"""
The most atomic way to train and run inference for a GPT in pure, dependency-free Python.
This file is the complete algorithm.
Everything else is just efficiency.

@karpathy
"""

import os       # os.path.exists
import math     # math.log, math.exp
import random   # random.seed, random.choices, random.gauss, random.shuffle
random.seed(42) # Let there be order among chaos

# Let there be a Dataset `docs`: list[str] of documents (e.g. a list of names)
if not os.path.exists('input.txt'):
    import urllib.request
    names_url = 'https://raw.githubusercontent.com/karpathy/makemore/988aa59/names.txt'
    urllib.request.urlretrieve(names_url, 'input.txt')
docs = [line.strip() for line in open('input.txt') if line.strip()]
random.shuffle(docs)
print(f"num docs: {len(docs)}")

# Let there be a Tokenizer to translate strings to sequences of integers ("tokens") and back
uchars = sorted(set(''.join(docs))) # unique characters in the dataset become token ids 0..n-1
BOS = len(uchars) # token id for a special Beginning of Sequence (BOS) token
vocab_size = len(uchars) + 1 # total number of unique tokens, +1 is for BOS
print(f"vocab size: {vocab_size}")

# Let there be Autograd to recursively apply the chain rule through a computation graph
class Value:
    __slots__ = ('data', 'grad', '_children', '_local_grads') # Python optimization for memory usage

    def __init__(self, data, children=(), local_grads=()):
        self.data = data                # scalar value of this node calculated during forward pass
        self.grad = 0                   # derivative of the loss w.r.t. this node, calculated in backward pass
        self._children = children       # children of this node in the computation graph
        self._local_grads = local_grads # local derivative of this node w.r.t. its children

    def __add__(self, other):
        other = other if isinstance(other, Value) else Value(other)
        return Value(self.data + other.data, (self, other), (1, 1))

    def __mul__(self, other):
        other = other if isinstance(other, Value) else Value(other)
        return Value(self.data * other.data, (self, other), (other.data, self.data))

    def __pow__(self, other): return Value(self.data**other, (self,), (other * self.data**(other-1),))
    def log(self): return Value(math.log(self.data), (self,), (1/self.data,))
    def exp(self): return Value(math.exp(self.data), (self,), (math.exp(self.data),))
    def relu(self): return Value(max(0, self.data), (self,), (float(self.data > 0),))
    def __neg__(self): return self * -1
    def __radd__(self, other): return self + other
    def __sub__(self, other): return self + (-other)
    def __rsub__(self, other): return other + (-self)
    def __rmul__(self, other): return self * other
    def __truediv__(self, other): return self * other**-1
    def __rtruediv__(self, other): return other * self**-1

    def backward(self):
        topo = []
        visited = set()
        def build_topo(v):
            if v not in visited:
                visited.add(v)
                for child in v._children:
                    build_topo(child)
                topo.append(v)
        build_topo(self)
        self.grad = 1
        for v in reversed(topo):
            for child, local_grad in zip(v._children, v._local_grads):
                child.grad += local_grad * v.grad

# ===== [追加] 途中の数値を表示するための設定と関数 =====
TRACE_STEP = 0      # [追加] 何ステップ目（0始まり）の学習を表示するか
TRACE = None        # [追加] 表示中は記録用の辞書になる
PAUSE = True        # [追加] True: Enter で1段ずつ進む

def pause():
    global PAUSE
    if PAUSE:
        a = input("  ---- [Enter] 次へ  /  [a+Enter] 最後まで止まらずに表示 ---- ")
        if a.strip().lower() == 'a':
            PAUSE = False
    print()

def val(x):
    return x.data if isinstance(x, Value) else x

def show(title, M, rows, cols, fmt="{:+.3f}", w=7, mask=None):
    """行列 M（リストのリスト）を行ラベル・列ラベル付きで2次元に表示する"""
    print(f"■ {title}   （{len(M)} 行 × {len(M[0])} 列）")
    lw = max(len(r) for r in rows) + 1
    print(" " * lw + "".join(f"{c:>{w}}" for c in cols))
    for i, (r, row) in enumerate(zip(rows, M)):
        cells = []
        for j, x in enumerate(row):
            if mask is not None and mask(i, j):
                cells.append(f"{'·':>{w}}")
            else:
                cells.append(f"{fmt.format(val(x)):>{w}}")
        print(f"{r:<{lw}}" + "".join(cells))
    print()

def T(M):
    return [list(c) for c in zip(*M)]

