nouiz / transformerengine Goto Github PK

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 own 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 the introduction of 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 today.

TE addresses the problem of FP8 support by providing APIs that integrate with popular Large Language Model (LLM) libraries. It provides python layer consisting of modules to easily build Transformer layer as well as 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 for the users.

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, interval=1, 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 jax
import jax.numpy as jnp
import transformer_engine.jax as te
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.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 = variables.pop('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)

• Easy-to-use modules enabling building of the Transformer layers with FP8 support on H100 GPUs.
• Optimizations (e.g. fused kernels) for Transformer models across all precisions and NVIDIA GPU architectures.

Transformer Engine comes preinstalled in the pyTorch container on NVIDIA GPU Cloud (versions 22.09 and later).

Clone the repository and inside it type:

NVTE_FRAMEWORK=all pip install .     # Building with all frameworks.
NVTE_FRAMEWORK=pytorch pip install . # Building with pyTorch only.
NVTE_FRAMEWORK=jax pip install .     # Building with JAX only.

For examples, tutorials and API reference please refer to the User Guide.

While the more granular modules in Transformer Engine allow building any Transformer architecture, the TransformerLayer API of Transformer Engine is flexible enough to build multiple major variations of Transformers.

NOTE: For simplicity, we only show pyTorch examples below. For the usage of TransformerLayer of all supported frameworks, refer to examples.

GPT architecture has LayerNorm at the input side (before QKV Gemm) and the residual connection is taken from the input of that LayerNorm. In TE this can be achieved by setting the following arguments in the TransformerLayer API.

transformer_engine.pytorch.TransformerLayer(
        ...,
        ...,
        apply_residual_connection_post_layernorm=False,
        output_layernorm=False,
        layer_type="encoder",
)

BERT architecture has LayerNorm at the output side (after the final BiasDropoutAdd) and the residual connection is taken from the output of that LayerNorm. In TE this can be achieved by setting the following arguments in the TransformerLayer API.

transformer_engine.pytorch.TransformerLayer(
        ...,
        ...,
        apply_residual_connection_post_layernorm=True,
        output_layernorm=True,
        layer_type="encoder",
)

T5 architecture has an additional cross-attention + BiasDropoutAdd + LayerNorm block before the MLP layer. In TE this can be added by setting the layer_type to decoder in the TransformerLayer API.

transformer_engine.pytorch.TransformerLayer(
        ...,
        ...,
        layer_type="decoder",
)

We welcome contributions to Transformer Engine. To contribute to TE and make pull requests, follow the guidelines outlined in the CONTRIBUTING.rst document.

• Attention original paper
• Megatron-LM tensor parallel
• Megatron-LM sequence parallel