File size: 6,379 Bytes
44810a8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 | """
Benchmark: Complexity and Scaling Analysis (Paper Experiment 2 & 3)
Evaluates:
1. Offline BFS Build Time (ms) vs Number of States |S|
2. Memory Footprint (MB) vs Number of States |S|
3. Online Token Masking Latency (us) vs Number of States |S| (Empirical O(1) Verification)
Generates publication-quality figure: 'paper_figure_scaling.png'
"""
import os
import time
import torch
import matplotlib.pyplot as plt
from tabulate import tabulate
from typing import List, Dict
from core.fsm_builder import ReachabilityFSM
from core.logit_processor import GoalReachabilityLogitsProcessor
def run_scaling_benchmark(output_plot_path: str = "paper_figure_scaling.png"):
print("\n" + "=" * 80)
print(" [EXPERIMENT 2 & 3] FSM Complexity & O(1) Latency Scaling Benchmark")
print("=" * 80)
device = "cpu" # Run on CPU for strict baseline consistency
vocab_sizes = [32000, 151643] # Standard 32k vs Qwen 151k
state_counts = [10, 50, 200, 1000, 5000, 10000]
max_budget = 50
num_latency_trials = 500
results: Dict[int, Dict[str, List]] = {
v: {"states": [], "build_time_ms": [], "memory_mb": [], "online_latency_us": []}
for v in vocab_sizes
}
table_rows = []
for vocab_size in vocab_sizes:
print(f"\n--- Testing Vocab Size: {vocab_size:,} ---")
for num_states in state_counts:
# 1. Build FSM structure
fsm = ReachabilityFSM(num_states=num_states, vocab_size=vocab_size, device=device)
# Add chain transitions + some branches
for s in range(num_states - 1):
fsm.add_transition(s, token_id=s % vocab_size, to_state=s + 1)
# Add branching transition
if s + 2 < num_states:
fsm.add_transition(s, token_id=(s * 7 + 13) % vocab_size, to_state=s + 2)
fsm.set_goal_states([num_states - 1])
# 2. Measure Offline BFS Build Time
start_b = time.perf_counter()
fsm.build_reachability(max_steps=max_budget)
end_b = time.perf_counter()
build_time_ms = (end_b - start_b) * 1000
# 3. Measure Memory Footprint
mem_bytes = fsm.memory_footprint_bytes()
mem_mb = mem_bytes / (1024 * 1024)
# 4. Measure Online 1-Token Masking Latency
processor = GoalReachabilityLogitsProcessor(fsm=fsm, max_budget=max_budget)
dummy_input = torch.tensor([[0]], dtype=torch.long, device=device)
dummy_scores = torch.randn((1, vocab_size), device=device)
# Warmup
for _ in range(50):
_ = processor(dummy_input, dummy_scores.clone())
start_l = time.perf_counter()
for _ in range(num_latency_trials):
_ = processor(dummy_input, dummy_scores)
end_l = time.perf_counter()
online_latency_us = ((end_l - start_l) / num_latency_trials) * 1e6
results[vocab_size]["states"].append(num_states)
results[vocab_size]["build_time_ms"].append(build_time_ms)
results[vocab_size]["memory_mb"].append(mem_mb)
results[vocab_size]["online_latency_us"].append(online_latency_us)
table_rows.append([
f"{vocab_size:,}",
f"{num_states:,}",
f"{build_time_ms:.2f} ms",
f"{mem_mb:.2f} MB",
f"{online_latency_us:.2f} us ({online_latency_us/1000:.4f} ms)",
])
headers = [
"Vocab Size",
"States |S|",
"Offline BFS Time",
"Memory Footprint",
"Online 1-Token Latency",
]
print("\n" + tabulate(table_rows, headers=headers, tablefmt="grid"))
# -------------------------------------------------------------
# Generate Publication-Quality Figures (Matplotlib)
# -------------------------------------------------------------
print(f"\n[PLOT] Generating publication-quality figures -> {output_plot_path}...")
plt.style.use("seaborn-v0_8-whitegrid" if "seaborn-v0_8-whitegrid" in plt.style.available else "default")
fig, axes = plt.subplots(1, 3, figsize=(18, 5), dpi=300)
colors = {32000: "#1f77b4", 151643: "#ff7f0e"}
labels = {32000: "Vocab: 32,000 (LLaMA/Mistral)", 151643: "Vocab: 151,643 (Qwen2.5)"}
# Plot 1: Offline BFS Build Time
for v in vocab_sizes:
axes[0].plot(
results[v]["states"],
results[v]["build_time_ms"],
marker="o",
linewidth=2,
color=colors[v],
label=labels[v],
)
axes[0].set_title("(a) Offline BFS Precomputation Time", fontsize=12, fontweight="bold")
axes[0].set_xlabel("FSM State Count |S|", fontsize=11)
axes[0].set_ylabel("Build Time (ms)", fontsize=11)
axes[0].set_xscale("log")
axes[0].legend(fontsize=9)
axes[0].grid(True, linestyle="--", alpha=0.6)
# Plot 2: Memory Footprint
for v in vocab_sizes:
axes[1].plot(
results[v]["states"],
results[v]["memory_mb"],
marker="s",
linewidth=2,
color=colors[v],
label=labels[v],
)
axes[1].set_title("(b) Memory Footprint (VRAM / RAM)", fontsize=12, fontweight="bold")
axes[1].set_xlabel("FSM State Count |S|", fontsize=11)
axes[1].set_ylabel("Memory (MB)", fontsize=11)
axes[1].set_xscale("log")
axes[1].legend(fontsize=9)
axes[1].grid(True, linestyle="--", alpha=0.6)
# Plot 3: Online 1-Token Latency (O(1) Verification)
for v in vocab_sizes:
axes[2].plot(
results[v]["states"],
results[v]["online_latency_us"],
marker="^",
linewidth=2,
color=colors[v],
label=labels[v],
)
axes[2].set_title("(c) Online Runtime Latency: Strict O(1)", fontsize=12, fontweight="bold")
axes[2].set_xlabel("FSM State Count |S|", fontsize=11)
axes[2].set_ylabel("Latency per Token (us)", fontsize=11)
axes[2].set_xscale("log")
axes[2].legend(fontsize=9)
axes[2].grid(True, linestyle="--", alpha=0.6)
plt.tight_layout()
plt.savefig(output_plot_path, bbox_inches="tight")
plt.close()
print(f"[SUCCESS] Figure successfully saved to: {os.path.abspath(output_plot_path)}\n")
if __name__ == "__main__":
run_scaling_benchmark()
|