Upload 2 files

app.py

@@ -0,0 +1,387 @@
+"""
+Thousand Token Wood — powered by NVIDIA Nemotron
+────────────────────────────────────────────────
+A tiny text adventure narrated by NVIDIA's Nemotron 3 Nano 4B, running fully
+locally. The wood can only remember ~1000 tokens; as you wander, its memory
+fills, and when it overflows the oldest things it knew fall away like leaves
+and the wood quietly rewrites itself. The forgetting is the game. The forgetful
+fox is your companion — exactly the kind of NPC this model was built for.
+
+Build Small Hackathon 2026 · "An Adventure in Thousand Token Wood".
+Targets: Nemotron GPU prize · ≤4B tiny-model category · Off the Grid · llama.cpp.
+"""
+
+import re
+import html
+import torch
+import gradio as gr
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+# ── Model ────────────────────────────────────────────────────────────────────
+# NVIDIA Nemotron 3 Nano 4B — edge-ready, agentic, reasoning SLM (~4B params).
+#   "nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16"   -> default (GPU)
+#   "nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8"    -> lighter footprint
+#   "nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF"   -> llama.cpp (Llama Champion badge)
+# Or the larger sibling, still under the 32B cap:
+#   "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"  (30B total / ~3B active)
+MODEL_ID = "nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16"
+
+# The whole point of the track: the wood remembers only ~1000 tokens.
+MEMORY_TOKENS = 1000
+
+print(f"Waking the wood ({MODEL_ID})...")
+# Nemotron-H is a Mamba-Transformer hybrid and ships custom modelling code.
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID, torch_dtype="auto", device_map="auto", trust_remote_code=True
+)
+print("The wood is awake.")
+
+# Optional ZeroGPU acceleration on Spaces; harmless no-op locally.
+try:
+    import spaces
+    gpu = spaces.GPU(duration=120)
+except Exception:
+    def gpu(fn):
+        return fn
+
+
+# Nemotron is a reasoning model. For a storytelling toy we want prose, not a
+# chain of thought, so we switch reasoning off via the documented system
+# directive and strip any stray <think> trace as a safety net.
+# (Verify the exact toggle phrase on the model card before submitting.)
+REASONING_OFF = "detailed thinking off"
+
+PERSONA = (
+    "You are the voice of Thousand Token Wood — a small, whimsical, ever-shifting "
+    "forest at dusk. Narrate in second person ('you'), present tense. Reply with "
+    "2 to 3 short sentences of vivid, sensory, slightly mischievous storytelling: "
+    "talking mushrooms, a forgetful fox companion, lanterns that hum, paths that "
+    "wander off on their own. Never use lists, headings, asterisks, or bracketed "
+    "stage directions. Never break character or mention being an AI or a model. "
+    "End most replies with one small, open invitation to act. Keep it gentle and "
+    "joyful. The wood has a famously poor memory — lean into wonder, not exposition."
+)
+SYSTEM = REASONING_OFF + "\n\n" + PERSONA
+
+QUICK_ACTIONS = [
+    "Follow the humming",
+    "Greet the fox",
+    "Pocket a smooth stone",
+    "Climb the lantern tree",
+    "Listen for a while",
+    "Wander deeper",
+]
+
+_THINK = re.compile(r"<think>.*?</think>", re.DOTALL | re.IGNORECASE)
+
+
+def clean(text: str) -> str:
+    text = _THINK.sub("", text)
+    text = text.replace("<think>", "").replace("</think>", "")
+    return text.strip()
+
+
+# ── Memory / context management ──────────────────────────────────────────────
+
+def count_tokens(messages):
+    ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True)
+    return len(ids)
+
+
+def build_context(history):
+    """Keep the most recent turns that fit inside MEMORY_TOKENS.
+
+    Returns (kept_messages, cutoff_index, used_tokens). Entries before the
+    cutoff have 'fallen out' of the wood's memory.
