Skip to content
Huzaifa Qadeer
Back to all work

Self-Hosted Video Generation

Running open video-generation models on infrastructure I control.

Role
Solo — inference and serving
Status
UnfinishedNot hosted
Context

Why it exists

Generated shots in these pipelines run through a third-party video API, so the serving stack behind them is not mine to tune. Self-hosting moves that onto infrastructure I control. This one is in progress: I am optimizing inference, not reporting a result.

Approach

How it works

  • Self-host open video-generation models instead of calling a third-party API.
  • Working on inference optimization — FP8 quantization and serving-level optimizations.
  • Benchmarking on rented B200 GPUs through RunPod.
  • No performance claims yet. A figure gets published only once the same run repeats on a fresh machine.
Decisions

Trade-offs

  • Self-host open models

    Instead of Keep calling the third-party video API

    Quantization and serving become things I can change. The cost is that uptime, model updates and GPU provisioning are now mine to handle, and the hosted API handled all three.

  • Benchmark on rented GPUs

    Instead of Buy or lease dedicated hardware

    The hardware is released between runs instead of sitting idle. The cost is that every benchmark starts from a cold machine and depends on whichever host the provider allocates, which makes runs harder to compare.

Constraints

Limits and how they are handled

Nothing here is measured yet, so no performance number is published.
The figures worth measuring are already named; each is held back until a cold-start rental reproduces it.
Stack

Built with

  • RunPod
  • B200 GPUs
  • FP8 quantization
Links

The repository is private. Ask for a walkthrough.