/// Mistral LLM, https://github.com/mistralai/mistral-src
use crate::models::{
    mistral::Config,
    with_tracing::{linear_no_bias, Linear, RmsNorm},
};
use crate::utils::repeat_kv;
use candle::{DType, Device, Module, Result, Tensor};
use candle_nn::{Activation, VarBuilder};
use std::sync::Arc;

#[derive(Debug, Clone)]
struct RotaryEmbedding {
    sin: Tensor,
    cos: Tensor,
}

impl RotaryEmbedding {
    fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
        let rope_theta = cfg.rope_theta as f32;
        let dim = cfg.hidden_size / cfg.num_attention_heads;
        let max_seq_len = cfg.max_position_embeddings;
        let inv_freq: Vec<_> = (0..dim)
            .step_by(2)
            .map(|i| 1f32 / rope_theta.powf(i as f32 / dim as f32))
            .collect();
        let inv_freq_len = inv_freq.len();
        let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), dev)?.to_dtype(dtype)?;
        let t = Tensor::arange(0u32, max_seq_len as u32, dev)?
            .to_dtype(dtype)?
            .reshape((max_seq_len, 1))?;
        let freqs = t.matmul(&inv_freq)?;
        Ok(Self {
            sin: freqs.sin()?,
            cos: freqs.cos()?,
        })
    }

    fn apply_rotary_emb_qkv(
        &self,
        q: &Tensor,
        k: &Tensor,
        seqlen_offset: usize,
    ) -> Result<(Tensor, Tensor)> {
        let (_b_sz, _h, seq_len, _n_embd) = q.dims4()?;
        let cos = self.cos.narrow(0, seqlen_offset, seq_len)?;
        let sin = self.sin.narrow(0, seqlen_offset, seq_len)?;
        let q_embed = candle_nn::rotary_emb::rope(q, &cos, &sin)?;
        let k_embed = candle_nn::rotary_emb::rope(k, &cos, &sin)?;
        Ok((q_embed, k_embed))
    }
}

#[derive(Debug, Clone)]
#[allow(clippy::upper_case_acronyms)]
struct MLP {
    gate_proj: Linear,
    up_proj: Linear,
    down_proj: Linear,
    act_fn: Activation,
}

impl MLP {
    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
        let hidden_sz = cfg.hidden_size;
        let intermediate_sz = cfg.intermediate_size;
        let gate_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("gate_proj"))?;
        let up_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("up_proj"))?;
        let down_proj = linear_no_bias(intermediate_sz, hidden_sz, vb.pp("down_proj"))?;
        Ok(Self {
            gate_proj,
            up_proj,
            down_proj,
            act_fn: cfg.hidden_act,
        })
    }
}

impl Module for MLP {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let lhs = xs.apply(&self.gate_proj)?.apply(&self.act_fn)?;
        let rhs = xs.apply(&self.up_proj)?;
        (lhs * rhs)?.apply(&self.down_proj)
    }
}

#[derive(Debug, Clone)]
struct Attention {
    q_proj: Linear,
    k_proj: Linear,
    v_proj: Linear,
    o_proj: Linear,
    num_heads: usize,
    num_kv_heads: usize,
    num_kv_groups: usize,
    head_dim: usize,
    hidden_size: usize,
    rotary_emb: Arc<RotaryEmbedding>,
}

impl Attention {
    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
        let hidden_sz = cfg.hidden_size;
        let num_heads = cfg.num_attention_heads;
        let num_kv_heads = cfg.num_key_value_heads;
        let num_kv_groups = num_heads / num_kv_heads;
        let head_dim = hidden_sz / num_heads;
        let q_proj = linear_no_bias(hidden_sz, num_heads * head_dim, vb.pp("q_proj"))?;
        let k_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("k_proj"))?;
        let v_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("v_proj"))?;
        let o_proj = linear_no_bias(num_heads * head_dim, hidden_sz, vb.pp("o_proj"))?;
        Ok(Self {
            q_proj,
            k_proj,
            v_proj,
            o_proj,
            num_heads,
            num_kv_heads,
            num_kv_groups,
            head_dim,
            hidden_size: hidden_sz,
            rotary_emb,
        })
    }

    fn forward(
        &mut self,
        xs: &Tensor,
        attention_mask: Option<&Tensor>,
        seqlen_offset: usize,
    ) -> Result<Tensor> {
        let (b_sz, q_len, _) = xs.dims3()?;

        let query_states = self.q_proj.forward(xs)?;
        let key_states = self.k_proj.forward(xs)?;
        let value_states = self.v_proj.forward(xs)?;

        let query_states = query_states
            .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
            .transpose(1, 2)?
            .contiguous()?;

        let key_states = key_states
            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
            .transpose(1, 2)?
            .contiguous()?;
        let value_states = value_states
            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
            .transpose(1, 2)?;

