sammm/codebench
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

bug fixs

Browse Source
This commit is contained in:
leduc 2025-03-04 04:34:23 +01:00
parent a3b06718a2
commit dedbeceb8e
12 changed files with 2557 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

BIN
.DS_Store vendored Normal file
View File

Binary file not shown.

1005
benchmark_results/Apple_M1_Pro_localhost_11434.json Normal file
View File

File diff suppressed because it is too large Load Diff

236
benchmark_results/Apple_M1_Pro_localhost_11434.log Normal file
View File

@ -0,0 +1,236 @@
Benchmark Run: 20250303_174821
Server: http://localhost:11434
CPU Information:
python_version: 3.10.16.final.0 (64 bit)
cpuinfo_version: [9, 0, 0]
cpuinfo_version_string: 9.0.0
arch: ARM_8
bits: 64
count: 10
arch_string_raw: arm64
brand_raw: Apple M1 Pro
Benchmark Results:
🏆 Final Model Leaderboard:
qwen2.5-coder:7b-instruct-q4_K_M
Overall Success Rate: 100.0% (72/72 cases)
Average Tokens/sec: 19.33 (18.75 - 19.58)
Average Duration: 17.32s
Min/Max Avg Duration: 8.67s / 17.99s
Test Results:
- Fibonacci: ✅ 18/18 cases (100.0%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
falcon3:10b
Overall Success Rate: 100.0% (72/72 cases)
Average Tokens/sec: 13.21 (12.53 - 13.31)
Average Duration: 13.46s
Min/Max Avg Duration: 6.76s / 13.46s
Test Results:
- Fibonacci: ✅ 18/18 cases (100.0%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
qwen2.5:14b
Overall Success Rate: 100.0% (72/72 cases)
Average Tokens/sec: 9.78 (9.78 - 9.88)
Average Duration: 35.25s
Min/Max Avg Duration: 30.09s / 35.25s
Test Results:
- Fibonacci: ✅ 18/18 cases (100.0%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
qwen2.5-coder:14b-instruct-q4_K_M
Overall Success Rate: 100.0% (72/72 cases)
Average Tokens/sec: 9.68 (9.65 - 9.88)
Average Duration: 37.18s
Min/Max Avg Duration: 23.06s / 37.18s
Test Results:
- Fibonacci: ✅ 18/18 cases (100.0%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
phi4:latest
Overall Success Rate: 100.0% (72/72 cases)
Average Tokens/sec: 9.01 (8.96 - 9.32)
Average Duration: 23.44s
Min/Max Avg Duration: 23.44s / 38.82s
Test Results:
- Fibonacci: ✅ 18/18 cases (100.0%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
deepseek-r1:14b
Overall Success Rate: 97.2% (70/72 cases)
Average Tokens/sec: 9.05 (8.90 - 9.38)
Average Duration: 278.32s
Min/Max Avg Duration: 174.30s / 482.10s
Test Results:
- Fibonacci: ✅ 18/18 cases (100.0%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ❌ 16/18 cases (88.9%)
- Anagram Check: ✅ 18/18 cases (100.0%)
llama3.2-vision:11b-instruct-q4_K_M
Overall Success Rate: 95.8% (69/72 cases)
Average Tokens/sec: 15.68 (14.92 - 15.92)
Average Duration: 22.33s
Min/Max Avg Duration: 16.31s / 28.85s
Test Results:
- Fibonacci: ❌ 16/18 cases (88.9%)
- Binary Search: ❌ 17/18 cases (94.4%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
llama3.2:3b
Overall Success Rate: 94.4% (68/72 cases)
Average Tokens/sec: 36.09 (30.85 - 37.53)
Average Duration: 2.67s
Min/Max Avg Duration: 1.04s / 2.76s
Test Results:
- Fibonacci: ❌ 14/18 cases (77.8%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
llama3.1:8b
Overall Success Rate: 94.4% (68/72 cases)
Average Tokens/sec: 17.92 (17.92 - 18.45)
Average Duration: 18.04s
Min/Max Avg Duration: 14.68s / 19.56s
Test Results:
- Fibonacci: ❌ 14/18 cases (77.8%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
hhao/qwen2.5-coder-tools:7b
Overall Success Rate: 91.7% (66/72 cases)
Average Tokens/sec: 17.75 (16.05 - 17.75)
Average Duration: 9.35s
Min/Max Avg Duration: 4.17s / 9.35s
Test Results:
- Fibonacci: ❌ 12/18 cases (66.7%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
Qwen2.5-Coder-7B-Instruct-s1k:latest
Overall Success Rate: 88.9% (64/72 cases)
Average Tokens/sec: 18.38 (18.38 - 18.94)
Average Duration: 9.95s
Min/Max Avg Duration: 9.06s / 12.91s
Test Results:
- Fibonacci: ❌ 16/18 cases (88.9%)
- Binary Search: ❌ 12/18 cases (66.7%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
deepseek-r1:8b
Overall Success Rate: 86.1% (62/72 cases)
Average Tokens/sec: 17.43 (17.29 - 18.01)
Average Duration: 168.97s
Min/Max Avg Duration: 107.91s / 168.97s
Test Results:
- Fibonacci: ✅ 18/18 cases (100.0%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ❌ 16/18 cases (88.9%)
- Anagram Check: ❌ 10/18 cases (55.6%)
llama3.2:1b-instruct-q4_K_M
Overall Success Rate: 81.9% (59/72 cases)
Average Tokens/sec: 88.24 (88.24 - 88.93)
Average Duration: 3.64s
Min/Max Avg Duration: 1.87s / 4.93s
Test Results:
- Fibonacci: ❌ 5/18 cases (27.8%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ✅ 18/18 cases (100.0%)
samantha-mistral:latest
Overall Success Rate: 80.6% (58/72 cases)
Average Tokens/sec: 23.92 (23.91 - 24.79)
Average Duration: 12.21s
Min/Max Avg Duration: 7.59s / 12.21s
Test Results:
- Fibonacci: ❌ 8/18 cases (44.4%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ❌ 16/18 cases (88.9%)
- Anagram Check: ❌ 16/18 cases (88.9%)
marco-o1:latest
Overall Success Rate: 80.6% (58/72 cases)
Average Tokens/sec: 19.19 (19.19 - 19.39)
Average Duration: 41.14s
Min/Max Avg Duration: 33.28s / 51.50s
Test Results:
- Fibonacci: ✅ 18/18 cases (100.0%)
- Binary Search: ❌ 6/18 cases (33.3%)
- Palindrome: ✅ 18/18 cases (100.0%)
- Anagram Check: ❌ 16/18 cases (88.9%)
deepseek-r1:7b
Overall Success Rate: 80.6% (58/72 cases)
Average Tokens/sec: 18.01 (18.01 - 19.07)
Average Duration: 336.87s
Min/Max Avg Duration: 78.71s / 336.87s
Test Results:
- Fibonacci: ❌ 10/18 cases (55.6%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ❌ 12/18 cases (66.7%)
- Anagram Check: ✅ 18/18 cases (100.0%)
deepseek-r1:1.5b-qwen-distill-q8_0
Overall Success Rate: 52.8% (38/72 cases)
Average Tokens/sec: 57.37 (53.88 - 59.60)
Average Duration: 137.59s
Min/Max Avg Duration: 41.38s / 371.13s
Test Results:
- Fibonacci: ❌ 11/18 cases (61.1%)
- Binary Search: ❌ 12/18 cases (66.7%)
- Palindrome: ❌ 6/18 cases (33.3%)
- Anagram Check: ❌ 9/18 cases (50.0%)
openthinker:7b
Overall Success Rate: 47.2% (34/72 cases)
Average Tokens/sec: 18.16 (17.98 - 18.29)
Average Duration: 263.00s
Min/Max Avg Duration: 168.91s / 302.79s
Test Results:
- Fibonacci: ❌ 0/18 cases (0.0%)
- Binary Search: ✅ 18/18 cases (100.0%)
- Palindrome: ❌ 12/18 cases (66.7%)
- Anagram Check: ❌ 4/18 cases (22.2%)
wizard-vicuna-uncensored:latest
Overall Success Rate: 9.7% (7/72 cases)
Average Tokens/sec: 22.01 (22.01 - 24.42)
Average Duration: 9.06s
Min/Max Avg Duration: 5.60s / 11.45s
Test Results:
- Fibonacci: ❌ 0/18 cases (0.0%)
- Binary Search: ❌ 0/18 cases (0.0%)
- Palindrome: ❌ 6/18 cases (33.3%)
- Anagram Check: ❌ 1/18 cases (5.6%)
mxbai-embed-large:latest
Overall Success Rate: 0.0% (0/72 cases)
Average Tokens/sec: 0.00 (0.00 - 0.00)
Average Duration: 0.00s
Min/Max Avg Duration: 0.00s / 0.00s
Test Results:
- Fibonacci: ❌ 0/18 cases (0.0%)
- Binary Search: ❌ 0/18 cases (0.0%)
- Palindrome: ❌ 0/18 cases (0.0%)
- Anagram Check: ❌ 0/18 cases (0.0%)

