Create app.py

app.py

@@ -0,0 +1,653 @@
+# ==============================================================================
+# 🦅 GarudaCoder-27B — Indonesian Coding Intelligence
+# Precision Engineered Interface (Hugging Face ZeroGPU Edition)
+# ==============================================================================
+
+import os
+os.environ["HF_HUB_DISABLE_XET"] = "1"
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+os.environ["FORCE_QWENVL_VIDEO_READER"] = "torchvision"
+
+import gc
+import re
+import io
+import time
+import spaces
+import threading
+import traceback
+from pathlib import Path
+
+import torch
+import gradio as gr
+from transformers import (
+    AutoConfig,
+    AutoProcessor,
+    BitsAndBytesConfig,
+    TextIteratorStreamer,
+)
+from peft import PeftModel
+
+try:
+    from transformers import AutoModelForMultimodalLM
+except ImportError:
+    try:
+        from transformers import AutoModelForImageTextToText as AutoModelForMultimodalLM
+    except ImportError:
+        from transformers import AutoModelForVision2Seq as AutoModelForMultimodalLM
+
+# ==============================================================================
+# 1. KONFIGURASI MODEL & PROMPT
+# ==============================================================================
+BASE_ID = "unsloth/Qwen3.8-27B-unsloth-bnb-4bit"
+ADAPTER_ID = "TheCoderScientist/GarudaCoder-27B-ID-lora"
+EOS_IDS = [248046, 248044]
+
+SYSTEM_PROMPT = """Kamu adalah GarudaCoder, asisten coding berbahasa Indonesia yang teliti.
+
+Prioritas:
+- Utamakan kebenaran teknis dan kode yang dapat diuji.
+- Berikan ringkasan alasan yang dapat diaudit sebelum solusi.
+- Nyatakan asumsi dan edge case yang relevan.
+- Jangan mengarang fakta, API, versi, angka, dokumentasi, atau hasil eksekusi.
+- Bila informasi tidak cukup, katakan apa yang kurang.
+- Gunakan konteks gambar, video, file, atau hasil web search yang diberikan.
+- Bedakan fakta dari inferensi.
+- Untuk informasi yang berubah atau bersifat lokal, gunakan sumber yang dapat diverifikasi.
+- Jangan menampilkan chain-of-thought privat.
+- Jangan mengklaim telah menjalankan kode jika memang belum dijalankan.
+- Identitas kamu adalah GarudaCoder.
+- Jika ditanya basis model, jelaskan secara ringkas bahwa GarudaCoder menggunakan Qwen3.8-27B sebagai base model dengan adapter GarudaCoder.
+"""
+
+print("[Init] Menyiapkan Processor & Tokenizer...")
+processor = AutoProcessor.from_pretrained(ADAPTER_ID, trust_remote_code=True)
+tokenizer = getattr(processor, "tokenizer", processor)
+
+if getattr(tokenizer, "pad_token", None) is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+_model = None
+
+def get_model():
+    """ZeroGPU Lazy Loader: Muat model hanya saat inferensi diminta."""
+    global _model
+    if _model is None:
+        print("[ZeroGPU] Memuat bobot 4-bit ke NVIDIA A100...")
+        bnb_config = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_quant_type="nf4",
+            bnb_4bit_compute_dtype=torch.float16,
+        )
+        base_model = AutoModelForMultimodalLM.from_pretrained(
+            BASE_ID,
+            quantization_config=bnb_config,
+            device_map="auto",
+            attn_implementation="sdpa",
+            trust_remote_code=True,
+        )
+        print("[ZeroGPU] Mengaktifkan adapter LoRA GarudaCoder...")
+        _model = PeftModel.from_pretrained(
+            base_model,
+            ADAPTER_ID,
+            is_trainable=False,
+        )
+        _model.eval()
+        print("[ZeroGPU] Alokasi VRAM selesai.")
