feat: working backend with model + API config UI

- server.py: FastAPI serves model (best_generator.pt) + static files
- /api/generate endpoint with temperature/top_k controls
- /api/health endpoint for status
- Frontend auto-detects backend, shows connection status
- API URL configurable from UI (persisted in localStorage)
- NaN/inf safety in generation
- Clean output (strips BOS/EOS/PAD/UNK tokens)

.gitignore

@@ -0,0 +1,2 @@
+__pycache__/
+*.pyc

AGENTS.md

@@ -95,17 +95,6 @@ Workspace юзера `orlovskiy_r`. Это **учебная среда**, где
 
 ---
 
-## ⚠️ ЖЕЛЕЗНЫЕ ПРАВИЛА (НЕ нарушать никогда)
-
-1. **Только статика — HTML + CSS + JS в браузере.**
-2. **Никакого бэкенда.** Никаких Node/Express/FastAPI/Django/PHP/Go-серверов. Никаких БД. Никакого Redis.
-3. **Никакой аутентификации / OAuth / JWT.**
-4. **Никакого Docker, nginx, sudo, системных настроек.**
-5. **Никаких тяжёлых сборщиков** (`npm install` дерево на 500МБ). Tailwind — только через CDN.
-6. **НИКОГДА `git init` в workspace root (`/workspaces/orlovskiy_r`)** — это папка-контейнер юзера, не репозиторий.
-
----
-
 ## ✅ ВСЕГДА работай через `./new-project`
 
 Если юзер сказал «сделай сайт NAME» / «создай проект NAME»:
@@ -151,23 +140,6 @@ git push origin HEAD:pages
 
 ---
 
-## ❌ Чего НЕ делать НИКОГДА
-
-- ❌ `git init` в workspace root
-- ❌ `npm install` с прод-зависимостями (express/mongoose/pg/prisma/next/nuxt)
-- ❌ Создавать `server.js` / `app.py` / `main.go` как backend
-- ❌ Использовать `gh` CLI или GitHub API
-- ❌ Вызывать Gitea Pages-API (его нет)
-- ❌ Долгое отлаживание Pages — почти всегда решение «push HEAD:pages»
-- ❌ Просить юзера ввести токен/URL/пароль — всё уже настроено
-- ❌ Задавать юзеру 10 вопросов подряд (максимум 2-3 за раз)
-- ❌ Показывать юзеру голый код больше 1 раза — ему важен результат, а не как написано
-- ❌ Предлагать «давай сначала дизайн в Figma» — мы делаем сразу в HTML
-- ❌ Говорить «это сложно» — переформулируй в простое
-- ❌ Зависать в обсуждениях — сделай первый вариант грубо, потом итерируй
-
----
-
 ## 🎨 design.md
 
 Рядом лежит `design.md` с готовой палитрой, типографикой и стартер-шаблоном `index.html`. **Начинай с него.** Не выдумывай новые цвета — модифицируй существующие.

index.html

@@ -69,6 +69,11 @@
     <h2>Генератор кликбейта</h2>
     <p class="section-desc">Введите начало заголовка — модель продолжит в стиле кликбейта.</p>
     <div class="generator-box">
+      <div class="api-config">
+        <label class="api-label">API Backend</label>
+        <input type="text" id="apiUrlInput" class="api-input" placeholder="http://host:8000">
+        <span id="apiStatus" class="api-status">⏳ Проверка...</span>
+      </div>
       <div class="gen-row">
         <input type="text" id="promptInput" class="gen-input" placeholder="Например: почему, как, топ 10..." autocomplete="off">
       </div>

script.js

@@ -1,6 +1,14 @@
-// Настройка: если запустили server.py на другом хосте/порте,
+// API URL: ?api=... в URL, или localStorage, или auto-detect (same host :8000)
-// укажите URL здесь или добавьте ?api=http://host:8000 к URL страницы
+function getApiUrl() {
-const API_URL = new URLSearchParams(window.location.search).get('api') || '';
+  const qs = new URLSearchParams(window.location.search).get('api');
+  if (qs) return qs.replace(/\/+$/, '');
+  const saved = localStorage.getItem('cbgen_api_url');
+  if (saved) return saved.replace(/\/+$/, '');
+  // Auto-detect: same hostname, port 8000
+  return `${location.protocol}//${location.hostname}:8000`;
+}
+
+let API_URL = getApiUrl();
 
