use std::collections::HashMap;
use half::bf16;
use xserv_kernels::*;
use xserv_tensor::{DType, Device, Tensor};

use crate::config::ModelConfig;
use crate::gpt2::KVCache;

pub struct Qwen3 {
    pub config: ModelConfig,
    embed_tokens: Tensor,
    layers: Vec<Qwen3Block>,
    norm: Tensor,
    lm_head_t: Tensor,  // precomputed transpose
    rope_cache: RopeCache,
}

struct Qwen3Block {
    input_norm: Tensor,       // [hidden]
    q_proj_wt: Tensor,        // TRANSPOSED: [hidden, num_heads*head_dim]
    k_proj_wt: Tensor,        // TRANSPOSED: [hidden, num_kv_heads*head_dim]
    v_proj_wt: Tensor,
    o_proj_wt: Tensor,        // TRANSPOSED: [num_heads*head_dim, hidden]
    q_norm: Tensor,           // [head_dim]
    k_norm: Tensor,           // [head_dim]
    post_norm: Tensor,        // [hidden]
    gate_proj_wt: Tensor,     // TRANSPOSED: [hidden, intermediate]
    up_proj_wt: Tensor,
    down_proj_wt: Tensor,     // TRANSPOSED: [intermediate, hidden]
}

impl Qwen3 {
    pub fn from_weights(config: ModelConfig, mut w: HashMap<String, Tensor>) -> Self {
        let take = |w: &mut HashMap<String, Tensor>, name: &str| -> Tensor {
            w.remove(name).unwrap_or_else(|| panic!("missing weight: {name}"))
        };

        let embed_tokens = take(&mut w, "model.embed_tokens.weight");
        let norm = take(&mut w, "model.norm.weight");
        let lm_head_raw = take(&mut w, "lm_head.weight");

        let rope_cache = RopeCache::new(
            config.max_seq_len().min(8192), // limit for memory
            config.head_dim(),
            config.rope_theta.unwrap_or(1_000_000.0) as f32,
        );

        // Precompute transposed weights: [out, in] → [in, out] so we can do x @ wt directly
        let transpose_w = |t: Tensor| -> Tensor {
            t.transpose(0, 1).contiguous()
        };

        let num_layers = config.num_layers();
        let mut layers = Vec::with_capacity(num_layers);
        eprintln!("Transposing weights for {} layers...", num_layers);
        for i in 0..num_layers {
            let p = format!("model.layers.{i}");
            layers.push(Qwen3Block {
                input_norm: take(&mut w, &format!("{p}.input_layernorm.weight")),
                q_proj_wt: transpose_w(take(&mut w, &format!("{p}.self_attn.q_proj.weight"))),
                k_proj_wt: transpose_w(take(&mut w, &format!("{p}.self_attn.k_proj.weight"))),
                v_proj_wt: transpose_w(take(&mut w, &format!("{p}.self_attn.v_proj.weight"))),
                o_proj_wt: transpose_w(take(&mut w, &format!("{p}.self_attn.o_proj.weight"))),
                q_norm: take(&mut w, &format!("{p}.self_attn.q_norm.weight")),
                k_norm: take(&mut w, &format!("{p}.self_attn.k_norm.weight")),
                post_norm: take(&mut w, &format!("{p}.post_attention_layernorm.weight")),
                gate_proj_wt: transpose_w(take(&mut w, &format!("{p}.mlp.gate_proj.weight"))),
                up_proj_wt: transpose_w(take(&mut w, &format!("{p}.mlp.up_proj.weight"))),
                down_proj_wt: transpose_w(take(&mut w, &format!("{p}.mlp.down_proj.weight"))),
            });
        }

        let lm_head_t = transpose_w(lm_head_raw);
        Self { config, embed_tokens, layers, norm, lm_head_t, rope_cache }
    }

    pub fn forward_with_cache(&self, token_ids: &[u32], cache: &mut KVCache) -> Tensor {
        let new_tokens = token_ids.len();
        let pos_offset = cache.seq_len();
        let hidden = self.config.hidden();
        let num_heads = self.config.num_heads();
        let num_kv_heads = self.config.num_kv_heads();
        let head_dim = self.config.head_dim();
        let eps = self.config.rms_norm_eps.unwrap_or(1e-6) as f32;

        let mut x = embedding(&self.embed_tokens, token_ids);
        let positions: Vec<u32> = (pos_offset..pos_offset + new_tokens).map(|p| p as u32).collect();

        for (layer_idx, layer) in self.layers.iter().enumerate() {
            let residual = x.clone();
            let normed = rmsnorm(&x, &layer.input_norm, eps);

            // Q/K/V projections (pre-transposed weights, x @ wt)
            let q = matmul_2d(&normed, &layer.q_proj_wt);
            let k = matmul_2d(&normed, &layer.k_proj_wt);
            let v = matmul_2d(&normed, &layer.v_proj_wt);