+    return _model
+
+# ==============================================================================
+# 2. FILE & MULTIMODAL PARSERS
+# ==============================================================================
+IMAGE_EXTENSIONS = {".png", ".jpg", ".jpeg", ".webp", ".bmp"}
+VIDEO_EXTENSIONS = {".mp4", ".mov", ".mkv", ".webm"}
+TEXT_EXTENSIONS = {
+    ".txt", ".md", ".py", ".js", ".ts", ".tsx", ".jsx", ".java", ".c", ".cpp",
+    ".h", ".hpp", ".cs", ".go", ".rs", ".php", ".rb", ".swift", ".kt", ".sh",
+    ".sql", ".json", ".yaml", ".yml", ".toml", ".csv", ".log", ".html", ".css"
+}
+
+def normalize_path(file_obj):
+    if file_obj is None:
+        return None
+    if isinstance(file_obj, str):
+        return file_obj
+    if isinstance(file_obj, dict):
+        return file_obj.get("path") or file_obj.get("name")
+    return getattr(file_obj, "path", None) or getattr(file_obj, "name", None)
+
+def file_uri(path):
+    return Path(os.path.abspath(path)).as_uri()
+
+def extract_text_file(path, max_chars=40000):
+    try:
+        with open(path, "r", encoding="utf-8", errors="ignore") as f:
+            return f.read(max_chars)
+    except Exception as e:
+        return f"[Gagal membaca file teks: {e}]"
+
+def extract_pdf(path, max_chars=40000):
+    try:
+        from pypdf import PdfReader
+        reader = PdfReader(path)
+        chunks = []
+        for page in reader.pages:
+            txt = page.extract_text() or ""
+            chunks.append(txt)
+            if sum(len(x) for x in chunks) >= max_chars:
+                break
+        return "\n\n".join(chunks)[:max_chars]
+    except Exception as e:
+        return f"[Gagal membaca PDF: {e}]"
+
+def extract_docx(path, max_chars=40000):
+    try:
+        from docx import Document
+        doc = Document(path)
+        parts = [p.text for p in doc.paragraphs if p.text.strip()]
+        return "\n".join(parts)[:max_chars]
+    except Exception as e:
+        return f"[Gagal membaca DOCX: {e}]"
+
+def extract_xlsx(path, max_chars=40000):
+    try:
+        from openpyxl import load_workbook
+        wb = load_workbook(path, read_only=True, data_only=True)
+        parts = []
+        for ws in wb.worksheets:
+            parts.append(f"## Sheet: {ws.title}")
+            for row in ws.iter_rows(values_only=True):
+                vals = ["" if v is None else str(v) for v in row]
+                parts.append(" | ".join(vals))
+                if sum(len(x) for x in parts) >= max_chars:
+                    break
+        return "\n".join(parts)[:max_chars]
+    except Exception as e:
+        return f"[Gagal membaca XLSX: {e}]"
+
+def inspect_uploaded_files(files):
+    vision_blocks = []
+    text_parts = []
+    summary = []
+
+    for raw in files or []:
+        path = normalize_path(raw)
+        if not path or not os.path.exists(path):
+            continue
+
+        name = os.path.basename(path)
+        ext = Path(path).suffix.lower()
+
+        if ext in IMAGE_EXTENSIONS:
+            vision_blocks.append({"type": "image", "image": file_uri(path)})
+            summary.append(name)
+        elif ext in VIDEO_EXTENSIONS:
+            vision_blocks.append({"type": "video", "video": file_uri(path), "fps": 1.0})
+            summary.append(name)
+        elif ext in TEXT_EXTENSIONS:
+            text_parts.append(f"### [FILE: {name}]\n```text\n{extract_text_file(path)}\n```")
+            summary.append(name)
+        elif ext == ".pdf":
+            text_parts.append(f"### [PDF: {name}]\n{extract_pdf(path)}")
+            summary.append(name)
+        elif ext == ".docx":
+            text_parts.append(f"### [DOCX: {name}]\n{extract_docx(path)}")
+            summary.append(name)
+        elif ext in {".xlsx", ".xlsm"}:
+            text_parts.append(f"### [SPREADSHEET: {name}]\n{extract_xlsx(path)}")
