Quickstart | Installation | User Guide | Examples | FP8 Convergence | Integrations | Release notes
- [03/2024] Turbocharged Training: Optimizing the Databricks Mosaic AI stack with FP8
- [03/2024] FP8 Training Support in SageMaker Model Parallelism Library
- [12/2023] New NVIDIA NeMo Framework Features and NVIDIA H200
- [11/2023] Inflection-2: The Next Step Up
- [11/2023] Unleashing The Power Of Transformers With NVIDIA Transformer Engine
- [11/2023] Accelerating PyTorch Training Workloads with FP8
- [09/2023] Transformer Engine added to AWS DL Container for PyTorch Training
- [06/2023] Breaking MLPerf Training Records with NVIDIA H100 GPUs
- [04/2023] Benchmarking Large Language Models on NVIDIA H100 GPUs with CoreWeave (Part 1)
Transformer Engine (TE) is a library for accelerating Transformer models on NVIDIA GPUs, including using 8-bit floating point (FP8) precision on Hopper, Ada, and Blackwell GPUs, to provide better performance with lower memory utilization in both training and inference. TE provides a collection of highly optimized building blocks for popular Transformer architectures and an automatic mixed precision-like API that can be used seamlessly with your framework-specific code. TE also includes a framework agnostic C++ API that can be integrated with other deep learning libraries to enable FP8 support for Transformers.
As the number of parameters in Transformer models continues to grow, training and inference for architectures such as BERT, GPT and T5 become very memory and compute-intensive. Most deep learning frameworks train with FP32 by default. This is not essential, however, to achieve full accuracy for many deep learning models. Using mixed-precision training, which combines single-precision (FP32) with lower precision (e.g. FP16) format when training a model, results in significant speedups with minimal differences in accuracy as compared to FP32 training. With Hopper GPU architecture FP8 precision was introduced, which offers improved performance over FP16 with no degradation in accuracy. Although all major deep learning frameworks support FP16, FP8 support is not available natively in frameworks today.
TE addresses the problem of FP8 support by providing APIs that integrate with popular Large Language Model (LLM) libraries. It provides a Python API consisting of modules to easily build a Transformer layer as well as a framework-agnostic library in C++ including structs and kernels needed for FP8 support. Modules provided by TE internally maintain scaling factors and other values needed for FP8 training, greatly simplifying mixed precision training for users.
- Easy-to-use modules for building Transformer layers with FP8 support
- Optimizations (e.g. fused kernels) for Transformer models
- Support for FP8 on NVIDIA Hopper, Ada, and Blackwell GPUs
- Support for optimizations across all precisions (FP16, BF16) on NVIDIA Ampere GPU architecture generations and later
import torch
import transformer_engine.pytorch as te
from transformer_engine.common import recipe
# Set dimensions.
in_features = 768
out_features = 3072
hidden_size = 2048
# Initialize model and inputs.
model = te.Linear(in_features, out_features, bias=True)
inp = torch.randn(hidden_size, in_features, device="cuda")
# Create an FP8 recipe. Note: All input args are optional.
fp8_recipe = recipe.DelayedScaling(margin=0, fp8_format=recipe.Format.E4M3)
# Enable autocasting for the forward pass
with te.fp8_autocast(enabled=True, fp8_recipe=fp8_recipe):
out = model(inp)
loss = out.sum()
loss.backward()import flax
import jax
import jax.numpy as jnp
import transformer_engine.jax as te
import transformer_engine.jax.flax as te_flax
from transformer_engine.common import recipe
BATCH = 32
SEQLEN = 128
HIDDEN = 1024
# Initialize RNG and inputs.
rng = jax.random.PRNGKey(0)
init_rng, data_rng = jax.random.split(rng)
inp = jax.random.normal(data_rng, [BATCH, SEQLEN, HIDDEN], jnp.float32)
