#!/usr/bin/env python3 """Speculative Decoding animation for @fminxyz Series 9 Post 2.""" import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt import matplotlib.patches as patches from matplotlib.animation import FuncAnimation import numpy as np fig, ax = plt.subplots(figsize=(10, 10), facecolor='#0d1117') ax.set_xlim(0, 10) ax.set_ylim(0, 10) ax.axis('off') BG = '#0d1117' BLUE = '#58a6ff' GREEN = '#3fb950' ORANGE = '#f78166' PURPLE = '#bc8cff' GRAY = '#8b949e' WHITE = '#e6edf3' YELLOW = '#e3b341' RED = '#ff7b72' def clear_ax(): ax.clear() ax.set_xlim(0, 10) ax.set_ylim(0, 10) ax.axis('off') ax.set_facecolor(BG) fig.patch.set_facecolor(BG) def add_title(title, subtitle=None, y=9.3): ax.text(5, y, title, ha='center', va='center', fontsize=22, fontweight='bold', color=WHITE, fontfamily='monospace') if subtitle: ax.text(5, y - 0.65, subtitle, ha='center', va='center', fontsize=13, color=GRAY) def draw_block(x, y, w, h, color, text=None, fontsize=11, alpha=1.0): rect = patches.FancyBboxPatch((x, y), w, h, boxstyle="round,pad=0.05", facecolor=color, edgecolor='none', alpha=alpha) ax.add_patch(rect) if text: ax.text(x + w/2, y + h/2, text, ha='center', va='center', fontsize=fontsize, color=WHITE, fontweight='bold') total_frames = 150 def animate(frame): clear_ax() phase = frame / total_frames # Phase 1: Problem — autoregressive bottleneck (0-0.22) if phase < 0.22: t = phase / 0.22 add_title('Speculative Decoding', 'Латентность авторегрессии — главная проблема', y=9.3) # Show sequential token generation ax.text(5, 8.3, 'Наивная генерация: 1 forward → 1 токен', ha='center', fontsize=13, color=ORANGE) tokens = ['The', 'quick', 'brown', 'fox', 'jumps', 'over', 'the', 'lazy'] tok_w = 0.9 tok_x = 0.5 n_shown = min(8, max(1, int(t * 12))) for i in range(n_shown): alpha_val = 1.0 if i < n_shown - 1 else min(1.0, (t * 12 - i) * 2) draw_block(tok_x + i * (tok_w + 0.1), 6.8, tok_w, 0.7, BLUE if i < n_shown - 1 else GREEN, tokens[i], fontsize=10, alpha=alpha_val) if i < n_shown - 1: ax.text(tok_x + i * (tok_w + 0.1) + tok_w + 0.05, 7.15, '→', ha='center', fontsize=14, color=GRAY, alpha=alpha_val) # Time cost if t > 0.5: ax.text(5, 5.8, f'Каждый токен = 1 forward pass ≈ 50 мс', ha='center', fontsize=13, color=RED, alpha=min(1.0, (t-0.5)*4)) if t > 0.7: ax.text(5, 5.1, f'1000 токенов → 50 секунд ожидания 😱', ha='center', fontsize=13, color=RED, fontweight='bold', alpha=min(1.0, (t-0.7)*5)) if t > 0.85: ax.text(5, 4.2, 'GPU занят sequential операциями — нельзя параллелить', ha='center', fontsize=11, color=GRAY, style='italic', alpha=min(1.0, (t-0.85)*7)) # Phase 2: Key insight — draft + verify (0.22-0.50) elif phase < 0.50: t = (phase - 0.22) / 0.28 add_title('Идея: черновик + проверка', 'Маленькая модель пишет, большая верифицирует', y=9.3) # Draft model (small) draft_alpha = min(1.0, t * 4) rect = patches.FancyBboxPatch((0.5, 6.5), 3.5, 1.5, boxstyle="round,pad=0.1", facecolor='#1f2937', edgecolor=BLUE, linewidth=2, alpha=draft_alpha) ax.add_patch(rect) ax.text(2.25, 7.55, '📝 Draft Model', ha='center', va='center', fontsize=14, color=BLUE, fontweight='bold', alpha=draft_alpha) ax.text(2.25, 7.0, 'Llama-68M · 5 мс/токен', ha='center', va='center', fontsize=11, color=GRAY, alpha=draft_alpha) # Target model (large) target_alpha = min(1.0, max(0, t * 4 - 0.5)) rect2 = patches.FancyBboxPatch((5.5, 6.5), 3.5, 1.5, boxstyle="round,pad=0.1", facecolor='#1f2937', edgecolor=GREEN, linewidth=2, alpha=target_alpha) ax.add_patch(rect2) ax.text(7.25, 7.55, '✅ Target Model', ha='center', va='center', fontsize=14, color=GREEN, fontweight='bold', alpha=target_alpha) ax.text(7.25, 7.0, 'Llama-70B · 50 мс/шаг', ha='center', va='center', fontsize=11, color=GRAY, alpha=target_alpha) # Step visualization if t > 0.35: step_alpha = min(1.0, (t - 0.35) * 5) # Draft proposes γ=4 