FYI if you want a smoother/more compact progress bar you can use fractional blocks. Example implementation for training:

@JohannesGaessler was thinking docker-style because of portability, do you know if this works on Windows?

Hey :)

I'm currently in the middle of a larger rework of this exact section, with the goal of removing the cURL dependency completely. You can see the work in progress here: #16185

And also made a very simple progressbar 😆

static void print_progress(size_t current, size_t total) { // TODO isatty
    if (!total) {
        return;
    }

    size_t width = 50;
    size_t pct = (100 * current) / total;
    size_t pos = (width * current) / total;

    std::cout << "["
              << std::string(pos, '=')
              << (pos < width ? ">" : "")
              << std::string(width - pos, ' ')
              << "] " << std::setw(3) << pct << "%  ("
              << current / (1024 * 1024) << " MB / "
              << total / (1024 * 1024) << " MB)\r";
    std::cout.flush();
}

To avoid us getting tangled in merge conflicts, how about we get my refactor merged first to create a clean base, and then we can circle back to integrate the improvements from your patch?

Let me know what you think

#16185

Let me know when the PR is ready for review, would like to test and review.

I'll take time to review in the next few days. It's quite important as multi-thread downloading can significantly improve download speed.

In the meantime, @ericcurtin could you confirm if the user-reported bugs from original PR are resolved?

I'm also thinking of a method to test this, maybe extend server test and hide this new test behind a flag, so it doesn't run automatically on CI.

I think for simplification we can adopt this style:

[ #####                   ] 30%

It's just plain ASCII so should be ok on all systems. Other style can be implemented in another PR.

I'll take time to review in the next few days. It's quite important as multi-thread downloading can significantly improve download speed.

Yeah, I kinda abandoned this for now until @angt was happy. Yeah, it can improve performance quite a bit I think @xenoscopic saw like a 25% performance increase with similar code for DockerHub.

A lot of people see zero improvement also, but I guess it depends on your networking setup, you can simply be throttled by bandwidth. I'm my apartment I see zero improvement.

In the meantime, @ericcurtin could you confirm if the user-reported bugs from original PR are resolved?

I think I caught everything except for maybe, the file closing one before rename, don't remember addressing that one.

I'm also thinking of a method to test this, maybe extend server test and hide this new test behind a flag, so it doesn't run automatically on CI.

Sounds like a great idea. Or just use llama-pull to test pull in an isolated way.