Llama-3.1-8B-Instruct-4bit keeps looping at the end.

[chigkim]

I'm on mlx-lm v0.19.1.

Running the following command with 4bit produced a bug where it would just generate the full 1000 max-tokens and just repeat the last two paragraphs over and over.

mlx_lm.generate --model mlx-community/Meta-Llama-3.1-8B-Instruct-4bit --max-kv-size 33000 --max-tokens 1000 --temp 0.0 --top-p 0.9 --seed 1000 --prompt -<../text/portugal.txt;say done

Here is my Full prompt from Wikipedia article.

Just to see what happens, I increased --max-kv-size to 34k and --max-tokens to 2000, and it generated 2k tokens with the loop.

Running the exact same command replacing 4bit with 8bit generated the correct text with 811 tokens and stopped at the end without the loop.

[awni]

I ran the same command, below is the output I get which looks pretty reasonable. Can you share what you are seeing? Also could you share the machine, the OS, and the version of MLX as well?

The text provides a comprehensive overview of Portugal's history, geography, government, economy, demographics, culture, and more. Here are the main points:

**History**

* Portugal is a country located on the Iberian Peninsula, in Southwestern Europe.
* The country has a rich history, with evidence of human presence dating back to prehistoric times.
* The territory was inhabited by various peoples, including the Celts, Iberians, and Romans.
* Portugal gained independence from the Kingdom of León in 1143 and became a major economic and political power in the 15th and 16th centuries.
* The country was a colonial power, with colonies in Africa, Asia, and the Americas.
* Portugal voluntarily entered a dynastic union with Spain in 1580, which lasted until 1640.
* The country experienced a period of decline in the 19th and 20th centuries, but has since recovered and become a developed country.

**Geography**

* Portugal occupies an area on the Iberian Peninsula and two archipelagos in the Atlantic Ocean: Madeira and the Azores.
* The country has a diverse geography, with mountains, plains, and coastlines.
* The Tagus River is the main river in Portugal and flows through the capital city of Lisbon.
* The country has a Mediterranean climate, with warm summers and mild winters.

**Government**

* Portugal is a semi-presidential representative democratic republic.
* The government is divided into four branches: the President, the Government, the Assembly of the Republic, and the Courts.
* The President is elected to a five-year term and has executive powers.
* The Assembly of the Republic is a unicameral body composed of up to 230 deputies.
* The Government is headed by the Prime Minister and includes Ministers and Secretaries of State.

**Economy**

* Portugal is a developed and high-income country.
* The country has a diverse economy, with a strong focus on services, industry, and agriculture.
* The country has a significant fishing industry and is a major producer of cork and carob.
* Portugal has a high level of foreign debt and has received a bailout from the European Union and the International Monetary Fund.
* The country has a strong tradition of trade and commerce, with a significant presence of foreign companies.

**Demographics**

* The population of Portugal is approximately 10.5 million people.
* The country has a relatively homogeneous population, with a single language and a single religion.
* The population is aging, with a high percentage of people over 65.
* The country has a low fertility rate, with an average of 1.5 children per woman.
* The country has a significant number of immigrants, with around 7.5% of the population being foreign-born.

**Culture**

* Portugal has a rich cultural heritage, with a strong tradition of music, dance, and art.
* The country has a significant number of UNESCO World Heritage Sites, including the Jerónimos Monastery and the Tower of Belém.
* The country has a strong tradition of folk music and dance, with a significant number of festivals and events throughout the year.
* The country has a significant number of museums and art galleries, including the National Museum of Ancient Art and the Calouste Gulbenkian Museum.

**Sport**

* Football is the most popular sport in Portugal, with a significant number of fans and a strong national team.
* The country has a significant number of sports clubs, including Benfica, Sporting CP, and FC Porto.
* The country has a strong tradition of athletics, with a significant number of Olympic medals and world championships.
* The country has a significant number of water sports, including surfing, windsurfing, and kitesurfing.

**Visual Arts**

* Portugal has a rich history in painting, with a significant number of famous painters, including Nuno Gonçalves and Vasco Fernandes.
* The country has a strong tradition of modern art, with a significant number of famous artists, including Amadeo de Souza-Cardoso and Almada Negreiros.
* The country has a significant number of contemporary artists, including Helena Almeida, Joana Vasconcelos, and Julião Sarmento.