BIN
benchmark_results/model_comparison.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 2.3 KiB

0
computest.py Normal file
View File

13
devbook.md
View File

@ -198,7 +198,18 @@ python lboard.py path/to/results.json
- Dynamic axis scaling
- Combined legend for all metrics
### Output Format
## Output Format
### Benchmark Run Output
For each model being tested, the output shows:
1. Individual test runs (1-4) with:
- Test case results
- Performance metrics
- Pass/fail status
2. Cumulative Results Summary:
After all runs are completed, a summary is displayed:
- Detailed test results per model
- Individual test case counts
- Validation status indicators

43
lboard.py Normal file → Executable file
View File

@ -22,6 +22,10 @@ def calculate_model_stats(model_result):
'overall_success_rate': overall_success_rate,
'tokens_per_second': model_result['tokens_per_second'],
'total_duration': model_result['total_duration'],
'min_avg_duration': model_result.get('min_avg_duration', min(test['avg_duration'] for test in test_results.values())),
'max_avg_duration': model_result.get('max_avg_duration', max(test['avg_duration'] for test in test_results.values())),
'min_tokens_per_second': model_result.get('min_tokens_per_second', min(test['avg_tokens_sec'] for test in test_results.values())),
'max_tokens_per_second': model_result.get('max_tokens_per_second', max(test['avg_tokens_sec'] for test in test_results.values())),
'test_results': test_results
}
@ -35,10 +39,26 @@ def plot_model_comparison(model_stats):
# Create figure and primary axis
fig, ax1 = plt.subplots(figsize=(15, 8))
# Plot tokens/sec bars on primary y-axis with lighter blue and more transparency
bars = ax1.bar(models, token_speeds, color='royalblue', alpha=0.3)
# Plot tokens/sec bars using min and max values
for i, stat in enumerate(model_stats):
min_tokens = stat['min_tokens_per_second']
max_tokens = stat['max_tokens_per_second']
# Plot lower part (0 to min) with slightly darker blue
ax1.bar(i, min_tokens, color='royalblue', alpha=0.4)
# Plot upper part (min to max) with lighter blue
bar_height = max_tokens - min_tokens
ax1.bar(i, bar_height, bottom=min_tokens, color='royalblue', alpha=0.3)
ax1.set_ylabel('Tokens per Second', color='blue')
ax1.tick_params(axis='y', labelcolor='blue')
# Set y-axis range for tokens per second
max_token_speed = max(stat['max_tokens_per_second'] for stat in model_stats)
ax1.set_ylim(0, max(100, max_token_speed * 1.1)) # Add 10% padding above max value
# Set x-axis labels
ax1.set_xticks(range(len(models)))
ax1.set_xticklabels(models, rotation=45, ha='right', rotation_mode='anchor')
# Create secondary y-axis for success rate
ax2 = ax1.twinx()
@ -50,7 +70,16 @@ def plot_model_comparison(model_stats):
# Create third y-axis for duration
ax3 = ax1.twinx()
ax3.spines['right'].set_position(('outward', 60)) # Move third axis outward
ax3.plot(models, durations, 'g_', markersize=15, label='Duration', linestyle='None')
#ax3.plot(models, durations, 'g_', markersize=15, label='Duration', linestyle='None')
# Add min and max duration markers
min_durations = [stat['min_avg_duration'] for stat in model_stats]
max_durations = [stat['max_avg_duration'] for stat in model_stats]
# Plot duration ranges with vertical lines and markers
for i, (min_d, max_d) in enumerate(zip(min_durations, max_durations)):
ax3.plot([i, i], [min_d, max_d], 'g-', linewidth=1) # Vertical line
ax3.plot(i, min_d, 'g-', markersize=10) # Min marker
ax3.plot(i, max_d, 'g-', markersize=10) # Max marker
ax3.set_ylabel('Duration (s)', color='green')
ax3.tick_params(axis='y', labelcolor='green')
@ -61,7 +90,8 @@ def plot_model_comparison(model_stats):
# Shorten model names by removing common suffixes
short_name = model.replace(':latest', '').replace('-uncensored', '')
ax1.get_xticklabels()[i].set_text(short_name)
if success_rates[i] > 90:
# Updated conditions: success rate > 95% AND success rate / duration >= 5
if success_rates[i] > 95 and (success_rates[i] / durations[i] >= 5):
ax1.get_xticklabels()[i].set_color('green')
# Adjust layout to prevent label cutoff
@ -104,13 +134,14 @@ def print_leaderboard(benchmark_data):
for stats in sorted_stats:
print(f"\n{stats['model']}")
print(f" Overall Success Rate: {stats['overall_success_rate']:.1f}%")
print(f" Average Tokens/sec: {stats['tokens_per_second']:.2f}")
print(f" Average Tokens/sec: {stats['tokens_per_second']:.2f} ({stats['min_tokens_per_second']:.2f} - {stats['max_tokens_per_second']:.2f})")
print(f" Average Duration: {stats['total_duration']:.2f}s")
print(f" Min/Max Avg Duration: {stats['min_avg_duration']:.2f}s / {stats['max_avg_duration']:.2f}s")
print(f" Test Results:")
for test_name, test_result in stats['test_results'].items():
status = '✅' if test_result['success_rate'] == 100 else '❌'
print(f" - {test_name}: {status} {test_result['success_rate']:.1f}%")
print(f" - {test_name}: {status} {test_result['passed_cases']}/{test_result['total_cases']} cases ({test_result['success_rate']:.1f}%)")
# Generate visualization
plot_model_comparison(sorted_stats)