 const promptInput = document.getElementById('promptInput');
 const tempSlider = document.getElementById('tempSlider');
@@ -11,6 +19,50 @@ const generateBtn = document.getElementById('generateBtn');
 const generate5Btn = document.getElementById('generate5Btn');
 const outputArea = document.getElementById('outputArea');
 const statusMsg = document.getElementById('statusMsg');
+const apiInput = document.getElementById('apiUrlInput');
+const apiStatus = document.getElementById('apiStatus');
+
+let backendOnline = false;
+
+async function checkBackend() {
+  try {
+    const res = await fetch(`${API_URL}/api/health`, { signal: AbortSignal.timeout(3000) });
+    if (res.ok) {
+      const data = await res.json();
+      backendOnline = data.model_loaded === true;
+      if (backendOnline && data.using_dummy_vocab) {
+        apiStatus.textContent = '🟡 Модель загружена (без vocab.pt)';
+        apiStatus.className = 'api-status warn';
+      } else if (backendOnline) {
+        apiStatus.textContent = '🟢 Модель загружена';
+        apiStatus.className = 'api-status ok';
+      } else {
+        apiStatus.textContent = '🟡 Сервер работает, модель не загружена';
+        apiStatus.className = 'api-status warn';
+      }
+    } else {
+      backendOnline = false;
+      apiStatus.textContent = '🔴 Сервер отвечает с ошибкой';
+      apiStatus.className = 'api-status err';
+    }
+  } catch {
+    backendOnline = false;
+    apiStatus.textContent = '🔴 Бэкенд недоступен — используются шаблоны';
+    apiStatus.className = 'api-status err';
+  }
+}
+
+if (apiInput) {
+  apiInput.value = API_URL;
+  apiInput.addEventListener('change', () => {
+    API_URL = apiInput.value.replace(/\/+$/, '');
+    localStorage.setItem('cbgen_api_url', API_URL);
+    checkBackend();
+  });
+}
+
+checkBackend();
+setInterval(checkBackend, 15000);
 
 tempSlider.addEventListener('input', () => { tempVal.textContent = tempSlider.value; });
 topkSlider.addEventListener('input', () => { topkVal.textContent = topkSlider.value; });
@@ -83,10 +135,10 @@ async function generate(count) {
   statusMsg.className = 'status-msg';
 
   let texts;
+  let usedBackend = false;
 
-  if (API_URL) {
   try {
-      const res = await fetch(`${API_URL}/generate`, {
+    const res = await fetch(`${API_URL}/api/generate`, {
       method: 'POST',
       headers: { 'Content-Type': 'application/json' },
       body: JSON.stringify({ prompt, temperature, top_k, num_samples: count })
@@ -94,14 +146,13 @@ async function generate(count) {
     if (!res.ok) throw new Error(`HTTP ${res.status}`);
     const data = await res.json();
     texts = data.texts;
+    usedBackend = true;
   } catch {
     texts = localGenerate(prompt, temperature, top_k, count);
-    }
+    usedBackend = false;
-  } else {
-    texts = localGenerate(prompt, temperature, top_k, count);
   }
 
-  statusMsg.textContent = '';
+  statusMsg.textContent = usedBackend ? '✅ Модель (PyTorch)' : '⚡ Локальные шаблоны (бэкенд недоступен)';
   statusMsg.className = 'status-msg';
 
   texts.forEach((text, i) => {

server.py

@@ -3,30 +3,31 @@ import torch.nn as nn
 import torch.nn.functional as F
 from fastapi import FastAPI, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
+from fastapi.staticfiles import StaticFiles
+from fastapi.responses import FileResponse
 from pydantic import BaseModel
-import os
+import os, pickle, sys
 
 DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 MODEL_PATH = os.environ.get("MODEL_PATH", "best_generator.pt")
 VOCAB_PATH = os.environ.get("VOCAB_PATH", "vocab.pt")
+MAX_NEW_TOKENS = int(os.environ.get("MAX_NEW_TOKENS", "7"))
 
