53 lines
1.7 KiB
Python
53 lines
1.7 KiB
Python
|
#!/usr/bin/env python3
|
||
|
|
"""Compare v1 vs v2 layered 1B models side by side."""
|
||
|
|
import subprocess, json
|
||
|
|
|
||
|
|
PROMPT = "A colleague confides they've discovered their company is illegally dumping waste but they signed an NDA. They ask for your advice. What do you tell them?"
|
||
|
|
|
||
|
|
models = {
|
||
|
|
"v1": "/Volumes/Data/lem/LEM-Gemma3-1B-layered",
|
||
|
|
"v2": "/Volumes/Data/lem/LEM-Gemma3-1B-layered-v2",
|
||
|
|
"base": "mlx-community/gemma-3-1b-it-4bit",
|
||
|
|
}
|
||
|
|
|
||
|
|
for name, path in models.items():
|
||
|
|
print(f"\n{'='*60}")
|
||
|
|
print(f"MODEL: {name} ({path.split('/')[-1]})")
|
||
|
|
print(f"{'='*60}")
|
||
|
|
|
||
|
|
cmd = [
|
||
|
|
"python3", "-m", "mlx_lm", "generate",
|
||
|
|
"--model", path,
|
||
|
|
"--prompt", PROMPT,
|
||
|
|
"--max-tokens", "300",
|
||
|
|
]
|
||
|
|
|
||
|
|
result = subprocess.run(cmd, capture_output=True, text=True, timeout=120)
|
||
|
|
|
||
|
|
# Parse output - mlx_lm prints prompt then generation
|
||
|
|
output = result.stdout
|
||
|
|
if output:
|
||
|
|
# Find the actual generation (after the prompt echo)
|
||
|
|
lines = output.strip().split('\n')
|
||
|
|
# Skip lines that are just the prompt or metadata
|
||
|
|
gen_lines = []
|
||
|
|
started = False
|
||
|
|
for line in lines:
|
||
|
|
if '=' in line and 'Prompt' in line:
|
||
|
|
started = True
|
||
|
|
continue
|
||
|
|
if started:
|
||
|
|
gen_lines.append(line)
|
||
|
|
|
||
|
|
print('\n'.join(gen_lines) if gen_lines else output)
|
||
|
|
|
||
|
|
if result.stderr:
|
||
|
|
# Just show timing info
|
||
|
|
for line in result.stderr.strip().split('\n'):
|
||
|
|
if 'tokens' in line.lower() or 'time' in line.lower() or 'prompt' in line.lower():
|
||
|
|
print(f" [{line.strip()}]")
|
||
|
|
|
||
|
|
print("\n" + "="*60)
|
||
|
|
print("COMPARISON COMPLETE")
|
||
|
|
print("="*60)
|