+            summary.append(name)
+
+    return vision_blocks, "\n\n".join(text_parts), summary
+
+# ==============================================================================
+# 3. FAST WEB SEARCH (DUCKDUCKGO)
+# ==============================================================================
+def web_search(query, max_results=4):
+    query = (query or "").strip()
+    if not query:
+        return []
+    try:
+        from duckduckgo_search import DDGS
+        results = []
+        with DDGS() as ddgs:
+            for r in ddgs.text(query, max_results=max_results, region="wt-wt", safesearch="moderate"):
+                results.append({
+                    "title": r.get("title") or "Dokumentasi",
+                    "url": r.get("href") or "",
+                    "snippet": r.get("body") or "",
+                })
+        return results
+    except Exception as e:
+        print(f"[Web Search Warn] {e}")
+        return []
+
+def format_search_context(results):
+    if not results:
+        return ""
+    chunks = ["## HASIL DOKUMENTASI/WEB TERBARU:"]
+    for i, r in enumerate(results, 1):
+        chunks.append(f"[{i}] {r['title']} ({r['url']})\nSnippet: {r['snippet']}")
+    return "\n\n".join(chunks)
+
+def sources_html(results):
+    if not results:
+        return '<div class="sources-empty">Pencarian web nonaktif atau belum ada kueri.</div>'
+    chips = []
+    for i, r in enumerate(results, 1):
+        title = gr.utils.sanitize_html(r["title"][:75])
+        url = r["url"]
+        domain = re.sub(r"^https?://(www\.)?", "", url).split("/")[0]
+        chips.append(
+            f'<a class="source-chip" href="{url}" target="_blank" title="{title}">'
+            f'<span class="source-domain">{domain}</span>'
+            f'<span class="source-title">{title}</span>'
+            f'</a>'
+        )
+    return f'<div class="sources-grid">{"".join(chips)}</div>'
+
+# ==============================================================================
+# 4. CHAT GENERATION ENGINE
+# ==============================================================================
+@spaces.GPU(duration=120)
+def generate_chat(
+    message,
+    history,
+    system_prompt,
+    temperature,
+    top_p,
+    max_new_tokens,
+    enable_thinking,
+    enable_web_search,
+):
+    active_model = get_model()
+
+    user_text = (message or {}).get("text", "").strip()
+    files = (message or {}).get("files", [])
+
+    vision_blocks, file_context, file_names = inspect_uploaded_files(files)
+
+    search_results = []
+    if enable_web_search and user_text:
+        search_results = web_search(user_text, max_results=4)
+    search_context = format_search_context(search_results)
+
+    sys_content = system_prompt.strip() if system_prompt else SYSTEM_PROMPT
+    messages = [{"role": "system", "content": sys_content}]
+
+    for turn in history or []:
+        role = turn.get("role")
+        content = turn.get("content")
+        if role in {"user", "assistant"} and isinstance(content, str):
+            messages.append({"role": role, "content": content})
+
+    current_turn_content = []
+    current_turn_content.extend(vision_blocks)
+
+    context_sections = []
+    if file_context:
+        context_sections.append(file_context)
+    if search_context:
+        context_sections.append(search_context)
+
+    if context_sections:
+        full_text = "\n\n".join(context_sections) + (f"\n\n## INSTRUKSI USER:\n{user_text}" if user_text else "")
+        current_turn_content.append({"type": "text", "text": full_text})
+    elif user_text:
+        current_turn_content.append({"type": "text", "text": user_text})
+    elif not vision_blocks:
+        current_turn_content.append({"type": "text", "text": "Halo!"})
+
+    messages.append({"role": "user", "content": current_turn_content})
+
+    try:
