Skip to content

Latest commit

 

History

History
72 lines (54 loc) · 2.63 KB

api.md

File metadata and controls

72 lines (54 loc) · 2.63 KB
Raw

API

The API is a web application to host generation. It runs as a standalone server and is built on top of Flask.

Currently, the API supports two endpoints:

  • /generate - which generates for a fixed number of tokens with hardcoded generation parameters.
  • /completions - which allows for setting sampling/beam parameters.

Launching the API

Prerequisites

Complete all of the setup as mentioned in the Setup doc.

Prepare checkpoints

  • Reshard the FSDP checkpoints using the script metaseq/scripts/reshard_fsdp.py. For example, we can merge all FSDP shards within each of the 8 model parallel parts of OPT-175B using the following command:

    for j in {0..7}; do
    python -m metaseq.scripts.reshard_fsdp \
    --input "/path/to/raw/checkpoints/checkpoint_last-model_part-$j-shard*.pt" \
    --output "/path/to/resharded/checkpoints/reshard-model_part-$j.pt" \
    --num-output-shards 1 --skip-optimizer-state True --unflatten-weights True
done

  • Update the paths CHECKPOINT_FOLDER, MODEL_FILE, BPE_MERGES, BPE_VOCAB as well as the configs MODEL_PARALLEL, TOTAL_WORLD_SIZE defined in metaseq/service/constants.py. For example,

    CHECKPOINT_FOLDER = "/path/to/resharded/checkpoints"
MODEL_FILE = os.path.join(CHECKPOINT_FOLDER, "reshard.pt")
BPE_MERGES = os.path.join(CHECKPOINT_FOLDER, "gpt2-merges.txt")
BPE_VOCAB = os.path.join(CHECKPOINT_FOLDER, "gpt2-vocab.json")
MODEL_PARALLEL = 8
TOTAL_WORLD_SIZE = 8

Run locally on a worker

metaseq-api-local

Note that when you ctrl-C, you may need to also run "killall python" or otherwise manually kill the processes. This is due to the batching thread, which is not properly ended on an interrupt.

Launching a worker from SLURM

srun --ntasks-per-node 1 --gpus-per-node $MODEL_PARALLEL --nodes 1 --cpus-per-task 8 --mem 400gb \
    --quit-on-interrupt --job-name genwork \
    python3 -m metaseq.cli.interactive_hosted

FAQ

How fast is generation?

Slow. As of 2022-04-7, QPS is about 40 with 3 workers.

Where can I run this?

Right now only on Azure, as it requires the 80GB A100s. To enable it on other locations, we need to either try CPU offloading, or we need to use MP 16. FSDP should not be used because some workers will only be used for parameter hosting, and will not actually perform computations.

Alternatively, one can run OPT-175B via the integration provided by the Alpa project, which enables serving OPT-175B with more flexible parallelisms on older generations of GPUs, such as 40GB A100, V100, T4, M60, etc.