Qwen2.5: A Party of Foundation Models | Hacker News Reader

Qwen2.5: A Party of Foundation Models

qwenlm.github.io

168 pointsapsec11238 comments

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

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.

That's a great explanation, thank you!

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.

They definitely have to be trained to reinforce longer outputs, but I do not believe this adequately explains the low-ish generation limits.

We are starting to see models with longer and longer generation limits (gpt-4o-mini having 16k, the o1 models going up to 64k), as well as longer and longer context limits (often 128k, google offering a million).

I find it very unlikely they are actually training with inputs or outputs near these maximums.

If you want to convince yourself, do the attention calculation math for these sequence lengths.

You can also see how openai restricts the sequence length for fine tuning to 64k - almost certainly bound by available GPU sizes

I suspect the 4096 limits have been set as a "reasonable" limit for a myriad of reasons.

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.

Chunked prefill or some similar technique is also necessary for serving long context requests where there is not enough GPU memory available, regardless of concerns about latency.

For example, consider a prompt sent to Llama 3.1 405B that uses 128k input tokens.

The KV cache will be 123GB. No matter how many GPUs you shard the model across, you are not fitting that KV cache in GPU memory (a H100 has 80GB)

You can do tensor parallelism 8 ways (8 KV heads). You can also do pipeline parallelism (there is 126 layers). Either way would work. A million tokens is possible just very slow.

Also, 405b has 8 KV heads of 128 size (hidden_size/num_attention_heads) times 126 layers [0] times 2 (K and V) times 2 bytes (bf16) is 504k per token. At FP8 it's 252k.

[0] https://huggingface.co/meta-llama/Meta-Llama-3.1-405B/blob/m...

Besides technical details, in a normal usage scenario, the context length is also much greater than the generation length.

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.

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!)

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.

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

405b even at low quants would have very low tokens generation speed, so even if you got the 192GB it would probably not be a good experience. I think 405b is the kind of model that only makes sense to run in clusters of A100/H100.

IMO it is not worth it, 70b models at q8 are already pretty darn good, and 128gb is more than enough for those.

Exactly! Have you tried the Phi models? To me, they indicate that we can get much more efficient models. In a few years, 70b on gold standard synthetic data + RL might run circles around SotA. It's such an exciting time to be alive.

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)

I run q5s usually, since it's a 40% haircut on model size, with nearly no PPL loss. (Presuming an 8b native model like Qwen.)

> 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.

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.

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.

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

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...

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.

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.

14B with Q4_K_M quantization is about 9 GB.

Remarkable that it is at all comparable to Sonnet 3.5

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.

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

Ctrl F - Code Reasoning:

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

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

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

and was considered too dangerous to be released publicly.

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

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

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

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

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

I remember when Ps2 chips were considered too advanced the US government banned shipment of PlayStations to China lest the PLA gets hold of them

I'm impressed by the scope of this drop. The raw intelligence of open models seems to be falling behind closed. But I think that's because frontier models from openai and anthropic are not just raw models, but probably include stuff like COT, 'best of N', or control vectors.

> 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.

Reply on news.ycombinator.com