# Copyright © 2025 Poolside AI / MLX port by Théophile Lafargue (@eauchs, x.com/eauchs)
#
# Laguna M.1 — MLX implementation
# Architecture: MoE (256 experts, top-16, sigmoid routing)
# Attention: GQA with QK-norm + softplus output gating (g_proj)
# RoPE: YaRN (factor=32, original_max=4096, beta_fast=64, beta_slow=1)

from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Union

import mlx.core as mx
import mlx.nn as nn

from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
from .rope_utils import initialize_rope
from .switch_layers import SwitchGLU


@dataclass
class ModelArgs(BaseModelArgs):
    model_type: str = "laguna"
    vocab_size: int = 100352
    hidden_size: int = 4096
    intermediate_size: int = 16384
    num_hidden_layers: int = 70
    num_attention_heads: int = 64
    num_key_value_heads: int = 8
    head_dim: int = 128
    max_position_embeddings: int = 131072
    rms_norm_eps: float = 1e-6
    rope_theta: float = 500000.0
    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
    attention_bias: bool = False
    attention_dropout: float = 0.0
    tie_word_embeddings: bool = False
    # MoE
    num_experts: int = 256
    num_experts_per_tok: int = 16
    moe_intermediate_size: int = 1024
    shared_expert_intermediate_size: int = 1024
    norm_topk_prob: bool = True
    moe_routed_scaling_factor: float = 1.0
    # Layer types: "dense" or "sparse" per layer
    mlp_layer_types: Optional[List[str]] = None
    # Gating: "per-element" enables softplus output gating on attention
    gating: Optional[str] = "per-element"


class LagunaAttention(nn.Module):
    """Laguna attention: GQA + QK-norm + softplus output gating."""

    def __init__(self, args: ModelArgs):
        super().__init__()
        dim = args.hidden_size
        self.n_heads = args.num_attention_heads
        self.n_kv_heads = args.num_key_value_heads
        self.head_dim = args.head_dim
        self.scale = self.head_dim ** -0.5

        self.q_proj = nn.Linear(dim, self.n_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(dim, self.n_kv_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(dim, self.n_kv_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.n_heads * self.head_dim, dim, bias=False)

        # Laguna-specific: output gating projection
        self.g_proj = nn.Linear(dim, self.n_heads * self.head_dim, bias=False)

        # QK normalization (per-head RMSNorm before RoPE)
        self.q_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)
        self.k_norm = nn.RMSNorm(self.head_dim, eps=args.rms_norm_eps)

        self.rope = initialize_rope(
            dims=self.head_dim,
            base=args.rope_theta,
            traditional=False,
            max_position_embeddings=args.max_position_embeddings,
            scaling_config=args.rope_scaling,
        )

    def __call__(
        self,
        x: mx.array,
        mask: Optional[mx.array] = None,
        cache: Optional[Any] = None,
    ) -> mx.array:
        B, L, D = x.shape

        queries = self.q_proj(x)
        keys = self.k_proj(x)
        values = self.v_proj(x)

        # Reshape to [B, n_heads, L, head_dim]
        queries = self.q_norm(
            queries.reshape(B, L, self.n_heads, -1)
        ).transpose(0, 2, 1, 3)
        keys = self.k_norm(
            keys.reshape(B, L, self.n_kv_heads, -1)
        ).transpose(0, 2, 1, 3)
        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)

        # RoPE
        if cache is not None:
            queries = self.rope(queries, offset=cache.offset)
            keys = self.rope(keys, offset=cache.offset)
            keys, values = cache.update_and_fetch(keys, values)
        else:
            queries = self.rope(queries)
            keys = self.rope(keys)

        # Attention
        output = scaled_dot_product_attention(
            queries, keys, values, cache=cache, scale=self.scale, mask=mask
        )
        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)

        # Laguna-specific: softplus gating BEFORE o_proj
        gate = nn.softplus(self.g_proj(x))
        output = output * gate

        return self.o_proj(output)


class LagunaMLP(nn.Module):
    def __init__(self, dim: int, hidden_dim: int):
        super().__init__()
        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)

    def __call__(self, x: mx.array) -> mx.array:
        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))


class LagunaRouter(nn.Module):
    """Sigmoid-based top-k router (not softmax)."""

    def __init__(self, args: ModelArgs):
        super().__init__()
        self.top_k = args.num_experts_per_tok
        self.num_experts = args.num_experts
        self.norm_topk_prob = args.norm_topk_prob
        self.weight = mx.zeros((self.num_experts, args.hidden_size))
        self.e_score_correction_bias = mx.zeros((self.num_experts,))

    def __call__(self, x: mx.array):
        # Sigmoid scoring in float32 for precision
        router_logits = x @ self.weight.T
        scores = mx.sigmoid(router_logits.astype(mx.float32))

        # Apply correction bias for expert selection (not for final weights)
        corrected_scores = scores + self.e_score_correction_bias

        # Top-k selection on corrected scores
        k = self.top_k
        inds = mx.argpartition(-corrected_scores, kth=k - 1, axis=-1)[..., :k]
        # Use original scores (without bias) for weighting
        routing_weights = mx.take_along_axis(scores, inds, axis=-1)

        if self.norm_topk_prob:
            routing_weights = routing_weights / mx.sum(
                routing_weights, axis=-1, keepdims=True
            )

        return inds, routing_weights.astype(x.dtype)


class LagunaSparseMoeBlock(nn.Module):
    """Laguna MoE: sigmoid router + SwitchGLU experts + shared expert."""