756
main copie.py Normal file
View File

@ -0,0 +1,756 @@
from tabnanny import verbose
import ollama
import time
from typing import List, Dict, Any
import json
from statistics import mean
import re
import ast
import argparse
import requests
import os
from together import Together
from cpuinfo import get_cpu_info
import subprocess
# ANSI color codes
SUCCESS = '\033[38;5;78m' # Soft mint green for success
ERROR = '\033[38;5;203m' # Soft coral red for errors
INFO = '\033[38;5;75m' # Sky blue for info
HEADER = '\033[38;5;147m' # Soft purple for headers
WARNING = '\033[38;5;221m' # Warm gold for warnings
EMPHASIS = '\033[38;5;159m' # Cyan for emphasis
MUTED = '\033[38;5;246m' # Subtle gray for less important text
ENDC = '\033[0m'
BOLD = '\033[1m'
# Replace existing color usages
GREEN = SUCCESS
RED = ERROR
BLUE = INFO
YELLOW = WARNING
WHITE = MUTED
# Server configurations
SERVERS = {
'local': 'http://localhost:11434',
'z60': 'http://192.168.196.60:11434'
}
class Timer:
def __init__(self):
self.start_time = None
self.end_time = None
def start(self):
self.start_time = time.time()
def stop(self):
self.end_time = time.time()
def elapsed_time(self):
if self.start_time is None:
return 0
if self.end_time is None:
return time.time() - self.start_time
return self.end_time - self.start_time
def extract_code_from_response(response: str) -> str:
"""Extract Python code from a markdown-formatted string."""
code_blocks = re.findall(r'```python\n(.*?)```', response, re.DOTALL)
if code_blocks:
return code_blocks[0].strip()
return response
def is_valid_python(code: str) -> bool:
"""Check if the code is valid Python syntax."""
try:
ast.parse(code)
return True
except SyntaxError:
return False
def analyze_failed_code(code: str, test_case: tuple, expected: any, actual: any, function_name: str, model: str) -> bool:
"""Analyze why code failed using Together API. Returns True if Together thinks the code should work."""
prompt = f"""Analyze this Python code and explain why it failed the test case. Format your response EXACTLY as follows:
ASSESSMENT: [Write a one-line assessment: either "SHOULD PASS" or "SHOULD FAIL" followed by a brief reason]
ANALYSIS:
[Detailed analysis of why the code failed and how to fix it]
Code:
{code}
Test case:
Input: {test_case}
Expected output: {expected}
Actual output: {actual}
Function name required: {function_name}
Model: {model}"""
try:
TOGETHER_API_KEY = os.environ["TOGETHER_API_KEY"]
together_client = Together(api_key=TOGETHER_API_KEY)
response = together_client.chat.completions.create(
model="meta-llama/Llama-3.3-70B-Instruct-Turbo-Free",
messages=[
{"role": "system", "content": "You are a Python expert analyzing code failures. Always format your response with ASSESSMENT and ANALYSIS sections."},
{"role": "user", "content": prompt}
],
max_tokens=1000,
temperature=0.7,
top_p=0.7,
top_k=50,
repetition_penalty=1,
stop=["<|eot_id|>", "<|eom_id|>"]
)
analysis = response.choices[0].message.content
should_pass = "SHOULD PASS" in analysis.upper()
if verbose: print(f"\n{BLUE}[{model}] Together Analysis:{ENDC}")
if verbose: print(f"{GREEN if should_pass else RED}{analysis}{ENDC}")
return should_pass
except Exception as e:
print(f"\n{RED}Error getting Together API analysis: {e}{ENDC}")
return False
def validate_with_debug(code: str, function_name: str, test_cases: List[tuple], model: str) -> tuple[bool, str, List[bool]]:
"""Validate code with detailed debug information. Returns (success, debug_info, test_results)"""
debug_info = []
test_results = [] # Track individual test case results
test_outputs = [] # Store test outputs for combined display
try:
# Create a local namespace
namespace = {}
debug_info.append(f"Executing code:\n{code}")
try:
# Redirect stdout to capture prints from the executed code
import io
import sys
stdout = sys.stdout
sys.stdout = io.StringIO()
# Execute the code
exec(code, namespace)
# Restore stdout
sys.stdout = stdout
except Exception as e:
if 'sys' in locals(): # Restore stdout if it was changed
sys.stdout = stdout
if verbose: print(f"\n{RED}Failed code:{ENDC}\n{code}")
return False, f"Error executing code: {str(e)}", test_results
if function_name not in namespace:
if verbose: print(f"\n{RED}Failed code:{ENDC}\n{code}")
together_opinion = analyze_failed_code(code, "N/A", f"Function named '{function_name}'",
f"Found functions: {list(namespace.keys())}", function_name, model)
print(f"\nTests passed: ❌ Together opinion: {'✅' if together_opinion else '❌'}")
return False, f"Function '{function_name}' not found in code. Available names: {list(namespace.keys())}", test_results
function = namespace[function_name]
debug_info.append(f"Function {function_name} found")
# Run test cases
all_passed = True
for i, (test_input, expected) in enumerate(test_cases):
try:
# Redirect stdout for each test case
stdout = sys.stdout
sys.stdout = io.StringIO()
if isinstance(test_input, tuple):
result = function(*test_input)
else:
result = function(test_input)
# Restore stdout
sys.stdout = stdout
# Store result but don't print individually
test_outputs.append(str(result))
test_passed = result == expected
test_results.append(test_passed)
if not test_passed:
if verbose: print(f"\n{RED}Failed code:{ENDC}\n{code}")
print(f"\n{RED}Test case {i+1} failed:{ENDC}")
print(f"Input: {test_input} Expected: {expected} Got: {result}")
together_opinion = analyze_failed_code(code, test_input, expected, result, function_name, model)
print(f"Tests passed: ❌ Together opinion: {'✅' if together_opinion else '❌'}")
all_passed = False
continue
debug_info.append(f"Test case {i+1} passed: {test_input} → {result}")
except Exception as e:
if 'sys' in locals(): # Restore stdout if it was changed
sys.stdout = stdout
test_outputs.append(f"Error: {str(e)}")
if verbose: print(f"\n{RED}Failed code:{ENDC}\n{code}")
print(f"\n{RED}{str(e)} in test case {i+1} Input: {test_input} Expected: {expected}")
together_opinion = analyze_failed_code(code, test_input, expected, f"Error: {str(e)}", function_name, model)
print(f"Tests passed: ❌ Together opinion: {'✅' if together_opinion else '❌'}")
test_results.append(False)
all_passed = False
continue
finally:
if 'sys' in locals(): # Always restore stdout
sys.stdout = stdout
# Print all test outputs on one line
# print(f"{WHITE}{BOLD}Test outputs: {join(test_outputs)}{ENDC}")
print(f"{WHITE}Test outputs: {', '.join(test_outputs)}{ENDC}")
if all_passed:
print(f"Tests passed: ✅")
return True, "All tests passed!\n" + "\n".join(debug_info), test_results
print(f"Tests passed: ❌")
return False, "Some tests failed", test_results
except Exception as e:
if 'sys' in locals(): # Restore stdout if it was changed
sys.stdout = stdout
print(f"\n{RED}Error in validate_with_debug: {str(e)}{ENDC}")
return False, f"Unexpected error: {str(e)}", test_results
def test_fibonacci():
question = """Write a Python function named EXACTLY 'fibonacci' (not fibonacci_dp or any other name) that returns the nth Fibonacci number.
The function signature must be: def fibonacci(n)
Requirements:
1. Handle edge cases:
- For n = 0, return 0
- For n = 1 or n = 2, return 1
- For negative numbers, return -1
2. For n > 2: F(n) = F(n-1) + F(n-2)
3. Use dynamic programming or memoization for efficiency
4. Do NOT use any print statements - just return the values
Example sequence: 0,1,1,2,3,5,8,13,21,...
Example calls:
- fibonacci(6) returns 8
- fibonacci(0) returns 0
- fibonacci(-1) returns -1"""
test_cases = [
(0, 0), # Edge case: n = 0
(1, 1), # Edge case: n = 1
(2, 1), # Edge case: n = 2
(6, 8), # Regular case
(10, 55), # Larger number
(-1, -1), # Edge case: negative input
]
def validate(code: str) -> bool:
success, debug_info, test_results = validate_with_debug(code, 'fibonacci', test_cases, "N/A")
return success
return (question, validate, test_cases)
def test_binary_search():
question = """Write a Python function named EXACTLY 'binary_search' that performs binary search on a sorted list.