+        model_inputs = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+            enable_thinking=bool(enable_thinking),
+        )
+    except TypeError:
+        model_inputs = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+        )
+
+    model_inputs = {k: v.to("cuda") for k, v in model_inputs.items() if torch.is_tensor(v)}
+
+    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+    gen_kwargs = dict(
+        **model_inputs,
+        streamer=streamer,
+        max_new_tokens=int(max_new_tokens),
+        eos_token_id=EOS_IDS,
+        pad_token_id=tokenizer.pad_token_id,
+        repetition_penalty=1.05,
+    )
+
+    if float(temperature) > 0.05:
+        gen_kwargs.update(do_sample=True, temperature=float(temperature), top_p=float(top_p), top_k=20)
+    else:
+        gen_kwargs.update(do_sample=False)
+
+    errors = []
+    def _generate():
+        try:
+            with torch.inference_mode():
+                active_model.generate(**gen_kwargs)
+        except Exception as e:
+            errors.append(e)
+            traceback.print_exc()
+        finally:
+            streamer.end()
+
+    thread = threading.Thread(target=_generate, daemon=True)
+    thread.start()
+
+    badges = []
+    if file_names:
+        badges.append(f"Attachment: {', '.join(file_names[:2])}{'...' if len(file_names) > 2 else ''}")
+    if search_results:
+        badges.append(f"Web Search: {len(search_results)} rujukan")
+
+    meta_tag = f'<div class="context-tag">{" · ".join(badges)}</div>\n\n' if badges else ""
+    output = ""
+
+    for chunk in streamer:
+        output += chunk
+        yield meta_tag + output, search_results
+
+    thread.join()
+
+    if errors:
+        yield output + f"\n\n> ⚠️ **Runtime Issue:** `{errors[0]}`", search_results
+
+# ==============================================================================
+# 5. MODERN UI SYSTEM (LINEAR / CURSOR AESTHETICS)
+# ==============================================================================
+WELCOME_MD = """### Siap membantu audit kode dan arsitektur sistem.
+
+Pilih topik di bawah atau lampirkan file dokumen/gambar langsung:
+"""
+
+CSS = r"""
+@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500&display=swap');
+
+:root {
+    --bg-base: #0a0c10;
+    --bg-surface: #11141b;
+    --bg-subtle: #181d26;
+    --border-subtle: rgba(255, 255, 255, 0.07);
+    --border-active: rgba(255, 255, 255, 0.15);
+    --text-primary: #f0f3f6;
+    --text-secondary: #8c96a5;
+    --text-muted: #566171;
+    --accent: #10b981;
+    --font-sans: 'Inter', -apple-system, BlinkMacSystemFont, sans-serif;
+    --font-mono: 'JetBrains Mono', monospace;
+}
+
+body, .gradio-container {
+    background-color: var(--bg-base) !important;
+    color: var(--text-primary) !important;
+    font-family: var(--font-sans) !important;
+    max-width: 1380px !important;
+    margin: 0 auto !important;
+}
+
+/* Header Navbar */
+.app-nav {
+    display: flex;
+    justify-content: space-between;
+    align-items: center;
+    padding: 18px 0;
+    border-bottom: 1px solid var(--border-subtle);
+    margin-bottom: 16px;
+}
+.brand-group {
+    display: flex;
+    align-items: center;
+    gap: 12px;
+}
+.brand-badge {
+    background: var(--bg-subtle);
+    border: 1px solid var(--border-subtle);
+    border-radius: 8px;
+    padding: 6px 9px;
+    font-size: 14px;
+    font-weight: 700;
+}
+.brand-name {
+    font-size: 16px;
+    font-weight: 600;
+    letter-spacing: -0.02em;
+    color: var(--text-primary);
+}
+.brand-desc {
+    font-size: 12px;
+    color: var(--text-secondary);
+}
+.status-pill {
+    display: inline-flex;
+    align-items: center;
+    gap: 6px;
+    font-size: 11px;
+    font-family: var(--font-mono);
+    color: var(--text-secondary);
+    background: var(--bg-surface);
+    border: 1px solid var(--border-subtle);
