Release Notes#

cuTensorNet v2.15.0#

Changes

  • Aligned the behavior of state projection MPS (cutensornetCreateStateProjectionMPS()) for MPS-finalized states with that of other state properties, such as accessors, marginals, expectations and samplers. Previously, the behavior was ambiguously documented and depended on details of the state:

    • For states with only matrix product operators applied to their initial MPS, the operators were previously contracted exactly, ignoring the state computation. Now, the previously computed MPS representation of the state is taken into account. If truncation happened during the MPS output state computation, the output of cutensornetStateProjectionMPSComputeTensorEnv() differs from previous releases.

    • Creating a state projection MPS for a state with tensor operators that was not computed after its last change returns CUTENSORNET_STATUS_INVALID_VALUE. Its prepare and environment computation return CUTENSORNET_STATUS_INVALID_VALUE after a change of the state until it is computed again, and after a computation with other bond extents, canonical center or gauge option, or after cutensornetStateFinalizeMPS() was called again (create it again).

    • Creating a symmetric state projection MPS for an MPS-finalized state with tensor operators returns CUTENSORNET_STATUS_NOT_SUPPORTED.

    To project the tensor operators exactly, use a separate cutensornetState_t with the same initial state and tensor operators that is not finalized with cutensornetStateFinalizeMPS().

  • The Prepare calls of the tensor network state API and of the MPS projection API set the required scratch workspace sizes for both the device and the host memory space in the workspace descriptor (query the host part with CUTENSORNET_MEMSPACE_HOST). Users are expected to attach a host scratch buffer of at least the reported size before the subsequent compute calls; the documentation and the samples have been updated accordingly. With cuSOLVER 12.3.4 or later, providing the host scratch workspace also enables a faster SVD path for CUTENSORNET_TENSOR_SVD_ALGO_GESVD.

  • Improved the accuracy of MPS simulations that apply a matrix product operator (MPO) with cutensornetStateApplyNetworkOperator() and CUTENSORNET_STATE_MPO_APPLICATION_INEXACT (the default).

  • Improved the accuracy of MPS simulations of states whose qudits have different dimensions, when operators (tensor operators, controlled operators or MPOs) act on non-adjacent qudits and an SVD truncation option is set or an MPO is applied with CUTENSORNET_STATE_MPO_APPLICATION_INEXACT.

Bugs fixed

  • Fixed a functionality issue where MPS simulation with the default CUTENSORNET_TENSOR_SVD_ALGO_GESVD algorithm failed with CUTENSORNET_STATUS_INTERNAL_ERROR in cutensornetStateCompute() and cutensornetStateProjectionMPSInsertTensor() (and hence in the dependent state property computations) when run with cuSOLVER 12.3.4 (CUDA 13.4 Update 1) or later, whose SVD requests a host workspace: the host scratch workspace attached to the workspace descriptor was not forwarded to the SVD. The same issue made CUTENSORNET_TENSOR_SVD_ALGO_GESVDR, whose host workspace is always required, unusable in MPS simulation with every cuSOLVER version.

  • The deprecated cutensornetWorkspaceGetSize() now reports the same size as cutensornetWorkspaceGetMemorySize(). Previously it reported a smaller size, so a workspace buffer of exactly that size could be rejected with CUTENSORNET_STATUS_INSUFFICIENT_WORKSPACE depending on its alignment. The API has been deprecated since cuTensorNet v2.1.0 in favor of cutensornetWorkspaceGetMemorySize().

  • Fixed zero or inaccurate results from cutensornetTensorSVD() with CUTENSORNET_TENSOR_SVD_PARTITION_UV_EQUAL and from MPS simulations when singular values are below machine epsilon in absolute value, for example for MPS states whose norm becomes very small after non-unitary operators.

  • Fixed incorrect MPS results for states initialized with cutensornetStateInitializeMPS() using non-column-major strides and CUTENSORNET_STATE_MPS_GAUGE_FREE (the default), when some MPS tensors are not acted on by any operator.

  • Fixed out-of-bounds writes to the output MPS tensors when the two tensors sharing a bond specify different extents in cutensornetStateFinalizeMPS() or cutensornetStateInitializeMPS(); such extents are now rejected with CUTENSORNET_STATUS_INVALID_VALUE.

  • Fixed cutensornetStatePrepare() returning CUTENSORNET_STATUS_INVALID_VALUE when the MPS output extents requested with cutensornetStateFinalizeMPS() are close to the largest 64-bit integer.

  • Fixed a larger-than-necessary workspace size reported by cutensornetStatePrepare() for MPS simulations whose output tensors, as specified with cutensornetStateFinalizeMPS(), have extents above the maximal possible ones (present since cuTensorNet v2.12.0).

  • Fixed an issue in MPS simulations using CUTENSORNET_STATE_MPS_GAUGE_SIMPLE where establishing an explicitly requested canonical center on the output MPS could produce suboptimal truncation. Exact, non-truncating results were unaffected.

  • Fixed a correctness issue in MPS simulation of general quantum channels (cutensornetStateApplyGeneralChannel()): the probability used to select a Kraus operator was evaluated for the transposed operator (with the qudit order reversed for two-qudit channels) instead of the operator as given, so channels whose Kraus operators are not symmetric, such as amplitude damping with |0><1|, selected the wrong branch. On the pure-state MPS backend this could select a zero-probability branch and fail with CUTENSORNET_STATUS_INTERNAL_ERROR on the next channel. Channels with symmetric Kraus operators (Pauli channels, diagonal operators) and the mixed-state (density matrix) backend were unaffected.

