from the checkpoint by mlx-lm: they are read on demand by the
ExpertStore. `capacity` = number of resident experts per layer (num_experts = everything resident).
"""
from __future__ import annotations

import os
from typing import Dict, List, Optional

import mlx.core as mx
import mlx.nn as nn
from mlx_lm.utils import load_model, load_tokenizer

from .moe import TitanSwitchGLU
from .registry import ModelSpec, load_spec
from .slots import EvictionPolicy, LFUPolicy, SlotTable
from .store import ExpertStore

POLICIES = {"lru": EvictionPolicy, "lfu": LFUPolicy}


class TitanModel:
    def __init__(self, model_dir: str, capacity: Optional[int] = None, policy: str = "lru", nocache: bool = False,
                 on_route=None):
        self.model_dir = os.path.realpath(model_dir)
        self.spec: ModelSpec = load_spec(self.model_dir)
        self.store = ExpertStore(self.spec, nocache=nocache)
        self.capacity = capacity or self.spec.num_experts
        self.policy_name = policy
        # Load the host model lazily so the stacked expert tensors are never evaluated.
        from pathlib import Path
        self.model, cfg = load_model(Path(self.model_dir), lazy=True)
        self.tokenizer = load_tokenizer(Path(self.model_dir))
        q = cfg.get("quantization") or cfg.get("quantization_config")
        self.tables: List[SlotTable] = []
        self.blocks: List[TitanSwitchGLU] = []
        for l, layer in enumerate(self.model.model.layers):
            table = SlotTable(l, self.capacity, self.store, POLICIES[policy]())
            block = TitanSwitchGLU(table, q["group_size"], q["bits"], on_route=on_route)
            # drop the lazy stacked weights and splice in the TITAN block
            mlp = layer.mlp
            del mlp.switch_mlp  # removes the lazy stacked SwitchGLU (never evaluated)
            mlp.switch_mlp = block  # plain attribute: not an nn.Module, not in parameters()
            self.tables.append(table)
            self.blocks.append(block)
        mx.eval(self.model.parameters())  # dense weights only

    # convenience -----------------------------------------------------------------
    @property
    def layers(self):
        return self.model.layers

    def __call__(self, inputs, cache=None):
        return self.model(inputs, cache=cache)

    def warm_all(self):
        """Load every expert (only meaningful when capacity == num_experts)."""
        for t in self.tables:
            t.ensure(list(range(self.spec.num_experts)))
            t.reset_stats()
        self.store.reset_stats()

    def cache_stats(self) -> Dict:
        per = [t.stats() for t in self.tables]
        hits = sum(p["hits"] for p in per)
        misses = sum(p["misses"] for p in per)
        return {"layers": per, "hits": hits, "misses": misses,
                "hit_rate": hits / (hits + misses) if hits + misses else None,
                "stall_time_s": sum(p["stall_time_s"] for p in per), "store": self.store.stats()}

    def reset_stats(self):
        for t in self.tables:
            t.reset_stats()
        self.store.reset_stats()