Static Newsabout
apsec112 | 32 comments

jcoc611|next|

Probably an ignorant question, but could someone explain why the Context Length is much larger than the Generation Length?

dacox|parent|next|

When doing inference for an LLM, there are two stages.

The first phase is referred to as "prefill", where the input is processed to create the KV Cache.

After that phase, the "decode" phase is called auto-regressively. Each decode yields one new token.

This post on [Inference Memory Requirements](https://huggingface.co/blog/llama31#inference-memory-require...) is quite good.

These two phases have pretty different performance characteristics - prefill can really maximize GPU memory. For long contexts, its can be nigh impossible to do it all in a single pass - frameworks like vLLM use a technique called "chunked prefill".

The decode phase is compute intensive, but tends not to maximize GPU memory.

If you are serving these models, you really want to be able to have larger batch sizes during inference, which can only really come with scale - for a smaller app, you won't want to make the user wait that long.

So, long contexts only have to be processed _once_ per inference, which is basically a scheduling problem.

But the number of decode passes scales linearly with the output length. If it was unlimited, you could get some requests just _always_ present in an inference batch, reducing throughput for everyone.


mmoskal|root|parent|next|

Decode speed is generally memory bandwidth bound. Prefill is typically arithmetic bound. This is the reason for mixed batches (both decode and prefill) - it let's you saturate both memory and arithmetic.

Chunked prefill is for minimizing latency for decode entries in the same batch. It's not needed if you have only one request - in that case it's the fastest to just prefill in one chunk.

I'm pretty sure the sibling comment is right about different length limits - it's because of training and model talking nonsense if you let too long.


easygenes|root|parent|prev|next|

It is also a training issue. The model has to be trained to reinforce longer outputs, which has a quadratic train-time cost and requires suitable long-context response training data.

jcoc611|root|parent|prev|next|

That's a great explanation, thank you!

freeqaz|prev|next|

32B is a nice size for 2x 3090s. That comfortably fits on the GPU with minimal quantization and still leaves extra memory for the long context length.

70B is just a littttle rough trying to run without offloading some layers to the CPU.


a_wild_dandan|parent|next|

70B+ models typically run great with my MacBook's 96GB of (V)RAM. I want a Mac Studio to run e.g. llama-405B, but I can't justify the marginal model quality ROI for like $7k or whatever. (But I waaant iiit!)

tarruda|root|parent|next|

You can get refurbished Mac Studio M1 Ultra with 128GB VRAM for ~ $3k on ebay. M1 ultra has 800GB/s memory bandwidth, same as the M2 ultra.

Not sure if 128GB VRAM is enough for running 405b (maybe at 3-bit quant?), but it seems to offer great value for running 70B models at 8-bit.


a_wild_dandan|root|parent|next|

Yeah, I would want the 192GB Mac for attempting such hefty models. But I have such basic bitch needs that 405B is overkill haha.

diggan|root|parent|prev|next|

> run great

How many tokens/second is that approx?

For reference, Qwen 2.5 32B on CPU (5950X) with GPU offloading (to RTX 3090ti) gets about 8.5 token/s, while 14B (fully on GPU) gets about ~64 tokens/s.


a_wild_dandan|root|parent|next|

For 70B models, I usually get 15-25 t/s on my laptop. Obviously that heavily depends on which quant, context length, etc. I usually roll with q5s, since the loss is so minuscule.

BaculumMeumEst|root|parent|prev|next|

what quant are you running for that rig? i've been running q4, not sure if I can bump that up to q5 across the board (or if it's worth it in general)

Flux159|prev|next|

It would be nice to have comparisons to Claude 3.5 for the coder model, only comparing to open source models isn’t super helpful because I would want to compare to the model I’m currently using for development work.

imjonse|parent|next|

Aider will probably have some numbers at https://aider.chat/docs/leaderboards/

Deathmax|root|parent|next|

They've posted their own run of the Aider benchmark [1] if you want to compare, it achieved 57.1%.

[1]: https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5/Qwen...


reissbaker|root|parent|next|

Oof. I'm really not sure why companies keep releasing these mini coding models; 57.1% is worse than gpt-3.5-turbo, and running it locally will be slower than OpenAI's API. I guess you could use it if you took your laptop into the woods, but with such poor coding ability, would you even want to?

The Qwen2.5-72B model seems to do pretty well on coding benchmarks, though — although no word about Aider yet.


diggan|parent|prev|next|

Here is a comparison of the prompt "I want to create a basic Flight simulator in Bevy and Rust. Help me figure out the core properties I need for take off, in air flight and landing" between Claude Sonnet 3.5 and Qwen2.5-14B-Instruct-Q4_K_M.gguf:

https://gist.github.com/victorb/7749e76f7c27674f3ae36d791e20...

AFAIK, there isn't any (micro)benchmark comparisons out yet.


yourMadness|root|parent|next|

14B with Q4_K_M quantization is about 9 GB.

Remarkable that it is at all comparable to Sonnet 3.5


diggan|root|parent|next|

Comparable, I guess. But the result is a lot worse compared to Sonnet for sure. Parts of the example code doesn't make much sense. Meanwhile Sonnet seems to have the latest API of Bevy considered, and mostly makes sense.

Sn0wCoder|parent|prev|next|

This might be what you are asking for... https://qwenlm.github.io/blog/qwen2.5-coder/

Ctrl F - Code Reasoning:


ekojs|prev|next|

Actually really impressive. They went up from 7T tokens to 18T tokens. Curious to see how they perform after finetuning.

cateye|prev|next|

> we are inspired by the recent advancements in reinforcement learning (e.g., o1)

It is interesting to see what the future will bring when models incorporate chain of thought approaches and whether o1 will get outperformed by open source models.


GaggiX|prev|next|

>our latest large-scale dataset, encompassing up to 18 trillion tokens

I remember when GPT-3 was trained on 300B tokens.


imjonse|parent|next|

and was considered too dangerous to be released publicly.

baq|root|parent|next|

they are dangerous... for folks who need to scrape the web for low background tokens to train their transformers.

abc-1|root|parent|prev|next|

Nobody ever really believed this, the truth is rarely in vogue.

GaggiX|root|parent|prev|next|

The larger GPT-2s were also considered too dangerous to release publicly at first.

Workaccount2|root|parent|next|

I remember be very understanding of it too after seeing the incredible (but absolutely terrible in retrospect) outputs.

GaggiX|root|parent|next|

I wasn't really compelled at the time, nothing has changed.