import argparse
import json
import math
from pathlib import Path
from contextlib import nullcontext

import torch
import torch.nn as nn
import torch.nn.functional as F
import sentencepiece as spm

torch.set_num_threads(12)
torch.set_num_interop_threads(12)

# ---------------------------------------------------------------------------
# Model definition — must match the training script exactly, or the
# checkpoint's state_dict won't line up with the module structure.
# ---------------------------------------------------------------------------

class RMSNorm(nn.Module):
    def __init__(self, dim, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(dim))
        self.eps = eps
    def forward(self, x):
        return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps) * self.weight

def rope_cache(seq_len, head_dim, device, dtype):
    inv_freq = 1.0 / (10000 ** (torch.arange(0, head_dim, 2, device=device, dtype=torch.float32) / head_dim))
    pos = torch.arange(seq_len, device=device, dtype=torch.float32)
    freqs = torch.outer(pos, inv_freq)
    cos = freqs.cos().to(dtype=dtype)[None, None, :, :]
    sin = freqs.sin().to(dtype=dtype)[None, None, :, :]
    return cos, sin

def apply_rope(x, cos, sin):
    x1 = x[..., ::2]
    x2 = x[..., 1::2]
    out = torch.empty_like(x)
    out[..., ::2] = x1 * cos - x2 * sin
    out[..., 1::2] = x1 * sin + x2 * cos
    return out

class CausalSelfAttention(nn.Module):
    def __init__(self, dim, n_head, dropout):
        super().__init__()
        assert dim % n_head == 0
        self.n_head = n_head
        self.head_dim = dim // n_head
        assert self.head_dim % 2 == 0
        self.qkv = nn.Linear(dim, 3 * dim)
        self.proj = nn.Linear(dim, dim)
        self.dropout = dropout
    def forward(self, x):
        B, T, C = x.shape
        q, k, v = self.qkv(x).chunk(3, dim=-1)
        q = q.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
        k = k.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
        v = v.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
        cos, sin = rope_cache(T, self.head_dim, x.device, x.dtype)
        q = apply_rope(q, cos, sin)
        k = apply_rope(k, cos, sin)
        if hasattr(F, "scaled_dot_product_attention"):
            a = F.scaled_dot_product_attention(q, k, v, dropout_p=0.0, is_causal=True)
        else:
            att = (q @ k.transpose(-2, -1)) / math.sqrt(self.head_dim)
            mask = torch.triu(torch.ones(T, T, device=x.device, dtype=torch.bool), diagonal=1)
            att = att.masked_fill(mask, float("-inf"))
            att = F.softmax(att, dim=-1)
            a = att @ v
        a = a.transpose(1, 2).contiguous().view(B, T, C)
        return self.proj(a)

class SwiGLU(nn.Module):
    def __init__(self, dim, dropout):
        super().__init__()
        hidden = 4 * dim
        self.fc = nn.Linear(dim, hidden * 2)
        self.proj = nn.Linear(hidden, dim)
        self.drop = nn.Dropout(dropout)
    def forward(self, x):
        x1, x2 = self.fc(x).chunk(2, dim=-1)
        return self.drop(self.proj(F.silu(x1) * x2))

class Block(nn.Module):
    def __init__(self, dim, n_head, dropout):
        super().__init__()
        self.n1 = RMSNorm(dim)
        self.attn = CausalSelfAttention(dim, n_head, dropout)
        self.n2 = RMSNorm(dim)
        self.mlp = SwiGLU(dim, dropout)
    def forward(self, x):
        x = x + self.attn(self.n1(x))
        x = x + self.mlp(self.n2(x))
        return x

class GPT(nn.Module):
    def __init__(self, vocab_size, block_size, n_layer, n_head, n_embd, dropout=0.0):
        super().__init__()
        self.block_size = block_size
        self.tok_emb = nn.Embedding(vocab_size, n_embd)
        self.drop = nn.Dropout(dropout)
        self.blocks = nn.ModuleList([Block(n_embd, n_head, dropout) for _ in range(n_layer)])
        self.norm_f = RMSNorm(n_embd)
        self.lm_head = nn.Linear(n_embd, vocab_size, bias=False)
        self.lm_head.weight = self.tok_emb.weight
    def forward(self, idx):
        B, T = idx.shape
        if T > self.block_size:
            idx = idx[:, -self.block_size:]
        x = self.tok_emb(idx)
        x = self.drop(x)
        for block in self.blocks:
            x = block(x)
        x = self.norm_f(x)
        logits = self.lm_head(x)
        return logits

def get_stop_ids(sp):
    ids = set()
    for piece in ("<|user|>", "<|system|>"):
        pid = sp.piece_to_id(piece)
        if pid != sp.unk_id():
            ids.add(pid)
    return ids

