Distributed: HF Accelerate
This tutorial demonstrates how to use Runhouse with HuggingFace accelerate to launch distributed code on your own remote hardware. We also show how one can reproducibly perform hardware dependency autosetup, to ensure that your code runs smoothly every time.
You can run this on your own cluster, or through a standard cloud account (AWS, GCP, Azure, LambdaLabs). If you do not have any compute or cloud accounts set up, we recommend creating a LambdaLabs account for the easiest setup path.
Install dependencies
!pip install accelerate
!pip install runhouse
import runhouse as rh
INFO | 2023-03-20 17:56:13,023 | No auth token provided, so not using RNS API to save and load configs
INFO | 2023-03-20 17:56:14,334 | NumExpr defaulting to 2 threads.
Setting up the Cluster
On-Demand Cluster (AWS, Azure, GCP, or LambdaLabs)
For instructions on setting up cloud access for on-demand clusters, please refer to Cluster Setup.
# single V100 GPU
# gpu = rh.ondemand_cluster(name="rh-v100", instance_type="V100:1").up_if_not()
# multigpu: 4 V100s
gpu = rh.ondemand_cluster(name="rh-4-v100", instance_type="V100:4").up_if_not()
# Set GPU to autostop after 60 min of inactivity (default is 30 min)
gpu.keep_warm(60) # or -1 to keep up indefinitely
On-Premise Cluster
For an on-prem cluster, you can instantaite it as follows, filling in the IP address, ssh user and private key path.
# For an existing cluster
# gpu = rh.cluster(ips=['<ip of the cluster>'],
# ssh_creds={'ssh_user': '...', 'ssh_private_key':'<path_to_key>'},
# name='rh-cluster')
Setting up Functions on Remote Hardware
Training Function
For simplicity, let’s use the training_function from accelerate/examples/nlp_example.py to demonstrate how to run this function remotely.
In this case, because the function is available on GitHub, we can pass in a string pointing to the GitHub function.
For local functions, for instance if we had nlp_example.py in our
directory, we can also simply import the function.
# if nlp_example.py is in local directory
# from nlp_example import training_function
# if function is available on GitHub, use it's string representation
training_function = "https://github.com/huggingface/accelerate/blob/v0.15.0/examples/nlp_example.py:training_function"
Next, define the dependencies necessary to run the imported training function using accelerate.
reqs = ['pip:./accelerate', 'transformers', 'datasets', 'evaluate','tqdm', 'scipy', 'scikit-learn', 'tensorboard',
'torch --upgrade --extra-index-url https://download.pytorch.org/whl/cu117']
Now, we can put together the above components (gpu cluster, training function, and dependencies) to create our train function on remote hardware.
train_function_gpu = rh.function(
fn=training_function,
system=gpu,
reqs=reqs,
)
INFO | 2023-03-20 21:01:46,942 | Setting up Function on cluster.
INFO | 2023-03-20 21:01:46,951 | Installing packages on cluster rh-v100: ['GitPackage: https://github.com/huggingface/accelerate.git@v0.15.0', 'pip:./accelerate', 'transformers', 'datasets', 'evaluate', 'tqdm', 'scipy', 'scikit-learn', 'tensorboard', 'torch --upgrade --extra-index-url https://download.pytorch.org/whl/cu117']
INFO | 2023-03-20 21:02:02,988 | Function setup complete.
train_function_gpu is a callable that can be used just like the
original training_function function in the NLP example, except that
it runs the function on the specified cluster/system instead.
Launch Helper Function
Here we define a helper function for launching accelerate training, and then send the function to run on our GPU as well
def launch_training(training_function, *args):
from accelerate.utils import PrepareForLaunch, patch_environment
import torch
num_processes = torch.cuda.device_count()
print(f'Device count: {num_processes}')
with patch_environment(world_size=num_processes, master_addr="127.0.01", master_port="29500",
mixed_precision=args[1].mixed_precision):
launcher = PrepareForLaunch(training_function, distributed_type="MULTI_GPU")
torch.multiprocessing.start_processes(launcher, args=args, nprocs=num_processes, start_method="spawn")
launch_training_gpu = rh.function(fn=launch_training).to(gpu)
INFO | 2023-03-20 19:56:15,257 | Writing out function function to /content/launch_training_fn.py as functions serialized in notebooks are brittle. Please make sure the function does not rely on any local variables, including imports (which should be moved inside the function body).