tokens ax.text(5, 5.8, 'Шаг 1: Draft предлагает γ=4 токена', ha='center', fontsize=13, color=BLUE, alpha=step_alpha) draft_tokens = ['quick', 'brown', 'fox', 'jumps'] for i, tok in enumerate(draft_tokens): draw_block(1.0 + i * 1.2, 4.8, 1.0, 0.6, BLUE, tok, fontsize=10, alpha=step_alpha) if t > 0.6: verify_alpha = min(1.0, (t - 0.6) * 5) ax.text(5, 4.1, 'Шаг 2: Target верифицирует ВСЕ 4 токена за 1 pass', ha='center', fontsize=13, color=GREEN, alpha=verify_alpha) # Show accept/reject results = ['✅', '✅', '✅', '❌'] result_colors = [GREEN, GREEN, GREEN, RED] for i, (r, rc) in enumerate(zip(results, result_colors)): ax.text(1.5 + i * 1.2, 3.4, r, ha='center', fontsize=18, color=rc, alpha=verify_alpha) if t > 0.82: final_alpha = min(1.0, (t - 0.82) * 6) ax.text(5, 2.6, '3 токена приняты + 1 от Target = 4 за цену 1!', ha='center', fontsize=13, color=YELLOW, fontweight='bold', alpha=final_alpha) # Phase 3: Math and acceptance (0.50-0.75) elif phase < 0.75: t = (phase - 0.50) / 0.25 add_title('Математика принятия', 'Почему результат идентичен target-only?', y=9.3) # Acceptance formula formulas = [ ('Вероятность принятия токена:', 'α = min(1, p_target(x) / p_draft(x))', BLUE), ('Среднее токенов за step:', 'E[n] = (1 − α^(γ+1)) / (1 − α)', GREEN), ('Speedup:', 'S = E[n] · T_target / (γ · T_draft + T_target)', PURPLE), ] for i, (label, formula, color) in enumerate(formulas): y = 7.5 - i * 2.2 alpha_val = min(1.0, max(0, t * 4 - i * 0.8)) rect = patches.FancyBboxPatch((0.3, y - 0.7), 9.4, 1.5, boxstyle="round,pad=0.1", facecolor='#161b22', edgecolor=color, linewidth=1.5, alpha=alpha_val) ax.add_patch(rect) ax.text(0.7, y + 0.25, label, ha='left', va='center', fontsize=11, color=GRAY, alpha=alpha_val) ax.text(5, y - 0.2, formula, ha='center', va='center', fontsize=13, color=color, fontweight='bold', fontfamily='monospace', alpha=alpha_val) # Key insight box if t > 0.65: insight_alpha = min(1.0, (t - 0.65) * 3) rect = patches.FancyBboxPatch((0.5, 0.3), 9.0, 0.9, boxstyle="round,pad=0.1", facecolor='#1a2730', edgecolor=YELLOW, linewidth=2, alpha=insight_alpha) ax.add_patch(rect) ax.text(5, 0.75, '🔑 Если α ≈ 0.9, γ=4 → E[n] ≈ 3.5 → Speedup ≈ 2.5–3×', ha='center', va='center', fontsize=13, color=YELLOW, alpha=insight_alpha) # Phase 4: Real-world results (0.75-1.0) else: t = (phase - 0.75) / 0.25 add_title('Результаты в production', 'Speculative Decoding 2024–2026', y=9.4) results = [ ('Llama-3-70B', 'Llama-68M', '2.3–3.1×', 'Codeium (код)', BLUE), ('Gemini 1.5 Pro', 'Gemini Nano', '2.8×', 'Google prod.', GREEN), ('Self-speculative', 'Layer skip', '1.5–2×', 'без draft модели', PURPLE), ('Medusa heads', 'N draft heads', '2.4×', 'parallel heads', ORANGE), ] for i, (big, small, speedup, note, color) in enumerate(results): y = 8.0 - i * 1.7 alpha_val = min(1.0, max(0, t * 4 - i * 0.5)) rect = patches.FancyBboxPatch((0.3, y - 0.6), 9.4, 1.2, boxstyle="round,pad=0.1", facecolor='#161b22', edgecolor=color, linewidth=2, alpha=alpha_val) ax.add_patch(rect) ax.text(0.8, y, big, ha='left', va='center', fontsize=12, color=WHITE, fontweight='bold', alpha=alpha_val) ax.text(3.5, y, f'+ {small}', ha='center', va='center', fontsize=11, color=GRAY, alpha=alpha_val) ax.text(6.5, y, speedup, ha='center', va='center', fontsize=20, color=color, fontweight='bold', alpha=alpha_val) ax.text(9.5, y, note, ha='right', va='center', fontsize=10, color=GRAY, alpha=alpha_val) if t > 0.75: ax.text(5, 0.6, '@fminxyz · Series 9: LLM Efficiency', ha='center', fontsize=11, color=GRAY, alpha=min(1.0, (t-0.75)*4)) ani = FuncAnimation(fig, animate, frames=total_frames, interval=1000/30, blit=False) ani.save('/root/Strategy/content/drafts/speculative_decoding_animation.mp4', writer='ffmpeg', fps=30, dpi=108, savefig_kwargs={'facecolor': BG}) plt.close() print("speculative_decoding_animation.mp4 saved")