DEFAULT_CONFIG = {
    "vocab_size": 32000,
    "block_size": 512,
    "n_layer": 10,
    "n_head": 8,
    "n_embd": 576,
}


def pick_device():
    if torch.cuda.is_available():
        return torch.device("cuda")
    if torch.backends.mps.is_available():
        return torch.device("mps")
    return torch.device("cpu")

_MODULE_DIR = Path(__file__).resolve().parent

DEFAULT_CKPT = _MODULE_DIR / "qWisp-base-v1.pt"
DEFAULT_TOKENIZER = _MODULE_DIR / "qWisp.model"

class Wisp:
    def __init__(self):
        self.device = pick_device()
        self.model = None
        self.tokenizer = None
        self.stop_ids = None
        self.config = DEFAULT_CONFIG.copy()

    def load(
        self,
        ckpt=str(DEFAULT_CKPT),
        tokenizer=str(DEFAULT_TOKENIZER),
        device=None,
    ):
        if device is not None:
            self.device = torch.device(device)

        self.tokenizer = spm.SentencePieceProcessor()
        self.tokenizer.load(tokenizer)

        self.model = GPT(
            vocab_size=self.config["vocab_size"],
            block_size=self.config["block_size"],
            n_layer=self.config["n_layer"],
            n_head=self.config["n_head"],
            n_embd=self.config["n_embd"],
            dropout=0.0,
        )

        obj = torch.load(ckpt, map_location=self.device)

        state_dict = obj["model"] if isinstance(obj, dict) and "model" in obj else obj

        self.model.load_state_dict(state_dict, strict=True)
        self.model.to(self.device)
        self.model.eval()

        self.stop_ids = get_stop_ids(self.tokenizer)

        return self

    def unload(self):
        self.model = None
        self.tokenizer = None
        self.stop_ids = None

        if torch.cuda.is_available():
            torch.cuda.empty_cache()

    def encode(self, text):
        if self.tokenizer is None:
            raise RuntimeError("Model not loaded.")
        return self.tokenizer.encode(text, out_type=int)

    def decode(self, ids):
        if self.tokenizer is None:
            raise RuntimeError("Model not loaded.")
        return self.tokenizer.decode(ids)

    @torch.no_grad()
    def generate(
        self,
        prompt,
        max_new_tokens=200,
        temperature=0.8,
        top_k=50,
    ):
        if self.model is None:
            raise RuntimeError("Model not loaded.")

        ids = self.tokenizer.encode(prompt, out_type=int)

        if len(ids) == 0:
            ids = [self.tokenizer.bos_id()]

        x = torch.tensor([ids], dtype=torch.long, device=self.device)

        self.model.eval()

        for _ in range(max_new_tokens):
            x_cond = x[:, -self.model.block_size :]
            logits = self.model(x_cond)
            logits = logits[:, -1] / max(temperature, 1e-6)

            if top_k is not None and top_k > 0:
                values, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                logits = torch.where(
                    logits < values[:, [-1]],
                    torch.full_like(logits, float("-inf")),
                    logits,
                )

            probs = F.softmax(logits, dim=-1)

            next_token = torch.multinomial(probs, 1)
            token = int(next_token.item())

            if token == self.tokenizer.eos_id() or token in self.stop_ids:
                break

            x = torch.cat((x, next_token), dim=1)

        return self.tokenizer.decode(x[0].tolist())

    def chat(
        self,
        message,
        max_new_tokens=200,
        temperature=0.8,
        top_k=50,
        syspr="Answer to what user have said."
    ):
        prompt = f"<|system|> {syspr}\n<|user|> {message}\n<|assistant|>"
        return self.generate(
            prompt,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            top_k=top_k,
        )

    def __call__(
        self,
        message,
        max_new_tokens=200,
        temperature=0.8,
        top_k=50,
    ):
        return self.chat(
            message,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            top_k=top_k,
        )

    def __repr__(self):
        loaded = self.model is not None
        return (
            f"Wisp("
            f"loaded={loaded}, "
            f"device='{self.device}', "
            f"layers={self.config['n_layer']}, "
            f"hidden={self.config['n_embd']}, "
            f"vocab={self.config['vocab_size']}"
            f")"
        )