mirror of
https://github.com/ParisNeo/lollms_hub.git
synced 2026-05-04 03:01:01 -04:00
2d2bc728c0
Summary of Changes: - [INFRA] Added `space_hunter.py` to the project root for new system utilities. - [CONFIG] Updated `.gitignore` to include `*.db-wal` files and exclude `docs/article` directory. - [UI] Refreshed admin templates (`models_manager.html`, `settings.html`) to reflect new architectural capabilities. - [CORE] Modified `architect_manager.py` to support image generation capabilities (`supports_images`). - [CORE] Updated `memory_manager.py` to improve context formatting and category handling. - [API] Implemented new route logic and rate limiting dependencies in `proxy.py` and `admin.py`. - [MODELS] Updated `models.py` to support new fields and `migrations.py` to handle schema evolution. Technical Notes: - The `architect_manager` now manages assets including image nodes. - Memory formatting has been streamlined to handle multiple titles per category. - Admin settings and model management pages have been updated to visualize these new capabilities. BREAKING CHANGE: None.
85 lines
2.5 KiB
Python
85 lines
2.5 KiB
Python
import random
|
|
import numpy as np
|
|
import matplotlib.pyplot as plt
|
|
|
|
# Simulate B-IRRA Constants
|
|
ALPHA_PRIOR = 10.0
|
|
BETA_PRIOR = 1.0
|
|
DECAY = 0.95
|
|
BASELINE_TPS = 50.0
|
|
|
|
class MockServer:
|
|
def __init__(self, name, true_success_rate, true_avg_tps):
|
|
self.name = name
|
|
self.true_success_rate = true_success_rate
|
|
self.true_avg_tps = true_avg_tps
|
|
self.alpha = ALPHA_PRIOR
|
|
self.beta = BETA_PRIOR
|
|
|
|
def sample(self):
|
|
# Thompson Sampling
|
|
return random.betavariate(self.alpha, self.beta)
|
|
|
|
def process_request(self):
|
|
success = random.random() < self.true_success_rate
|
|
if not success:
|
|
return False, 0
|
|
# TPS with some noise
|
|
tps = max(1, self.true_avg_tps + np.random.normal(0, 5))
|
|
return True, tps
|
|
|
|
def update(self, success, tps=0):
|
|
# Apply Decay
|
|
self.alpha = max(1.0, self.alpha * DECAY)
|
|
self.beta = max(1.0, self.beta * DECAY)
|
|
|
|
if success:
|
|
reward = min(1.0, tps / BASELINE_TPS)
|
|
self.alpha += reward
|
|
self.beta += (1.0 - reward)
|
|
else:
|
|
self.beta += 1.0
|
|
|
|
def run_benchmark(iterations=200):
|
|
servers = [
|
|
MockServer("High-End GPU", 0.99, 85.0),
|
|
MockServer("Stable CPU", 0.98, 12.0),
|
|
MockServer("Unstable Cloud", 0.60, 100.0)
|
|
]
|
|
|
|
history = {s.name: [] for s in servers}
|
|
selections = {s.name: 0 for s in servers}
|
|
|
|
for i in range(iterations):
|
|
# 1. Selection Phase
|
|
samples = [(s.sample(), s) for s in servers]
|
|
# B-IRRA score (Lower is better)
|
|
# Assuming all priorities are equal (10)
|
|
best_server = min(samples, key=lambda x: 10 + (1.0 - x[0]) * 100)[1]
|
|
|
|
# 2. Execution Phase
|
|
success, tps = best_server.process_request()
|
|
best_server.update(success, tps)
|
|
|
|
# 3. Log State
|
|
selections[best_server.name] += 1
|
|
for s in servers:
|
|
history[s.name].append(s.alpha / (s.alpha + s.beta))
|
|
|
|
# --- Plotting ---
|
|
plt.figure(figsize=(10, 6))
|
|
for name, vals in history.items():
|
|
plt.plot(vals, label=f"{name} (P_success)")
|
|
|
|
plt.title("B-IRRA Posterior Convergence")
|
|
plt.xlabel("Requests")
|
|
plt.ylabel("Expected Success Probability (Alpha / Total)")
|
|
plt.legend()
|
|
plt.grid(True, alpha=0.3)
|
|
plt.savefig("benchmark_results.png")
|
|
|
|
print("Benchmark complete. Results saved to benchmark_results.png")
|
|
print(f"Selection Stats: {selections}")
|
|
|
|
if __name__ == "__main__":
|
|
run_benchmark() |