# Create an FP8 recipe. Note: All input args are optional.
fp8_recipe = recipe.DelayedScaling(margin=0, fp8_format=recipe.Format.HYBRID)
# Enable autocasting for the forward pass
with te.fp8_autocast(enabled=True, fp8_recipe=fp8_recipe):
model = te_flax.DenseGeneral(features=HIDDEN)
def loss_fn(params, other_vars, inp):
out = model.apply({'params':params, **other_vars}, inp)
return jnp.mean(out)
# Initialize models.
variables = model.init(init_rng, inp)
other_variables, params = flax.core.pop(variables, 'params')
# Construct the forward and backward function
fwd_bwd_fn = jax.value_and_grad(loss_fn, argnums=(0, 1))
for _ in range(10):
loss, (param_grads, other_grads) = fwd_bwd_fn(params, other_variables, inp)- Hardware: Blackwell, Hopper, Grace Hopper/Blackwell, Ada, Ampere
- OS: Linux (official), WSL2 (limited support)
- Software:
- CUDA: 12.1+ (Hopper/Ada/Ampere), 12.8+ (Blackwell) with compatible NVIDIA drivers
- cuDNN: 9.3+
- Compiler: GCC 9+ or Clang 10+ with C++17 support
- Python: 3.12 recommended
- Source Build Requirements: CMake 3.18+, Ninja, Git 2.17+, pybind11 2.6.0+
- Notes: FP8 features require Compute Capability 8.9+ (Ada/Hopper/Blackwell)
The quickest way to get started with Transformer Engine is by using Docker images on NVIDIA GPU Cloud (NGC) Catalog. For example to use the NGC PyTorch container interactively,
docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:25.01-py3Where 25.01 (corresponding to January 2025 release) is the container version.
Benefits of using NGC containers:
- All dependencies pre-installed with compatible versions and optimized configurations
- NGC PyTorch 23.08+ containers include FlashAttention-2
Prerequisites for pip installation:
- A compatible C++ compiler
- CUDA Toolkit with cuDNN and NVCC (NVIDIA CUDA Compiler) installed
To install the latest stable version with pip:
# For PyTorch integration
pip install --no-build-isolation transformer_engine[pytorch]
# For JAX integration
pip install --no-build-isolation transformer_engine[jax]
# For both frameworks
pip install --no-build-isolation transformer_engine[pytorch,jax]Alternatively, install directly from the GitHub repository:
pip install git+https://github.com/NVIDIA/TransformerEngine.git@stableWhen installing from GitHub, you can explicitly specify frameworks using the environment variable:
NVTE_FRAMEWORK=pytorch,jax pip install git+https://github.com/NVIDIA/TransformerEngine.git@stableThese environment variables can be set before installation to customize the build process:
- CUDA_PATH: Path to CUDA installation
- CUDNN_PATH: Path to cuDNN installation
- CXX: Path to C++ compiler
- NVTE_FRAMEWORK: Comma-separated list of frameworks to build for (e.g.,
pytorch,jax) - MAX_JOBS: Limit number of parallel build jobs (default varies by system)
- NVTE_BUILD_THREADS_PER_JOB: Control threads per build job
Transformer Engine supports both FlashAttention-2 and FlashAttention-3 in PyTorch for improved performance. FlashAttention-3 was added in release v1.11 and is prioritized over FlashAttention-2 when both are present in the environment.
You can verify which FlashAttention version is being used by setting these environment variables:
NVTE_DEBUG=1 NVTE_DEBUG_LEVEL=1 python your_script.pyIt is a known issue that FlashAttention-2 compilation is resource-intensive and requires a large amount of RAM (see bug), which may lead to out of memory errors during the installation of Transformer Engine. Please try setting MAX_JOBS=1 in the environment to circumvent the issue.
Common Issues and Solutions:
ABI Compatibility Issues:
- Symptoms:
ImportErrorwith undefined symbols when importing transformer_engine - Solution: Ensure PyTorch and Transformer Engine are built with the same C++ ABI setting. Rebuild PyTorch from source with matching ABI.
- Context: If you're using PyTorch built with a different C++ ABI than your system's default, you may encounter these undefined symbol errors. This is particularly common with pip-installed PyTorch outside of containers.
