text-generation-inference VS vllm

I wanted to write that TGI inference engine is not Open Source anymore, but they have reverted the license back to Apache 2.0 for the new version TGI v2.0: https://github.com/huggingface/text-generation-inference/rel...

Good news!

Setting up a server for multiple users is very different from setting up LLM for yourself. A safe bet would be to just use TGI, which supports continuous batching and is very easy to run via Docker on your server. https://github.com/huggingface/text-generation-inference

Okay, I actually got local co-pilot set up. You will need these 4 things.

1) CodeLlama 13B or another FIM model https://huggingface.co/codellama/CodeLlama-13b-hf. You want "Fill in Middle" models because you're looking at context on both sides of your cursor.

2) HuggingFace llm-ls https://github.com/huggingface/llm-ls A large language mode Language Server (is this making sense yet)

3) HuggingFace inference framework. https://github.com/huggingface/text-generation-inference At least when I tested you couldn't use something like llama.cpp or exllama with the llm-ls, so you need to break out the heavy duty badboy HuggingFace inference server. Just config and run. Now config and run llm-ls.

4) Okay, I mean you need an editor. I just tried nvim, and this was a few weeks ago, so there may be better support. My expereicen was that is was full honest to god copilot. The CodeLlama models are known to be quite good for its size. The FIM part is great. Boilerplace works so much easier with the surrounding context. I'd like to see more models released that can work this way.

A simple microservice would be https://github.com/huggingface/text-generation-inference .

Works flawlessly in Docker on my Windows machine, which is extremely shocking.

I am wondering what is the best / most cost-efficient way to serve llama V2. - llama.cpp (is it production ready or just for playing around?) ? - Triton inference server ? - HF text generation inference ?

I guess the root cause of my claim is that OpenAI won't tell us whether or not GPT-3.5 is an MoE model, and I assumed it wasn't. Since GPT-3.5 is clearly nondeterministic at temp=0, I believed the nondeterminism was due to FPU stuff, and this effect was amplified with GPT-4's MoE. But if GPT-3.5 is also MoE then that's just wrong.

What makes this especially tricky is that small models are truly 100% deterministic at temp=0 because the relative likelihoods are too coarse for FPU issues to be a factor. I had thought 3.5 was big enough that some of its token probabilities were too fine-grained for the FPU. But that's probably wrong.

On the other hand, it's not just GPT, there are currently floating-point difficulties in vllm which significantly affect the determinism of any model run on it: https://github.com/vllm-project/vllm/issues/966 Note that a suggested fix is upcasting to float32. So it's possible that GPT-3.5 is using an especially low-precision float and introducing nondeterminism by saving money on compute costs.

Sadly I do not have the money[1] to actually run a test to falsify any of this. It seems like this would be a good little research project.

[1] Or the time, or the motivation :) But this stuff is expensive.

The easiest is to use vllm (https://github.com/vllm-project/vllm) to run it on a Couple of A100's, and you can benchmark this using this library (https://github.com/EleutherAI/lm-evaluation-harness)

Great question! scheduling workloads onto GPUs in a way where VRAM is being utilised efficiently was quite the challenge.

What we found was the IO latency for loading model weights into VRAM will kill responsiveness if you don't "re-use" sessions (i.e. where the model weights remain loaded and you run multiple inference sessions over the same loaded weights).

Obviously projects like https://github.com/vllm-project/vllm exist but we needed to build out a scheduler that can run a fleet of GPUs for a matrix of text/image vs inference/finetune sessions.

disclaimer: I work on Helix

FYI, vLLM also just added experiment multi-lora support: https://github.com/vllm-project/vllm/releases/tag/v0.3.0

Also check out the new prefix caching, I see huge potential for batch processing purposes there!