INFO | 2023-03-20 19:56:15,262 | Setting up Function on cluster.
INFO | 2023-03-20 19:56:15,265 | Copying local package content to cluster <rh-v100>
INFO | 2023-03-20 19:56:20,623 | Installing packages on cluster rh-v100: ['./']
INFO | 2023-03-20 19:56:20,753 | Function setup complete.
Launch Distributed Training
Now, we’re ready to launch distributed training on our self-hosted hardware!
import argparse
# define basic train args and hyperparams
train_args = argparse.Namespace(cpu=False, mixed_precision='fp16')
hps = {"lr": 2e-5, "num_epochs": 3, "seed": 42, "batch_size": 16}
launch_training_gpu(train_function_gpu, hps, train_args, stream_logs=True)
INFO | 2023-03-20 20:11:45,415 | Running launch_training via gRPC
INFO | 2023-03-20 20:11:45,718 | Time to send message: 0.3 seconds
INFO | 2023-03-20 20:11:45,720 | Submitted remote call to cluster. Result or logs can be retrieved
with run_key "launch_training_20230320_201145", e.g.
rh.cluster(name="~/rh-v100").get("launch_training_20230320_201145", stream_logs=True) in python
runhouse logs "rh-v100" launch_training_20230320_201145 from the command line.
or cancelled with
rh.cluster(name="~/rh-v100").cancel("launch_training_20230320_201145") in python or
runhouse cancel "rh-v100" launch_training_20230320_201145 from the command line.
:task_name:launch_training
:task_name:launch_training
INFO | 2023-03-20 20:11:46,328 | Loading config from local file /home/ubuntu/runhouse/runhouse/builtins/config.json
INFO | 2023-03-20 20:11:46,328 | No auth token provided, so not using RNS API to save and load configs
Device count: 1
INFO | 2023-03-20 20:11:49,486 | Loading config from local file /home/ubuntu/runhouse/runhouse/builtins/config.json
INFO | 2023-03-20 20:11:49,486 | No auth token provided, so not using RNS API to save and load configs
INFO | 2023-03-20 20:11:49,844 | Appending /home/ubuntu/accelerate/examples to sys.path
INFO | 2023-03-20 20:11:49,844 | Importing module nlp_example
Downloading builder script: 0%| | 0.00/5.75k [00:00<?, ?B/s]
Downloading builder script: 100%|██████████| 5.75k/5.75k [00:00<00:00, 8.19MB/s]
Downloading (…)okenizer_config.json: 0%| | 0.00/29.0 [00:00<?, ?B/s]
Downloading (…)okenizer_config.json: 100%|██████████| 29.0/29.0 [00:00<00:00, 65.9kB/s]
Downloading (…)lve/main/config.json: 0%| | 0.00/570 [00:00<?, ?B/s]
Downloading (…)lve/main/config.json: 100%|██████████| 570/570 [00:00<00:00, 1.22MB/s]
Downloading (…)solve/main/vocab.txt: 0%| | 0.00/213k [00:00<?, ?B/s]
Downloading (…)solve/main/vocab.txt: 100%|██████████| 213k/213k [00:00<00:00, 1.05MB/s]
Downloading (…)solve/main/vocab.txt: 100%|██████████| 213k/213k [00:00<00:00, 1.05MB/s]
Downloading (…)/main/tokenizer.json: 0%| | 0.00/436k [00:00<?, ?B/s]
Downloading (…)/main/tokenizer.json: 100%|██████████| 436k/436k [00:00<00:00, 1.61MB/s]
Downloading (…)/main/tokenizer.json: 100%|██████████| 436k/436k [00:00<00:00, 1.60MB/s]
Downloading builder script: 0%| | 0.00/28.8k [00:00<?, ?B/s]
Downloading builder script: 100%|██████████| 28.8k/28.8k [00:00<00:00, 380kB/s]
Downloading metadata: 0%| | 0.00/28.7k [00:00<?, ?B/s]
Downloading metadata: 100%|██████████| 28.7k/28.7k [00:00<00:00, 422kB/s]
Downloading readme: 0%| | 0.00/27.9k [00:00<?, ?B/s]
Downloading readme: 100%|██████████| 27.9k/27.9k [00:00<00:00, 412kB/s]
Downloading data files: 0%| | 0/3 [00:00<?, ?it/s]Downloading and preparing dataset glue/mrpc to /home/ubuntu/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad...
