Skip to content

Latest commit

 

History

History
105 lines (78 loc) · 4.77 KB

parallel.md

File metadata and controls

105 lines (78 loc) · 4.77 KB
Raw

Multithreading and parallelism

Intra-op multithreading on CPU

Most model operations (matmul, softmax, etc.) are using multiple threads on CPU. The number of threads can be configured with the parameter intra_threads (the default value is 4):

translator = ctranslate2.Translator(model_path, device="cpu", intra_threads=8)

This multithreading is generally implemented with OpenMP so the threads behavior can also be customized with the different OMP_* environment variables.

When OpenMP is disabled (which is the case for example in the Python ARM64 wheels for macOS), the multithreading is implemented with BS::thread_pool.

Data parallelism

Objects running models such as the Translator and Generator can be configured to process multiple batches in parallel, including on multiple GPUs:

# Create a CPU translator with 4 workers each using 1 intra-op thread:
translator = ctranslate2.Translator(model_path, device="cpu", inter_threads=4, intra_threads=1)

# Create a GPU translator with 4 workers each running on a separate GPU:
translator = ctranslate2.Translator(model_path, device="cuda", device_index=[0, 1, 2, 3])

# Create a GPU translator with 4 workers each using a different CUDA stream:
# (Note: depending on the workload and GPU specifications this may not improve the global throughput.)
translator = ctranslate2.Translator(model_path, device="cuda", inter_threads=4)

When the workers are running on the same device, the model weights are shared to save on memory.

Multiple batches should be submitted concurrently to enable this parallelization. Parallel executions are enabled in the following cases:

  • When calling methods from multiple Python threads.
  • When calling methods multiple times with asynchronous=True.
  • When calling file-based or stream-based methods.
  • When setting max_batch_size: the input will be split according to max_batch_size and each sub-batch will be executed in parallel.
Parallelization with multiple Python threads is possible because all computation methods release the [Python GIL](https://wiki.python.org/moin/GlobalInterpreterLock).

Model and tensor parallelism

Models used with Translator and Generator can be split into multiple GPUs. This is very useful when the model is too big to be loaded in only 1 GPU.

translator = ctranslate2.Translator(model_path, device="cuda", tensor_parallel=True)

Setup environment:

  • Install open-mpi
  • Configure open-mpi by creating the config file like hostfile:
[ipaddress or dns] slots=nbGPU1
[other ipaddress or dns] slots=NbGPU2

Run:

  • Run the application in multiprocess to use tensor parallel:
mpirun -np nbGPUExpected -hostfile hostfile python3 script

If you're trying to use tensor parallelism in multiple machines, some additional configuration is needed:

  • Make sure Master and Slave can connect to each other as a pair with ssh + pubkey
  • Export all necessary environment variables from Master to Slave like the example below:
mpirun -x VIRTUAL_ENV_PROMPT -x PATH -x VIRTUAL_ENV -x _ -x LD_LIBRARY_PATH -np nbGPUExpected -hostfile hostfile python3 script

Read more open-mpi docs for more information.

  • In this mode, the application will run in multiprocess. We can filter out the master process by using:
if ctranslate2.MpiInfo.getCurRank() == 0:
    print(...)
Running model in tensor parallel mode in one machine can boost the performance but if the model shared between multiple machines
could be slower because of the latency in the connectivity.

In mode tensor parallel, `inter_threads` is always supported to run multiple workers. Otherwise, `device_index` no longer has any effect
because tensor parallel mode will check only for available gpus on the system and the number of gpus you want to use.

Asynchronous execution

Some methods can run asynchronously with asynchronous=True. In this mode, the method returns immediately and the result can be retrieved later:

async_results = []
for batch in batch_generator():
    async_results.extend(translator.translate_batch(batch, asynchronous=True))

for async_result in async_results:
    print(async_result.result())  # This method blocks until the result is available.
Instances supporting asynchronous execution have a limited queue size by default. When the queue of batches is full, the method will block even with `asynchronous=True`. See the parameter `max_queued_batches` in their constructor to configure the queue size.