Huggingface_accelerate

Loading large Huggingface models with constrained resources using accelerate

This document briefs on serving large HG models with limited resource using accelerate. This option can be activated with low_cpu_mem_usage=True. The model is first created on the Meta device (with empty weights) and the state dict is then loaded inside it (shard by shard in the case of a sharded checkpoint).

Step 1: Download model

Login into huggingface hub with token by running the below command

huggingface-cli login

paste the token generated from huggingface hub.

python Download_model.py --model_name bigscience/bloom-7b1

The script prints the path where the model is downloaded as below.

model/models--bigscience-bloom-7b1/snapshots/5546055f03398095e385d7dc625e636cc8910bf2/

The downloaded model is around 14GB.

Step 2: Compress downloaded model

NOTE: Install Zip cli tool

Navigate to the path got from the above script. In this example it is

cd model/models--bigscience-bloom-7b1/snapshots/5546055f03398095e385d7dc625e636cc8910bf2/
zip -r /home/ubuntu/serve/examples/Huggingface_Largemodels//model.zip *
cd -

Step 3: Generate MAR file

Navigate up to Huggingface_Largemodels directory.

torch-model-archiver --model-name bloom --version 1.0 --handler custom_handler.py --extra-files model.zip,setup_config.json -r requirements.txt

Note: Modifying setup_config.json

• Enable low_cpu_mem_usage to use accelerate
• Recommended max_memory in setup_config.json is the max size of shard.
• Refer: https://huggingface.co/docs/transformers/main_classes/model#large-model-loading

Step 4: Add the mar file to model store

mkdir model_store
mv bloom.mar model_store

Step 5: Start torchserve

Update config.properties and start torchserve

torchserve --start --ncs --ts-config config.properties --disable-token-auth  --enable-model-api

Step 5: Run inference

curl -v "http://localhost:8080/predictions/bloom" -T sample_text.txt