Downloading data: 0.00B [00:00, ?B/s][A
Downloading data: 6.22kB [00:00, 11.1MB/s]
Downloading data files: 33%|███▎ | 1/3 [00:00<00:00, 4.26it/s]
Downloading data: 0.00B [00:00, ?B/s][A
Downloading data: 1.05MB [00:00, 55.0MB/s]
Downloading data files: 67%|██████▋ | 2/3 [00:00<00:00, 5.30it/s]
Downloading data: 0.00B [00:00, ?B/s][A
Downloading data: 441kB [00:00, 44.3MB/s]
Downloading data files: 100%|██████████| 3/3 [00:00<00:00, 5.87it/s]
Downloading data files: 100%|██████████| 3/3 [00:00<00:00, 5.56it/s]
Generating train split: 0%| | 0/3668 [00:00<?, ? examples/s]
Generating train split: 79%|███████▉ | 2898/3668 [00:00<00:00, 28934.98 examples/s]
Generating validation split: 0%| | 0/408 [00:00<?, ? examples/s]
Generating test split: 0%| | 0/1725 [00:00<?, ? examples/s]
0%| | 0/3 [00:00<?, ?it/s]
100%|██████████| 3/3 [00:00<00:00, 1296.81it/s]
Map: 0%| | 0/3668 [00:00<?, ? examples/s]
Map: 100%|██████████| 3668/3668 [00:00<00:00, 33587.18 examples/s]
Map: 0%| | 0/408 [00:00<?, ? examples/s]
Map: 0%| | 0/1725 [00:00<?, ? examples/s]
Dataset glue downloaded and prepared to /home/ubuntu/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad. Subsequent calls will reuse this data.
Downloading pytorch_model.bin: 0%| | 0.00/436M [00:00<?, ?B/s]
Downloading pytorch_model.bin: 2%|▏ | 10.5M/436M [00:00<00:04, 95.8MB/s]
Downloading pytorch_model.bin: 5%|▍ | 21.0M/436M [00:00<00:04, 97.1MB/s]
Downloading pytorch_model.bin: 7%|▋ | 31.5M/436M [00:00<00:04, 93.2MB/s]
Downloading pytorch_model.bin: 10%|▉ | 41.9M/436M [00:00<00:04, 91.3MB/s]
Downloading pytorch_model.bin: 12%|█▏ | 52.4M/436M [00:00<00:04, 92.6MB/s]
Downloading pytorch_model.bin: 14%|█▍ | 62.9M/436M [00:00<00:04, 86.0MB/s]
Downloading pytorch_model.bin: 17%|█▋ | 73.4M/436M [00:00<00:04, 89.9MB/s]
Downloading pytorch_model.bin: 19%|█▉ | 83.9M/436M [00:00<00:03, 90.2MB/s]
Downloading pytorch_model.bin: 22%|██▏ | 94.4M/436M [00:01<00:03, 91.5MB/s]
Downloading pytorch_model.bin: 24%|██▍ | 105M/436M [00:01<00:03, 93.3MB/s]
Downloading pytorch_model.bin: 26%|██▋ | 115M/436M [00:01<00:03, 86.5MB/s]
Downloading pytorch_model.bin: 29%|██▉ | 126M/436M [00:01<00:03, 86.9MB/s]
Downloading pytorch_model.bin: 31%|███▏ | 136M/436M [00:01<00:03, 87.2MB/s]
Downloading pytorch_model.bin: 34%|███▎ | 147M/436M [00:01<00:03, 88.6MB/s]
Downloading pytorch_model.bin: 36%|███▌ | 157M/436M [00:01<00:03, 90.7MB/s]
Downloading pytorch_model.bin: 38%|███▊ | 168M/436M [00:01<00:02, 90.4MB/s]
Downloading pytorch_model.bin: 41%|████ | 178M/436M [00:02<00:03, 82.5MB/s]
Downloading pytorch_model.bin: 43%|████▎ | 189M/436M [00:02<00:02, 84.6MB/s]
Downloading pytorch_model.bin: 46%|████▌ | 199M/436M [00:02<00:02, 81.3MB/s]
Downloading pytorch_model.bin: 48%|████▊ | 210M/436M [00:02<00:02, 84.4MB/s]
Downloading pytorch_model.bin: 51%|█████ | 220M/436M [00:02<00:02, 83.4MB/s]