# Initialize the parameters, to store the knowledge of the model
n_layer = 1     # depth of the transformer neural network (number of layers)
n_embd = 16     # width of the network (embedding dimension)
block_size = 16 # maximum context length of the attention window (note: the longest name is 15 characters)
n_head = 4      # number of attention heads
head_dim = n_embd // n_head # derived dimension of each head
matrix = lambda nout, nin, std=0.08: [[Value(random.gauss(0, std)) for _ in range(nin)] for _ in range(nout)]
state_dict = {'wte': matrix(vocab_size, n_embd), 'wpe': matrix(block_size, n_embd), 'lm_head': matrix(vocab_size, n_embd)}
for i in range(n_layer):
    state_dict[f'layer{i}.attn_wq'] = matrix(n_embd, n_embd)
    state_dict[f'layer{i}.attn_wk'] = matrix(n_embd, n_embd)
    state_dict[f'layer{i}.attn_wv'] = matrix(n_embd, n_embd)
    state_dict[f'layer{i}.attn_wo'] = matrix(n_embd, n_embd)
    state_dict[f'layer{i}.mlp_fc1'] = matrix(4 * n_embd, n_embd)
    state_dict[f'layer{i}.mlp_fc2'] = matrix(n_embd, 4 * n_embd)
params = [p for mat in state_dict.values() for row in mat for p in row] # flatten params into a single list[Value]
print(f"num params: {len(params)}")

# Define the model architecture: a function mapping tokens and parameters to logits over what comes next
# Follow GPT-2, blessed among the GPTs, with minor differences: layernorm -> rmsnorm, no biases, GeLU -> ReLU
def linear(x, w):
    return [sum(wi * xi for wi, xi in zip(wo, x)) for wo in w]

def softmax(logits):
    max_val = max(val.data for val in logits)
    exps = [(val - max_val).exp() for val in logits]
    total = sum(exps)
    return [e / total for e in exps]

def rmsnorm(x):
    ms = sum(xi * xi for xi in x) / len(x)
    scale = (ms + 1e-5) ** -0.5
    return [xi * scale for xi in x]

def gpt(token_id, pos_id, keys, values):
    tok_emb = state_dict['wte'][token_id] # token embedding
    pos_emb = state_dict['wpe'][pos_id] # position embedding
    x = [t + p for t, p in zip(tok_emb, pos_emb)] # joint token and position embedding
    x = rmsnorm(x) # note: not redundant due to backward pass via the residual connection
    if TRACE is not None:  # [追加]
        TRACE['tok'].append(tok_emb); TRACE['pos'].append(pos_emb); TRACE['x0'].append(x)

    for li in range(n_layer):
        # 1) Multi-head Attention block
        x_residual = x
        x = rmsnorm(x)
        q = linear(x, state_dict[f'layer{li}.attn_wq'])
        k = linear(x, state_dict[f'layer{li}.attn_wk'])
        v = linear(x, state_dict[f'layer{li}.attn_wv'])
        if TRACE is not None:  # [追加]
            TRACE['xn1'].append(x); TRACE['q'].append(q); TRACE['k'].append(k); TRACE['v'].append(v)
        keys[li].append(k)
        values[li].append(v)
        x_attn = []
        for h in range(n_head):
            hs = h * head_dim
            q_h = q[hs:hs+head_dim]
            k_h = [ki[hs:hs+head_dim] for ki in keys[li]]
            v_h = [vi[hs:hs+head_dim] for vi in values[li]]
            attn_logits = [sum(q_h[j] * k_h[t][j] for j in range(head_dim)) / head_dim**0.5 for t in range(len(k_h))]
            attn_weights = softmax(attn_logits)
            if TRACE is not None:  # [追加]
                TRACE['score'][h].append(attn_logits); TRACE['weight'][h].append(attn_weights)
            head_out = [sum(attn_weights[t] * v_h[t][j] for t in range(len(v_h))) for j in range(head_dim)]
            x_attn.extend(head_out)
        x = linear(x_attn, state_dict[f'layer{li}.attn_wo'])
        if TRACE is not None:  # [追加]
            TRACE['wo'].append(x)
        x = [a + b for a, b in zip(x, x_residual)]
        # 2) MLP block
        if TRACE is not None:  # [追加]
            TRACE['attn_cat'].append(x_attn); TRACE['x1'].append(x)
        x_residual = x
        x = rmsnorm(x)
        x = linear(x, state_dict[f'layer{li}.mlp_fc1'])
        x = [xi.relu() for xi in x]
        if TRACE is not None:  # [追加]
            TRACE['hidden'].append(x)
        x = linear(x, state_dict[f'layer{li}.mlp_fc2'])
        if TRACE is not None:  # [追加]
            TRACE['fc2'].append(x)
        x = [a + b for a, b in zip(x, x_residual)]

    logits = linear(x, state_dict['lm_head'])
    if TRACE is not None:  # [追加]
        TRACE['x2'].append(x)
    return logits