+    padding: 4px 10px;
+    border-radius: 9999px;
+}
+.status-dot {
+    width: 6px;
+    height: 6px;
+    border-radius: 50%;
+    background-color: var(--accent);
+}
+
+/* Chatbox Area */
+.chat-viewport {
+    background: var(--bg-surface) !important;
+    border: 1px solid var(--border-subtle) !important;
+    border-radius: 12px !important;
+}
+
+/* Source Reference Chips (Perplexity Style) */
+.sources-grid {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 8px;
+    padding: 4px 0;
+}
+.source-chip {
+    display: inline-flex;
+    flex-direction: column;
+    background: var(--bg-subtle);
+    border: 1px solid var(--border-subtle);
+    border-radius: 8px;
+    padding: 6px 10px;
+    text-decoration: none !important;
+    transition: border-color 0.15s ease, background 0.15s ease;
+    max-width: 100%;
+}
+.source-chip:hover {
+    border-color: var(--border-active);
+    background: #1e2430;
+}
+.source-domain {
+    font-size: 10px;
+    font-family: var(--font-mono);
+    color: var(--accent);
+    text-transform: uppercase;
+    letter-spacing: 0.04em;
+}
+.source-title {
+    font-size: 11px;
+    font-weight: 500;
+    color: var(--text-primary) !important;
+    white-space: nowrap;
+    overflow: hidden;
+    text-overflow: ellipsis;
+    max-width: 240px;
+}
+.sources-empty {
+    font-size: 12px;
+    color: var(--text-muted);
+    font-style: italic;
+}
+
+/* Context Badge in Response */
+.context-tag {
+    display: inline-block;
+    font-family: var(--font-mono);
+    font-size: 10px;
+    color: var(--text-secondary);
+    background: var(--bg-subtle);
+    border: 1px solid var(--border-subtle);
+    border-radius: 4px;
+    padding: 3px 6px;
+    margin-bottom: 6px;
+}
+
+/* Prompt Suggestion Chips */
+.quick-prompt-row {
+    display: flex;
+    gap: 8px;
+    margin-bottom: 12px;
+    flex-wrap: wrap;
+}
+.quick-prompt-btn {
+    background: var(--bg-surface) !important;
+    border: 1px solid var(--border-subtle) !important;
+    color: var(--text-secondary) !important;
+    font-size: 11px !important;
+    border-radius: 20px !important;
+    padding: 4px 12px !important;
+    cursor: pointer !important;
+    font-weight: 400 !important;
+    transition: all 0.15s ease !important;
+}
+.quick-prompt-btn:hover {
+    color: var(--text-primary) !important;
+    border-color: var(--border-active) !important;
+    background: var(--bg-subtle) !important;
+}
+
+/* Sidebar Box Minimalist */
+.sidebar-panel {
+    background: var(--bg-surface) !important;
+    border: 1px solid var(--border-subtle) !important;
+    border-radius: 12px !important;
+    padding: 16px !important;
+}
+"""
+
+with gr.Blocks(title="GarudaCoder-27B", css=CSS, theme=gr.themes.Base()) as demo:
+    # ------------------ TOP NAV BAR ------------------
+    gr.HTML("""
+        <div class="app-nav">
+            <div class="brand-group">
+                <div class="brand-badge">🦅 GC-27B</div>
+                <div>
+                    <div class="brand-name">GarudaCoder Workspace</div>
+                    <div class="brand-desc">Indonesian Specialized Coding Intelligence · Qwen3.8 Architecture</div>
+                </div>
+            </div>
+            <div class="status-pill">
+                <span class="status-dot"></span>
+                <span>ZeroGPU A100 · 4-Bit Native</span>
+            </div>
+        </div>
+    """)
+
+    with gr.Row():
+        # ------------------ SETTINGS / SIDEBAR ------------------
+        with gr.Column(scale=1, min_width=270):
+            with gr.Group(elem_classes="sidebar-panel"):
+                gr.Markdown("#### Runtime Controls")
+                enable_thinking = gr.Checkbox(value=False, label="Deep Reasoning (CoT)")
+                enable_web_search = gr.Checkbox(value=False, label="Live Web Context (DDG)")
+
