Cutting inference cold starts by 40x with LP, FUSE, C/R, and CUDA-checkpoint(modal.com)
58 pointsby charles_irl5 hours ago5 comments
  • a_t486 minutes ago
    charles, amit, can you go into more about the path based caching? Particularly "shared bytes aren’t guaranteed to be in the exact same container image layer"? I've built something that solves issues around sharing data between layers, and am interested to see if it fits usecases like Modal's.
  • aidanhs2 hours ago
    I've recently been going down the rabbit hole of creating a "fast start dev env" and it's interesting to see how this article differs from other approaches (codesandbox has some fantastic blogs, the fly.io blog on sprites has interesting pointers, e2b and daytona are related open source tools). Everyone has a different solution based on their tradeoffs.

    I thought the memory snapshotting part in particular was clever since most container based systems don't bother (VM/firecracker based ones can use UFFD and call it a day), but by having emulated syscalls you can actually do single-process restore pretty well.

    I am a bit dubious of the use of fuse (though it clearly works well!), and I wonder if ublk (what I ended up using) might alleviate some of the pain/magic in fuse tuning. I'd personally also be looking at forking gvisor to take a memfd which you enable UFFD on for the page loading (I have some firecracker patches where I do the same). It's nice because you can optimistically push pages, rather than waiting for the requests to come in. The series of three codesandbox blog posts are good background reading.

    • amitprasad2 hours ago
      Our system somewhat predates ublk; at the time we wrote it, FUSE was the most reasonable option. Moving over to ublk would require re-architecting around block devices rather than filesystems, but is indeed something we're looking at!
      • aidanhs20 minutes ago
        Fair enough, would love to see another writeup on the performance you observe even if it fails - in practice numbers are hard to come by.
  • sluongng2 hours ago
    There are plenty of cool advancements in reducing inference cold start when I was meeting with folks in person at FOSDEM this year. However, I still struggle to understand: why would folks care about this?

    Major AI Labs all have secured their own compute in the form of hardware, data center, and power generation. That means their resource pool is fixed, and they can do all sorts of tricks to pre-load, pre-allocate, etc... to improve on inference latency.

    Cold start is usually a solution for "cloud" environment when your pool is flexible, and you only pay for what you use. Its effectiveness lowered in bare-metal settings as folks do not care about scaling up and down as much.

    So my question is: who is this for? AWS and GCP running Anthropic models?

    • artisin21 minutes ago
      At least folks like me care about it. My local hardware is more than enough to handle my app, but given Spectrum's internet service is as fickle as a broken fiddle I'm forced to rent a dedicated cloud gpu that sits idle most days. However, I would save a serious chunk of change if I could boot up a GPU snapshot in ~10s. I evaluated various options a while back and, while modal.com was the fastest, it still took around a minute-ish. Granted, my use case is unique, but I imagine this could be a decent solution for gpu-poor ComfyUI users.
  • kgeist3 hours ago
    So, does this snapshotting optimization support arbitrary containers?

    I'm currently planning to deploy using Amazon SageMaker, but a cold start takes a whopping ~9 minutes: 6 minutes for instance provisioning + 3 minutes for PyTorch initialization. My Docker image is ~14 GB, and the weights are a few GB. How long would it take to cold start this configuration on Modal?

    SageMaker's performance makes it pretty much useless without many warm instances around (= tens of thousands of dollars per month), because users won't be happy if they have to randomly wait 9 minutes

    • charles_irl3 hours ago
      Yep! That should start in ten seconds or so -- about a second per gigabyte of weights, plus a second to start the container and a few seconds to load the memory snapshot.

      There are a few limitations with snapshotting, e.g. it generally fails when using multiple GPUs, which we document here: https://modal.com/docs/guide/memory-snapshots.

  • iLoveOncall4 hours ago
    What is "cutting by 40x" supposed to mean?
    • charles_irl4 hours ago
      Cutting latencies by 40x! Unfortunately couldn't fit the whole title in the character limit :<
      • aaronblohowiak3 hours ago
        How can you cut latency by more than 1x? I am no intending to be snarky, it just doesn’t fit my brain how you can reduce a measure time by more than the original starting time.
        • aaronblohowiak3 hours ago
          Put differently, 1/40 is not the same as 1x - 40x. I’d phrase as Reduced by 97.5% or 0.975x
        • bfeynman3 hours ago
          probably just AI slop and using wrong semantics, they mean speedup ratio.