XLA (Accelerated Linear Algebra) is a domain-specific compiler for linear algebra that can accelerate TensorFlow models with potentially no source code changes.
The results are improvements in speed and memory usage: most internal benchmarks run ~1.15x faster after XLA is enabled. The dataset below is evaluated on a single NVidia V100 GPU:
When a TensorFlow program is run, all of the operations are executed individually by the TensorFlow executor. Each TensorFlow operation has a precompiled GPU kernel implementation that the executor dispatches to.
XLA provides an alternative mode of running models: it compiles the TensorFlow graph into a sequence of computation kernels generated specifically for the given model. Because these kernels are unique to the model, they can exploit model-specific information for optimization. For example, let's look at an optimization XLA does in the context of a simple TensorFlow computation:
def model_fn(x, y, z): return tf.reduce_sum(x + y * z)
Run without XLA, the graph launches three kernels: one for the multiplication,
one for the addition and one for the reduction. However, XLA can optimize the
graph so that it computes the result in a single kernel launch. It does this by
"fusing" the addition, multiplication and reduction into a single GPU kernel.
Moreover, this fused operation does not write out the intermediate values
x+y*z to memory; instead it "streams" the results of
these intermediate computations directly to their users while keeping them
entirely in GPU registers. Fusion is XLA's single most important optimization.
Memory bandwidth is typically the scarcest resource on hardware accelerators, so
removing memory operations is one of the best ways to improve performance.
Enable XLA for TensorFlow models
A simplest way to start using XLA in TensorFlow models is to enable
auto-clustering, which automatically finds clusters (connected subgraphs)
within the TensorFlow graph which can be compiled and executed using XLA.
Auto-clustering on GPU can be enabled by either modifying the
$ TF_XLA_FLAGS=--tf_xla_auto_jit=2 path/to/your/tf/program
Or by setting a configuration value within the program:
import tensorflow as tf tf.config.optimizer.set_jit(True) # ... the rest of your program ...
Auto-clustering is currently optimized for GPU workloads, but it can also be
enabled on CPU by additionally using the flag
$ TF_XLA_FLAGS="--tf_xla_auto_jit=2 --tf_xla_cpu_global_jit" path/to/your/program
For a detailed usage example, see the auto-clustering tutorial colab.
Explicit compilation API offers a more fine-grained control for choosing which functions should be compiled with XLA. However, it requires restructuring source code, as not all TensorFlow operations can be represented in XLA. That is, using explicit compilation on API on functions which can not be represented in XLA results in an exception.
Optimizing sections of the program using
tf.function is a
standard approach for
of TF2 programs. You can enable compilation with XLA by setting the
experimental_compile argument of
If you are using TF1, you can use the
xla.compile API for explicit compilation
using XLA. See the tutorial colab for usage
AOT (Ahead-of-time) compilation for CPU with
You can also use a standalone
which converts TensorFlow graph into executable code (for CPU only).
Inspect compiled programs
XLA provides introspection facilities which let you inspect the generated
programs. To dump the generated programs, use the environment variable
$ XLA_FLAGS="--dump_hlo_as_text --xla_dump_to=/tmp/generated" TF_XLA_FLAGS="--tf_xla_auto_jit=2" my/tensorflow/program
After the dumping is performed, you can find the following files in
module_XXXX.*_optimizations.txtGenerated XLA programs, one per each compiled cluster. Attaching those when submitting XLA bug reports is extremely helpful!
module_XXXX.ptxGenerated PTX files.
You can also dump the graph visualizing the embedding of XLA clusters inside of the TensorFlow graph with:
$ TF_DUMP_GRAPH_PREFIX=/tmp/generated TF_XLA_FLAGS="--tf_xla_clustering_debug"
Auto-clustering is supported on NVIDIA GPUs, and ahead-of-time compilation is supported on x86-64 CPUs. Auto-clustering support on multi-GPU environments and on a CPU is experimental.
Generating great bug reports
A bug report is much easier to reproduce if it includes dumps for the generated XLA programs and the used auto-clustering embedding. To generate them for a TensorFlow program running with auto-clustering, launch:
$ TF_DUMP_GRAPH_PREFIX=/tmp/generated \ TF_XLA_FLAGS="--tf_xla_clustering_debug --tf_xla_auto_jit=2" \ XLA_FLAGS="--xla_dump_hlo_as_text --xla_dump_to=/tmp/generated" \ my/tensorflow/program"
When filing bugs, attach the contents of the
If possible, try to isolate
a bug to a single XLA program by using the
and iteratively running it on generated programs.
Apart from TensorFlow, XLA programs can be generated by:
- JAX: Composable transformations of Python+NumPy programs
- Julia: The Julia language for scientific computing
- PyTorch: PyTorch framework