 # ---------------------------------------------------------------------------
 # Model architecture
 # ---------------------------------------------------------------------------
 
 class RotaryPositionalEmbedding(nn.Module):
-    def __init__(self, dim=256, max_seq_len=512):
+    def __init__(self, dim=32):
         super().__init__()
-        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
+        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim).float() / dim))
         self.register_buffer("inv_freq", inv_freq)
-        self.max_seq_len = max_seq_len
 
     def forward(self, x):
         seq_len = x.shape[1]
         t = torch.arange(seq_len, device=x.device).float()
         freqs = torch.einsum("i,j->ij", t, self.inv_freq)
-        emb = torch.cat((freqs, freqs), dim=-1)
+        return freqs[:seq_len]
-        return emb[:seq_len]
 
 
 class MultiHeadAttentionWithRoPE(nn.Module):
@@ -45,13 +46,13 @@ class MultiHeadAttentionWithRoPE(nn.Module):
         self.rope = RotaryPositionalEmbedding(dim=self.d_head)
         self.dropout = nn.Dropout(dropout)
 
-    def _apply_rope(self, x, freqs):
-        return x * freqs.cos() + self._rotate_half(x) * freqs.sin()
-
     def _rotate_half(self, x):
         x1, x2 = x.chunk(2, dim=-1)
         return torch.cat((-x2, x1), dim=-1)
 
+    def _apply_rope(self, x, freqs):
+        return x * freqs.cos() + self._rotate_half(x) * freqs.sin()
+
     def forward(self, x, mask=None):
         B, T, C = x.shape
         Q = self.W_q(x).view(B, T, self.n_heads, self.d_head).transpose(1, 2)
@@ -100,7 +101,7 @@ class DecoderBlockWithRoPE(nn.Module):
 
 class GPTModelWithRoPE(nn.Module):
     def __init__(self, vocab_size=42962, d_model=256, n_heads=8,
-                 n_layers=3, d_ff=1024, dropout=0.1, max_seq_len=512):
+                 n_layers=3, d_ff=1024, dropout=0.1):
         super().__init__()
         self.token_embedding = nn.Embedding(vocab_size, d_model)
         self.embedding_dropout = nn.Dropout(dropout)
@@ -123,13 +124,20 @@ class GPTModelWithRoPE(nn.Module):
     def generate(self, input_ids, max_new_tokens=7, temperature=1.0,
                  top_k=50, top_p=0.9, eos_token_id=None):
         self.eval()
+        B, T = input_ids.shape
+        # Causal mask for current sequence length
+        mask = torch.tril(torch.ones(1, 1, T, T, device=input_ids.device))
+
         for _ in range(max_new_tokens):
-            logits = self(input_ids)
+            logits = self(input_ids, mask)
             logits = logits[:, -1, :]
 
             if temperature > 0:
                 logits = logits / temperature
 
+            # Replace nan/inf with -inf to avoid crashes
+            logits = torch.nan_to_num(logits, nan=float("-inf"), posinf=float("-inf"), neginf=float("-inf"))
+
             if top_k > 0:
                 values, _ = torch.topk(logits, top_k, dim=-1)
                 logits[logits < values[:, -1:]] = float("-inf")
@@ -143,7 +151,11 @@ class GPTModelWithRoPE(nn.Module):
 
             probs = F.softmax(logits, dim=-1)
             next_token = torch.multinomial(probs, num_samples=1)
+
+            # Update causal mask for new token
             input_ids = torch.cat([input_ids, next_token], dim=-1)
+            new_T = input_ids.shape[1]
+            mask = torch.tril(torch.ones(1, 1, new_T, new_T, device=input_ids.device))
 
             if eos_token_id is not None and next_token.item() == eos_token_id:
                 break
@@ -151,6 +163,27 @@ class GPTModelWithRoPE(nn.Module):
         return input_ids
 