# ===== [追加] 表示の本体 =====
def lab(t):
    return 'BOS' if t == BOS else uchars[t]

def show_forward(step, doc, tokens, n, loss):
    P = [f"{i}:{lab(tokens[i])}" for i in range(n)]           # 行ラベル＝位置:入力の文字
    D = [f"d{j}" for j in range(n_embd)]
    print("\n" + "=" * 70)
    print(f"学習ステップ {step+1} の計算を表示します（学習率などは元のまま）")
    print("=" * 70 + "\n")
    print(f"■ 0. 入力：この1ステップで使う名前 = '{doc}'")
    print("  トークン番号 :", tokens)
    print("  入力 → 正解（次の文字）:")
    for i in range(n):
        print(f"    位置{i}: {lab(tokens[i]):>3} → {lab(tokens[i+1])}")
    print("  各行は1つの位置（位置番号:その位置に入力された文字）を表します。\n")
    pause()
    show("1a. トークン埋め込み wte の行（文字ごとのベクトル）", TRACE['tok'], P, D); pause()
    show("1b. 位置埋め込み wpe の行（位置ごとのベクトル）", TRACE['pos'], P, D); pause()
    show("1c. 和を RMSNorm したもの＝ブロックへの入力 x", TRACE['x0'], P, D); pause()
    show("2a. クエリ q = Wq·RMSNorm(x)", TRACE['q'], P, D); pause()
    show("2b. キー   k = Wk·RMSNorm(x)", TRACE['k'], P, D); pause()
    show("2c. バリュー v = Wv·RMSNorm(x)", TRACE['v'], P, D)
    print(f"  ※ 16次元を4次元ずつに分け，d0–d3 がヘッド0，d4–d7 がヘッド1，… に使われます。\n"); pause()
    for h in range(n_head):
        S = [row + [0.0] * (n - len(row)) for row in TRACE['score'][h]]
        W = [row + [0.0] * (n - len(row)) for row in TRACE['weight'][h]]
        fut = lambda i, j: j > i
        print(f"--- ヘッド{h}（次元 d{h*head_dim}–d{h*head_dim+head_dim-1}）---")
        show(f"3a. 注意の点数 q·k/√{head_dim}（行＝クエリの位置，列＝キーの位置）", S, P, P, mask=fut)
        show("3b. 注意の重み softmax(点数)（各行の和＝1）", W, P, P, fmt="{:.3f}", mask=fut)
        print("  ※ '·' はまだ計算されていない未来の位置（キーがまだ存在しない）です。\n")
        pause()
    show("4a. 4ヘッドの出力をつないだもの（重みでバリューを混ぜたもの）", TRACE['attn_cat'], P, D); pause()
    show("4b. Wo を掛けたもの", TRACE['wo'], P, D); pause()
    show("4c. 残差接続：4b ＋ ブロックへの入力 x（1c）＝注意ブロックの出力", TRACE['x1'], P, D); pause()
    H = [f"h{j}" for j in range(4 * n_embd)]
    show("5a. MLP の隠れ層 ReLU(W1·RMSNorm(x))（64次元なので転置：行＝成分，列＝位置）",
         T(TRACE['hidden']), H, P, fmt="{:.3f}", mask=lambda i, j: val(TRACE['hidden'][j][i]) == 0)
    print("  ※ '·' は ReLU で 0 になった成分です。\n"); pause()
    show("5b. W2 を掛けたもの", TRACE['fc2'], P, D); pause()
    show("5c. 残差接続：5b ＋ 4c ＝ MLP ブロックの出力（最終層の出力）", TRACE['x2'], P, D); pause()
    V = [lab(i) for i in range(vocab_size)]
    show("6a. ロジット（次の文字の点数）（転置：行＝文字，列＝位置）", T(TRACE['logits']), V, P); pause()
    show("6b. 確率 softmax(ロジット)（転置：行＝文字，列＝位置，各列の和＝1）", T(TRACE['probs']), V, P, fmt="{:.3f}")
    print("■ 6c. 正解の文字の確率と損失 −ln p")
    for i in range(n):
        tgt = tokens[i + 1]
        print(f"    位置{i}: 正解 {lab(tgt):>3}  p = {val(TRACE['probs'][i][tgt]):.4f}   −ln p = {val(TRACE['loss'][i]):.4f}")
    print(f"  損失（平均） L = {loss.data:.4f}    （参考：でたらめに当てた場合 ln {vocab_size} = {math.log(vocab_size):.4f}）\n")
    pause()

def show_backward(loss):
    D = [f"d{j}" for j in range(n_embd)]
    print("■ 7. 逆伝播 loss.backward() が終わりました。各パラメータの .grad に ∂L/∂w が入っています。\n")
    G = [[p.grad for p in row] for row in state_dict['layer0.attn_wq']]
    show("7a. Wq の勾配 ∂L/∂Wq（行＝q の成分，列＝入力の成分）", G, [f"q{j}" for j in range(n_embd)], D,
         fmt="{:+.4f}", w=8)
    pause()