            // Reshape to [1, heads, seq, head_dim]
            let q = reshape_heads(&q, new_tokens, num_heads, head_dim);
            let k = reshape_heads(&k, new_tokens, num_kv_heads, head_dim);
            let v = reshape_heads(&v, new_tokens, num_kv_heads, head_dim);

            // QK normalization (per-head RMSNorm)
            let q = head_rmsnorm(&q, &layer.q_norm, eps);
            let k = head_rmsnorm(&k, &layer.k_norm, eps);

            // RoPE — kernel expects [S, H, D], our tensors are [1, H, S, D]
            // Transpose to [1, S, H, D] → reshape to [S, H, D] for RoPE
            let q = transpose_for_rope(&q, new_tokens, num_heads, head_dim);
            let k = transpose_for_rope(&k, new_tokens, num_kv_heads, head_dim);
            rope_inplace(&q, &self.rope_cache, &positions);
            rope_inplace(&k, &self.rope_cache, &positions);
            // Transpose back to [1, H, S, D]
            let q = transpose_from_rope(&q, new_tokens, num_heads, head_dim);
            let k = transpose_from_rope(&k, new_tokens, num_kv_heads, head_dim);

            // KV cache
            let k_cpu = k.to_device(Device::Cpu);
            let v_cpu = v.to_device(Device::Cpu);
            cache.append_kv_tensor(layer_idx, &k_cpu, &v_cpu, new_tokens);
            let (k_full, v_full) = cache.get_kv_tensors(layer_idx);

            // GQA: repeat K/V
            let n_rep = num_heads / num_kv_heads;
            let k_full = repeat_kv(&k_full, n_rep);
            let v_full = repeat_kv(&v_full, n_rep);

            // Attention
            let attn_out = attention(&q, &k_full, &v_full, true);
            let attn_merged = merge_heads_any(&attn_out, new_tokens, hidden);
            let attn_proj = matmul_2d(&attn_merged, &layer.o_proj_wt);
            x = add_any(&residual, &attn_proj);

            // SwiGLU FFN
            let residual = x.clone();
            let normed = rmsnorm(&x, &layer.post_norm, eps);
            let gate = matmul_2d(&normed, &layer.gate_proj_wt);
            let up = matmul_2d(&normed, &layer.up_proj_wt);
            let gate_activated = silu(&gate);
            let hidden_states = mul_any(&gate_activated, &up);
            let down = matmul_2d(&hidden_states, &layer.down_proj_wt);
            x = add_any(&residual, &down);
        }

        let x = rmsnorm(&x, &self.norm, eps);
        matmul_2d(&x, &self.lm_head_t)
    }
}

// --- Helpers ---

fn matmul_2d(a: &Tensor, b: &Tensor) -> Tensor {
    assert_eq!(a.ndim(), 2);
    assert_eq!(b.ndim(), 2);
    matmul(a, b, GemmBackend::CuBlas)
}

fn reshape_heads(x: &Tensor, seq_len: usize, num_heads: usize, head_dim: usize) -> Tensor {
    let x_cpu = x.to_device(Device::Cpu);
    let hidden = num_heads * head_dim;
    let src = x_cpu.as_slice::<bf16>();
    let mut out = vec![bf16::ZERO; num_heads * seq_len * head_dim];
    for s in 0..seq_len {
        for h in 0..num_heads {
            let si = s * hidden + h * head_dim;
            let di = (h * seq_len + s) * head_dim;
            out[di..di + head_dim].copy_from_slice(&src[si..si + head_dim]);
        }
    }
    Tensor::from_slice(&out, &[1, num_heads, seq_len, head_dim]).to_device(x.device())
}

fn merge_heads_any(x: &Tensor, seq_len: usize, hidden: usize) -> Tensor {
    let num_heads = x.shape()[1];
    let head_dim = x.shape()[3];
    let x_cpu = x.to_device(Device::Cpu);
    let src = x_cpu.as_slice::<bf16>();
    let mut out = vec![bf16::ZERO; seq_len * hidden];
    for s in 0..seq_len {
        for h in 0..num_heads {
            let si = (h * seq_len + s) * head_dim;
            let di = s * hidden + h * head_dim;
            out[di..di + head_dim].copy_from_slice(&src[si..si + head_dim]);
        }
    }
    Tensor::from_slice(&out, &[seq_len, hidden]).to_device(x.device())
}