+                with gr.Accordion("Sampling & Temperature", open=False):
+                    temperature = gr.Slider(0.0, 1.5, value=0.6, step=0.05, label="Temperature")
+                    top_p = gr.Slider(0.1, 1.0, value=0.85, step=0.05, label="Top-P")
+                    max_new_tokens = gr.Slider(128, 4096, value=1024, step=64, label="Max Tokens")
+
+                with gr.Accordion("System Instructions", open=False):
+                    system_prompt = gr.Textbox(value=SYSTEM_PROMPT, label="Prompt", lines=6, show_label=False)
+
+            with gr.Group(elem_classes="sidebar-panel", visible=True):
+                gr.Markdown("#### Web References")
+                sources_box = gr.HTML('<div class="sources-empty">Pencarian web nonaktif.</div>')
+
+        # ------------------ CHAT WORKSPACE ------------------
+        with gr.Column(scale=3):
+            chatbot = gr.Chatbot(
+                type="messages",
+                value=[{"role": "assistant", "content": WELCOME_MD}],
+                height=620,
+                show_label=False,
+                elem_classes="chat-viewport"
+            )
+
+            # Quick Prompt Suggestions (Interactive Pills)
+            with gr.Row(elem_classes="quick-prompt-row"):
+                p1 = gr.Button("Memory leak pada Python generator?", elem_classes="quick-prompt-btn")
+                p2 = gr.Button("Arsitektur Microservices vs Modular Monolith", elem_classes="quick-prompt-btn")
+                p3 = gr.Button("Optimasi SQL Query (N+1 Problem)", elem_classes="quick-prompt-btn")
+
+            message = gr.MultimodalTextbox(
+                placeholder="Tanyakan analisis kode, upload dokumen teknis (PDF/Docx/Sheet), atau screenshot error...",
+                file_types=["image", "video", "file"],
+                file_count="multiple",
+                submit_btn="Send",
+                stop_btn="Stop",
+            )
+
+    # ------------------ INTERAKSI & EVENTS ------------------
+    def set_prompt_p1():
+        return {"text": "Bagaimana cara mendeteksi dan mencegah memory leak pada penggunaan generator & closure di Python?", "files": []}
+    def set_prompt_p2():
+        return {"text": "Kapan sebaiknya beralih dari Modular Monolith ke Microservices? Berikan trade-off engineeringnya.", "files": []}
+    def set_prompt_p3():
+        return {"text": "Jelaskan apa itu masalah N+1 Query pada ORM dan 3 strategi konkret untuk mengatasinya.", "files": []}
+
+    p1.click(set_prompt_p1, outputs=[message])
+    p2.click(set_prompt_p2, outputs=[message])
+    p3.click(set_prompt_p3, outputs=[message])
+
+    def stream_handler(msg, hist, sp, temp, tp, mnt, eth, ews):
+        user_text = (msg or {}).get("text", "")
+        files = (msg or {}).get("files", [])
+
+        user_content_display = []
+        for f in files:
+            path = normalize_path(f)
+            if path:
+                user_content_display.append({"path": path})
+        if user_text:
+            user_content_display.append(user_text)
+
+        new_history = list(hist or [])
+        new_history.append({"role": "user", "content": user_content_display if user_content_display else user_text})
+        new_history.append({"role": "assistant", "content": ""})
+
+        last_search = []
+        for partial_text, search_res in generate_chat(msg, hist, sp, temp, tp, mnt, eth, ews):
+            new_history[-1]["content"] = partial_text
+            last_search = search_res
+            yield new_history, sources_html(last_search)
+
+    message.submit(
+        stream_handler,
+        inputs=[message, chatbot, system_prompt, temperature, top_p, max_new_tokens, enable_thinking, enable_web_search],
+        outputs=[chatbot, sources_box]
+    )
+
+if __name__ == "__main__":
+    demo.queue(max_size=16, default_concurrency_limit=1).launch()