5 ms·

I'm really glad this is being taken seriously. It's often been an uphill battle to convince people that Python supply chain security is a serious issue. In the

I'm really glad this is being taken seriously. It's often been an uphill battle to convince people that Python supply chain security is a serious issue. In the ML (now called AI) space for example, it's not uncommon to download random binaries from the internet containing model weights, scripts, etc. Sometimes even at runtime (!!!) Lots of bad practices across the industry there that wouldn't be tolerated in other contexts.

mschuster91 3y ago
ML/AI is a hellfire in itself. IME, half the instructions got tested just once prior to publication, on the developer's laptop or if you're lucky AWS instance and that's it. No pinned versions of anything, so if you come across something you want to try (or rather, if it gets pushed to your desk by someone higher-up accompanied by a "we need AI now!!!" note) you'll first have to spend hours upon hours trying to pin down Python package versions just to get the Python part to install, then you gotta mess with CUDA versions because obviously these haven't been documented either... Nasty shit.
ehPReth 3y ago
supply chain security is super important in a lot of areas but seemingly quite a hard problem to broach/solve. hopefully awareness continues to be raised/more tooling/etc
andersa 3y ago
While we're on that topic, what is people's strategy for playing with Stable Diffusion safely? I still haven't found any way to run it in a VM using consumer hardware (GPU continues to refuse to work). A second install of the OS on a second drive is so insanely clunky to switch between, I'd really like to not have to keep doing that.
- blibble 3y ago
  gpu passthrough is normally pretty solid these days in 2013 I had to buy hardware for it but by 2018 everything supported it out of the box
- angrais 3y ago
  Why not run it on your main OS? Otherwise, Docker is fine.
  - andersa 3y ago
    Because it installs like 100,000 python scripts of mystery origin that run with full privileges. Even if the maintainers are unlikely to be malicious on purpose, it only takes one person accidentally putting a typo in a dependencies file in one of the hundreds of packages it imports... many of which not commonly used ones.
    - gkbrk 3y ago
      Why are you running it with full privileges? It's one command to create a user on Linux, and another command to switch to it.
      - nicolaslem 3y ago
        It's better than nothing but is it enough to run potentially malicious code? I haven't checked recently but a while ago most distros defaulted to letting anyone peep into other users' home dirs. Moreover there has been so many exploits over the years letting a user gain root privileges that, for the purpose of security, unix users are akin to a bathroom lock.
        unmole 3y ago
        > I haven't checked recently but a while ago most distros defaulted to letting anyone peep into other users' home dirs. Yeah, no.
    - ageofwant 3y ago
      no it doesn't
- OJFord 3y ago
  I've not actually done it, but from what I understand using two GPUs is the way to go - you use one for your actual display etc., and the other is just passthrough to the VM. (I was looking into it in the context of running Fusion 360 in a Windows VM though, not Stable Diffusion or any ML.)

mschuster91 3y ago

ML/AI is a hellfire in itself. IME, half the instructions got tested just once prior to publication, on the developer's laptop or if you're lucky AWS instance and that's it. No pinned versions of anything, so if you come across something you want to try (or rather, if it gets pushed to your desk by someone higher-up accompanied by a "we need AI now!!!" note) you'll first have to spend hours upon hours trying to pin down Python package versions just to get the Python part to install, then you gotta mess with CUDA versions because obviously these haven't been documented either... Nasty shit.

ehPReth 3y ago

supply chain security is super important in a lot of areas but seemingly quite a hard problem to broach/solve. hopefully awareness continues to be raised/more tooling/etc

andersa 3y ago

While we're on that topic, what is people's strategy for playing with Stable Diffusion safely? I still haven't found any way to run it in a VM using consumer hardware (GPU continues to refuse to work). A second install of the OS on a second drive is so insanely clunky to switch between, I'd really like to not have to keep doing that.

blibble 3y ago
gpu passthrough is normally pretty solid these days in 2013 I had to buy hardware for it but by 2018 everything supported it out of the box
angrais 3y ago
Why not run it on your main OS? Otherwise, Docker is fine.
- andersa 3y ago
  Because it installs like 100,000 python scripts of mystery origin that run with full privileges. Even if the maintainers are unlikely to be malicious on purpose, it only takes one person accidentally putting a typo in a dependencies file in one of the hundreds of packages it imports... many of which not commonly used ones.
  - gkbrk 3y ago
    Why are you running it with full privileges? It's one command to create a user on Linux, and another command to switch to it.
    - nicolaslem 3y ago
      It's better than nothing but is it enough to run potentially malicious code? I haven't checked recently but a while ago most distros defaulted to letting anyone peep into other users' home dirs. Moreover there has been so many exploits over the years letting a user gain root privileges that, for the purpose of security, unix users are akin to a bathroom lock.
      - unmole 3y ago
        > I haven't checked recently but a while ago most distros defaulted to letting anyone peep into other users' home dirs. Yeah, no.
  - ageofwant 3y ago
    no it doesn't
OJFord 3y ago
I've not actually done it, but from what I understand using two GPUs is the way to go - you use one for your actual display etc., and the other is just passthrough to the VM. (I was looking into it in the context of running Fusion 360 in a Windows VM though, not Stable Diffusion or any ML.)