    def __init__(self, args: ModelArgs):
        super().__init__()
        self.gate = LagunaRouter(args)
        self.switch_mlp = SwitchGLU(
            args.hidden_size,
            args.moe_intermediate_size,
            args.num_experts,
        )
        self.shared_expert = LagunaMLP(
            args.hidden_size, args.shared_expert_intermediate_size
        )

    def __call__(self, x: mx.array) -> mx.array:
        inds, scores = self.gate(x)

        # Routed experts via SwitchGLU
        y = self.switch_mlp(x, inds)
        y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)

        # Add shared expert
        y = y + self.shared_expert(x)

        return y


class LagunaDecoderLayer(nn.Module):
    def __init__(self, args: ModelArgs, layer_idx: int):
        super().__init__()
        self.self_attn = LagunaAttention(args)
        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
        self.post_attention_layernorm = nn.RMSNorm(
            args.hidden_size, eps=args.rms_norm_eps
        )

        # Dense or sparse MLP based on mlp_layer_types
        is_sparse = (
            args.mlp_layer_types is not None
            and layer_idx < len(args.mlp_layer_types)
            and args.mlp_layer_types[layer_idx] == "sparse"
        )
        if is_sparse:
            self.mlp = LagunaSparseMoeBlock(args)
        else:
            self.mlp = LagunaMLP(args.hidden_size, args.intermediate_size)

    def __call__(
        self,
        x: mx.array,
        mask: Optional[mx.array] = None,
        cache: Optional[Any] = None,
    ) -> mx.array:
        r = self.self_attn(self.input_layernorm(x), mask, cache)
        h = x + r
        r = self.mlp(self.post_attention_layernorm(h))
        return h + r


class LagunaModel(nn.Module):
    def __init__(self, args: ModelArgs):
        super().__init__()
        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
        self.layers = [
            LagunaDecoderLayer(args, i) for i in range(args.num_hidden_layers)
        ]
        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)

    def __call__(
        self,
        inputs: mx.array,
        cache=None,
    ) -> mx.array:
        h = self.embed_tokens(inputs)

        if cache is None:
            cache = [None] * len(self.layers)

        mask = create_attention_mask(h, cache[0])

        for layer, c in zip(self.layers, cache):
            h = layer(h, mask, c)

        return self.norm(h)


class Model(nn.Module):
    def __init__(self, args: ModelArgs):
        super().__init__()
        self.args = args
        self.model_type = args.model_type
        self.model = LagunaModel(args)
        if not args.tie_word_embeddings:
            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)

    def __call__(
        self,
        inputs: mx.array,
        cache=None,
    ):
        out = self.model(inputs, cache)
        if self.args.tie_word_embeddings:
            return self.model.embed_tokens.as_linear(out)
        return self.lm_head(out)

    def sanitize(self, weights):
        # Dequantize FP8 compressed-tensors: uint8 + weight_scale → bfloat16
        new_weights = {}
        for k, v in weights.items():
            if k.endswith("weight_scale"):
                continue  # handled with corresponding weight
            if k.endswith(".weight") and f"{k}_scale" in weights:
                scale = weights[f"{k}_scale"]
                new_weights[k] = self._dequant_fp8_block(v, scale)
            else:
                new_weights[k] = v
        weights = new_weights

        # Remap e_score_correction_bias:
        # HF: model.layers.X.mlp.experts.e_score_correction_bias
        # MLX: model.layers.X.mlp.gate.e_score_correction_bias
        remapped = {}
        for k, v in weights.items():
            if "mlp.experts.e_score_correction_bias" in k:
                new_key = k.replace(
                    "mlp.experts.e_score_correction_bias",
                    "mlp.gate.e_score_correction_bias",
                )
                remapped[new_key] = v
            else:
                remapped[k] = v
        weights = remapped

        # Stack per-expert weights into SwitchGLU format
        for l in range(self.args.num_hidden_layers):
            prefix = f"model.layers.{l}"
            for n in ["up_proj", "down_proj", "gate_proj"]:
                key0 = f"{prefix}.mlp.experts.0.{n}.weight"
                if key0 in weights:
                    to_join = [
                        weights.pop(f"{prefix}.mlp.experts.{e}.{n}.weight")
                        for e in range(self.args.num_experts)
                    ]
                    weights[f"{prefix}.mlp.switch_mlp.{n}.weight"] = mx.stack(
                        to_join
                    )

        # Remove unused keys
        weights = {
            k: v
            for k, v in weights.items()
            if "rotary_emb.inv_freq" not in k
        }

        return weights

    @staticmethod
    def _dequant_fp8_block(weight: mx.array, scale: mx.array) -> mx.array:
        """Dequantize FP8 block-quantized weight (uint8 + scale → bfloat16).

        Block size is 128x128 (inferred from scale shape vs weight shape).
        """
        m, n = weight.shape
        sm, sn = scale.shape
        bs_m = m // sm  # block size along rows
        bs_n = n // sn  # block size along cols

        # Reinterpret uint8 as float8 E4M3, then cast to bfloat16
        weight_bf16 = mx.from_fp8(weight, mx.bfloat16)

        # Apply block-wise scale
        weight_bf16 = weight_bf16.reshape(sm, bs_m, sn, bs_n)
        weight_bf16 = weight_bf16 * scale[:, None, :, None]
        weight_bf16 = weight_bf16.reshape(m, n)

        return weight_bf16

    @property
    def layers(self):
        return self.model.layers

    @property
    def quant_predicate(self):
        def predicate(path, _):
            # Don't quantize router gate weights — keep full precision
            if "mlp.gate.weight" in path:
                return {"group_size": 64, "bits": 8}
            return True

        return predicate