- Symptoms:
Missing Headers or Libraries:
Symptoms: CMake errors about missing headers (
cudnn.h,cublas_v2.h,filesystem, etc.)Solution: Install missing development packages or set environment variables to point to correct locations:
export CUDA_PATH=/path/to/cuda export CUDNN_PATH=/path/to/cudnn
If CMake can't find a C++ compiler, set the
CXXenvironment variable.Ensure all paths are correctly set before installation.
Build Resource Issues:
Symptoms: Compilation hangs, system freezes, or out-of-memory errors
Solution: Limit parallel builds:
MAX_JOBS=1 NVTE_BUILD_THREADS_PER_JOB=1 pip install ...
Verbose Build Logging:
For detailed build logs to help diagnose issues:
cd transformer_engine pip install -v -v -v --no-build-isolation .
In an effort to unify the definition and usage of the attention mask across all three frameworks in Transformer Engine, the padding mask has changed from True meaning inclusion of the corresponding position in attention to exclusion of that position in our PyTorch implementation. Since v1.7, all attention mask types follow the same definition where True means masking out the corresponding position and False means including that position in attention calculation.
An example of this change is,
# for a batch of 3 sequences where `a`s, `b`s and `c`s are the useful tokens
# and `0`s are the padding tokens,
[a, a, a, 0, 0,
b, b, 0, 0, 0,
c, c, c, c, 0]
# the padding mask for this batch before v1.7 is,
[ True, True, True, False, False,
True, True, False, False, False,
True, True, True, True, False]
# and for v1.7 onwards it should be,
[False, False, False, True, True,
False, False, True, True, True,
False, False, False, False, True]FP8 has been tested extensively across different model architectures and configurations and we found no significant difference between FP8 and BF16 training loss curves. FP8 has also been validated for accuracy on downstream LLM tasks (e.g. LAMBADA and WikiText). Below are examples of models tested for convergence across different frameworks.
| Model | Framework | Source |
|---|---|---|
| T5-770M | JAX/T5x | https://github.com/NVIDIA/JAX-Toolbox/tree/main/rosetta/rosetta/projects/t5x#convergence-and-performance |
| MPT-1.3B | Mosaic Composer | https://www.mosaicml.com/blog/coreweave-nvidia-h100-part-1 |
| GPT-5B | JAX/Paxml | https://github.com/NVIDIA/JAX-Toolbox/tree/main/rosetta/rosetta/projects/pax#h100-results |
| GPT-5B | NeMo Framework | Available on request |
| LLama2-7B | Alibaba Pai | https://mp.weixin.qq.com/s/NQT0uKXLbXyh5031zBdeBQ |
| T5-11B | JAX/T5x | Available on request |
| MPT-13B | Mosaic Composer | https://www.databricks.com/blog/turbocharged-training-optimizing-databricks-mosaic-ai-stack-fp8 |
| GPT-22B | NeMo Framework | Available on request |
| LLama2-70B | Alibaba Pai | https://mp.weixin.qq.com/s/NQT0uKXLbXyh5031zBdeBQ |
| GPT-175B | JAX/Paxml | https://github.com/NVIDIA/JAX-Toolbox/tree/main/rosetta/rosetta/projects/pax#h100-results |
Transformer Engine has been integrated with popular LLM frameworks such as:
- DeepSpeed
- Hugging Face Accelerate
- Lightning
- MosaicML Composer
- NVIDIA JAX Toolbox
- NVIDIA Megatron-LM
- NVIDIA NeMo Framework
- Amazon SageMaker Model Parallel Library
- Levanter
- GPT-NeoX
- Hugging Face Nanotron - Coming soon!
- Colossal-AI - Coming soon!
- PeriFlow - Coming soon!
We welcome contributions to Transformer Engine! To contribute to Transformer Engine and make pull requests, follow the guidelines outlined in the CONTRIBUTING.rst guide.
- Attention original paper
- Megatron-LM tensor parallel
- Megatron-LM sequence parallel
- FP8 Formats for Deep Learning
![@pre-commit-ci[bot]](https://avatars.githubusercontent.com/in/68672?s=64&v=4){kind=small}