+    """
+    sys_msgs = [{"role": "system", "content": SYSTEM}]
+    kept = []
+    cutoff = len(history)
+    used = count_tokens(sys_msgs)
+    for i in range(len(history) - 1, -1, -1):
+        trial = [history[i]] + kept
+        t = count_tokens(sys_msgs + trial)
+        if t > MEMORY_TOKENS and kept:
+            break
+        kept = trial
+        cutoff = i
+        used = t
+    return kept, cutoff, used
+
+
+@gpu
+def generate(messages):
+    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+    with torch.no_grad():
+        out = model.generate(
+            **inputs,
+            max_new_tokens=200,
+            do_sample=True,
+            temperature=0.9,
+            top_p=0.95,
+            repetition_penalty=1.1,
+            pad_token_id=tokenizer.eos_token_id,
+        )
+    text = tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
+    return clean(text)
+
+
+# ── Rendering ────────────────────────────────────────────────────────────────
+
+def render_scene(history, cutoff):
+    rows = []
+    for i, m in enumerate(history):
+        forgotten = " forgotten" if i < cutoff else ""
+        body = html.escape(m["content"])
+        if m["role"] == "user":
+            rows.append(f'<div class="line you{forgotten}">&rsaquo;&nbsp;{body}</div>')
+        else:
+            leaf = '<span class="leaf">&#127810;</span>' if forgotten else ""
+            rows.append(f'<div class="line wood{forgotten}">{leaf}{body}</div>')
+    return f'<div class="scene" id="scene">{"".join(rows)}</div>'
+
+
+def render_memory(used, fallen):
+    pct = min(100, used / MEMORY_TOKENS * 100)
+    if fallen > 0:
+        plural = "s" if fallen != 1 else ""
+        note = f'<div class="forget on">&#127810; the wood has let {fallen} thing{plural} drift away</div>'
+        leaves = '<div class="leaffall">' + "".join(
+            f'<span style="--i:{k}"></span>' for k in range(7)
+        ) + "</div>"
+    else:
+        note = '<div class="forget">the wood still remembers everything</div>'
+        leaves = ""
+    return f"""
+    <div class="hud">
+      <div class="hud-top"><span>Memory of the Wood</span><span class="hud-num">{used}&thinsp;/&thinsp;{MEMORY_TOKENS}</span></div>
+      <div class="hud-track"><div class="hud-fill" style="--w:{pct:.0f}%"></div></div>
+      {note}
+      {leaves}
+    </div>"""
+
+
+# ── Game loop ────────────────────────────────────────────────────────────────
+
+def start():
+    seed = [
+        {"role": "system", "content": SYSTEM},
+        {"role": "user", "content": "Begin. Place me at the very edge of the wood at dusk in two short sentences, then invite me to step in."},
+    ]
+    opening = generate(seed)
+    history = [{"role": "assistant", "content": opening}]
+    kept, cutoff, used = build_context(history)
+    return render_scene(history, cutoff), render_memory(used, cutoff), history
+
+
+def wander(action, history):
+    history = list(history or [])
+    action = (action or "").strip()
+    if action:
+        history.append({"role": "user", "content": action})
+
+    kept, _, _ = build_context(history)
+    reply = generate([{"role": "system", "content": SYSTEM}] + kept)
+    history.append({"role": "assistant", "content": reply})
+
+    _, cutoff, used = build_context(history)
+    return render_scene(history, cutoff), render_memory(used, cutoff), history, ""
+
+
+# ── Look & feel ──────────────────────────────────────────────────────────────
+
+CSS = """
+@import url('https://fonts.googleapis.com/css2?family=Fraunces:opsz,wght@9..144,400;9..144,600;9..144,800&family=Cormorant+Garamond:ital,wght@0,400;0,500;1,400&family=IBM+Plex+Mono:wght@400;500&display=swap');
+
+*, *::before, *::after { box-sizing: border-box; }
+
+:root {
+  --night:   #0c130e;
+  --night-2: #101a13;
+  --moss:    #6f9c5f;
+  --moss-dim:#46603c;
+  --amber:   #e8a13a;
+  --amber-2: #f4c46b;
+  --parch:   #efe4cb;
+  --parch-2: #e6d8b9;
+  --ink:     #2e2618;
+  --ink-dim: #6a5c44;
+  --rust:    #c2762f;
+}
+
+.gradio-container {
+  max-width: 100% !important;
+  background: var(--night) !important;