/// Per-head RMSNorm: apply RMSNorm to each [head_dim] slice independently.
/// x: [1, H, S, D], norm_weight: [D]
fn head_rmsnorm(x: &Tensor, norm_weight: &Tensor, eps: f32) -> Tensor {
    let num_heads = x.shape()[1];
    let seq_len = x.shape()[2];
    let head_dim = x.shape()[3];
    // Reshape to [H*S, D], apply rmsnorm, reshape back
    let total_rows = num_heads * seq_len;
    let flat = x.reshape(&[total_rows, head_dim]);
    let normed = rmsnorm(&flat, norm_weight, eps);
    normed.reshape(&[1, num_heads, seq_len, head_dim])
}

/// [1, H, S, D] → [S, H, D] for RoPE kernel
fn transpose_for_rope(x: &Tensor, seq_len: usize, num_heads: usize, head_dim: usize) -> Tensor {
    let x_cpu = x.to_device(Device::Cpu);
    let src = x_cpu.as_slice::<bf16>();
    let mut out = vec![bf16::ZERO; seq_len * num_heads * head_dim];
    for h in 0..num_heads {
        for s in 0..seq_len {
            let si = (h * seq_len + s) * head_dim;
            let di = (s * num_heads + h) * head_dim;
            out[di..di + head_dim].copy_from_slice(&src[si..si + head_dim]);
        }
    }
    Tensor::from_slice(&out, &[seq_len, num_heads, head_dim]).to_device(x.device())
}

/// [S, H, D] → [1, H, S, D] after RoPE
fn transpose_from_rope(x: &Tensor, seq_len: usize, num_heads: usize, head_dim: usize) -> Tensor {
    let x_cpu = x.to_device(Device::Cpu);
    let src = x_cpu.as_slice::<bf16>();
    let mut out = vec![bf16::ZERO; num_heads * seq_len * head_dim];
    for s in 0..seq_len {
        for h in 0..num_heads {
            let si = (s * num_heads + h) * head_dim;
            let di = (h * seq_len + s) * head_dim;
            out[di..di + head_dim].copy_from_slice(&src[si..si + head_dim]);
        }
    }
    Tensor::from_slice(&out, &[1, num_heads, seq_len, head_dim]).to_device(x.device())
}

fn repeat_kv(x: &Tensor, n_rep: usize) -> Tensor {
    if n_rep == 1 { return x.clone(); }
    let kv_heads = x.shape()[1];
    let seq_len = x.shape()[2];
    let head_dim = x.shape()[3];
    let x_cpu = x.to_device(Device::Cpu);
    let src = x_cpu.as_slice::<bf16>();
    let new_heads = kv_heads * n_rep;
    let mut out = vec![bf16::ZERO; new_heads * seq_len * head_dim];
    let chunk = seq_len * head_dim;
    for kv_h in 0..kv_heads {
        for r in 0..n_rep {
            let dst_h = kv_h * n_rep + r;
            out[dst_h * chunk..(dst_h + 1) * chunk]
                .copy_from_slice(&src[kv_h * chunk..(kv_h + 1) * chunk]);
        }
    }
    Tensor::from_slice(&out, &[1, new_heads, seq_len, head_dim]).to_device(x.device())
}

fn add_any(a: &Tensor, b: &Tensor) -> Tensor {
    assert_eq!(a.shape(), b.shape());
    let a_cpu = a.to_device(Device::Cpu);
    let b_cpu = b.to_device(Device::Cpu);
    let ad = a_cpu.as_slice::<bf16>();
    let bd = b_cpu.as_slice::<bf16>();
    let r: Vec<bf16> = ad.iter().zip(bd)
        .map(|(x, y)| bf16::from_f32(x.to_f32() + y.to_f32()))
        .collect();
    Tensor::from_slice(&r, a.shape()).to_device(a.device())
}

fn mul_any(a: &Tensor, b: &Tensor) -> Tensor {
    assert_eq!(a.shape(), b.shape());
    let a_cpu = a.to_device(Device::Cpu);
    let b_cpu = b.to_device(Device::Cpu);
    let ad = a_cpu.as_slice::<bf16>();
    let bd = b_cpu.as_slice::<bf16>();
    let r: Vec<bf16> = ad.iter().zip(bd)
        .map(|(x, y)| bf16::from_f32(x.to_f32() * y.to_f32()))
        .collect();
    Tensor::from_slice(&r, a.shape()).to_device(a.device())
}

pub fn sample_greedy(logits: &Tensor) -> u32 {
    assert_eq!(logits.ndim(), 2);
    let logits_cpu = logits.to_device(Device::Cpu);
    let vocab_size = logits.shape()[1];
    let seq_len = logits.shape()[0];
    let data = logits_cpu.as_slice::<bf16>();
    let last = &data[(seq_len - 1) * vocab_size..seq_len * vocab_size];
    last.iter().enumerate()
        .max_by(|a, b| a.1.to_f32().partial_cmp(&b.1.to_f32()).unwrap())
        .map(|(i, _)| i as u32).unwrap()
}