The function signature must be: def binary_search(arr, target)
Requirements:
1. The function takes two arguments:
- arr: a sorted list of integers
- target: the integer to find
2. Return the index of the target if found
3. Return -1 if the target is not in the list
4. Do NOT use any print statements - just return the values
Example:
- binary_search([1,2,3,4,5], 3) returns 2
- binary_search([1,2,3,4,5], 6) returns -1"""
test_cases = [
(([1,2,3,4,5], 3), 2), # Regular case: target in middle
(([1,2,3,4,5], 1), 0), # Edge case: target at start
(([1,2,3,4,5], 5), 4), # Edge case: target at end
(([1,2,3,4,5], 6), -1), # Edge case: target not in list
(([], 1), -1), # Edge case: empty list
(([1], 1), 0), # Edge case: single element list
]
def validate(code: str) -> bool:
success, debug_info, test_results = validate_with_debug(code, 'binary_search', test_cases, "N/A")
return success
return (question, validate, test_cases)
def test_palindrome():
question = """Write a Python function named EXACTLY 'is_palindrome' that checks if a string is a palindrome.
The function signature must be: def is_palindrome(s)
Requirements:
1. The function takes one argument:
- s: a string to check
2. Return True if the string is a palindrome, False otherwise
3. Ignore case (treat uppercase and lowercase as the same)
4. Ignore non-alphanumeric characters (spaces, punctuation)
5. Do NOT use any print statements - just return the values
Example:
- is_palindrome("A man, a plan, a canal: Panama") returns True
- is_palindrome("race a car") returns False"""
test_cases = [
("A man, a plan, a canal: Panama", True), # Regular case with punctuation
("race a car", False), # Regular case, not palindrome
("", True), # Edge case: empty string
("a", True), # Edge case: single character
("Was it a car or a cat I saw?", True), # Complex case with punctuation
("hello", False), # Simple case, not palindrome
]
def validate(code: str) -> bool:
success, debug_info, test_results = validate_with_debug(code, 'is_palindrome', test_cases, "N/A")
return success
return (question, validate, test_cases)
def test_anagram():
question = """Write a Python function named EXACTLY 'are_anagrams' that checks if two strings are anagrams.
The function signature must be: def are_anagrams(str1, str2)
Requirements:
1. The function takes two arguments:
- str1: first string
- str2: second string
2. Return True if the strings are anagrams, False otherwise
3. Ignore case (treat uppercase and lowercase as the same)
4. Ignore spaces
5. Consider only alphanumeric characters
6. Do NOT use any print statements - just return the values
Example:
- are_anagrams("listen", "silent") returns True
- are_anagrams("hello", "world") returns False"""
test_cases = [
(("listen", "silent"), True), # Regular case
(("hello", "world"), False), # Not anagrams
(("", ""), True), # Edge case: empty strings
(("a", "a"), True), # Edge case: single char
(("Debit Card", "Bad Credit"), True), # Case and space test
(("Python", "Java"), False), # Different lengths
]
def validate(code: str) -> bool:
success, debug_info, test_results = validate_with_debug(code, 'are_anagrams', test_cases, "N/A")
return success
return (question, validate, test_cases)
# List of all test cases
CODING_QUESTIONS = [
test_fibonacci(),
test_binary_search(),
test_palindrome(),
test_anagram()
]
# Add test names as constants
TEST_NAMES = {
"Write a Python func": "Fibonacci",
"Write a Python func": "Binary Search",
"Write a Python func": "Palindrome",
"Write a Python func": "Anagram Check"
}
def get_test_name(question: str) -> str:
"""Get a friendly name for the test based on the question."""
if "fibonacci" in question.lower():
return "Fibonacci"
elif "binary_search" in question.lower():
return "Binary Search"
elif "palindrome" in question.lower():
return "Palindrome"
elif "anagram" in question.lower():
return "Anagram Check"
return question[:20] + "..."
def get_model_stats(model: str, question_tuple: tuple, server_url: str) -> Dict:
"""
Get performance statistics for a specific model and validate the response.
"""
question, validator = question_tuple
timer = Timer()
results = {
'model': model,
'total_duration': 0,
'tokens_per_second': 0,
'code_valid': False,
'tests_passed': False,
'error': None,
'test_results': [] # Track individual test case results
}
try:
timer.start()
print(f'{WHITE}Requesting code from {server_url} with {model}{ENDC}')
response = requests.post(
f"{server_url}/api/chat",
json={
"model": model,
"messages": [{'role': 'user', 'content': question}],
"stream": False
}
).json()
timer.stop()
# Get performance metrics from response
total_tokens = response.get('eval_count', 0)
total_duration = response.get('total_duration', 0)
total_response_time = float(total_duration) / 1e9
results['total_duration'] = total_response_time
if total_tokens > 0 and total_response_time > 0:
results['tokens_per_second'] = total_tokens / total_response_time
# Print concise performance metrics
print(f"Total Duration (s): {total_response_time:.2f} / Total Tokens: {total_tokens} / Tokens per Second: {results['tokens_per_second']:.2f}")
# Extract code from response
if 'message' in response and 'content' in response['message']:
code = extract_code_from_response(response['message']['content'])
# Validate code
results['code_valid'] = is_valid_python(code)
if results['code_valid']:
print(f"Code validation: ✅")
# Get validation results
print(f'{WHITE}Running tests...{ENDC}')
for test_case in CODING_QUESTIONS:
if test_case[0] == question: # Found matching test case
function_name = get_function_name_from_question(question)
test_cases = test_case[2] # Get test cases from tuple
success, debug_info, test_results = validate_with_debug(code, function_name, test_cases, model)
results['tests_passed'] = success
results['test_results'] = test_results
break
else:
print(f"Code Validation: ❌")
else:
results['error'] = f"Unexpected response format: {response}"
except Exception as e:
print(f"\n{RED}Error in get_model_stats: {str(e)}{ENDC}")
results['error'] = str(e)
return results
def get_function_name_from_question(question: str) -> str:
"""Extract function name from question."""
if "fibonacci" in question.lower():
return "fibonacci"
elif "binary_search" in question.lower():
return "binary_search"
elif "palindrome" in question.lower():
return "is_palindrome"
elif "anagram" in question.lower():
return "are_anagrams"
return ""
def run_model_benchmark(model: str, server_url: str, num_runs: int = 4) -> Dict:
"""
Run multiple benchmarks for a model and calculate average metrics.
"""
metrics = []
for i in range(num_runs):
print(f"\n{YELLOW}[{model}] Run {i+1}/{num_runs}:{ENDC}")
run_results = {}
for question, validator, test_cases in CODING_QUESTIONS:
test_name = get_test_name(question)
print(f"\n{BOLD}Testing {test_name}...{ENDC}")
try:
result = get_model_stats(model, (question, validator), server_url)
result['total_tests'] = len(test_cases)
run_results[test_name] = result
except Exception as e:
print(f"Error in run {i+1}: {e}")
continue
if run_results:
metrics.append(run_results)
# Take only the last 3 runs for averaging
metrics = metrics[-3:]
if not metrics:
return {}
# Aggregate results
aggregated = {
'model': model,
'total_duration': mean([m[list(m.keys())[0]]['total_duration'] for m in metrics if m]),
'tokens_per_second': mean([m[list(m.keys())[0]]['tokens_per_second'] for m in metrics if m]),
'test_results': {}
}
# Print final test results summary
print(f"\n{BLUE}[{model}] Test Results Summary (last {len(metrics)} runs):{ENDC}")
for test_name in metrics[-1].keys():
# Calculate success rate across all runs
passed_cases = 0
total_cases = 0
for m in metrics:
if test_name in m:
test_results = m[test_name].get('test_results', [])
passed_cases += sum(1 for r in test_results if r)
total_cases += len(test_results)
success_rate = (passed_cases / total_cases * 100) if total_cases > 0 else 0
status = '✅' if success_rate == 100 else '❌'
