#!/usr/bin/env python3
"""Custom loader for Compactbot/swordies-22m.

Swordies-22M is a from-scratch BPE GPT (NOT a transformers model). This file
reconstructs the architecture from config.json and loads model.safetensors.

Usage:
    from load_model import load_model
    model = load_model("model.safetensors")
    logits = model(token_ids)          # token_ids: int64 [B, T], vocab 8192
    probs  = torch.softmax(logits, -1)

Tensor layout (57 tensors, F32, weight-tied):
    tok.weight                 [8192, 448]   (also the lm_head, tied)
    pos.weight                 [512,  448]
    blocks.{0..8}.ln1.w        [448]
    blocks.{0..8}.ln2.w        [448]
    blocks.{0..8}.qkv.weight  [448, 1344]    (fused q|k|v, no bias)
    blocks.{0..8}.proj.weight  [448, 448]
    blocks.{0..8}.fc1.weight   [448, 1408]
    blocks.{0..8}.fc2.weight   [1408, 448]
    ln_f.w                     [448]
"""
import json, os
import torch
import torch.nn as nn
import torch.nn.functional as F
from safetensors import safe_open

VOCAB = 8192
D = 448
L = 9
H = 7
FFN = 1408
SEQ = 512


class RMSNorm(nn.Module):
    def __init__(self, d):
        super().__init__()
        self.w = nn.Parameter(torch.ones(d))

    def forward(self, x):
        return self.w * x * torch.rsqrt(x.float().pow(2).mean(-1, keepdim=True) + 1e-6)


class Block(nn.Module):
    def __init__(self, d, h):
        super().__init__()
        self.ln1 = RMSNorm(d)
        self.ln2 = RMSNorm(d)
        self.qkv = nn.Linear(d, 3 * d, bias=False)
        self.proj = nn.Linear(d, d, bias=False)
        self.fc1 = nn.Linear(d, FFN, bias=False)
        self.fc2 = nn.Linear(FFN, d, bias=False)
        self.h, self.d = h, d

    def forward(self, x):
        B, T, Dd = x.shape
        h = self.ln1(x)
        qkv = self.qkv(h).view(B, T, 3, self.h, Dd // self.h).transpose(2, 1)
        q, k, v = qkv[:, 0], qkv[:, 1], qkv[:, 2]
        q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
        att = F.scaled_dot_product_attention(q, k, v, is_causal=True)
        att = att.transpose(1, 2).reshape(B, T, Dd)
        x = x + self.proj(att)
        x = x + self.fc2(F.gelu(self.fc1(self.ln2(x))))
        return x


class SwordiesGPT(nn.Module):
    def __init__(self):
        super().__init__()
        self.tok = nn.Embedding(VOCAB, D)
        self.pos = nn.Embedding(SEQ, D)
        self.blocks = nn.ModuleList([Block(D, H) for _ in range(L)])
        self.ln_f = RMSNorm(D)

    def forward(self, idx, targets=None):
        B, T = idx.shape
        x = self.tok(idx) + self.pos(torch.arange(T, device=idx.device))
        for b in self.blocks:
            x = b(x)
        x = self.ln_f(x)
        logits = x @ self.tok.weight.t()
        if targets is not None:
            return F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
        return logits


def load_model(path, device="cpu"):
    """Load model.safetensors into a SwordiesGPT and return it (eval mode)."""
    model = SwordiesGPT().to(device)
    with safe_open(path, framework="pt") as f:
        state = {k: f.get_tensor(k) for k in f.keys()}
    missing, unexpected = model.load_state_dict(state, strict=True)
    model.eval()
    n = sum(p.numel() for p in model.parameters())
    assert n == 22487360, f"param mismatch: {n}"
    return model


if __name__ == "__main__":
    here = os.path.dirname(os.path.abspath(__file__))
    m = load_model(os.path.join(here, "model.safetensors"))
    x = torch.randint(0, VOCAB, (1, 64), dtype=torch.int64)
    with torch.no_grad():
        lg = m(x)
    print("loaded OK; params =", sum(p.numel() for p in m.parameters()))
    print("logits shape", tuple(lg.shape), "finite:", bool(torch.isfinite(lg).all()))