""" BitNet b1.58 2B4T — CPU-Only Inference Explorer (bitnet.cpp Edition) ==================================================================== Powered by bitnet.cpp's optimized ternary kernels for 4-10x faster inference. Uses llama-server with OpenAI-compatible API for streaming generation. Paper: https://arxiv.org/abs/2504.12285 Model: https://huggingface.co/microsoft/bitnet-b1.58-2B-4T """ import os import time import psutil import gradio as gr from openai import OpenAI # ─── Configuration ─────────────────────────────────────────────────────────── SERVER_URL = "http://127.0.0.1:8080/v1" MODEL_NAME = "bitnet-b1.58-2B-4T" # Connect to local llama-server client = OpenAI(base_url=SERVER_URL, api_key="bitnet") # ─── System Info ───────────────────────────────────────────────────────────── cpu_count = psutil.cpu_count(logical=True) total_ram = psutil.virtual_memory().total / 1024**3 proc = psutil.Process(os.getpid()) def get_system_info(): mem = proc.memory_info().rss / 1024**3 return f"""### System | Metric | Value | |---|---| | CPU cores | {cpu_count} | | Total RAM | {total_ram:.1f} GB | | Process RSS | {mem:.2f} GB | | Inference engine | bitnet.cpp (I2_S kernel) | | Weights | 1.58-bit ternary ({{-1, 0, +1}}) | | Activations | 8-bit integer | | Context | 4096 tokens | | Backend | llama-server (OpenAI API) | """ # ─── Paper benchmark table ─────────────────────────────────────────────────── PAPER_TABLE = """### Published Benchmarks (from the paper) | Benchmark | LLaMA 3.2 1B | Gemma-3 1B | Qwen2.5 1.5B | SmolLM2 1.7B | **BitNet 2B** | |---|---|---|---|---|---| | **Memory** | 2 GB | 1.4 GB | 2.6 GB | 3.2 GB | **0.4 GB** | | **CPU Latency** | 48ms | 41ms | 65ms | 67ms | **29ms** | | **Energy/token** | 0.258J | 0.186J | 0.347J | 0.425J | **0.028J** | | ARC-Challenge | 37.8 | 38.4 | 46.7 | 43.5 | **49.9** | | WinoGrande | 59.5 | 58.5 | 62.8 | 69.0 | **71.9** | | GSM8K | 38.2 | 31.2 | 56.8 | 45.1 | **58.4** | | MMLU | 45.6 | 39.9 | **60.3** | 49.2 | 53.2 | | HumanEval+ | 31.1 | 37.2 | **50.6** | 28.0 | 38.4 | | **Average** | 44.9 | 43.7 | **55.2** | 48.7 | 54.2 | *BitNet uses 5-13× less memory and 6-9× less energy than comparable models.* > ✅ This demo uses **bitnet.cpp** with the optimized I2_S kernel — the same > engine that achieves the 29ms/token latency shown above. """ # ─── Architecture explainer ────────────────────────────────────────────────── ARCHITECTURE_MD = """### How BitNet b1.58 Works ``` Standard Transformer → BitNet b1.58 ───────────────────── ───────────────── FP16/BF16 weights (16 bits) → Ternary weights: {-1, 0, +1} (1.58 bits) FP16 activations → INT8 activations (absmax per-token) nn.Linear → BitLinear (absmean quantization) SwiGLU activation → Squared ReLU (ReLU²) LayerNorm → SubLN normalization Standard MatMul → Additions only (no multiplications!) ``` **Key Insight:** Since weights are only -1, 0, or +1, matrix multiplication becomes pure addition/subtraction. This is why CPUs can run BitNet models so efficiently — you don't need floating-point multiply hardware at all. **bitnet.cpp Kernels:** - **I2_S** (Int2 with Scale): MAD-based, lossless, 2 bits/weight storage - **TL1/TL2** (Ternary Lookup): LUT-based, lossless, sub-2-bit storage - Both achieve **4-6× speedup** over FP16 llama.cpp on the same CPU **Training:** The model was trained **from scratch** with this quantization, not post-training quantized. This is crucial — native 1-bit training preserves quality far better than quantizing a pre-trained FP16 model down to 1-bit. **3-Stage Training Pipeline:** 1. **Pre-training** on 4T tokens (text, code, synthetic math) 2. **SFT** on instruction-following datasets 3. **DPO** for alignment with human preferences """ # ─── Generation functions ──────────────────────────────────────────────────── def chat_respond(message, history, system_prompt, max_new_tokens, temperature, top_p): """Streaming chat via bitnet.cpp llama-server.""" messages = [{"role": "system", "content": system_prompt}] for item in history: messages.append(item) messages.append({"role": "user", "content": message}) t0 = time.perf_counter() tok_count = 0 response = "" try: stream = client.chat.completions.create( model=MODEL_NAME, messages=messages, max_tokens=int(max_new_tokens), temperature=float(temperature) if temperature > 0 else 0.0, top_p=float(top_p), stream=True, ) for chunk in stream: if chunk.choices[0].delta.content: token_text = chunk.choices[0].delta.content response += token_text tok_count += 1 elapsed = time.perf_counter() - t0 tps = tok_count / elapsed if elapsed > 0 else 0 stats = f"\n\n---\n*⚡ {tok_count} tokens · {tps:.1f} tok/s · {elapsed:.1f}s · bitnet.cpp I2_S*" yield response + stats except Exception as e: yield f"**Error:** {str(e)}\n\nIs the llama-server running on port 8080?" def single_benchmark(prompt, max_new_tokens): """Run a single non-streaming generation with detailed stats.""" messages = [ {"role": "system", "content": "You are a helpful AI assistant."