Data Generation of random negative number throws error occasionally

vvagias commented 2 days ago

just wanted to point this out. Occasionally you get the following error when using inference benchmark :

2024-09-17 07:56 INFO     User selected random dataset. Generating prompt and output lengths from distributions
Traceback (most recent call last):
  File "/Users/vv/Documents/code_local/flexible-inference-bench-main/flex-env/bin/inference-benchmark", line 8, in <module>
    sys.exit(main())
  File "/Users/vv/Documents/code_local/flexible-inference-bench-main/flex-env/lib/python3.10/site-packages/flexible_inference_benchmark/main.py", line 240, in main
    requests_prompts = generate_prompts(args, size)
  File "/Users/vv/Documents/code_local/flexible-inference-bench-main/flex-env/lib/python3.10/site-packages/flexible_inference_benchmark/main.py", line 88, in generate_prompts
    data = prompt_cls.generate_data(size_adjusted)
  File "/Users/vv/Documents/code_local/flexible-inference-bench-main/flex-env/lib/python3.10/site-packages/flexible_inference_benchmark/engine/data.py", line 193, in generate_data
    data = list(self.token_distribution.generate_distribution(lengths[i] + self.num_trials))
  File "/Users/vv/Documents/code_local/flexible-inference-bench-main/flex-env/lib/python3.10/site-packages/flexible_inference_benchmark/engine/distributions.py", line 47, in generate_distribution
    return [int(elem) for elem in np.random.randint(self.low, self.high, size)]
  File "numpy/random/mtrand.pyx", line 798, in numpy.random.mtrand.RandomState.randint
  File "numpy/random/_bounded_integers.pyx", line 1343, in numpy.random._bounded_integers._rand_int64
ValueError: negative dimensions are not allowed

It seems to be in the random generation creating a negative number some times and truly is random because just running the command again will work perfect. Not sure if this is a bug or what. Just wanted to bring it up.

johncalesp commented 2 days ago

can you provide the config file that you use? seems like a normal distribution with high variance

vvagias commented 1 day ago

{
    "backend": "cserve-debug",
    "base_url": "http://a100-llama8b.user-1404.gcp.centml.org",
    "endpoint": "/cserve/v1/generate",
    "dataset_name": "random",
    "input_token_distribution": ["normal", 20, 10],
    "output_token_distribution": ["uniform", 200, 201],
    "request_distribution": ["poisson", 5],
    "num_of_req": 100,
    "model": "meta-llama/Meta-Llama-3-8B-Instruct",
    "tokenizer": "meta-llama/Meta-Llama-3-8B-Instruct",
    "https_ssl": false,
    "no_prefix": true
}

johncalesp commented 1 day ago

I see, the input token has normal dist with mean 20 and std 10, so it goes between 10 and 30, at the lower end there may be some issues, will take a look later, for now maybe try with lower std values like 2 or 5

CentML / flexible-inference-bench

Data Generation of random negative number throws error occasionally #64