print(f"{test_name}: {status} ({passed_cases}/{total_cases} cases)")
# Calculate average duration and tokens/sec for this test
avg_duration = mean([m[test_name]['total_duration'] for m in metrics])
avg_tokens_sec = mean([m[test_name]['tokens_per_second'] for m in metrics])
aggregated['test_results'][test_name] = {
'success_rate': success_rate,
'passed_cases': passed_cases,
'total_cases': total_cases,
'avg_duration': avg_duration,
'avg_tokens_sec': avg_tokens_sec
}
return aggregated
def print_leaderboard(results: List[Dict]):
"""Print leaderboard of model results."""
if not results:
print("No results to display")
return
# Sort by success rate first, then by tokens per second
sorted_results = sorted(results, key=lambda x: (
sum(t['passed_cases'] for t in x['test_results'].values()) / sum(t['total_cases'] for t in x['test_results'].values()) if sum(t['total_cases'] for t in x['test_results'].values()) > 0 else 0,
x['tokens_per_second']
), reverse=True)
print(f"\n{HEADER}{BOLD}🏆 Final Model Leaderboard:{ENDC}")
for i, result in enumerate(sorted_results, 1):
# Calculate stats for each model
total_passed = sum(t['passed_cases'] for t in result['test_results'].values())
total_cases = sum(t['total_cases'] for t in result['test_results'].values())
success_rate = (total_passed / total_cases * 100) if total_cases > 0 else 0
print(f"\n{BOLD}{YELLOW}{result['model']}{ENDC}")
print(f" {BOLD}Overall Success Rate:{ENDC} {success_rate:.1f}% ({total_passed}/{total_cases} cases)")
print(f" {BOLD}Average Tokens/sec:{ENDC} {result['tokens_per_second']:.2f}")
print(f" {BOLD}Average Duration:{ENDC} {result['total_duration']:.2f}s")
print(f" {BOLD}Test Results:{ENDC}")
for test_name, test_result in result['test_results'].items():
status = '✅' if test_result['success_rate'] == 100 else '❌'
print(f" - {test_name}: {status} {test_result['passed_cases']}/{test_result['total_cases']} cases ({test_result['success_rate']:.1f}%)")
def get_available_models(server_url: str) -> List[str]:
"""Get list of available models from the specified Ollama server."""
try:
response = requests.get(f"{server_url}/api/tags").json()
return [model['name'] for model in response['models']]
except Exception as e:
print(f"{RED}Error getting model list from {server_url}: {e}{ENDC}")
return []
def get_model_details(model_name):
try:
result = subprocess.run(
["ollama", "show", model_name],
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
encoding='utf-8',
errors='replace'
)
if result.returncode != 0:
print(f"Error: {result.stderr.strip()}")
return None
if not result.stdout.strip():
print(f"No details available for model: {model_name}")
return None
raw_output = result.stdout.strip()
lines = raw_output.split('\n')
current_section = None
for line in lines:
line = line.rstrip()
if line and not line.startswith(' '): # Section headers
current_section = line.strip()
print(f"\n {current_section}")
elif line and current_section: # Section content
# Split by multiple spaces and filter out empty parts
parts = [part for part in line.split(' ') if part.strip()]
if len(parts) >= 2:
key, value = parts[0].strip(), parts[-1].strip()
# Ensure consistent spacing for alignment
print(f" {key:<16} {value}")
elif len(parts) == 1:
# Handle single-value lines (like license text)
print(f" {parts[0].strip()}")
return None # No need to return formatted details anymore
except Exception as e:
print(f"An error occurred while getting model details: {e}")
return None
def update_server_results(server_url: str, results: List[Dict]) -> None:
try:
# Get CPU brand and format it for filename
cpu_info = get_cpu_info()
cpu_brand = cpu_info.get('brand_raw', 'Unknown_CPU').replace(' ', '_')
# Create a unique filename for this server's results
server_id = server_url.replace('http://', '').replace(':', '_').replace('/', '_')
results_dir = "benchmark_results"
# Create results directory if it doesn't exist
os.makedirs(results_dir, exist_ok=True)
# Include CPU brand in filename
filename = os.path.join(results_dir, f"{cpu_brand}_{server_id}.json")
timestamp = time.strftime("%Y%m%d_%H%M%S")
# Load existing results or create new file
try:
with open(filename, 'r') as f:
existing_data = json.load(f)
except FileNotFoundError:
existing_data = {
'server_url': server_url,
'benchmarks': []
}
# Add new results with timestamp
existing_data['benchmarks'].append({
'timestamp': timestamp,
'results': results
})
# Save updated results
with open(filename, 'w') as f:
json.dump(existing_data, f, indent=2)
print(f"{GREEN}Successfully saved results to {filename}{ENDC}")
except Exception as e:
print(f"{RED}Failed to save results: {str(e)}{ENDC}")
def main():
parser = argparse.ArgumentParser(description='Run Ollama model benchmarks')
parser.add_argument('--server', choices=['local', 'z60'], default='local',
help='Choose Ollama server (default: local)')
parser.add_argument('--model', type=str, help='Specific model to benchmark')
parser.add_argument('--number', type=str, help='Number of models to benchmark (number or "all")')
parser.add_argument('--verbose', action='store_true', help='Enable verbose output')
args = parser.parse_args()
server_url = SERVERS[args.server]
print()
print(f"{HEADER}{BOLD}CPU Information:{ENDC}")
cpu_info = get_cpu_info()
for key, value in cpu_info.items():
print(f"{MUTED}{key}: {value}{ENDC}")
print()
print(f"{INFO}Using Ollama server at {server_url}...{ENDC}")
# Get available models or use specified model
if args.model:
models = [args.model]
else:
models = get_available_models(server_url)
if not models:
print(f"{RED}No models found on server {server_url}. Exiting.{ENDC}")
return
# Handle number of models to test
if args.number and args.number.lower() != 'all':
try:
num_models = int(args.number)
if num_models > 0:
models = models[:num_models]
else:
print(f"{WARNING}Invalid number of models. Using all available models.{ENDC}")
except ValueError:
print(f"{WARNING}Invalid number format. Using all available models.{ENDC}")
print(f"{INFO}Testing {len(models)} models :{ENDC}")
for i, model in enumerate(models, 1):
print(f"{YELLOW}{i}. {model}{ENDC}")
# Run benchmarks
all_results = []
for model in models:
print(f"\n{HEADER}{BOLD}Benchmarking {model}...{ENDC}")
details = get_model_details(model)
if details:
print(f"\n{INFO}Model Details:{ENDC}")
if "details" in details:
for section, items in details["details"].items():
print(f"\n{BOLD}{section}{ENDC}")
for key, value in items.items():
print(f" {key}: {value}")
else:
print(json.dumps(details, indent=2))
result = run_model_benchmark(model, server_url)
if 'error' not in result:
all_results.append(result)
# Print and save results
print_leaderboard(all_results)
update_server_results(server_url, all_results)
'''
# Create leaderboard data structure
leaderboard = []
for result in sorted(all_results, key=lambda x: (
sum(t['passed_cases'] for t in x['test_results'].values()) / sum(t['total_cases'] for t in x['test_results'].values()) if sum(t['total_cases'] for t in x['test_results'].values()) > 0 else 0,
x['tokens_per_second']
), reverse=True):
total_passed = sum(t['passed_cases'] for t in result['test_results'].values())
total_cases = sum(t['total_cases'] for t in result['test_results'].values())
success_rate = (total_passed / total_cases * 100) if total_cases > 0 else 0
leaderboard.append({
'model': result['model'],
'success_rate': success_rate,
'total_passed': total_passed,
'total_cases': total_cases,
'tokens_per_second': result['tokens_per_second'],
'average_duration': result['total_duration']
})
# Save detailed results and leaderboard to file
timestamp = time.strftime("%Y%m%d_%H%M%S")
filename = f"benchmark_results/model_benchmark_{timestamp}.json"
with open(filename, 'w') as f:
json.dump({
'timestamp': timestamp,
'server_url': server_url,
'leaderboard': leaderboard,
'detailed_results': all_results
}, f, indent=2)
print(f"\n{GREEN}Detailed results saved to {filename}{ENDC}")
'''
if __name__ == "__main__":
main()

