Underloaded Neuron Cores with Llama3

[dlptv]

System Info

TGI Image: ghcr.io/huggingface/neuronx-tgi:0.0.23

Platform:

- Platform: Linux-5.15.0-1031-aws-x86_64-with-glibc2.35
- Python version: 3.10.12

Python packages:

- `optimum-neuron` version: 0.0.23
- `neuron-sdk` version: 2.18.0
- `optimum` version: 1.20.0
- `transformers` version: 4.41.1
- `huggingface_hub` version: 0.23.2
- `torch` version: 2.1.2+cu121
- `aws-neuronx-runtime-discovery` version: 2.9
- `libneuronxla` version: 2.0.965
- `neuronx-cc` version: 2.13.66.0+6dfecc895
- `neuronx-distributed` version: NA
- `neuronx-hwm` version: NA
- `torch-neuronx` version: 2.1.2.2.1.0
- `torch-xla` version: 2.1.2
- `transformers-neuronx` version: 0.10.0.21

Neuron Driver:

aws-neuronx-collectives/now 2.20.22.0-c101c322e amd64 [installed,local]
aws-neuronx-dkms/now 2.16.7.0 amd64 [installed,local]
aws-neuronx-runtime-lib/now 2.20.22.0-1b3ca6425 amd64 [installed,local]
aws-neuronx-tools/now 2.17.1.0 amd64 [installed,local]

Who can help?

@dacorvo

Information

• [ ] The official example scripts
• [X] My own modified scripts

Tasks

• [X] An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• [ ] My own task or dataset (give details below)

Reproduction (minimal, reproducible, runnable)

I use optimum-cli to export Llama3 model from the official repository with this command:

optimum-cli export neuron --model meta-llama/Meta-Llama-3-8B \ --batch_size 1 \ --sequence_length 8192 \ --auto_cast_type fp16 ` \ --num_cores 24 \ /data/llama3_neuron/

I use the same image for that as I use to run TGI: ghcr.io/huggingface/neuronx-tgi:0.0.23

Then I run the TGI container with this command: docker run -p 8080:80 \ --rm \ -it \ -v $(pwd)/data:/data \ --device=/dev/neuron0 \ --device=/dev/neuron1 \ --device=/dev/neuron2 \ --device=/dev/neuron3 \ --device=/dev/neuron4 \ --device=/dev/neuron5 \ --device=/dev/neuron6 \ --device=/dev/neuron7 \ --device=/dev/neuron8 \ --device=/dev/neuron9 \ --device=/dev/neuron10 \ --device=/dev/neuron11 \ --privileged \ -e HF_TOKEN=... \ -e HF_AUTO_CAST_TYPE="fp16" \ -e HF_NUM_CORES=24 \ ghcr.io/huggingface/neuronx-tgi:latest \ --model-id /data/llama3_neuron \ --max-batch-size 1 \ --max-input-length 3164 \ --max-total-tokens 8192

I run simple load tests with locust, emulating 10 users sending concurrent requests of variable length, and see that all neuron cores are loaded around 60% only and the performance is far from ideal (0.1-0.2 rps):

Expected behavior

Neuron cores are loaded closely to 100%.

[dacorvo]

@dlptv, thank you for your feedback. I see that you added a token in the code snippet you submitted: this is not safe to publicly share hugginface tokens, even with low rights, so please invalidate it as soon as possible.

The reason why your cores are not used at full capacity is because you are using a batch size of 1, meaning that requests are processed in sequence instead of being parallelized. Try with a batch_size of 10 and you will see a difference.

[github-actions[bot]]

This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.