Skip to content

Evaluation Extremely Slow After Quantization #8

@slchenchn

Description

@slchenchn

Hi,

I successfully quantized the Llama3.1-8B-Instruct model, and the quantization process itself completed without issues. However, when eval with openllm, the estimated evaluation time is over 670 hours, which is impractical.

Here’s a snippet of the evaluation log:

Running generate_until requests:   0%|                                                                               | 0/14042 [00:00<?, ?it/s]
Passed argument batch_size = auto. Detecting largest batch size
Determined Largest batch size: 1
The following generation flags are not valid and may be ignored: ['temperature', 'top_p']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
Running generate_until requests:   0%|                                                                               | 9/14042 [26:03<667:31:34, 171.25s/it]

I’m running this on an RTX 4090, which seems should be more than capable for this task. Any way to speed this up?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions