This repository enables Transformer Engine (TE) on ROCm as a library to accelerate Transformer models on AMD GPUs, including using 8-bit floating point (FP8) precision on MI300 GPUs, to provide better performance with lower memory utilization in both training and inference. One of the missions is to provide an alternative to accelerate Transformer models that were previously run on NVIDIA GPUs like Hopper with best efforts to make the migration frictionless. Moreover, we add optimizations specific to AMD GPUs to get the best performance benefits out of AMD GPUs.
- Activation, cast, fused softmax, layernorm, rmsnorm, transpose, fused rope, fp8 recipe, HipRTC: fully supported
- GEMM: partially supported with following input/output types: (fp32/fp32), (fp16/fp16), (bf16/bf16), (fp8, bf8/fp16, bf16, fp32)
- Attention (Flash Attention, Fused Multihead Attention): partially supported: Fused Attention with AOTriton and CK backends
- HipGraph, HipTX: partially supported
- Tensor Parallelism, Sequence Parallelism, Context Parallelism: supported
Execute the following commands to install ROCm Transformer Engine from source on AMDGPUs:
# Clone TE repo and submodules
git clone --recursive https://github.com/ROCm/TransformerEngine.git
cd TransformerEngine
export NVTE_FRAMEWORK=pytorch,jax #optionally set framework, currently only support pytorch and jax; if not set will try to detect installed frameworks
export NVTE_ROCM_ARCH=gfx942 # CK fused attn only support MI200 and MI300 and fp8 features are only supported on MI300
pip install .The default installation above supports both rocBlas and hipBlasLt in GEMM computation. Building with single backend support can be done by setting NVTE_USE_HIPBLASLT or NVTE_USE_ROCBLAS environment variable before pip install as:
export NVTE_USE_HIPBLASLT=1
export NVTE_USE_ROCBLAS=1When both GEMM backends are supported the aforementioned env variables can be used to select which backend to use. If none is set hipBlasLt is used by default. The hipBlasLt backed has not yet supported all the GEMM configurations in the pytorch unit tests.
After a successful Transformer Engine installation via pip install, execute the following commands to build and test the framework agnostic C++ library:
cd tests/cpp
mkdir build
cd build
cmake ../
make
make testNote that some of operator unit tests fail in hipBLASLt config due to limited input data configurations support
Pytorch integration pytests under tests/pytorch/ and tests/pytorch/fused_attn/ are supported Except the following tests that are not supported in rocBLAS configuration
- tests/pytorch/test_cuda_graph.py
- tests/pytorch/test_sanity.py::test_gpt_guda_graph
Env ROCBLAS_STREAM_ORDER_ALLOC=1 should be used when run tests in pytorch-rocblas configuration.
Also test_onnx_export.py does not support FP8 dues to absence of custom QDQ operatrs library
All JAX pytests are supported.
cd examples/pytorch/mnist
python main.py
python main.py --use-te # Linear layers from TransformerEngine
python main.py --use-fp8 # FP8 + TransformerEngine for Linear layerscd examples/pytorch/minGPT
python gptSort.py --use-te # Linear and layernorm from TransformerEngine
python gptSort.py --use-te --ln-mlp # In addition, use LayernormMLP from transformer engine
python gptSort.py --use-te --ln-mlp --use-fp8 # In addition, use fp8import 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, interval=1, 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)
# Update FP8 metas
other_variables = te.update_fp8_metas(other_grads)cd examples/jax/mnist
python test_single_gpu_mnist.py # Use Flax to train MNIST with BF16 as usual
python test_single_gpu_mnist.py --use-te # Use `te.DenseGeneral` provided by Transformer Engine to train MNIST with BF16
python test_single_gpu_mnist.py --use-fp8 # Use `te.DenseGeneral` provided by Transformer Engine to train MNIST and enable FP8 training and evaluation.cd examples/jax/encoder
python test_single_gpu_encoder.py
python test_single_gpu_encoder.py --use-fp8When using GEMM with hipBlasLt, TE provides an ability to manually or automatically select GPU algorithm to use from a list generated by hipBlasLt. Selected algorithms info can be stored to file and read on further applications run. This ability is controlled by environment variables when call GEMM operation with specific config for the first time.
- TE_HIPBLASLT_ALGO_SELECTION - algorithm index to use in the list returned by hipBlasLt for the config or the first algorithm to select from if auto-selection is enabled; default=0.
- TE_HIPBLASLT_TUNING_RUN_COUNT - number of profiling loops for algorithm auto-selection; default=0 which means no auto-selection. For small tasks where run-to-run time variation is relatively high, using higher number of loops may give better auto-selection results.
- TE_HIPBLASLT_TUNING_ALGO_COUNT - maximal number of algorithms to check when auto-selection is enabled; default=16.
