tensor_parallel

Run large PyTorch models on multiple GPUs in one line of code.

import transformers
import tensor_parallel as tp
tokenizer = transformers.AutoTokenizer.from_pretrained("facebook/opt-13b")
model = transformers.AutoModelForCausalLM.from_pretrained("facebook/opt-13b")  # use opt-125m for testing

model = tp.tensor_parallel(model, ["cuda:0", "cuda:1"])  # <- each GPU has half the weights

inputs = tokenizer("A cat sat", return_tensors="pt")["input_ids"].to("cuda:0")
outputs = model.generate(inputs, num_beams=5)
print(tokenizer.decode(outputs[0])) # A cat sat on my lap for a few minutes ...

model(input_ids=inputs, labels=inputs).loss.backward()  # training works as usual

Installation

Latest stable version (recommended):

pip install tensor_parallel

Bleeding edge version:

pip install https://github.com/BlackSamorez/tensor_parallel/archive/main.zip

Usage

Simply wrap your PyTorch model with tp.tensor_parallel and use it normally. For best memory efficiency, call tp.tensor_parallel while the model is still on CPU.

Here are a few use cases:

• examples/training_flan-t5-xl.ipynb - fine-tune full FLAN-T5 model on text summarization
• TBA - inferencing a large language model with LLM.8bit + tensor_parallel
• TBA - defining custom parallelism strategy

Advanced parameters to tensor_parallel:

• device_ids: List[device] - which devices to use; defaults to all available GPUs
• output_device: device - model outputs will have this device
• config: tp.Config - use custom parallelism strategy, see slicing_configs.py
• distributed: bool - if True, use torch.distributed backend instead of threading (requires torchrun)
• sharded: bool - if True, find all trainable parameters that weren't split by Tensor Parallelism and split them using ZeRO-3 algorithm.
  • weights will be split between GPUs and re-assembled before each forward pass
  • TL;DR use this when training to avoid duplicate parameters (enabled by default!)
  • sharded_param_names: List[str] - parameter names that should be sharded this way, default = found automatically

Saving the model

To save a model such that it could be used in a non tensor_parallel context, you should use a save_tensor_parallel context wrapper.

import torch
import transformers
import tensor_parallel as tp

model = tp.tensor_parallel(
    transformers.AutoModelForCausalLM.from_pretrained("facebook/opt-13b"), 
)

# A whole lot of trainig...

with tp.save_tensor_parallel(model):
    torch.save(model.state_dict(), "/tmp/")
    # or 
    model.save_pretrained("/tmp/")

Such code saves a model as if it was never split. It works by gathering model parts during state_dict creation.

Memory efficient dispatch

Normally, to normally create and dispatch a tensor_parallel model, one needs the whole model in memory. This can be troublesome, but there is another way.

It's possible to create a tensor_parallel on a machine with enough RAM, save it preserving distributed state, and then reuse the save files to dispatch model shards straight to GPUs on any other machine.

The code to save distributed state should look like this:

import transformers
import tensor_parallel as tp

RAM_LIMIT = "6GB" # we want to deploy the model on machines with as little as 6GB of RAM

model = tp.TensorParallelPreTrainedModel(
    transformers.AutoModelForCausalLM.from_pretrained("facebook/opt-13b"), 
    device_ids=["cpu", "cpu"] # split the model but it load into RAM
)

model.save_pretrained("opt-13b-tensor-parallel", max_shard_size=RAM_USAGE_LIMIT) # save the model's distributed state

It normally produces many files containing model weights as well as model index. Those files then can be put on a machine for training/inference and loaded as follows:

import transformers
import tensor_parallel as tp

from accelerate import init_empty_weights, load_checkpoint_in_model

# Initialize a weightless model
with init_empty_weights():
    model = tp.TensorParallelPreTrainedModel(
        transformers.AutoModelForCausalLM.from_config(
            AutoConfig.from_pretrained("facebook/opt-13b")
        ),
        device_ids=[0, 1] # and prepare it to be put on GPUs 0 and 1
    )

device_map = tp.infer_sharded_device_map(model_tp) # assign parameter to devices

# Incrementally load the weights using your favorite loader
load_checkpoint_in_model(
    model,
    checkpoint="opt-13b-tensor-parallel/pytorch_model.bin.index.json",
    device_map=device_map,
)

Max RAM consumption of such loading is max_shard_size, which in this example was set to 6 GB.

FAQ

• Q: I don't have a multi-GPU server. Can I use tensor_parallel in Google Colab?

• A: Colab has a single GPU, so there's no point in tensor parallelism. However, Kaggle offers two T4 for free to all phone-verified accounts.

• Q: What is tensor parallelism?

• A: You split each layer's weights into parts, multiply each part on a separate GPU, then gather results. Read more here

• Q: Should I use TensorParallel or DataParallel?

• A: TensorParallel for large models, DataParallel for smaller ones

• Q: How does it compare against FullyShardedDataParallel and ZeRO?

• A: ZeRO is better if you can fit a large batch, TensorParallel is better for small batches

Why use tensor_parallel ...

• v.s. DeepSpeed and FairScale
  • DeepSpeed has many parallelization strategies, but requires careful configuration
  • tensor_parallel has one strategy that works with 1 line of code
  • tensor_parallel works in a jupyter notebook
• v.s. MegatronLM?
  • MegatronLM has great tensor parallelism for one model architecture
  • tensor_parallel has good parallelism for any architecture
  • tensor_parallel is way easier to install
• v.s. parallelformers?
  • parallelformers implements a fixed list of architectures
  • tensor_parallel works for any architecture automatically
  • parallelformers is inference-only, tensor_parallel supports training
• v.s. alpa
  • alpa is a powerful tool for automatic distributed training / inference in JAX
  • tensor_parallel works with PyTorch
• v.s. Model.parallelize()?
  • both are easy to use, both fit large models
  • in parallelize, one GPU works at a time
  • in tensor_parallel, GPUs work in parallel

In short, use tensor_parallel for quick prototyping on a single machine. Use DeepSpeed+Megatron or alpa for million-dollar training runs.

Troubleshooting

If you experience NCCL errors, or random hanging, you may have some code errors that are not displayed properly. To debug these errors, we recommend restarting with export TENSOR_PARALLEL_USE_NATIVE=1 or on a single device.

If you found a bug or encountered a problem, please report it to our issue tracker. We will do our best to help, but it may take some time before we get to it. Please create issues only if your problem is specifically with tensor_parallel. For example, if you need help installing transformers or optimizing your code, please seek it elsewhere.

Code style

We use black and isort for all pull requests. Before committing your code, simply run black . && isort . and you will be fine.