Fix a regression leading to a “not supported” error for unary (single-operand) contractions.
New experimental API
cutensornetComputeGradientsBackward()for computing gradients of a tensor network w.r.t. its input tensors.
Known limitations: operates on tensor networks with a single slice and no singleton modes, on a single GPU device.
New high-level APIs to facilitate definition of quantum circuit tensor states, compute arbitrary marginal distributions and perform sampling of those states (support of arbitrary tensor states is provided, for example, qudit-based tensor states).
See the introduction at High-level tensor network specification.
cutensornetWorkspacePurgeCache()to purge workspace cache.
New APIs to set/get network attributes.
New APIs to support more SVD algorithms including GESVD (default), GESVDJ, GESVDP and GESVDR. The SVD algorithm can be set via one call to
cutensornetTensorSVDConfigSetAttribute()with the attribute
CUTENSORNET_TENSOR_SVD_CONFIG_ALGO. For GESVDJ and GESVDR, user may further set algorithm specific parameters with the attribute
CUTENSORNET_TENSOR_SVD_CONFIG_ALGO_PARAMSusing new structs
New APIs to provide more runtime information for SVD execution in
cutensornetTensorSVDInfo_t. The SVD algorithm used can be accessed via one call to
cutensornetTensorSVDInfoGetAttribute()with the attribute
CUTENSORNET_TENSOR_SVD_INFO_ALGO. For GESVDJ and GESVDP, user may further query execution status with the attribute
CUTENSORNET_TENSOR_SVD_INFO_ALGO_STATUSusing new structs
New API, via the
CUTENSORNET_CONTRACTION_OPTIMIZER_CONFIG_CACHE_REUSE_NRUNSattribute, that enables the optimizer to factor in the constant input tensors benefit when a path is run multiple times.
New API to toggle “smart” optimization settings with attribute
CUTENSORNET_CONTRACTION_OPTIMIZER_CONFIG_SMART_OPTION. This option (turned on by default) will limit the pathfinder elapsed time by avoiding certain configurations as well as adjusting configuration on the fly. The path quality can differ from when the option is turned off. To restore the previous behavior, users should set this to off.
Improved performance of the contraction path optimization process (e.g., pathfinding). Speedup depends on the tensor network size. A large speedup >10x can be observed for large networks. For medium-size networks (hundreds of tensors) a speedup of almost 5x was observed.
Failed tensor network contraction involving constant input tensors, in some corner cases, when not enough cache memory was available.
cuTensorNet supports Ubuntu 20.04+.
cutensornetGateSplit()may potentially fail for certain input tensor operands when SVD algorithm is set to
Under single precision, when the input tensor/matrix has a low rank,
CUTENSORNET_TENSOR_SVD_ALGO_GESVDRbased tensor SVD may suffer from reduced accuracy.
When SVD algorithm is set to
CUTENSORNET_TENSOR_SVD_ALGO_GESVDP, user is responsible for checking the
CUTENSORNET_TENSOR_SVD_CONFIG_ALGO_PARAMSattribute from the SVDInfo object with corresponding struct
cutensornetGesvdpStatus_tto monitor the convergence.
Support for caching intermediate tensors for subsequent reuse in repeated tensor network contractions. This is a useful feature that results in a substantial speedup when users want to perform more than one execution of a tensor network contraction, where a large fraction of the input tensors stays constant, while the rest update their values. For example, computing amplitudes of individual bit-strings or small batches of bit-strings can benefit from this feature. We provide users with an opportunity to specify which tensors are constant. Subsequently, cuTensorNet will use this information to build internal data structures to cache constant intermediate tensors for their reuse in repeated executions of the tensor network contraction plan. Note that, if all input tensors are marked constant, the output tensor becomes constant as well, thus there is no benefit to contracting the network again, as such, the caching mechanism will not be triggered. Repeated contractions in this case will incur the same execution time.
cutensornetTensorQR()when users provide a customized memory pool to compute the QR factorization of double complex data with certain extent combinations.
Failed autotune in some corner cases with “insufficient workspace” error.
Failed execution of
cutensornetTensorSVD()when all singular values are trimmed out. For cuTensorNet v2.1.0, one singular value will be retained in the output for such cases. This behavior may be subject to change in a future release.
The cuTensorNet-MPI wrapper library (
libcutensornet_distributed_interface_mpi.so) needs to be linked to the MPI library
libmpi.so. If you use our conda-forge packages or cuQuantum Appliance container, or compile your own using the provided
activate_mpi.shscript, this is taken care for you.
Introduce support for CUDA 12.