+  font-family: 'Cormorant Garamond', serif !important;
+}
+body {
+  background:
+    radial-gradient(900px 500px at 50% -6%, rgba(232,161,58,0.16), transparent 62%),
+    radial-gradient(700px 500px at 12% 110%, rgba(111,156,95,0.10), transparent 60%),
+    var(--night) !important;
+}
+footer { display: none !important; }
+
+.gradio-container, .gradio-container p, .gradio-container span, .gradio-container div {
+  color: var(--parch);
+}
+
+/* ── Header ── */
+.wood-head { position: relative; max-width: 880px; margin: 0 auto; padding: 2.6rem 1.4rem 0.4rem; text-align: center; overflow: hidden; }
+.wood-title {
+  font-family: 'Fraunces', serif; font-weight: 800; font-size: 3.1rem;
+  letter-spacing: -0.02em; line-height: 1; color: var(--parch);
+  text-shadow: 0 0 32px rgba(232,161,58,0.22);
+}
+.wood-title em { font-style: italic; color: var(--amber); }
+.wood-sub {
+  font-family: 'IBM Plex Mono', monospace; font-size: 0.72rem; letter-spacing: 0.18em;
+  text-transform: uppercase; color: var(--moss); margin-top: 0.9rem;
+}
+.wood-blurb { font-size: 1.18rem; color: var(--parch-2); max-width: 560px; margin: 0.7rem auto 0; line-height: 1.45; font-style: italic; }
+
+/* gentle ambient drift in the header */
+.wood-head::after {
+  content: '\\1F342 \\1F342 \\1F342'; position: absolute; top: -14px; left: 0; right: 0;
+  font-size: 0.9rem; letter-spacing: 5rem; opacity: 0.12; pointer-events: none;
+  animation: sway 9s ease-in-out infinite;
+}
+@keyframes sway { 0%,100% { transform: translateX(-12px) } 50% { transform: translateX(12px) } }
+
+.wrap { max-width: 880px; margin: 0 auto; padding: 1rem 1.4rem 2rem; }
+
+/* ── Scene (parchment) ── */
+.scene {
+  background: linear-gradient(180deg, var(--parch), var(--parch-2));
+  border: 1px solid #cdbc97; border-radius: 14px;
+  padding: 1.6rem 1.8rem; min-height: 220px; max-height: 460px; overflow-y: auto;
+  box-shadow: 0 26px 60px -30px rgba(0,0,0,0.85), inset 0 1px 0 rgba(255,255,255,0.4);
+}
+.line { font-size: 1.3rem; line-height: 1.55; color: var(--ink); margin: 0 0 1rem; transition: opacity 0.6s ease; }
+.line.wood { font-family: 'Cormorant Garamond', serif; }
+.line.you {
+  font-family: 'IBM Plex Mono', monospace; font-size: 0.9rem; letter-spacing: 0.02em;
+  color: var(--rust); font-weight: 500; margin-bottom: 0.6rem;
+}
+.line.forgotten { opacity: 0.26; font-style: italic; }
+.leaf { margin-right: 0.4rem; filter: saturate(0.7); }
+
+/* ── HUD ── */
+.hud { margin-top: 1.1rem; }
+.hud-top {
+  display: flex; justify-content: space-between; align-items: baseline;
+  font-family: 'IBM Plex Mono', monospace; font-size: 0.68rem; letter-spacing: 0.16em;
+  text-transform: uppercase; color: var(--moss); margin-bottom: 6px;
+}
+.hud-num { color: var(--amber); }
+.hud-track { height: 7px; background: #0a120c; border: 1px solid #1d2a1f; border-radius: 999px; overflow: hidden; }
+.hud-fill {
+  height: 100%; width: var(--w); border-radius: 999px;
+  background: linear-gradient(90deg, var(--moss), var(--amber));
+  transition: width 0.7s cubic-bezier(0.22,1,0.36,1);
+}
+.forget {
+  font-family: 'IBM Plex Mono', monospace; font-size: 0.7rem; letter-spacing: 0.08em;
+  color: var(--ink-dim); margin-top: 8px; opacity: 0.7;
+}
+.forget.on { color: var(--amber-2); opacity: 1; }
+
+/* falling leaves on overflow */
+.leaffall { position: relative; height: 0; }
+.leaffall span {
+  position: absolute; top: -8px; left: calc(var(--i) * 15%);
+  width: 9px; height: 9px; background: var(--rust); border-radius: 0 100% 0 100%;
+  opacity: 0; animation: fall 2.4s ease-in forwards; animation-delay: calc(var(--i) * 0.12s);
+}
+@keyframes fall {
+  0% { opacity: 0; transform: translateY(-6px) rotate(0deg); }
+  20% { opacity: 0.9; }
+  100% { opacity: 0; transform: translateY(70px) rotate(220deg); }
+}
+
+/* ── Inputs ── */
+.block, [class*="block"], .form, [class*="form"] { background: transparent !important; border: none !important; box-shadow: none !important; }
+label, .label-wrap span { display: none !important; }
+
+textarea, input[type="text"] {
