this post was submitted on 31 Oct 2023
1 points (100.0% liked)

LocalLLaMA

3 readers
1 users here now

Community to discuss about Llama, the family of large language models created by Meta AI.

founded 1 year ago
MODERATORS
 

I'm only getting 0.8 tokens/second with my 3060 12gb using Zephyr 7b beta.

I'll admit I barely know what I'm doing, but was I wrong to expect a little more? I was hoping for something at least a quarter the speed of gpt-3.5...

top 6 comments
sorted by: hot top controversial new old
[–] Pashax22@alien.top 1 points 1 year ago

I get about 30 t/s on my 12Gb 4070Ti with Zephyr, so something is definitely borked. 0.8 is what I would expect from a 70b model running on CPU and system RAM. Make sure you're offloading as many layers to GPU as your system can handle (in this case, all of them).

[–] DarthNebo@alien.top 1 points 1 year ago (2 children)
[–] Aaaaaaaaaeeeee@alien.top 1 points 1 year ago (1 children)

What's the latest t/s on a 4bit model with TGI? is there a difference compared with HF transformer loader?

[–] DarthNebo@alien.top 1 points 1 year ago

The attention layers get replaced with flash attention 2, there's kv caching as well so you get way better batch1 & batchN results with continuous batching for every request

[–] dodo13333@alien.top 1 points 1 year ago

What is TGI?

[–] LostGoatOnHill@alien.top 1 points 1 year ago

Sounds like you are executing that with CPU. when you do nvidia-smi, do you see memory and GPU consumption?