}, {"role": "user", "content": prompt}, ] t0 = time.perf_counter() try: completion = client.chat.completions.create( model=MODEL_NAME, messages=messages, max_tokens=int(max_new_tokens), temperature=0.0, stream=False, ) elapsed = time.perf_counter() - t0 response = completion.choices[0].message.content n_generated = completion.usage.completion_tokens if completion.usage else len(response.split()) n_input = completion.usage.prompt_tokens if completion.usage else 0 tps = n_generated / elapsed if elapsed > 0 else 0 stats_md = f"""### ⚡ Benchmark Results (bitnet.cpp I2_S kernel) | Metric | Value | |---|---| | Input tokens | {n_input} | | Output tokens | {n_generated} | | Total time | {elapsed:.2f}s | | **Tokens/sec** | **{tps:.2f}** | | Avg ms/token | {(elapsed/max(n_generated,1)*1000):.1f}ms | | Engine | bitnet.cpp (lossless) | | Kernel | I2_S (MAD-based) | """ return response, stats_md except Exception as e: return f"Error: {str(e)}", "Server not responding" # ─── Build Gradio UI ───────────────────────────────────────────────────────── HEADER = """# 🧬 BitNet b1.58 2B4T — CPU-Only Inference Explorer **The first open-source native 1-bit LLM** by Microsoft Research — powered by **bitnet.cpp** optimized kernels. | | | |---|---| | 📄 [Paper](https://arxiv.org/abs/2504.12285) | 🤗 [Model](https://huggingface.co/microsoft/bitnet-b1.58-2B-4T) | | 💻 [bitnet.cpp](https://github.com/microsoft/BitNet) (38K+ ⭐) | ⚡ Ternary I2_S kernel · ~10 tok/s on CPU | """ with gr.Blocks( title="BitNet b1.58 2B4T — CPU Inference Explorer", ) as demo: gr.Markdown(HEADER) with gr.Tabs(): # ── Tab 1: Chat ────────────────────────────────────────────────── with gr.Tab("💬 Chat", id="chat"): chat = gr.ChatInterface( fn=chat_respond, description="Chat with BitNet b1.58 via bitnet.cpp on CPU. Live token/sec stats shown after each response.", additional_inputs=[ gr.Textbox( value="You are a helpful, concise AI assistant.", label="System Prompt", ), gr.Slider(1, 2048, value=256, step=1, label="Max New Tokens"), gr.Slider(0.0, 2.0, value=0.7, step=0.05, label="Temperature (0 = greedy)"), gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p"), ], examples=[ ["Explain what a 1-bit LLM is in 3 sentences."], ["Write a Python function to find the nth Fibonacci number."], ["What are the pros and cons of running AI on CPUs vs GPUs?"], ["Solve: If 3x + 7 = 22, what is x?"], ], cache_examples=False, ) # ── Tab 2: Benchmark ───────────────────────────────────────────── with gr.Tab("📊 Benchmark", id="bench"): gr.Markdown("### Run a single-shot benchmark (greedy decoding, bitnet.cpp)") with gr.Row(): with gr.Column(scale=2): bench_prompt = gr.Textbox( value="Write a detailed explanation of how transformer neural networks work, covering attention mechanisms, positional encoding, and the training process.", label="Prompt", lines=3, ) bench_tokens = gr.Slider(16, 512, value=128, step=16, label="Max New Tokens") bench_btn = gr.Button("🚀 Run Benchmark", variant="primary") with gr.Column(scale=1): bench_stats = gr.Markdown("*Click 'Run Benchmark' to start*") bench_output = gr.Textbox(label="Generated Text", lines=10, interactive=False) bench_btn.click( fn=single_benchmark, inputs=[bench_prompt, bench_tokens], outputs=[bench_output, bench_stats], ) # ── Tab 3: Paper Results ───────────────────────────────────────── with gr.Tab("📈 Paper Results", id="paper"): gr.Markdown(PAPER_TABLE) # ── Tab 4: Architecture ────────────────────────────────────────── with gr.Tab("🏗️ Architecture", id="arch"): gr.Markdown(ARCHITECTURE_MD) # ── Tab 5: System Info ─────────────────────────────────────────── with gr.Tab("⚙️ System", id="sys"): sys_info = gr.Markdown(get_system_info()) refresh_btn = gr.Button("🔄 Refresh") refresh_btn.click(fn=get_system_info, outputs=sys_info) if __name__ == "__main__": demo.launch(server_name="0.0.0.0", server_port=7860, theme=gr.themes.Soft())