Prompt eval time is counted twice

[sw]

Creating a new issue so this doesn't get forgotten:

@KASR posted a CSV of processing times in #603 (comment)

But the times don't add up: If you take the total time, and subtract the partial times that are supposed to add up to it, the result is all over the place:

The clue lies in the comment by @ggerganov :

@sw
After the mmap changes, the load time is incorrect:

https://github.com/ggerganov/llama.cpp/blob/6e7801d08d81c931a5427bae46f00763e993f54a/llama.cpp#L1681-L1685

Currently, the reported load time includes not only the page faults, but also the prompt eval time. So effectively, you get the negative number since the prompt eval time has been accounted 2 times.
We have to fix this.

Originally posted by @ggerganov in #603 (comment)

[goerch]

The timing computations looks correct for me, tested with ggml (main branch). Can this issue be closed then?

[ggerganov]

It's still broken in llama.cpp because we do the following:

https://github.com/ggerganov/llama.cpp/blob/0db14fef06836caaa13cc123c0a24dc598bdb9f0/llama.cpp#L3504-L3525

The reason we do it like this is because when using mmap, the loading of the data happens as we need it. I.e. it will happen during the first llama_eval_internal() call. But this call also involves computations which should not be counted towards the "load" time.

You don't see this in ggml because it does not have mmap support there and this effect cannot be observed.

[goerch]

The reason we do it like this is because when using mmap, the loading of the data happens as we need it.

I'm not sure this is the only problem. If I understand

std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(const gpt_params & params) {
    auto lparams = llama_context_params_from_gpt_params(params);

    llama_model * model  = llama_load_model_from_file(params.model.c_str(), lparams);
    if (model == NULL) {
        fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
        return std::make_tuple(nullptr, nullptr);
    }

    llama_context * lctx = llama_new_context_with_model(model, lparams);
    if (lctx == NULL) {
        fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
        llama_free_model(model);
        return std::make_tuple(nullptr, nullptr);
    }

    ...

    return std::make_tuple(model, lctx);
}

correctly, model loading is decoupled from context creation here and we cant' access the timings from this function even when not using mmap?

A simple workaround could be to remove the offending code in llama_eval and compute the load time in llama_print_timings as total_time - eval_time - prompt_eval_time - sample_time?

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.