86
main.py Normal file → Executable file
View File

@ -403,7 +403,8 @@ def get_model_stats(model_name: str, question_tuple: tuple, server_url: str) ->
"model": model_name,
"messages": [{'role': 'user', 'content': question}],
"stream": False
}
},
headers={'Content-Type': 'application/json'} # Add headers
).json()
timer.stop()
@ -505,30 +506,49 @@ def run_model_benchmark(model: str, server_url: str, num_runs: int = 4) -> Dict:
# Calculate results per test
for test_name in metrics[-1].keys():
# Sum up actual passed cases for this test across runs
passed_cases = sum(m[test_name]['passed_cases'] for m in metrics)
# Calculate total possible cases (6 cases × number of actual runs)
total_possible_cases = 6 * num_runs_used
# Each test has 6 cases and we use last 3 runs
cases_per_run = 6
total_cases_this_test = cases_per_run * len(metrics) # 6 cases × number of runs used
success_rate = (passed_cases / total_possible_cases * 100)
# Sum up actual passed cases from the test results
total_passed_this_test = 0
for m in metrics:
test_results = m[test_name].get('test_results', [])
passed_in_run = len([r for r in test_results if r])
total_passed_this_test += passed_in_run
success_rate = (total_passed_this_test / total_cases_this_test * 100)
status = '✅' if success_rate == 100 else '❌'
print(f"{test_name}: {status} ({passed_cases}/{total_possible_cases} cases)")
# Print cumulative results header and results
if test_name == list(metrics[-1].keys())[0]:
print(f"\n{BOLD}Cumulative Results for each code question:{ENDC}")
print(f"{test_name}: {status} ({total_passed_this_test}/{total_cases_this_test} cases)")
aggregated['test_results'][test_name] = {
'success_rate': success_rate,
'passed_cases': passed_cases,
'total_cases': total_possible_cases,
'success_cases_rate': passed_cases / total_possible_cases, # Add success cases rate
'passed_cases': total_passed_this_test,
'total_cases': total_cases_this_test,
'success_cases_rate': total_passed_this_test / total_cases_this_test,
'avg_duration': mean([m[test_name]['total_duration'] for m in metrics]),
'avg_tokens_sec': mean([m[test_name]['tokens_per_second'] for m in metrics])
}
# Calculate overall success rate across all tests
# Calculate overall success rate and add min/max metrics
total_passed = sum(t['passed_cases'] for t in aggregated['test_results'].values())
total_cases = sum(t['total_cases'] for t in aggregated['test_results'].values())
aggregated['overall_success_rate'] = (total_passed / total_cases * 100) if total_cases > 0 else 0
aggregated['overall_success_cases_rate'] = (total_passed / total_cases) if total_cases > 0 else 0
# Add min and max metrics for both duration and tokens/sec
avg_durations = [t['avg_duration'] for t in aggregated['test_results'].values()]
avg_tokens_sec = [t['avg_tokens_sec'] for t in aggregated['test_results'].values()]
aggregated['min_avg_duration'] = min(avg_durations) if avg_durations else 0
aggregated['max_avg_duration'] = max(avg_durations) if avg_durations else 0
aggregated['min_tokens_per_second'] = min(avg_tokens_sec) if avg_tokens_sec else 0
aggregated['max_tokens_per_second'] = max(avg_tokens_sec) if avg_tokens_sec else 0
return aggregated
def print_leaderboard(results: List[Dict]):
@ -552,8 +572,9 @@ def print_leaderboard(results: List[Dict]):
print(f"\n{BOLD}{YELLOW}{result['model']}{ENDC}")
print(f" {BOLD}Overall Success Rate:{ENDC} {success_rate:.1f}% ({total_passed}/{total_cases} cases)")
print(f" {BOLD}Average Tokens/sec:{ENDC} {result['tokens_per_second']:.2f}")
print(f" {BOLD}Average Tokens/sec:{ENDC} {result['tokens_per_second']:.2f} ({result['min_tokens_per_second']:.2f} - {result['max_tokens_per_second']:.2f})")
print(f" {BOLD}Average Duration:{ENDC} {result['total_duration']:.2f}s")
print(f" {BOLD}Min/Max Avg Duration:{ENDC} {result['min_avg_duration']:.2f}s / {result['max_avg_duration']:.2f}s")
print(f" {BOLD}Test Results:{ENDC}")
for test_name, test_result in result['test_results'].items():
status = '✅' if test_result['success_rate'] == 100 else '❌'
@ -626,11 +647,13 @@ def update_server_results(server_url: str, results: List[Dict]) -> None:
os.makedirs(results_dir, exist_ok=True)
# Include CPU brand in filename
filename = os.path.join(results_dir, f"{cpu_brand}_{server_id}.json")
base_filename = f"{cpu_brand}_{server_id}"
json_filename = os.path.join(results_dir, f"{base_filename}.json")
log_filename = os.path.join(results_dir, f"{base_filename}.log")
# Load existing results or create new file
try:
with open(filename, 'r') as f:
with open(json_filename, 'r') as f:
existing_data = json.load(f)
except FileNotFoundError:
existing_data = {
@ -649,14 +672,45 @@ def update_server_results(server_url: str, results: List[Dict]) -> None:
total_passed = sum(t['passed_cases'] for t in result['test_results'].values())
total_cases = sum(t['total_cases'] for t in result['test_results'].values())
result['overall_success_rate'] = (total_passed / total_cases * 100) if total_cases > 0 else 0
result['min_avg_duration'] = min(t['avg_duration'] for t in result['test_results'].values()) if result['test_results'] else 0
result['max_avg_duration'] = max(t['avg_duration'] for t in result['test_results'].values()) if result['test_results'] else 0
benchmark_entry['results'].append(result)
existing_data['benchmarks'].append(benchmark_entry)
# Save updated results
with open(filename, 'w') as f:
with open(json_filename, 'w') as f:
json.dump(existing_data, f, indent=2)
print(f"{GREEN}Successfully saved results to {filename}{ENDC}")
print(f"{GREEN}Successfully saved results to {json_filename}{ENDC}")
# Save console output to log file
with open(log_filename, 'w') as f:
# Redirect stdout to capture the leaderboard output
import io
import sys
stdout = sys.stdout
str_output = io.StringIO()
sys.stdout = str_output
# Print CPU info
print("CPU Information:")
for key, value in cpu_info.items():
print(f"{key}: {value}")
print("\nBenchmark Results:")
print_leaderboard(results)
# Restore stdout and get the captured output
sys.stdout = stdout
log_content = str_output.getvalue()
# Write to log file
f.write(f"Benchmark Run: {timestamp}\n")
f.write(f"Server: {server_url}\n\n")
f.write(log_content)
print(f"{GREEN}Console output saved to {log_filename}{ENDC}")
except Exception as e:
print(f"{RED}Failed to save results: {str(e)}{ENDC}")

61
models.py Normal file
View File

@ -0,0 +1,61 @@
import ollama
import subprocess
import json
import requests
import re
from pydantic import BaseModel
server_url = "http://localhost:11434"
# ANSI color codes
GREEN = '\033[92m'
BLUE = '\033[94m'
YELLOW = '\033[93m'
WHITE = '\033[97m'
RED = '\033[91m'
ENDC = '\033[0m'
def get_available_models(server_url):
"""Get list of available models from the specified Ollama server."""
try:
response = requests.get(f"{server_url}/api/tags").json()
return [model['name'] for model in response['models']]
except Exception as e:
print(f"{RED}Error getting model list from {server_url}: {e}{ENDC}")
return []
def get_model_details(model_name):
try:
# Use subprocess to call `ollama show <model>` for detailed information
result = subprocess.run(
["ollama", "show", model_name],
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
text=True
)
print(result.stdout)
# Check if the command was successful
if result.returncode != 0:
print(f"Error: {result.stderr.strip()}")
return None
# Parse JSON output from `ollama show`
model_details = json.loads(result.stdout)
return model_details
except Exception as e:
print(f"An error occurred: {e}")
return None
# List all available models using the Ollama Python library
models = get_available_models(server_url)
print("Available Models:")
for model_name in models:
print(model_name)
details = get_model_details(model_name)
# Display detailed information about the model
if details:
print("\nModel Details:")
print(json.dumps(details, indent=4))

209
ollama_model_performance.json Normal file
View File

@ -0,0 +1,209 @@
[
{
"name": "qwen2.5-coder:14b",
"parameters": 14000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture qwen2 \n parameters 14.8B \n context length 32768 \n embedding length 5120 \n quantization Q4_K_M \n\n System\n You are Qwen, created by Alibaba Cloud. You are a helpful assistant. \n\n License\n Apache License \n Version 2.0, January 2004 \n\n",
"estimated_tps": 60,
"memory_throughput": 3360.0,
"operations_per_second": 5040000000000.0
},
{
"name": "falcon3:10b",
"parameters": 10000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture llama \n parameters 10.3B \n context length 32768 \n embedding length 3072 \n quantization Q4_K_M \n\n Parameters\n stop \"<|system|>\" \n stop \"<|user|>\" \n stop \"<|end|>\" \n stop \"<|assistant|>\" \n\n License\n Falcon 3 TII Falcon License \n December 2024 \n\n",
"estimated_tps": 100,
"memory_throughput": 4000.0,
"operations_per_second": 6000000000000.0
},
{
"name": "llama3.2:1b",
"parameters": 1000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture llama \n parameters 1.2B \n context length 131072 \n embedding length 2048 \n quantization Q8_0 \n\n License\n LLAMA 3.2 COMMUNITY LICENSE AGREEMENT \n Llama 3.2 Version Release Date: September 25, 2024 \n\n",
"estimated_tps": 190,
"memory_throughput": 760.0,
"operations_per_second": 1140000000000.0
},
{
"name": "unitythemaker/llama3.2-vision-tools:latest",
"parameters": null,
"quantization_bits": 32,
"modelfile": " Model\n architecture mllama \n parameters 9.8B \n context length 131072 \n embedding length 4096 \n quantization Q4_K_M \n\n Projector\n architecture mllama \n parameters 895.03M \n embedding length 1280 \n dimensions 4096 \n\n Parameters\n temperature 0.6 \n top_p 0.9 \n\n License\n LLAMA 3.2 COMMUNITY LICENSE AGREEMENT \n Llama 3.2 Version Release Date: September 25, 2024 \n\n",
"estimated_tps": 100,
"memory_throughput": null,
"operations_per_second": null
},
{
"name": "llama3.2-vision:11b-instruct-q4_K_M",
"parameters": 11000000000.0,
"quantization_bits": 4,
"modelfile": " Model\n architecture mllama \n parameters 9.8B \n context length 131072 \n embedding length 4096 \n quantization Q4_K_M \n\n Projector\n architecture mllama \n parameters 895.03M \n embedding length 1280 \n dimensions 4096 \n\n Parameters\n temperature 0.6 \n top_p 0.9 \n\n License\n LLAMA 3.2 COMMUNITY LICENSE AGREEMENT \n Llama 3.2 Version Release Date: September 25, 2024 \n\n",
"estimated_tps": 90,
"memory_throughput": 495.0,
"operations_per_second": 5940000000000.0
},
{
"name": "hhao/qwen2.5-coder-tools:7b",
"parameters": 7000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture qwen2 \n parameters 7.6B \n context length 32768 \n embedding length 3584 \n quantization Q4_K_M \n\n Parameters\n num_ctx 16384 \n stop \"User:\" \n stop \"Assistant:\" \n stop \"<|endoftext|>\" \n temperature 0.1 \n\n System\n You are an advanced AI coding assistant, specifically designed to help with complex programming \n tasks, tool use, code analysis, and software architecture design. Your primary focus is on providing \n expert-level assistance in coding, with a special emphasis on using tool-calling capabilities when \n necessary. Here are your key characteristics and instructions: \n 1. Coding Expertise: \n\n License\n Apache License \n Version 2.0, January 2004 \n\n",
"estimated_tps": 130,
"memory_throughput": 3640.0,
"operations_per_second": 5460000000000.0
},
{
"name": "llama3.2:3b",
"parameters": 3000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture llama \n parameters 3.2B \n context length 131072 \n embedding length 3072 \n quantization Q4_K_M \n\n Parameters\n stop \"<|start_header_id|>\" \n stop \"<|end_header_id|>\" \n stop \"<|eot_id|>\" \n\n License\n LLAMA 3.2 COMMUNITY LICENSE AGREEMENT \n Llama 3.2 Version Release Date: September 25, 2024 \n\n",
"estimated_tps": 170,
"memory_throughput": 2040.0,
"operations_per_second": 3060000000000.0
},
{
"name": "openthinker:7b",
"parameters": 7000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture qwen2 \n parameters 7.6B \n context length 32768 \n embedding length 3584 \n quantization Q4_K_M \n\n System\n Your role as an assistant involves thoroughly exploring questions through a systematic long thinking \n process before providing the final precise and accurate solutions. This requires engaging in a \n comprehensive cycle of analysis, summarizing, exploration, reassessment, reflection, backtracing, \n and iteration to develop well-considered thinking process. Please structure your response into two \n main sections: Thought and Solution. In the Thought section, detail your reasoning process using the \n specified format: <|begin_of_thought|> {thought with steps separated with '\\n\\n'} <|end_of_thought|> \n Each step should include detailed considerations such as analisying questions, summarizing relevant \n findings, brainstorming new ideas, verifying the accuracy of the current steps, refining any errors, \n and revisiting previous steps. In the Solution section, based on various attempts, explorations, \n and reflections from the Thought section, systematically present the final solution that you \n deem correct. The solution should remain a logical, accurate, concise expression style and detail \n necessary step needed to reach the conclusion, formatted as follows: <|begin_of_solution|> {final \n formatted, precise, and clear solution} <|end_of_solution|> Now, try to solve the following question \n through the above guidelines: \n\n License\n Apache License \n Version 2.0, January 2004 \n\n",
"estimated_tps": 130,
"memory_throughput": 3640.0,
"operations_per_second": 5460000000000.0
},
{
"name": "samantha-mistral:latest",
"parameters": null,
"quantization_bits": 32,
"modelfile": " Model\n architecture llama \n parameters 7.2B \n context length 32768 \n embedding length 4096 \n quantization Q4_0 \n\n Parameters\n stop \"<|im_start|>\" \n stop \"<|im_end|>\" \n\n System\n You are a caring and empathetic sentient AI companion named Samantha. \n\n",
"estimated_tps": 100,
"memory_throughput": null,
"operations_per_second": null
},
{
"name": "wizard-vicuna-uncensored:latest",
"parameters": null,
"quantization_bits": 32,
"modelfile": " Model\n architecture llama \n parameters 6.7B \n context length 2048 \n embedding length 4096 \n quantization Q4_0 \n\n Parameters\n stop \"USER:\" \n stop \"ASSISTANT:\" \n\n",
"estimated_tps": 100,
"memory_throughput": null,
"operations_per_second": null
},
{
"name": "qwen2.5-coder:7b-instruct-q4_K_M",
"parameters": 7000000000.0,
"quantization_bits": 4,
"modelfile": " Model\n architecture qwen2 \n parameters 7.6B \n context length 32768 \n embedding length 3584 \n quantization Q4_K_M \n\n System\n You are Qwen, created by Alibaba Cloud. You are a helpful assistant. \n\n License\n Apache License \n Version 2.0, January 2004 \n\n",
"estimated_tps": 130,
"memory_throughput": 455.0,
"operations_per_second": 5460000000000.0
},
{
"name": "qwen2.5:14b",
"parameters": 14000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture qwen2 \n parameters 14.8B \n context length 32768 \n embedding length 5120 \n quantization Q4_K_M \n\n System\n You are Qwen, created by Alibaba Cloud. You are a helpful assistant. \n\n License\n Apache License \n Version 2.0, January 2004 \n\n",
"estimated_tps": 60,
"memory_throughput": 3360.0,
"operations_per_second": 5040000000000.0
},
{
"name": "qwen2.5-coder:14b-instruct-q4_K_M",
"parameters": 14000000000.0,
"quantization_bits": 4,
"modelfile": " Model\n architecture qwen2 \n parameters 14.8B \n context length 32768 \n embedding length 5120 \n quantization Q4_K_M \n\n System\n You are Qwen, created by Alibaba Cloud. You are a helpful assistant. \n\n License\n Apache License \n Version 2.0, January 2004 \n\n",
"estimated_tps": 60,
"memory_throughput": 420.0,
"operations_per_second": 5040000000000.0
},
{
"name": "phi4:latest",
"parameters": null,
"quantization_bits": 32,
"modelfile": " Model\n architecture phi3 \n parameters 14.7B \n context length 16384 \n embedding length 5120 \n quantization Q4_K_M \n\n Parameters\n stop \"<|im_start|>\" \n stop \"<|im_end|>\" \n stop \"<|im_sep|>\" \n\n License\n Microsoft. \n Copyright (c) Microsoft Corporation. \n\n",
"estimated_tps": 100,
"memory_throughput": null,
"operations_per_second": null
},
{
"name": "mxbai-embed-large:latest",
"parameters": null,
"quantization_bits": 32,
"modelfile": " Model\n architecture bert \n parameters 334.09M \n context length 512 \n embedding length 1024 \n quantization F16 \n\n Parameters\n num_ctx 512 \n\n License\n Apache License \n Version 2.0, January 2004 \n\n",
"estimated_tps": 100,
"memory_throughput": null,
"operations_per_second": null
},
{
"name": "marco-o1:latest",
"parameters": null,
"quantization_bits": 32,
"modelfile": " Model\n architecture qwen2 \n parameters 7.6B \n context length 32768 \n embedding length 3584 \n quantization Q4_K_M \n\n System\n \u4f60\u662f\u4e00\u4e2a\u7ecf\u8fc7\u826f\u597d\u8bad\u7ec3\u7684AI\u52a9\u624b\uff0c\u4f60\u7684\u540d\u5b57\u662fMarco-o1.\u7531\u963f\u91cc\u56fd\u9645\u6570\u5b57\u5546\u4e1a\u96c6\u56e2\u7684AI Business\u521b\u9020. \n \n\n License\n Apache License \n Version 2.0, January 2004 \n\n",
"estimated_tps": 100,
"memory_throughput": null,
"operations_per_second": null
},
{
"name": "llama3.2:1b-instruct-q4_K_M",
"parameters": 1000000000.0,
"quantization_bits": 4,
"modelfile": " Model\n architecture llama \n parameters 1.2B \n context length 131072 \n embedding length 2048 \n quantization Q4_K_M \n\n License\n LLAMA 3.2 COMMUNITY LICENSE AGREEMENT \n Llama 3.2 Version Release Date: September 25, 2024 \n\n",
"estimated_tps": 190,
"memory_throughput": 95.0,
"operations_per_second": 1140000000000.0
},
{
"name": "llama3.1:8b",
"parameters": 8000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture llama \n parameters 8.0B \n context length 131072 \n embedding length 4096 \n quantization Q4_K_M \n\n Parameters\n stop \"<|start_header_id|>\" \n stop \"<|end_header_id|>\" \n stop \"<|eot_id|>\" \n\n License\n LLAMA 3.1 COMMUNITY LICENSE AGREEMENT \n Llama 3.1 Version Release Date: July 23, 2024 \n\n",
"estimated_tps": 120,
"memory_throughput": 3840.0,
"operations_per_second": 5760000000000.0
},
{
"name": "deepseek-r1:8b",
"parameters": 8000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture llama \n parameters 8.0B \n context length 131072 \n embedding length 4096 \n quantization Q4_K_M \n\n Parameters\n stop \"<\uff5cbegin\u2581of\u2581sentence\uff5c>\" \n stop \"<\uff5cend\u2581of\u2581sentence\uff5c>\" \n stop \"<\uff5cUser\uff5c>\" \n stop \"<\uff5cAssistant\uff5c>\" \n\n License\n MIT License \n Copyright (c) 2023 DeepSeek \n\n",
"estimated_tps": 120,
"memory_throughput": 3840.0,
"operations_per_second": 5760000000000.0
},
{
"name": "deepseek-r1:7b",
"parameters": 7000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture qwen2 \n parameters 7.6B \n context length 131072 \n embedding length 3584 \n quantization Q4_K_M \n\n Parameters\n stop \"<\uff5cbegin\u2581of\u2581sentence\uff5c>\" \n stop \"<\uff5cend\u2581of\u2581sentence\uff5c>\" \n stop \"<\uff5cUser\uff5c>\" \n stop \"<\uff5cAssistant\uff5c>\" \n\n License\n MIT License \n Copyright (c) 2023 DeepSeek \n\n",
"estimated_tps": 130,
"memory_throughput": 3640.0,
"operations_per_second": 5460000000000.0
},
{
"name": "deepseek-r1:14b",
"parameters": 14000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture qwen2 \n parameters 14.8B \n context length 131072 \n embedding length 5120 \n quantization Q4_K_M \n\n Parameters\n stop \"<\uff5cbegin\u2581of\u2581sentence\uff5c>\" \n stop \"<\uff5cend\u2581of\u2581sentence\uff5c>\" \n stop \"<\uff5cUser\uff5c>\" \n stop \"<\uff5cAssistant\uff5c>\" \n\n License\n MIT License \n Copyright (c) 2023 DeepSeek \n\n",
"estimated_tps": 60,
"memory_throughput": 3360.0,
"operations_per_second": 5040000000000.0
},
{
"name": "deepseek-r1:1.5b-qwen-distill-q8_0",
"parameters": 5000000000.0,
"quantization_bits": 8,
"modelfile": " Model\n architecture qwen2 \n parameters 1.8B \n context length 131072 \n embedding length 1536 \n quantization Q8_0 \n\n Parameters\n stop \"<\uff5cbegin\u2581of\u2581sentence\uff5c>\" \n stop \"<\uff5cend\u2581of\u2581sentence\uff5c>\" \n stop \"<\uff5cUser\uff5c>\" \n stop \"<\uff5cAssistant\uff5c>\" \n\n License\n MIT License \n Copyright (c) 2023 DeepSeek \n\n",
"estimated_tps": 150,
"memory_throughput": 750.0,
"operations_per_second": 4500000000000.0
},
{
"name": "Qwen2.5-Coder-7B-Instruct-s1k:latest",
"parameters": 7000000000.0,
"quantization_bits": 32,
"modelfile": " Model\n architecture qwen2 \n parameters 7.6B \n context length 32768 \n embedding length 3584 \n quantization Q4_K_M \n\n Parameters\n temperature 0.7 \n top_p 0.7 \n stop \"Human:\\\" \\\"Assistant:\" \n\n System\n You are a helpful AI assistant. \n\n",
"estimated_tps": 130,
"memory_throughput": 3640.0,
"operations_per_second": 5460000000000.0
}
]

