Honestly, Heroku is basically a toy once you start loading heavy weights like PyTorch or Transformers into memory. Those slug limits exist because they are trying to pack too many containers onto one node. Its frustrating when you spend more time shrinking your requirements file than actually coding. Unfortunately, most managed platforms like Render or Railway have similar hidden walls that make them not as good as expected for data science projects. If you want to stay under that $15 mark, you really need to look at raw VPS providers where you control the swap space and disk partition. Most PaaS solutions will kill your process the moment it hits 512MB or 1GB of RAM, which happens instantly with most AI models.

• Vultr Cloud Compute 2GB RAM 55GB SSD is usually 12 dollars and gives you enough headroom to run a small model if you optimize your Docker image.
• Linode Shared CPU 2GB RAM 50GB Storage is another solid choice at 12 dollars, plus their technical documentation is actually useful for server hardening.
• AWS EC2 t3.small 2GB RAM 20GB EBS fits right in your budget and gives you access to the whole ecosystem if you need to scale later. The trick is using Docker. It bypasses the slug concept entirely and lets you pre-install dependencies into a static image. Just make sure you arent pulling the full CUDA versions of libraries if youre running on a CPU-only instance. That will eat up 4GB of disk space for no reason. Stick to the cpu specific wheels for torch to keep things lean and fast.

Been thinking about your build issues. Over the years, I have learned that RAM is the real killer for PyTorch. Quick question tho... how much memory does your model actually pull during inference? In my experience, these are safer bets:

• Hugging Face Spaces Pro is built specifically for AI and very stable.
• Render Starter 2GB RAM is way more flexible than Heroku. Lmk your usage and I can help more.

I honestly had the exact same issues with those one-click platforms and they were not as good as expected. Heroku is a total headache for machine learning because their slug limits are tiny and they dont tell you that until your build fails for the tenth time... Unfortunately, I wasted a week trying to make it work before switching to a real VPS. If you want reliability on a $15 budget, do this:

• Get a DigitalOcean Basic Droplet 2GB RAM 50GB SSD. It fits your budget and gives you actual root access.
• Use a swap file. This basically tricks the system into using disk space as extra RAM when your PyTorch model loads.
• Stick to Docker. It makes the environment consistent so you dont deal with dependency hell. Its way more stable than Heroku tho you gotta handle the linux setup yourself.