Overall, the text provides a comprehensive overview of Portugal's history, geography, government, economy, demographics, culture, and more.
==========
Prompt: 32188 tokens, 643.576 tokens-per-sec
Generation: 892 tokens, 31.103 tokens-per-sec
Peak memory: 12.535 GB

[chigkim]

Yea your output looks correct.
Here is my full log. It keeps looping after "The health system in Portugal is..."
https://pastebin.com/raw/XzgVh1Zc

MBp 16" with M3 Max 64GB
MacOS 15.0.1
Python 3.12.7 (main, Oct 1 2024, 02:05:46) [Clang 16.0.0 (clang-1600.0.26.3)] on darwin
Name: mlx
Version: 0.19.0
Name: mlx-lm
Version: 0.19.1

Thanks!

[chigkim]

Also I noticed that your prompt processing speed is 643.576, and mine is 417.707.
The format of the output looks very different. I saw that style when I run mlx-community/Llama-3.2-3B-Instruct-4bit.
I wonder I have like older weight with a bug or something. Is there way to check if I have the latest weight?

[barronalex]

I was able to reproduce this with the same long prompt on M1 Max.

I did a bit of bisecting and looks like after ml-explore/mlx#1509 I'm getting non deterministic outputs.
Weirdly it only seems to happen with long prompts.

[awni]

Ok I’ll take a look at that. But it’s not in MLX 0.19 so it can’t really be the same issue as above..

[awni]

@chigkim were you building the main branch of MLX from source or did you install MLX from PyPi?

[chigkim]

For MLX, I installed from Pip. For MLX-LM, I tried both from pip and git.

[awni]

I've tried to reproduce this on several machines (M1 Max, M2 Ultra, M1 Ultra, and M3 Max), so far not seeing any issues in the output. Some questions / suggestions:

• If you have a chance could you check mlx==0.19.1 to see if it still reproduces?
• Is the output always the same looping behavior or is it non-deterministic?

[chigkim]

I just tried mlx==0.19.1 as well as 0.20.0 along with mlx-lm==0.19.3.
The output always produces 1000 tokens with the same looping behavior 100% of time.
I wonder if there's any mistake in my command or prompt file?

[awni]

I'm really stumped by this one to be honest. I tried your exact command which works fine on several machines:

mlx_lm.generate --model mlx-community/Meta-Llama-3.1-8B-Instruct-4bit --max-kv-size 33000 --max-tokens 1000 --temp 0.0 --top-p 0.9 --seed 1000 --prompt -<../text/portugal.txt;say done

Given that it works for me, I don't think there is a problem with the command or the prompt (unless you've changed it).

Without the ability to reproduce, it's very difficult to debug. Anything you can do to fuzz around and see if there are conditions that the looping is sensitive to would be great to help us get to the bottom of this.

Some ideas:

• Try rebooting
• Try a new python env with a fresh install of MLX, MLX LM
• Try a different model. Maybe a smaller one like mlx-community/llama-3.2-1B-Instruct-4bit
• Anything else you can think of..

Would be really great if you have time for any of those and can share the results. Thanks!

[chigkim]

I deleted my environment and created a fresh one with python -m venv.
I deleted the model in the Huggingface cache with huggingface-cli.
I deleted mlx-examples from my local machine, and cloned from fresh.
I installed mlx-examples/llms with pip install -e ..
That's the only thing I installed in the new environment.
It still loops at the end. :(
However, mmlx-community/Llama-3.2-3B-Instruct-4bit works fine.
Also mlx-community/Meta-Llama-3.1-8B-Instruct-8bit works fine.
Only mlx-community/Meta-Llama-3.1-8B-Instruct-4bit creates a loop at the end.

[awni]

That is really curious..

There was a bug in one of our qmv kernels that was recently fixed. It might be possible (but I think pretty unlikely) that this would account for the looping behavior you are seeing.

Do you only notice looping for the very long prompt or do you also see it for shorter prompts?

Also you have a few arguments set. I'm wondering if you turn them off or change does the looping go away?

• --max-kv-size 33000 - don't set that, does it still loop?
• --temp 0.0 --top-p 0.9 - don't set that, does it still loop?
• --seed 1000 - change that to something else, does it still loop?

[chigkim]

I've tried without specifying max-kv-size, temp, top_p, seed, but they all looped exactly the same way.
Interestingly, if I delete the last section about visual art from the prompt, it doesn't loop!
That's only 295 tokens difference (32134-31839).

[awni]

Ok. Let's see if it's fixed after our next release (which includes a fix for quantization in some cases ml-explore/mlx#1577).