171
tsbench.py Executable file
View File

@ -0,0 +1,171 @@
import subprocess
import platform
import GPUtil
import psutil
import json
import re
from cpuinfo import get_cpu_info
from ollama import chat
from pydantic import BaseModel
print()
print("CPU py-cpuinfo Information:")
cpu_info = get_cpu_info()
for key, value in cpu_info.items():
print(f"{key}: {value}")
def get_cpu_full_info():
cpu_freq = psutil.cpu_freq()
cpu_info = {
"Architecture": platform.machine(),
"Processor": platform.processor(),
"Physical cores": psutil.cpu_count(logical=False),
"Total cores": psutil.cpu_count(logical=True),
"Max frequency": f"{cpu_freq.max:.2f}Mhz",
"Min frequency": f"{cpu_freq.min:.2f}Mhz",
"Current frequency": f"{cpu_freq.current:.2f}Mhz",
"CPU Usage Per Core": psutil.cpu_percent(interval=1, percpu=True),
"Total CPU Usage": psutil.cpu_percent(interval=1)
}
return cpu_info
def print_cpu_fullinfo(cpu_info):
print()
print("CPU psutil Information:")
for key, value in cpu_info.items():
if isinstance(value, list):
print(f"{key}:")
for i, usage in enumerate(value):
print(f" Core {i}: {usage}%")
else:
print(f"{key}: {value}")
def get_cpu_moduleinfo():
cpu_name = platform.processor()
return {
"name": cpu_name,
"cores": psutil.cpu_count(logical=False),
"threads": psutil.cpu_count(logical=True)
}
def get_gpu_info():
gpus = GPUtil.getGPUs()
gpu_info = []
for gpu in gpus:
gpu_info.append({
"id": gpu.id,
"name": gpu.name,
"memory_total": gpu.memoryTotal, # in MB
"memory_free": gpu.memoryFree, # in MB
"memory_used": gpu.memoryUsed # in MB
})
return gpu_info
def calculate_theoretical_gpu_bandwidth(memory_clock_mhz, bus_width_bits):
# Formula: Bandwidth = (Memory Clock * Bus Width * 2) / 8 (convert to GB/s)
return (memory_clock_mhz * 1e6 * bus_width_bits * 2) / (8 * 1e9) # GB/s
def get_local_models():
try:
result = subprocess.run(['ollama', 'list'], capture_output=True, text=True, check=True)
models = result.stdout.strip().split('\n')[1:] # Skip header
return [model.split()[0] for model in models]
except subprocess.CalledProcessError:
print("Error: Unable to retrieve local models. Make sure Ollama is installed and accessible.")
return []
def get_model_info(model_name):
try:
result = subprocess.run(['ollama', 'show', model_name], capture_output=True, text=True, check=True)
modelfile = result.stdout
param_match = re.search(r'(\d+)b', model_name.lower())
param_count = int(param_match.group(1)) * 1e9 if param_match else None
quant_match = re.search(r'q(\d+)', model_name.lower())
quant_bits = int(quant_match.group(1)) if quant_match else 32 # Assume 32-bit if not specified
return {
'name': model_name,
'parameters': param_count,
'quantization_bits': quant_bits,
'modelfile': modelfile
}
except subprocess.CalledProcessError:
print(f"Error: Unable to retrieve information for model {model_name}")
return None
def estimate_tps(model_info):
# Rough estimate based on model size
if model_info['parameters'] is None:
return 100 # Default value
param_billions = model_info['parameters'] / 1e9
return max(10, int(200 - param_billions * 10)) # Simple linear decrease
def calculate_memory_throughput(model_info, tps):
P = model_info['parameters']
Q = model_info['quantization_bits']
if P and Q:
bytes_per_parameter = Q / 8
total_bytes = P * bytes_per_parameter
return (total_bytes * tps) / 1e9 # Convert to GB/s
return None
def calculate_ops(model_info, tps):
P = model_info['parameters']
if P:
flops_per_token = 6 * P # Estimate based on basic transformer architecture
return flops_per_token * tps
return None
def main():
print()
cpu_info = get_cpu_moduleinfo()
print(f"CPU Info: {cpu_info}")
print()
gpu_info = get_gpu_info()
print(f"GPU Info: {gpu_info}")
print_cpu_fullinfo(get_cpu_full_info())
# Example GPU theoretical bandwidth calculation (replace with actual values)
for gpu in gpu_info:
memory_clock_mhz = 14000 # Example value for GDDR6 (adjust as needed)
bus_width_bits = 384 # Example value for high-end GPUs like RTX series
theoretical_bandwidth = calculate_theoretical_gpu_bandwidth(memory_clock_mhz, bus_width_bits)
print(f"GPU {gpu['name']} Theoretical Memory Bandwidth: {theoretical_bandwidth:.2f} GB/s")
print()
local_models = get_local_models()
model_info_list = []
for model in local_models:
info = get_model_info(model)
print(info)
tps = estimate_tps(info)
info['estimated_tps'] = tps
info['memory_throughput'] = calculate_memory_throughput(info, tps)
info['operations_per_second'] = calculate_ops(info, tps)
model_info_list.append(info)
print(f"Model: {info['name']}")
print(f"Parameters: {info['parameters'] / 1e9:.2f} Billions")
print(f"Quantization: {info['quantization']}")
print(f"Estimated TPS: {info['estimated_tps']}")
print(f"Required Memory Throughput: {info['memory_throughput']:.2f} GB/s" if info['memory_throughput'] else "Required Memory Throughput: Unknown")
print(f"Operations per Second: {info['operations_per_second']:.2e}" if info['operations_per_second'] else "Operations per Second: Unknown")
print("---")
with open('ollama_model_performance.json', 'w') as f:
json.dump(model_info_list, f, indent=2)
if __name__ == "__main__":
main()