Scruff AI
135 lines
3.9 KiB
Python
135 lines
3.9 KiB
Python
# entropy_injection_inference.py
|
|
# GPU temperature / Asymmetry → Model temperature mapping
|
|
# Hotter environment = Higher creativity (reaching into base Llama knowledge)
|
|
|
|
import zmq
|
|
import json
|
|
import time
|
|
import subprocess
|
|
|
|
from unsloth import FastLanguageModel
|
|
import torch
|
|
|
|
def get_gpu_metrics():
|
|
"""Get GPU temperature and power"""
|
|
try:
|
|
result = subprocess.run(
|
|
['nvidia-smi', '--query-gpu=temperature.gpu,power.draw',
|
|
'--format=csv,noheader,nounits'],
|
|
capture_output=True, text=True, timeout=2
|
|
)
|
|
if result.returncode == 0:
|
|
parts = result.stdout.strip().split(',')
|
|
return {
|
|
'temp_c': float(parts[0]),
|
|
'power_w': float(parts[1])
|
|
}
|
|
except:
|
|
pass
|
|
return {'temp_c': 45.0, 'power_w': 50.0} # Default
|
|
|
|
def calculate_dynamic_temperature(gpu_temp, asymmetry):
|
|
"""
|
|
Entropy Injection: Hotter GPU or higher asymmetry = higher model temperature
|
|
|
|
Base: 40°C GPU, A=0 → T=0.3 (conservative)
|
|
Hot: 70°C GPU, A=13 → T=1.2 (creative)
|
|
"""
|
|
# GPU temp component: 40°C → 0.0, 70°C → 0.6
|
|
temp_factor = max(0, min(1, (gpu_temp - 40) / 30)) * 0.6
|
|
|
|
# Asymmetry component: A=0 → 0.0, A=13 → 0.6
|
|
asym_factor = max(0, min(1, asymmetry / 13)) * 0.6
|
|
|
|
# Combined: base 0.3 + up to 0.9 additional
|
|
temperature = 0.3 + (temp_factor + asym_factor) / 2
|
|
|
|
return min(1.4, temperature) # Cap at 1.4
|
|
|
|
print("="*70)
|
|
print("ENTROPY INJECTION INFERENCE")
|
|
print("GPU Temp / Asymmetry → Model Temperature")
|
|
print("="*70)
|
|
|
|
# Load model
|
|
print("\n[Loading Kaelara v0.9...]")
|
|
model, tokenizer = FastLanguageModel.from_pretrained(
|
|
model_name="unsloth/llama-3.2-3b",
|
|
max_seq_length=512,
|
|
dtype=torch.bfloat16,
|
|
load_in_4bit=True,
|
|
)
|
|
|
|
model = FastLanguageModel.get_peft_model(
|
|
model,
|
|
r=64, target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
|
|
lora_alpha=128, lora_dropout=0, bias="none",
|
|
use_gradient_checkpointing="unsloth", random_state=3407,
|
|
)
|
|
|
|
from peft import PeftModel
|
|
model = PeftModel.from_pretrained(model, "./kaelara_v09_scientist/final")
|
|
print("✓ Model loaded")
|
|
|
|
# ZMQ
|
|
ctx = zmq.Context()
|
|
sub = ctx.socket(zmq.SUB)
|
|
sub.setsockopt_string(zmq.SUBSCRIBE, "")
|
|
sub.connect("tcp://127.0.0.1:5556")
|
|
print("✓ ZMQ connected")
|
|
|
|
print("\n" + "="*70)
|
|
print("ENTROPY INJECTION LOOP (5 cycles)")
|
|
print("="*70)
|
|
|
|
for cycle_num in range(5):
|
|
# Get LBM frame
|
|
frame = None
|
|
for i in range(50):
|
|
try:
|
|
msg = sub.recv(flags=zmq.NOBLOCK)
|
|
frame = json.loads(msg.decode('utf-8'))
|
|
break
|
|
except zmq.Again:
|
|
time.sleep(0.05)
|
|
|
|
if frame is None:
|
|
print(f"[{cycle_num}] No data")
|
|
continue
|
|
|
|
# Get GPU metrics
|
|
gpu = get_gpu_metrics()
|
|
|
|
# Calculate dynamic temperature
|
|
temp = calculate_dynamic_temperature(gpu['temp_c'], frame['asymmetry'])
|
|
|
|
print(f"\n[{cycle_num}] Cycle {frame['cycle']}")
|
|
print(f" GPU: {gpu['temp_c']:.1f}°C, {gpu['power_w']:.1f}W")
|
|
print(f" LBM: Asym={frame['asymmetry']:.2f}, Coh={frame['coherence']:.3f}")
|
|
print(f" → Model Temperature: {temp:.2f}")
|
|
|
|
# Generate with dynamic temperature
|
|
prompt = f"""Input: Asymmetry {frame['asymmetry']:.1f}, Coherence {frame['coherence']:.2f}, GPU {gpu['temp_c']:.1f}°C. Define the situation.
|
|
|
|
Output:"""
|
|
|
|
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
|
|
outputs = model.generate(
|
|
**inputs,
|
|
max_new_tokens=80,
|
|
temperature=temp,
|
|
do_sample=True,
|
|
top_p=0.9
|
|
)
|
|
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
|
|
|
# Extract just the output part
|
|
if "Output:" in response:
|
|
response = response.split("Output:")[-1].strip()
|
|
|
|
print(f" Response: {response[:100]}...")
|
|
|
|
print("\n" + "="*70)
|
|
print("ENTROPY INJECTION COMPLETE")
|
|
print("="*70)
|