A set of new wheels with suffix
-cu12are released on PyPI.org for CUDA 12 users.
pip install cutensornet-cu12for installing cuTensorNet compatible with CUDA 12.
cuquantumwheel (without the
-cuXXsuffix) is turned into an automated installer that will attempt to detect the current CUDA environment and install the appropriate wheels. Please note that this automated detection may encounter conditions under which detection is unsuccessful, especially in a CPU-only environment (such as CI/CD). If detection fails we assume that the target environment is CUDA 11 and proceed. This assumption may be changed in a future release, and in such cases we recommend that users explicitly (manually) install the correct wheels.
CUDA Lazy Loading is supported. This can significantly reduce memory footprint by deferring the loading of needed GPU kernels to the first call sites. This feature requires CUDA 11.8 (or above) and cuTENSOR 1.7.0 (or above). Please refer to the CUDA documentation for other requirements and details. Currently this feature requires users to opt in by setting the environment variable
CUDA_MODULE_LOADING=LAZY. In a future CUDA version, lazy loading may become the default.
cuTensorNet requires cuTENSOR 1.6.1 or above, but cuTENSOR 1.7.0 or above is recommended, for performance improvements, bug fixes, and the CUDA Lazy Loading support.
cuTensorNet supports Ubuntu 18.04+
In the next release, Ubuntu 18.04 will be dropped. The minimum supported Ubuntu version will be 20.04.
We are on NVIDIA/cuQuantum GitHub Discussions! For any questions regarding (or exciting works built upon) cuQuantum, please feel free to reach out to us on GitHub Discussions.
Bug reports should still go to our GitHub issue tracker.
A conda package is released on conda-forge:
conda install -c conda-forge cutensornet. Users can still obtain both cuTensorNet and cuStateVec with
conda install -c conda-forge cuquantum, as before.
A pip wheel is released on PyPI:
pip install cutensornet-cu11. Users can still obtain both cuTensorNet and cuStateVec with
pip install cuquantum, as before.
cuquantummeta-wheel points to the
cuquantum-cu11meta-wheel (which then points to
custatevec-cu11wheels). This may change in a future release when a new CUDA version becomes available. Using wheels with the
-cuXXsuffix is encouraged.
Initial support for Hopper users. This requires CUDA 11.8.
New APIs to create, query, and destroy tensor descriptor objects.
New APIs and functionalities for approximate tensor network algorithms. cuTensorNet now supports the computational primitives mentioned below to enable users to develop approximate tensor network simulators for quantum circuits including MPS, PEPS, and more:
Tensor decomposition via QR or SVD. Both exact and truncated SVD supported.
Application of a gate to a pair of connected tensors followed by compression.
New APIs to create, tune, query, and destroy tensor SVD truncation settings.
New APIs to create, query, and destroy runtime tensor SVD truncation information.
Automatic distributed execution: cuTensorNet API is extended to include functions enabling automated distributed parallelization of tensor network contractions across multiple GPUs. Once activated, the parallelization is applied to both tensor network contraction path finding (when hyper-sampling is enabled) and contraction execution, without making any changes to the original serial source code.
Functionalities introduced that break previous APIs:
Complex conjugation operator on input tensors (this adds an extra parameter that specifies tensor qualifiers in
Provide new API for users to specify the slices as (sliced modes, sliced extents) with one call to
cutensornetContractionOptimizerInfoSetAttribute()by using the attribute
CUTENSORNET_CONTRACTION_OPTIMIZER_INFO_SLICING_CONFIG. This will remove a limitation of the old API, which required users to set the modes first before the extents.
Removed the [in,out] alignment-requirement parameters from the
cutensornetCreateNetworkDescriptor()API. These are no longer required and are being inferred internally.
Some enum values are reordered. If your application stores any of cuTensorNet enum values as plain int, please make sure to rebuild your application.
Memory access error when running cuda-memcheck in a few corner cases.
Logging related bug upon setting some attributes.
Inaccurate flops computed by cuTensorNet with user-provided path & slicing.
“Undefined symbol” error when using cuTensorNet in the NVIDIA HPC SDK container.
Incorrect handling of extent-1 modes in the deprecated
Improved performance of the contraction path optimization process. On average, about 3X speedup was observed on many problems.
Improved performance of the contraction auto-tuning process.
Improved the quality of the slicing algorithm. We now select the configuration with the minimum number of slices that has the minimal flops overhead.
More auto-tuning heuristics added that improves tensor contraction performance.
GNU OpenMP Runtime (gomp) is no longer needed.
Two new APIs,
cutensornetGetTensorDetails(), replace the
cutensornetGetOutputTensorDetails()API, which is deprecated and will be removed in a future release.
New samples (samples/cutensornet/).
cuTensorNet requires cuTENSOR 1.6.1 or above, but cuTENSOR 1.6.2 or above is recommended, for performance improvements and bug fixes.
cuTensorNet requires CUDA 11.x, but CUDA 11.8 is recommended, for Hopper support, performance improvements, and bug fixes.