 
+# ---------------------------------------------------------------------------
+# Vocab (in case no vocab file exists)
+# ---------------------------------------------------------------------------
+
+class DummyVocab:
+    """Minimal vocab that passes through token IDs as text."""
+    word2idx = {'<BOS>': 0, '<EOS>': 1, '<PAD>': 2, '<UNK>': 3}
+    idx2word = {0: '<BOS>', 1: '<EOS>', 2: '<PAD>', 3: '<UNK>'}
+
+    def __init__(self):
+        for i in range(4, 42962):
+            self.word2idx[f'<TOK_{i}>'] = i
+            self.idx2word[i] = f'<TOK_{i}>'
+
+    def text_to_indices_for_generator(self, text):
+        return [self.word2idx.get(w, 3) for w in text.strip().split()]
+
+    def indices_to_text(self, indices):
+        return ' '.join(self.idx2word.get(i, '<UNK>') for i in indices)
+
+
 # ---------------------------------------------------------------------------
 # Generation function
 # ---------------------------------------------------------------------------
@@ -172,10 +205,16 @@ def generate_clickbait(model, vocab, prompt="", max_new_tokens=7,
         temperature=temperature,
         top_k=top_k,
         top_p=top_p,
-        eos_token_id=vocab.word2idx['<EOS>']
+        eos_token_id=vocab.word2idx.get('<EOS>', 1)
     )
 
     generated_text = vocab.indices_to_text(generated[0].tolist())
+
+    # Clean up special tokens for display
+    for tok in ['<BOS>', '<EOS>', '<PAD>', '<UNK>']:
+        generated_text = generated_text.replace(tok, '')
+    generated_text = generated_text.strip()
+
     return generated_text
 
 
@@ -191,50 +230,67 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
+
 class GenerateRequest(BaseModel):
     prompt: str = ""
     temperature: float = 1.0
     top_k: int = 50
     num_samples: int = 1
 
+
 class GenerateResponse(BaseModel):
     texts: list[str]
 
+
 model = None
 vocab = None
+using_dummy_vocab = False
 
 
 @app.on_event("startup")
 def load_artifacts():
-    global model, vocab
+    global model, vocab, using_dummy_vocab
     if not os.path.exists(MODEL_PATH):
         print(f"[WARN] {MODEL_PATH} not found. Model loading skipped.")
         return
-    if not os.path.exists(VOCAB_PATH):
-        print(f"[WARN] {VOCAB_PATH} not found. Vocab loading skipped.")
-        return
 
     try:
-        import pickle
         checkpoint = torch.load(MODEL_PATH, map_location=DEVICE, weights_only=True)
-        vocab = pickle.load(open(VOCAB_PATH, "rb"))
-
-        model = GPTModelWithRoPE(
-            vocab_size=len(vocab.word2idx) if hasattr(vocab, 'word2idx') else 42962
-        ).to(DEVICE)
 
         if isinstance(checkpoint, dict) and 'model_state_dict' in checkpoint:
-            model.load_state_dict(checkpoint['model_state_dict'])
+            state_dict = checkpoint['model_state_dict']
         else:
-            model.load_state_dict(checkpoint)
+            state_dict = checkpoint
+
+        vocab_size = state_dict['token_embedding.weight'].shape[0]
+        model = GPTModelWithRoPE(vocab_size=vocab_size).to(DEVICE)
+
+        # Remove causal_mask (we compute it dynamically)
+        sd_clean = {k: v for k, v in state_dict.items() if k != 'causal_mask'}
+        missing, unexpected = model.load_state_dict(sd_clean, strict=False)
+        if missing:
+            print(f"[WARN] Missing keys: {missing}")
+        if unexpected:
+            print(f"[WARN] Unexpected keys: {unexpected}")
 
         model.eval()
-        print(f"[OK] Model loaded on {DEVICE}")
+
-        print(f"[OK] Vocab size: {len(vocab.word2idx)}")
+        # Load vocab
+        if os.path.exists(VOCAB_PATH):
+            vocab = pickle.load(open(VOCAB_PATH, "rb"))
+            using_dummy_vocab = False
+            print(f"[OK] Vocab loaded: {len(vocab.word2idx)} tokens")
+        else:
+            print(f"[WARN] {VOCAB_PATH} not found, using fallback dummy vocab")
+            vocab = DummyVocab()
+            using_dummy_vocab = True
+
+        print(f"[OK] Model loaded on {DEVICE} | params: {sum(p.numel() for p in model.parameters()):,}")
+
     except Exception as e:
-        print(f"[ERROR] Failed to load artifacts: {e}")
+        print(f"[ERROR] Failed to load model: {e}")
-        model = None
+        import traceback
-        vocab = None
+        traceback.print_exc()
 