If it's not fixed, I will try fuzzing around a bit to see if it can be reproduced on our side.

[chigkim]

Do you guys need to requantize and update the model on HF, or can I just pull the main branch, install with pip install -e ., and test?

[awni]

No you can use the same model (no requantization needed). You can test by pulling and building the main branch. It would be great to know if that works for you or not.

[chigkim]

Oh that was merged 4 days ago, and I already pulled the latest main branch when I tested. So the fix didn't work. :(

[chigkim]

Ok, so I might have a potentially good news that could lead to something...
It looks like something happened between mlx 0.18.1 and 0.19.0.
It doesn't loop in 0.18.1, but it does loop in 0.19.0.

[awni]

Interesting.. you can see all the commits between 18.1 and 19.0 here: ml-explore/mlx@v0.18.1...v0.19.0

The commit in there that seems most likely to have changed something for LLM inference is the fused attention. Can you try building and the commits before and after to see if that is the case?

So concretely:

For including the fused attention:

git checkout 50d8bed4688e04d8ba4cc5b8e20a79f22e8e93ce
env CMAKE_BUILD_PARALLEL_LEVEL=10 pip install .

And the commit just before:

git checkout 9dd72cd421260ebc0f30e773f6b35fdf87555806
env CMAKE_BUILD_PARALLEL_LEVEL=10 pip install .

That would be my first guess as to a related cause but it would be good to check and see.

[chigkim]

Yep that was it! The commit 50d8bed loops, but 9dd72cd doesn't.

[awni]

Could you try running with Metal validation enabled to see if that gives us any clues? (Low probability but when it hits it hits well):

METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 mlx_lm.generate ...

[awni]

Also you can precompute the prompt cache to speed testing up:

mlx_lm.cache_prompt --prompt-cache-file prompt.safetensors --model mlx-community/Meta-Llama-3.1-8B-Instruct-4bit --prompt -<prompt.txt                

Then use that to generate:

mlx_lm.generate --prompt-cache-file prompt.safetensors --max-tokens 1024

[angeloskath]

@chigkim Adding on @awni 's last message. Can you run the following commands and report back the outputs?

mlx_lm.cache_prompt --prompt-cache-file prompt.safetensors --model mlx-community/Meta-Llama-3.1-8B-Instruct-4bit --prompt -<prompt.txt
mlx_lm.generate --prompt-cache-file prompt.safetensors --max-tokens 2048
mlx_lm.generate --prompt-cache-file prompt.safetensors --max-tokens 2048
mlx_lm.generate --prompt-cache-file prompt.safetensors --max-tokens 2048 --temp 0.1
mlx_lm.generate --prompt-cache-file prompt.safetensors --max-tokens 2048 --temp 0.1 --seed 100
mlx_lm.generate --prompt-cache-file prompt.safetensors --max-tokens 2048 --temp 0.2 
mlx_lm.generate --prompt-cache-file prompt.safetensors --max-tokens 2048 --temp 0.2 --seed 100

The first 2 should have exactly the same output, looping or not. The next 4 should all have different outputs.

[chigkim]

Hmm, the output doesn't seem to be any different. Am I supposed to look for something?

% METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 mlx_lm.generate --model mlx-community/Meta-Llama-3.1-8B-Instruct-4bit --max-kv-size 33000 --max-tokens 1000 --temp 0.0 --top-p 0.9 --seed 1000 --prompt -<../text/portugal.txt;say done
2024-11-13 22:46:18.124 python[16522:3704956] Metal API Validation Enabled
Fetching 6 files: 100%|████████████████████████| 6/6 [00:00<00:00, 45507.82it/s]
==========
Prompt: <|begin_of_text|><|start_header_id|>user<|end_header_id|>

Summarize the following:
......
<|eot_id|><|start_header_id|>assistant<|end_header_id|>


Portugal is a country located on the Iberian Peninsula in southwestern Europe.
.....
==========
Prompt: 32134 tokens, 414.286 tokens-per-sec
Generation: 1000 tokens, 32.263 tokens-per-sec
Peak memory: 12.535 GB

[awni]

Hmm, the output doesn't seem to be any different. Am I supposed to look for something?

Nope it would have been obvious if it threw a validation error. Thanks for checking.

[chigkim]

The first 2 should have exactly the same output, looping or not. The next 4 should all have different outputs.

Ooo, we're getting somewhere...
First 3 got loop, last 3 didn't.
Terminal Saved Output.txt