Downloading pytorch_model.bin: 53%|█████▎ | 231M/436M [00:02<00:02, 86.4MB/s]
Downloading pytorch_model.bin: 55%|█████▌ | 241M/436M [00:02<00:02, 88.9MB/s]
Downloading pytorch_model.bin: 58%|█████▊ | 252M/436M [00:02<00:02, 90.9MB/s]
Downloading pytorch_model.bin: 60%|██████ | 262M/436M [00:02<00:01, 91.6MB/s]
Downloading pytorch_model.bin: 63%|██████▎ | 273M/436M [00:03<00:01, 90.9MB/s]
Downloading pytorch_model.bin: 65%|██████▍ | 283M/436M [00:03<00:01, 90.8MB/s]
Downloading pytorch_model.bin: 67%|██████▋ | 294M/436M [00:03<00:01, 91.6MB/s]
Downloading pytorch_model.bin: 70%|██████▉ | 304M/436M [00:03<00:01, 92.1MB/s]
Downloading pytorch_model.bin: 72%|███████▏ | 315M/436M [00:03<00:01, 91.9MB/s]
Downloading pytorch_model.bin: 75%|███████▍ | 325M/436M [00:03<00:01, 91.0MB/s]
Downloading pytorch_model.bin: 77%|███████▋ | 336M/436M [00:03<00:01, 89.7MB/s]
Downloading pytorch_model.bin: 79%|███████▉ | 346M/436M [00:03<00:00, 90.2MB/s]
Downloading pytorch_model.bin: 82%|████████▏ | 357M/436M [00:03<00:00, 92.1MB/s]
Downloading pytorch_model.bin: 84%|████████▍ | 367M/436M [00:04<00:00, 93.5MB/s]
Downloading pytorch_model.bin: 87%|████████▋ | 377M/436M [00:04<00:00, 93.5MB/s]
Downloading pytorch_model.bin: 89%|████████▉ | 388M/436M [00:04<00:00, 92.9MB/s]
Downloading pytorch_model.bin: 91%|█████████▏| 398M/436M [00:04<00:00, 81.5MB/s]
Downloading pytorch_model.bin: 94%|█████████▍| 409M/436M [00:04<00:00, 83.7MB/s]
Downloading pytorch_model.bin: 96%|█████████▌| 419M/436M [00:04<00:00, 85.6MB/s]
Downloading pytorch_model.bin: 99%|█████████▊| 430M/436M [00:04<00:00, 80.6MB/s]
Downloading pytorch_model.bin: 100%|██████████| 436M/436M [00:04<00:00, 88.2MB/s]
Some weights of the model checkpoint at bert-base-cased were not used when initializing BertForSequenceClassification: ['cls.predictions.transform.LayerNorm.bias', 'cls.predictions.transform.dense.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.decoder.weight', 'cls.seq_relationship.weight', 'cls.predictions.bias']
- This IS expected if you are initializing BertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing BertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
Some weights of BertForSequenceClassification were not initialized from the model checkpoint at bert-base-cased and are newly initialized: ['classifier.weight', 'classifier.bias']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
You're using a BertTokenizerFast tokenizer. Please note that with a fast tokenizer, using the __call__ method is faster than using a method to encode the text followed by a call to the pad method to get a padded encoding.
epoch 0: {'accuracy': 0.7745098039215687, 'f1': 0.8557993730407523}
epoch 1: {'accuracy': 0.8406862745098039, 'f1': 0.8849557522123894}
epoch 2: {'accuracy': 0.8553921568627451, 'f1': 0.8981001727115717}
Terminate Cluster
Once you are done using the cluster, you can terminate it as follows:
gpu.teardown()