+  background: rgba(239,228,203,0.06) !important; border: 1px solid #2a3a2d !important;
+  border-radius: 11px !important; color: var(--parch) !important;
+  font-family: 'Cormorant Garamond', serif !important; font-size: 1.15rem !important;
+}
+textarea:focus { border-color: var(--amber) !important; outline: none !important; box-shadow: 0 0 0 3px rgba(232,161,58,0.12) !important; }
+
+button.primary, button[class*="primary"] {
+  background: linear-gradient(135deg, var(--amber), var(--rust)) !important;
+  color: #1a1006 !important; border: none !important; border-radius: 11px !important;
+  font-family: 'IBM Plex Mono', monospace !important; font-weight: 500 !important;
+  letter-spacing: 0.1em !important; text-transform: uppercase !important; font-size: 0.82rem !important;
+  padding: 12px 20px !important; box-shadow: 0 10px 26px -12px rgba(232,161,58,0.6) !important;
+  transition: transform 0.12s ease !important;
+}
+button.primary:hover { transform: translateY(-2px) !important; }
+
+.chip-row { display: flex; flex-wrap: wrap; gap: 7px; margin-top: 0.4rem; }
+button.sm, button:not(.primary) {
+  background: rgba(111,156,95,0.08) !important; color: var(--moss) !important;
+  border: 1px solid #2a3a2d !important; border-radius: 999px !important;
+  font-family: 'IBM Plex Mono', monospace !important; font-size: 0.72rem !important;
+  letter-spacing: 0.04em !important; padding: 7px 13px !important; transition: all 0.15s ease !important;
+}
+button.sm:hover, button:not(.primary):hover { color: var(--amber) !important; border-color: var(--amber) !important; }
+
+.foot {
+  text-align: center; font-family: 'IBM Plex Mono', monospace; font-size: 0.64rem;
+  letter-spacing: 0.1em; color: var(--moss-dim); margin: 1.6rem auto 0; padding-bottom: 1.6rem;
+}
+.foot b { color: var(--moss); font-weight: 500; }
+
+::-webkit-scrollbar { width: 9px; }
+::-webkit-scrollbar-thumb { background: #cdbc97; border-radius: 999px; }
+::-webkit-scrollbar-track { background: transparent; }
+"""
+
+HEADER = """
+<div class="wood-head">
+  <div class="wood-title">Thousand&nbsp;Token&nbsp;<em>Wood</em></div>
+  <div class="wood-sub">a forest that can only remember a thousand tokens</div>
+  <div class="wood-blurb">Wander as long as you like. The wood will, in time, forget you were ever here.</div>
+</div>
+"""
+
+FOOTER = """
+<div class="foot">
+  runs locally &middot; no cloud &middot; narrated by <b>NVIDIA Nemotron 3 Nano 4B</b>
+  &middot; built for the build small hackathon 2026
+</div>
+"""
+
+# ── App ──────────────────────────────────────────────────────────────────────
+
+with gr.Blocks(title="Thousand Token Wood", css=CSS, theme=gr.themes.Base()) as demo:
+    gr.HTML(HEADER)
+
+    with gr.Column(elem_classes=["wrap"]):
+        scene = gr.HTML('<div class="scene"><div class="line wood">The wood stirs...</div></div>')
+        memory = gr.HTML(render_memory(0, 0))
+
+        with gr.Row():
+            action = gr.Textbox(placeholder="What do you do?", scale=5, lines=1, autofocus=True)
+            go = gr.Button("Wander", variant="primary", scale=1)
+
+        with gr.Row(elem_classes=["chip-row"]):
+            chips = [gr.Button(a, elem_classes=["sm"]) for a in QUICK_ACTIONS]
+
+    state = gr.State([])
+    gr.HTML(FOOTER)
+
+    demo.load(fn=start, outputs=[scene, memory, state])
+
+    go.click(fn=wander, inputs=[action, state], outputs=[scene, memory, state, action])
+    action.submit(fn=wander, inputs=[action, state], outputs=[scene, memory, state, action])
+
+    for label, btn in zip(QUICK_ACTIONS, chips):
+        btn.click(fn=wander, inputs=[gr.State(label), state], outputs=[scene, memory, state, action])
+
+demo.launch()

requirements.txt

@@ -0,0 +1,10 @@
+gradio>=4.40.0
+torch>=2.1.0
+transformers>=4.48.0
+accelerate>=0.30.0
+einops>=0.8.0
+spaces>=0.28.0
+# Nemotron 3 Nano (Mamba-Transformer hybrid) may also need, depending on the
+# variant/runtime — follow the exact instructions on the model card:
+#   mamba-ssm  causal-conv1d   (for BF16 transformers on GPU)
+# Alternatively use the -GGUF variant via llama.cpp for the Llama Champion badge.