- TE_HIPBLASLT_ALGO_LOAD - filename of algorithm selection data saved by previous GEMM operation runs; if file does not exist, algorithm selection logic proceeds as if no filename were specified
- TE_HIPBLASLT_ALGO_SAVE - filename to save algorithm selection data to; can be the same as a filename to load in which case the file will be read first and then overwritten with updated results
It is not guaranteed that algorithm selection data file created with one version of TE or hipBlasLt will work with other versions. Even if it works, it is highly recommended to perform algorithm selection tuning again when switch to new libraries versions because new hipBLASLt may have new optimized algorithms.
Typical usage is the following:
- Run single iteration of training enabling algorithm selection autotuning and saving:
export TE_HIPBLASLT_TUNING_RUN_COUNT=20
export TE_HIPBLASLT_TUNING_ALGO_COUNT=400
export TE_HIPBLASLT_ALGO_SAVE=algo_tune.csv
some_training_app- Use resulting algo_tune.csv for further training runs
unset TE_HIPBLASLT_TUNING_RUN_COUNT TE_HIPBLASLT_TUNING_ALGO_COUNT TE_HIPBLASLT_ALGO_SAVE #these variables are not needed anymore
export TE_HIPBLASLT_ALGO_LOAD=algo_tune.csv
some_training_appIf you want to check that only previously tuned algorithms are used by your application, it can be done by keeping selection data saving enabled.
export TE_HIPBLASLT_ALGO_SAVE=algo_tune_check.csv
export TE_HIPBLASLT_ALGO_LOAD=algo_tune.csv
some_training_app
#If the files are different, some not previously cached algorithms are probably used
diff algo_tune.csv algo_tune_check.csvCurrently ROCm TE supports two backends, AOTriton and CK, for fused attention. To enable specific backends in compilation and/or in runtime, the following environment variables can be used:
- NVTE_FUSED_ATTN - enable the fused attention, default = 1;
- NVTE_FUSED_ATTN_CK - enable the CK backend, default = 1;
- NVTE_FUSED_ATTN_AOTRITON - enable the AOTriton backend, default = 1.
Setting env NVTE_FUSED_ATTN_<BACKEND>=0 in compilation will skip the build of the specific backend, which saves the overall building time. Setting env NVTE_FUSED_ATTN_<BACKEND>=0 in runtime provides the option to choose specific backends in runtime. Note that one backend can be enabled in compilation but disabled in runtime. However, if one backend is disabled in compilation, the same env NVTE_FUSED_ATTN_<BACKEND>=0 is required during runtime. Otherwise TE will error out that the specific backend is not compiled.
NVTE_FUSED_ATTN has higher priority than NVTE_FUSED_ATTN_CK and NVTE_FUSED_ATTN_AOTRITON. NVTE_FUSED_ATTN=0 will use the TE unfused attention even if NVTE_FUSED_ATTN_CK or NVTE_FUSED_ATTN_AOTRITON is set. Fused attention backends are chosen according to the match results between the actual problem config and the support matrix of the specific backend. For the scenario that both backends are enabled and match the problem configuration, the CK backend will be chosen with higher priority.
Most CUDA kernels in Transformer Engine are hipified to run on ROCm. While the hipifiled CUDA kernels are functional, they are not necessarily optimal on ROCm. We added some Triton kernels to TE ROCm to improve the performance over the hipified kernels. Currently, we have integrated Triton kernels for cast_transpose and cast_transpose_bgrad, which are commonly used in fp8 training. This feature is still experimental as it requires relatievely newer version of Pytorch+Triton, which is not available in ROCm 6.2 Pytorch release docker images. Also, it only works on Pytorch extension as JAX extension does not use it.
To use this feature, you will need to first uninstall Pytorch and get the nightly version:
pip3 uninstall -y torch pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/rocm6.2
Then at runtime, you can enable this feature using the environment variable:
export NVTE_USE_CAST_TRANSPOSE_TRITON=1
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 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 and NVIDIA Ada 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)- Linux x86_64
- CUDA 11.8+ for Hopper and CUDA 12.1+ for Ada
- NVIDIA Driver supporting CUDA 11.8 or later
- cuDNN 8.1 or later
- For fused attention, CUDA 12.1 or later, NVIDIA Driver supporting CUDA 12.1 or later, and cuDNN 8.9 or later.
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:23.10-py3Where 23.10 is the container version. For example, 23.10 for the October 2023 release.
To install the latest stable version of Transformer Engine,
pip install git+https://github.com/NVIDIA/TransformerEngine.git@stableThis will automatically detect if any supported deep learning frameworks are installed and build Transformer Engine support for them. To explicitly specify frameworks, set the environment variable NVTE_FRAMEWORK to a comma-separated list (e.g. NVTE_FRAMEWORK=jax,pytorch).
Transformer Engine release v0.11.0 adds support for FlashAttention-2 in PyTorch for improved performance.
It 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.
Note that NGC PyTorch 23.08+ containers include FlashAttention-2.
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
- Hugging Face Nanotron - Coming soon!
- Colossal-AI - Coming soon!
- PeriFlow - Coming soon!
- GPT-NeoX - 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