With CUDA 11.7 or lower,
cutensornetTensorQR()can potentially fail for certain extents.
cutensornetTensorQR()can potentially fail when users provide a customized memory pool to compute the QR factorization of double complex data with certain extent combinations.
With cuTENSOR 1.6.1 and Turing, broadcasting tensor modes with extent-1 might fail in certain cases.
The version constraint
cuTENSOR>=1.5,<2as promised elsewhere in the documentation was not correctly respected. Both the code and various package sources are now fixed.
New APIs and functionalities introduced:
A new API,
cutensornetContractionOptimizerInfoPackData(), that allows users to serialize/pack the optimizerInfo in order to broadcast it to other ranks. Similarly, another new API for unpacking is provided,
New APIs for creating and destroying slice group objects, which include
cutensornetDestroySliceGroup(). These APIs, when combined with the packing/unpacking APIs above, allow the users to employ the slicing technique to create independent tasks that be run on multiple GPUs.
An option to auto-tune intermediate modes through the
cutensornetContractionAutotune()API, which helps improve network contraction performance. The functionality of this API call can be controlled with the
An option to find a path that minimizes estimated time to solution (rather than FLOP count). This experimental feature can be controlled with the configuration attribute
An option to retrieve the mode labels for all intermediate tensors through the
CUTENSORNET_CONTRACTION_OPTIMIZER_INFO_INTERMEDIATE_MODESattribute of the contraction optimizer-info.
Since near optimal paths are easily found for small networks without simplification, and since simplification does not guarantee an optimal path, the simplification phase has been turned OFF by default when the simplified network is sufficiently small.
A new slicing algorithm has been developed, leading to potentially more efficient slicing solutions.
Improve contraction performance by optimizing intermediate mode-ordering.
Improve contraction performance of networks that have many singleton mode labels.
Previously, in rare circumstances, the slicing algorithm could fail to make progress toward finding a valid solution, resulting in an infinite loop. This has been fixed.
A bug in the deprecated
cutensornetContraction()API that accepted sliceId >= numSlices.
Provide a distributed (MPI-based) C sample that shows how easy it is to use cuTensorNet and create parallelism.
Update the (non-distributed) C sample by improving memory usage and employing the new contraction API
A workspace pointer alignment issue.
A potential path optimizer issue to avoid returning
This release improved the support for generalized einsum expression to provide a better contraction path.
Clarify in the documentation and sample that the contraction over slices needs to be done in ascending order, and that when parallelizing over the slices the output tensor should be zero-initialized.
Clarify in the documentation that the returned FLOP count assumes real-valued inputs.
Several issues in the C++ sample (samples/cutensornet/tensornet_example.cu) are fixed.
Greatly reduced the workspace memory size required.
Reduced the execution time of the pathfinder with multithreading and internal optimization.
Support for hyperedges in tensor networks.
Support for tensor networks described by generalized Einstein summation expressions.
Add new APIs and functionalities for:
Managing workspace (see Workspace Management API for details).
Binding a user-provided, stream-ordered memory pool to the library (see Memory Management API for details).
Query of the output tensor details (see
Set the number of threads for the hyperoptimizer (see Hyper-optimizer for details).
Setting a logger callback with user-provided data (see
cuTensorNet requires CUDA 11.x.
cuTensorNet requires cuTENSOR 1.5.0 or above.
cuTensorNet requires OpenMP runtime (GOMP).
cuTensorNet no longer requires NVIDIA HPC SDK.
If multiple slices are created, the order of contracting over slices using
cutensornetContraction()should be ascending starting from slice 0. If parallelizing over slices manually (in any fashion: streams, devices, processes, etc.), please make sure the output tensors (that are subject to a global reduction) are zero-initialized.
Initial public release
Add support for
Add new APIs and functionalities for:
Fine-tuning the slicing algorithm
Reconfiguring a tensor network
Simplifying a tensor network
Optimizing pathfinder parameters using the hyperoptimizer
Retrieving the optimizer configuration parameters
cutensornetContractionGetWorkspaceis renamed to
cutensornetContractionAutotune()’s function signature has changed
cuTensorNet requires cuTENSOR 1.4.0 or above
cuTensorNet requires NVIDIA HPC SDK 21.11 or above
Support Volta and Ampere architectures (compute capability 7.0+)
cuTensorNet requires CUDA 11.4 or above
cuTensorNet requires cuTENSOR 1.3.3 or above
cuTensorNet supports NVIDIA HPC SDK 21.7 or above
This release is optimized for NVIDIA A100 and V100 GPUs.