cog VS insanely-fast-whisper

View on GitHub

I'm in the same situation. I found this cog project to dockerise ML https://github.com/replicate/cog : you write just one python class and a yaml file, and it takes care of the "CUDA hell" and deps. It even creates a flask app in front of your model.

That helps keep your system clean, but someone with big $s please rewrite pytorch to golang or rust or even nodejs / typescript.

https://github.com/replicate/cog

Our thinking was just that a bunch of folks will want to fine-tune right away, then deploy the fine-tunes, so trying to make that easy... Or even just deploy the models-as-is on their own infra without dealing with CUDA insanity!

I have used this tool before: https://github.com/replicate/cog/tree/main

Then I discovered Cog: Containers for Machine Learning. Looks like a way more flexible solution to plug in the existing infrastructure: you write your custom code and Cog plugs it in a Docker image with FastAPI, no extra ecosystem complexity added.

Highly recommend using cog https://github.com/replicate/cog

Dokku Maintainer here.

Dokku also supports Dockerfiles, Docker Images, Tarballs (similar to heroku slugs), and Cloud Native Buildpacks. I'm also actively working on AWS Lambda support (both for simple usage without much config as well as SAM-based usage) and investigating Replicate's Cog[1] and Railways Nixpacks[2] functionalities for building apps.

There are quite a few options in the OSS space (as well as Commercial offerings from new startups and popular incumbents). It's an interesting space to be in, and its always fun to see how new offerings innovate on existing solutions.

  [1] https://github.com/replicate/cog

Hi HN! Since the launch of JigsawStack.com we've been trying to dive deeper into fully managed AI APIs built and fine tuned for specific use cases. Audio/video transcription was one of the more basic things and we wanted the best open source model and at this point it is OpenAI's whisper large v3 model based on the number languages it supports and accuracy.

The thing is, the model is huge and requires tons of GPU power for it to run efficiently at scale. Even OpenAI doesn't provide an API for their best transcription model while only providing whisper v2 at a pretty high price. I tried running the whisper large v3 model on multiple cloud providers from Modal.com, Replicate, Hugging faces dedicated interface and it takes a long time to transcribe any content about ~30mins long for 150mins of audio and this doesn't include the machine startup time for on demand GPUs. Keeping in mind at JigsawStack we aim to return any heavy computation under 25s or 2mins for async cases and any basic computation under 2s.

While exploring Replicate, I came across this project https://github.com/Vaibhavs10/insanely-fast-whisper by Vaibhav Srivastav which optimises the hell out of this whisper large v3 model with a variety of techniques like batching and using FlashAttention 2. This reduces computation time by almost 30x, check out the amazing repo for more stats! Open source wins again!!

First we tried using Replicates dedicated on-demand GPU service to run this model but that did not help, the cold startup/booting time alone of a GPU made the benefits of the optimised model pretty useless for our use case. Then tried Hugging face and modal.com and we got the same results, with a A100 80GB GPU, we were seeing around an average of ~2mins start up time to load the machine and model image. It didn't make sense for us to have a always on GPU running due to the crazy high cost. At this point I was inches away from giving up.

Next day I got an email from Fly.io: "Congrats, Yoeven D Khemlani has GPU access!" I totally forgot the Fly started providing GPUs and I'm a big fan of their infra reliability and ease to deploy. We also run a bunch of our GraphQL servers for JigsawStack on Fly's infra!

I quickly picked up some Python and Docker by referring to a bunch of other Github repos and Fly's GPU tutorials then wrote the API layer with the optimised version of whisper 3 and deployed on Fly's GPU machines.

And wow the results were pretty amazing, the start up time of the machine on average was ~20 seconds compared to the other providers at ~2mins with all the performance benefits from the optimised whisper. I've added some more stats in the Github repo. The more interesting thing to me is cost↓

Based on 10mins of audio:

There's a better parallel/batching that works on the 30s chunks resulting in 40X. From HF at https://github.com/Vaibhavs10/insanely-fast-whisper

This is again not native PyTorch so there's still room to have better RTFX numbers.

Founder of Replicate here. We open pull requests on models[0] to get them running on Replicate so people can try out a demo of the model and run them with an API. They're also packaged with Cog[1] so you can run them as a Docker image.

Somebody happened to stumble across our fork of the model and submitted it. We didn't submit it nor intend for it to be an ad. I hope the submission gets replaced with the upstream repo so the author gets full credit. :)

[0] https://github.com/Vaibhavs10/insanely-fast-whisper/pull/42