add Falcon 40B model support

[skirodev]

This PR has already passed the tests of the Falcon mini series models, but due to limitations of my device, I haven't tested it with the original Falcon series models. #293

[philpax]

Looks promising, I'll leave it to Lukas to check the specifics.

Is the model definition/file format beginning to stabilise a little? Part of the reason it's still experimental is because, as far as I can tell, upstream is still figuring out the specifics of the implementation.

[skirodev]

@philpax Yeah, perhaps it should remain in an experimental state until it undergoes all necessary tests and stabilizes.

[LLukas22]

@skirodev On a first glance this looks good. How did you create your ggmlv3 versions of the model? I would like to try 7B- and 40B-instruct before i merge this.

[skirodev]

@LLukas22 I just uploaded the conversion code (only for f16 and f32) to Hugging Face. The quantization code is not completed yet.

[LLukas22]

Thanks! You an probably just use llm to quantize your converted models, it should support the normal Q-quants if you implemented the hyperparams correctly.

[skirodev]

Thanks for the suggestion! I'll try using llm to quantize the converted models and see how it works.

[skirodev]

The performance of using the falcon-7b-instruct-ggml-q4_0.bin model is shown in the following:

cargo run --release -- infer --no-float16 -n 256 -a falcon -m "./models/falcon-7b-instruct-ggmlv3-q4_0.bin" --batch-size 512 -p "write a story about falcon" --stats

[LLukas22]

Alright, I used your script to create ggml versions of falcon-7B and falcon-40B and quantized them with llm to q4_0.

7B works as expected. Output:

write a story about falconry
About
Portfolio
Contacts
For all that is left in the forest, hunt down foxes, wolves, rabbits, and even birds of prey. The first is the hunt, the second is the capture of the bird, and the third is the training of the hawk. Falconers also breed hawks, and some birds may pass through more than one set of hands before they are considered ready for sale or use as a pet. However, you need to be an experienced bird owner before you start a hunt. There are three main types of hunting hawks: lure, live lure, and lure/live. Falconers are trained in many different types of birds, each with its own set of specialized handling techniques and training methods. Falconers hunt with two birds - one as a hawk and the other as a lure bird. For example, falconers use a lure hawk to hunt with live birds and lure hawks to hunt with lure birds. The lure hawk uses birds that are already wild and free to hunt for its own food, whereas the live hawk hunts live birds that are kept in captivity, either for the purpose of training or for hunting purposes. The birds that hunt with lure birds hunt birds like pigeons, doves, and prairie

Sadly 40B produces gibberish 😞. Output:

write a story about falcon/ it or  and  I your body, on the all of 
 not of the  it  new/  can a all you  the so  so. more- this. our past- so in: so I to so you  we'. in  we a  you  it a no.’ your people should a I of it,  by and, and  all the I  to, of

 that what is not all a. for  I can new
 for  in an and we you from all:  you may in of  all
/ to  - D

I don't know if the inference code is the problem or if the conversion-script/quantization corrupted some tensors.

[skirodev]

Sadly 40B produces gibberish 😞. Output:

write a story about falcon/ it or  and  I your body, on the all of 
 not of the  it  new/  can a all you  the so  so. more- this. our past- so in: so I to so you  we'. in  we a  you  it a no.’ your people > should a I of it,  by and, and  all the I  to, of

 that what is not all a. for  I can new
 for  in an and we you from all:  you may in of  all
/ to  - D

I don't know if the inference code is the problem or if the conversion-script/quantization corrupted some tensors.

My apologies, could you retest it now?

[LLukas22]

     Running `target\release\llm.exe infer --no-float16 -n 256 -a falcon -m C:\Users\lkreu\Desktop\falcon\falcon-40b-q4_0 --batch-size 512 -p "write a story about falcon" --stats -r tiiuae/falcon-40b`
⣟ Loading model...[2023-07-16T10:54:11Z INFO  cached_path::cache] Cached version of https://huggingface.co/tiiuae/falcon-40b/resolve/main/tokenizer.json is up-to-date
✓ Loaded 484 tensors (23.5 GB) after 750ms
write a story about falcon, our you this the . more to  not and
 on your so the- you.  I. it. we may a so it for that or any
 the and, and for more the and  we  I by no  so in and just   of our- a can of. all,? I it  at the 
 so,  is it of my personal to that. 
 the and by is it

It got a bit better but there probably still is something wrong. Maybe we should wait for the ggml update and revisit it then?

[skirodev]

It got a bit better but there probably still is something wrong. Maybe we should wait for the ggml update and revisit it then?

Absolutely. I agree with your suggestion. Let's await the ggml update and take another look at it when the time comes. Thank you for conducting the testing.

[LLukas22]

Could you try to pull the latest main into this? It should contain the latest ggml version.

[skirodev]

Could you try to pull the latest main into this? It should contain the latest ggml version.

Sure, I already pulled the latest main branch for the updated ggml version.

[LLukas22]

Falcon 40B still produces gibberish:

PS F:\Github\llm-main> cargo run --release --features falcon -- infer -a falcon -p "Tell me a story about a falcon" -m "C:\Users\lkreu\Desktop\falcon\falcon-40b-q4_0" --no-float16
    Finished release [optimized] target(s) in 0.24s
     Running `target\release\llm.exe infer -a falcon -p "Tell me a story about a falcon" -m C:\Users\lkreu\Desktop\falcon\falcon-40b-q4_0 --no-float16`
✓ Loaded 484 tensors (23.5 GB) after 114ms
Tell me a story about a falcon . more to or- they in, you your. de/ /   to  they.: 
  you, I so good more on to one. on a for all  I  so,.  i  or  and - and it  that  :: and la

Maybe i need to reconvert/quantize my model 🤔

[skirodev]

Falcon 40B is now capable of successful inference after my testing.

cargo run --release -- infer -a falcon -m "./models/falcon-40b-instruct-ggmlv3-q4_0.bin" --batch-size 512 -p "write a story about falcon" -r tiiuae/falcon-40b --stats

The majestic bird of prey soared through the sky, its wingspan stretching outwards as it searched for prey. Its sharp 
eyes scanned the horizon, and in an instant, it spotted movement below. With powerful strokes of its wings, it dove 
towards its target at incredible speeds before striking with lightning-fast precision. The falcon was a symbol of 
strength, agility, and intelligence – an awe-inspiring creature that commanded respect from all who saw it soar above.

However, the embedded tokenizer code still needs modification as Falcon does not require adding bos token id and has some special tokens, which may depend on the implementation of GGUF format.

[LLukas22]

Good Job 👍

I'll give this another look tomorrow and if everything works i'm gonna merge it.

[philpax]

Code looks good! Will leave it to Lukas to do final tests but OK from my end. Hopefully we can get the additional information from GGUF soon.

[LLukas22]

LGTM 👍

Now even works with the fp16 memory :D

[ghost]

Does it works with Metal?

[LLukas22]

@jempabroni Maybe, depends on if all necessary operations were already ported into metal shaders. You can try using it and if it gives you an invalid operation error it's not supported yet.