        let (query_states, key_states) =
            self.rotary_emb
                .apply_rotary_emb_qkv(&query_states, &key_states, seqlen_offset)?;

        let key_states = repeat_kv(key_states, self.num_kv_groups)?;
        let value_states = repeat_kv(value_states, self.num_kv_groups)?;

        let scale = 1f64 / f64::sqrt(self.head_dim as f64);
        let attn_weights = (query_states.matmul(&key_states.transpose(2, 3)?)? * scale)?;

        let attn_weights = match attention_mask {
            None => attn_weights,
            Some(mask) => attn_weights.broadcast_add(mask)?,
        };
        let attn_weights = candle_nn::ops::softmax_last_dim(&attn_weights)?;
        let attn_output = attn_weights.matmul(&value_states)?;

        attn_output
            .transpose(1, 2)?
            .reshape((b_sz, q_len, self.hidden_size))?
            .apply(&self.o_proj)
    }
}

#[derive(Debug, Clone)]
struct DecoderLayer {
    self_attn: Attention,
    mlp: MLP,
    input_layernorm: RmsNorm,
    post_attention_layernorm: RmsNorm,
}

impl DecoderLayer {
    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
        let self_attn = Attention::new(rotary_emb, cfg, vb.pp("self_attn"))?;
        let mlp = MLP::new(cfg, vb.pp("mlp"))?;
        let input_layernorm =
            RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
        let post_attention_layernorm = RmsNorm::new(
            cfg.hidden_size,
            cfg.rms_norm_eps,
            vb.pp("post_attention_layernorm"),
        )?;
        Ok(Self {
            self_attn,
            mlp,
            input_layernorm,
            post_attention_layernorm,
        })
    }

    fn forward(
        &mut self,
        xs: &Tensor,
        attention_mask: Option<&Tensor>,
        seqlen_offset: usize,
    ) -> Result<Tensor> {
        let residual = xs;
        let xs = self.input_layernorm.forward(xs)?;

        let xs = self.self_attn.forward(&xs, attention_mask, seqlen_offset)?;

        let xs = (xs + residual)?;
        let residual = &xs;
        let xs = xs.apply(&self.post_attention_layernorm)?.apply(&self.mlp)?;
        residual + xs
    }
}

#[derive(Debug, Clone)]
pub struct Model {
    embed_tokens: candle_nn::Embedding,
    layers: Vec<DecoderLayer>,
    norm: RmsNorm,
    pub cfg: Config,
}

impl Model {
    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
        let embed_tokens =
            candle_nn::embedding(cfg.vocab_size, cfg.hidden_size, vb.pp("embed_tokens"))?;
        let rotary_emb = Arc::new(RotaryEmbedding::new(vb.dtype(), cfg, vb.device())?);
        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
        let vb_l = vb.pp("layers");
        for layer_idx in 0..cfg.num_hidden_layers {
            let layer = DecoderLayer::new(rotary_emb.clone(), cfg, vb_l.pp(layer_idx))?;
            layers.push(layer)
        }
        let norm = RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("norm"))?;
        Ok(Self {
            embed_tokens,
            layers,
            norm,
            cfg: cfg.clone(),
        })
    }

    // Attn mask used to mask out padding tokens
    pub fn forward(
        &mut self,
        attn_mask: &Tensor,
        input_ids: &Tensor,
        dtype: DType,
    ) -> Result<Tensor> {
        let mut xs = self.embed_tokens.forward(input_ids)?;

        // Expand to 4d mask for sdpa
        let attn_mask = prepare_4d_attention_mask(attn_mask, dtype, None)?;

        for layer in self.layers.iter_mut() {
            xs = layer.forward(&xs, Some(&attn_mask), 0)?;
        }

        // Return hiddens instead of logits
        xs.apply(&self.norm)
    }
}

fn prepare_4d_attention_mask(
    mask: &Tensor,
    dtype: DType,
    tgt_len: Option<usize>,
) -> Result<Tensor> {
    let bsz = mask.dims()[0];
    let src_len = mask.dims()[1];
    let tgt_len = tgt_len.unwrap_or(src_len);

    let expanded_mask = mask
        .unsqueeze(1)?
        .unsqueeze(2)?
        .expand((bsz, 1, tgt_len, src_len))?
        .to_dtype(dtype)?;

    let inverted_mask = (1.0 - expanded_mask)?;

    (inverted_mask * get_dtype_min_val(dtype))?.to_dtype(dtype)
}

fn get_dtype_min_val(dtype: DType) -> f64 {
    match dtype {
        DType::F32 => f32::MIN as f64,
        DType::F64 => f64::MIN,
        _ => panic!("Unsupported data type"),
    }
}