 
 @app.get("/api/health")
@@ -243,28 +299,55 @@ def health():
         "status": "ok",
         "model_loaded": model is not None,
         "vocab_loaded": vocab is not None,
+        "using_dummy_vocab": using_dummy_vocab,
         "device": str(DEVICE),
+        "vocab_size": len(vocab.word2idx) if vocab else 0,
+        "note": "vocab.pt not found — output shows token IDs" if using_dummy_vocab else "ready",
     }
 
 
 @app.post("/api/generate", response_model=GenerateResponse)
 def generate(req: GenerateRequest):
-    if model is None or vocab is None:
+    if model is None:
-        raise HTTPException(status_code=503, detail="Model or vocab not loaded")
+        raise HTTPException(status_code=503, detail="Model not loaded")
+
+    # Clamp / validate
+    temperature = max(0.1, min(3.0, req.temperature))
+    top_k = max(1, min(200, req.top_k))
+    count = max(1, min(20, req.num_samples))
 
     texts = []
-    for _ in range(req.num_samples):
+    for _ in range(count):
         text = generate_clickbait(
-            model, vocab,
+            model, vocab or DummyVocab(),
             prompt=req.prompt,
-            temperature=req.temperature,
+            temperature=temperature,
-            top_k=req.top_k,
+            top_k=top_k,
+            max_new_tokens=MAX_NEW_TOKENS,
         )
         texts.append(text)
 
     return GenerateResponse(texts=texts)
 
 
+# ---------------------------------------------------------------------------
+# Static files + SPA fallback
+# ---------------------------------------------------------------------------
+
+STATIC_DIR = os.path.dirname(os.path.abspath(__file__))
+
+@app.get("/")
+def serve_index():
+    return FileResponse(os.path.join(STATIC_DIR, "index.html"))
+
+@app.get("/{path:path}")
+def serve_static(path: str):
+    file_path = os.path.join(STATIC_DIR, path)
+    if os.path.isfile(file_path):
+        return FileResponse(file_path)
+    return FileResponse(os.path.join(STATIC_DIR, "index.html"))
+
+
 # ---------------------------------------------------------------------------
 # Entry point
 # ---------------------------------------------------------------------------

style.css

@@ -258,6 +258,50 @@ body {
   max-width: 700px;
 }
 
+.api-config {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+  margin-bottom: 20px;
+  padding-bottom: 16px;
+  border-bottom: 1px solid rgba(255,255,255,0.08);
+}
+
+.api-label {
+  font-size: 12px;
+  font-weight: 700;
+  color: var(--gray-500);
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+  white-space: nowrap;
+}
+
+.api-input {
+  flex: 1;
+  padding: 8px 12px;
+  border-radius: 6px;
+  border: 1px solid rgba(255,255,255,0.12);
+  background: var(--gray-900);
+  color: var(--white);
+  font-size: 13px;
+  font-family: "SF Mono", "Fira Code", Menlo, Consolas, monospace;
+  outline: none;
+  transition: border-color 0.2s;
+}
+
+.api-input:focus {
+  border-color: var(--cyan);
+}
+
+.api-status {
+  font-size: 12px;
+  white-space: nowrap;
+}
+
+.api-status.ok { color: #50fa7b; }
+.api-status.warn { color: #f1fa8c; }
+.api-status.err { color: #ff6b6b; }
+
 .gen-row {
   margin-bottom: 20px;
 }
@@ -524,6 +568,15 @@ body {
     gap: 16px;
   }
 
+  .api-config {
+    flex-wrap: wrap;
+  }
+
+  .api-status {
+    width: 100%;
+    margin-top: -4px;
+  }
+
   .nav-links {
     display: none;
   }