That's unfortunate. Running the 65B Q4 on an AMD Epyc with 32 1.5ghz cores and 256 GB of ram I get around 3 tokens/sec, which is useable if not ideal. I wonder if the difference is related to the RAM or the number of CPUs?
Although there are multiple bottlenecks, my understanding (and why at a certain point, throwing more threads doesn't work) is that inference for dense LLMs are largely limited by memory bandwidth. Most desktop computers will have dual channel DDR4/DDR5 memory which will be hard pressed to get >60GB/s. A last-gen Epyc/Threadripper Pro should have 8 channel memory DDR4-3200 support, which should get you a theoretical max of 204.8 GB/s (benchmarking ends up more around 150GB/s in AIDA64).
The latest Genoa has 12 channel DDR5-4800 support (and boosted AVX-512) and I'd imagine should perform quite well, but if you primarily want to run inference on a quantized 65B model, I think you're best bang/buck (for local hardware) would be 2 x RTX 3090s (each of those has 24GB of GDDR6X w/ just shy of 1TB/s of memory bandwidth).
With my LLaMA AVX implementation on 32bit floats [0] there no performance gain after 2 threads, so remaining 14 threads available are of no use, there no memory bandwidth to load them with work :)
To the extent that you're memory bandwidth limited you should be able to do multiple inferences at once --- latency stays high but getting multiple samplings can be extremely useful for many uses and can cover up somewhat for high latency.
To an extent, but memory bandwidth soon becomes a bottleneck there too. The hidden state and the KV cache are large so it becomes a matter of how fast you can move data in and out of your L2 cache. If you don’t have a unified memory pool it gets even worse.
The two-channel DDR5 in desktops can't even do two channels very well -- if you try to put 64GB RAM in (two dual-rank 32GB DIMMs) then you lose around 50% of the bandwidth compared to a single rank DIMM (e.g. from 8GHz to 4GHz speeds, and increased latency).
I'm following the discussions on GitHub as well as their PRs closely.
The primary bottleneck for now is compute.
They've recently made a big improvement to performance by introducing partial gpu acceleration if you compile with a gpu accelerated variant of BLAS. Either cublas (Nvidia) or CLBlast (slightly slower but supports almost everything: Nvidia, Apple, AMD, mobile, raspberry pi etc)
3 tokens/sec is a lot faster than what I experienced. Even though your CPU has a lot more cores, I think llama.cpp was not being able to make good use of more than 8 threads.
When did you test this? Maybe llama.cpp had some improvements since I used it (which was at the start of the project).
It's not about threads number, it about memory bottleneck. Sweet spot for my M1 Pro laptop is around 6 threads and 4bit model - I've managed to get 20 tokens per sec, really impressive
slow could be useful if you do not want to chat with it, and instead you could code it to do a long running job, like code review your entire project like a code analysis tool. Or summarize a lot of content.
That's why it's not parallelized along the time axis but rather along the dimension of the embedding axis.
You split the big matrices into smaller matrices to dispatch the workload. But this means you have to add some communication overhead (roughly nblayers sequential synchronisation point per token). In official LLama implementation this is done transparently using RowParallelLinear, ColumnParallelLinear, ParallelEmbedding see https://github.com/facebookresearch/llama/blob/main/llama/mo...
Transformer have multiple attention heads, that can be computed independently and then summed together to produce the output of the layer. This allow to split the parameter space among machines without having to transfer them at each iteration.
I'm really curious how Meta, DeepMind and OpenAI make the big models work. The biggest A100 you can buy is just 80GB. And I assume the big companies use single precision floating point during training. Are they actually partitioning the big model across multiple GPU instances? If one had the hardware, how many GPUs does the biggest LLAMA take? These are systems issues and I have not read papers or blog posts on how this works. To me, this infra is very non-trivial.
The "standard" machine for these things has 8x80GB = 640GB memory (p4de instances here: https://aws.amazon.com/ec2/instance-types/p4/), with _very_ fast connections between GPUs. This fits even a large model comfortably. Nowadays probably most training use half precision ("bf16", not exactly float16, but still 2 bytes per parameter). However during training you easily get a 10-20x factor between the number of parameters and the bytes of memory needed, due to additional things you have to store in memory (activations, gradients, etc.). So in practice the largest models (70-175B parameters) can't be trained even on one of these beefy machines. And even if you could, it would be awfully slow.
In practice, they typically use servers with clusters of these machines, up to about 1000 GPUs in total (so around 80TB of memory, give or take a few?). This allows even the biggest models to be trained on large batches of several hundreds, or even thousands, of elements (the total memory usage is _not_ proportional to the product of number of parameters and the batch size, but it does increase as a function of both of them, a term of which being indeed the product of the two). It makes for some very tricky engineering choices to make just the right data travel across connections, trying to avoid as much as possible that you have to sync large amount of data between different machines (so "chunking" things to stay on the 640GB range) with strategies such as ZeRO being published every now and then. Plus of course the practical effort to make physical connections as fast as possible...
To get an idea of how hard these things are, take a look at how long the list of names in the published paper about BLOOM language model is :-)
While I do not have any A100 handy right now I have an instance running on Genesis Cloud with 4x RTX 3090.
A quick, very unscientific, test using the oobabooba/text-generation-webui with some models I tried earlier gives me:
* oasst-sft-7-llama-30b (spread over 4x GPU): Output generated in 28.26 seconds (5.77 tokens/s, 163 tokens, context 55, seed 1589698825)
* llama-30b-4bit-128g (only using 1 GPU as it is so small): Output generated in 12.88 seconds (6.29 tokens/s, 81 tokens, context 308, seed 1374806153)
* llama-65b-4bit-128g (only using 2 GPU): Output generated in 33.36 seconds (3.81 tokens/s, 127 tokens, context 94, seed 512503086)
* llama (vanilla, using 4x GPU): Output generated in 5.75 seconds (4.69 tokens/s, 27 tokens, context 160, seed 1561420693)
They all feel fast enough for interactive use. If you do not have an interface that streams the output (so you can see it progressing) it might feel a bit weird if you often have to wait ~30s to get the whole output chunk.