def show_update(before, lr_t, step):
    D = [f"d{j}" for j in range(n_embd)]
    R = [f"q{j}" for j in range(n_embd)]
    after = [[p.data for p in row] for row in state_dict['layer0.attn_wq']]
    diff = [[a - b for a, b in zip(ra, rb)] for ra, rb in zip(after, before['layer0.attn_wq'])]
    print(f"■ 8. Adam による更新（このステップの学習率 {lr_t:.5f}）\n")
    show("8a. 更新前の Wq", before['layer0.attn_wq'], R, D); pause()
    show("8b. 更新量（更新後 − 更新前）", diff, R, D, fmt="{:+.4f}", w=8)
    if step == 0:
        print("  ※ 最初のステップでは，Adam の補正後の値が m̂/√v̂ = 勾配の符号 になるため，")
        print("    更新量はどれも ±学習率 になります（勾配の大きさによらず，向きだけが効く）。\n")
    pause()
    show("8c. 更新後の Wq", after, R, D)
    print("これで学習の1ステップが終わりです。以降は元のプログラムどおり，表示せずに学習を続けます。\n")
    pause()

# Let there be Adam, the blessed optimizer and its buffers
learning_rate, beta1, beta2, eps_adam = 0.01, 0.85, 0.99, 1e-8
m = [0.0] * len(params) # first moment buffer
v = [0.0] * len(params) # second moment buffer

# Repeat in sequence
num_steps = 1000 # number of training steps
for step in range(num_steps):

    # Take single document, tokenize it, surround it with BOS special token on both sides
    doc = docs[step % len(docs)]
    tokens = [BOS] + [uchars.index(ch) for ch in doc] + [BOS]
    n = min(block_size, len(tokens) - 1)
    if step == TRACE_STEP:  # [追加] 記録の開始
        TRACE = {key: [] for key in ['tok','pos','x0','xn1','q','k','v','attn_cat','wo','x1','hidden','fc2','x2','logits','probs','loss']}
        TRACE['score'] = [[] for _ in range(n_head)]; TRACE['weight'] = [[] for _ in range(n_head)]

    # Forward the token sequence through the model, building up the computation graph all the way to the loss
    keys, values = [[] for _ in range(n_layer)], [[] for _ in range(n_layer)]
    losses = []
    for pos_id in range(n):
        token_id, target_id = tokens[pos_id], tokens[pos_id + 1]
        logits = gpt(token_id, pos_id, keys, values)
        probs = softmax(logits)
        loss_t = -probs[target_id].log()
        if TRACE is not None:  # [追加]
            TRACE['logits'].append(logits); TRACE['probs'].append(probs); TRACE['loss'].append(loss_t)
        losses.append(loss_t)
    loss = (1 / n) * sum(losses) # final average loss over the document sequence. May yours be low.
    if TRACE is not None:  # [追加] 順伝播の途中の値を表示
        show_forward(step, doc, tokens, n, loss)

    # Backward the loss, calculating the gradients with respect to all model parameters
    loss.backward()
    if TRACE is not None:  # [追加] 勾配を表示し，更新前の値を控える
        show_backward(loss)
        before = {name: [[p.data for p in row] for row in state_dict[name]] for name in ('layer0.attn_wq',)}

    # Adam optimizer update: update the model parameters based on the corresponding gradients
    lr_t = learning_rate * (1 - step / num_steps) # linear learning rate decay
    for i, p in enumerate(params):
        m[i] = beta1 * m[i] + (1 - beta1) * p.grad
        v[i] = beta2 * v[i] + (1 - beta2) * p.grad ** 2
        m_hat = m[i] / (1 - beta1 ** (step + 1))
        v_hat = v[i] / (1 - beta2 ** (step + 1))
        p.data -= lr_t * m_hat / (v_hat ** 0.5 + eps_adam)
        p.grad = 0
    if TRACE is not None:  # [追加] 更新の結果を表示して記録を終える
        show_update(before, lr_t, step)
        TRACE = None

    print(f"step {step+1:4d} / {num_steps:4d} | loss {loss.data:.4f}", end='\r')

# Inference: may the model babble back to us
temperature = 0.5 # in (0, 1], control the "creativity" of generated text, low to high
print("\n--- inference (new, hallucinated names) ---")
for sample_idx in range(20):
    keys, values = [[] for _ in range(n_layer)], [[] for _ in range(n_layer)]
    token_id = BOS
    sample = []
    for pos_id in range(block_size):
        logits = gpt(token_id, pos_id, keys, values)
        probs = softmax([l / temperature for l in logits])
        token_id = random.choices(range(vocab_size), weights=[p.data for p in probs])[0]
        if token_id == BOS:
            break
        sample.append(uchars[token_id])
    print(f"sample {sample_idx+1:2d}: {''.join(sample)}")