  • Fixed an out-of-bounds device memory access in cutensornetExpectationComputeWithGradientsBackward() when a tensor operator registered with cutensornetStateApplyTensorOperatorWithGradient() acts on qudits of different dimensions (for example, a (5, 2) pair). The gradient buffer was zeroed with strides that did not match its extents, which could fail with cudaErrorLaunchFailure. Operators on qudits of equal dimension were unaffected.

  • Fixed a correctness issue in cutensornetStateApplyDiagonalTensorOperator() for diagonal tensor operators acting on three or more state modes: the operator was applied as a diagonal combined with a cyclic permutation of the acted-on modes, giving silently wrong results for equal mode extents and CUTENSORNET_STATUS_INTERNAL_ERROR for unequal extents. Diagonal operators on one or two state modes were unaffected.

  • Fixed a segmentation fault when contracting or autotuning a tensor network with a NULL workspace descriptor after updating tensor storage.

  • Fixed corrupted output when the output tensor memory is provided before preparation: the output tensor strides were not applied correctly.

  • Fixed a bug where setting the output tensor twice kept the first setting and ignored the second.

  • Fixed a correctness issue for MPS-finalized states (cutensornetStateFinalizeMPS()): accessors, marginals, expectations and samplers could silently return results for an earlier state if the state was computed again after their creation (for example after cutensornetStateUpdateTensorOperator()), even when their prepare function was called again as described in the known issues of cuTensorNet v2.14.0, or if they were created after the state was changed but before it was computed again. They now reflect the latest state computation, which resolves that known issue. While the state has not been computed since a change, creating them and their functions return CUTENSORNET_STATUS_INVALID_VALUE; the same applies after the state was computed with other MPS bond extents (for example under value-based SVD truncation), canonical center or gauge option than at their creation, or after cutensornetStateFinalizeMPS() was called again, in which case they must be created again. See cutensornetStateFinalizeMPS() for details.

  • Fixed state projection MPS (cutensornetCreateStateProjectionMPS()) of MPS-finalized states silently returning environments of an earlier state computation, or of a captured MPS without the tensor operators applied afterwards. cutensornetCreateStateProjectionMPS() for an MPS-finalized state with tensor operators that was never computed previously returned CUTENSORNET_STATUS_INTERNAL_ERROR; it now returns CUTENSORNET_STATUS_INVALID_VALUE.

  • Fixed incorrect gradients returned by cutensornetExpectationComputeWithGradientsBackward() for MPS-finalized states with tensor operators applied by cutensornetStateApplyTensorOperatorWithGradient(). Gradients are not supported for MPS-finalized states; the function now returns CUTENSORNET_STATUS_NOT_SUPPORTED. Expectation values can still be computed through cutensornetExpectationCompute().

  • Fixed cutensornetGetOutputStateDetails() and cutensornetStateCaptureMPS() failing for an MPS-finalized state computed after cutensornetStateUpdateTensorOperator() without a new cutensornetStatePrepare(). Samples drawn with a fixed seed (CUTENSORNET_SAMPLER_CONFIG_DETERMINISTIC) from such a state, or right after cutensornetStateCaptureMPS(), may differ from previous releases.

  • Fixed a correctness issue in samplers of states given as an MPS (set with cutensornetStateInitializeMPS(), captured, or MPS-finalized and computed): after tensor operators were applied or the MPS was captured, they kept sampling the earlier state. cutensornetSamplerPrepare() and cutensornetSamplerSample() now return CUTENSORNET_STATUS_INVALID_VALUE in this case, as for all other samplers; create the sampler again.

  • Fixed out-of-bounds memory writes in cutensornetSamplerSample() for samplers of MPS states after a failed cutensornetSamplerPrepare() (for example with too small a workspace limit). Sampling now returns CUTENSORNET_STATUS_INVALID_VALUE until the sampler is prepared again.

  • cutensornetStateCompute() for an MPS-finalized state now returns CUTENSORNET_STATUS_INVALID_VALUE after a failed cutensornetStatePrepare() (for example with too small a workspace limit), until the state is prepared again; previously it ran operations whose preparation had not completed.

  • Fixed out-of-bounds memory writes in cutensornetStateCompute() after cutensornetStateFinalizeMPS() was called again with smaller extents on a prepared state. After calling it again, with any extents, computing now returns CUTENSORNET_STATUS_INVALID_VALUE until the state is prepared again.

  • Fixed issues with states whose initial MPS was set with cutensornetStateInitializeMPS() without state tensors (the vacuum state): cutensornetCreateSampler() returned CUTENSORNET_STATUS_INTERNAL_ERROR, and samplers created after the state or another object created from it was computed could return wrong samples. Setting state tensors afterwards with cutensornetStateInitializeMPS() at the same extents and strides (also on an MPS-finalized state prepared without an initial MPS) was silently ignored; it now takes effect, and, as after a change of extents, the state must be prepared again and objects created before must be created again.

  • Fixed a correctness issue in cutensornetStateProjectionMPSComputeTensorEnv(): 2-site environments could be wrong after cutensornetStateProjectionMPSInsertTensor() reduced a bond extent through SVD truncation, if the workspace limit was too small to compute the environment at once or the cuTensorNet handle was distributed over several processes.

  • cutensornetStateProjectionMPSComputeTensorEnv() now returns CUTENSORNET_STATUS_INVALID_VALUE when called for a symmetric state projection MPS without an input tensor (envTensorDataIn), which a symmetric projection requires; previously the process aborted.

  • Fixed host memory growth until cutensornetDestroy() with repeated cutensornetStatePrepare() calls for MPS-finalized states, repeated cutensornetSamplerPrepare() calls for samplers of MPS states, and Prepare calls that fail because the workspace limit is too small.

Known issues

  • Providing slice ids in non-ascending order for contracting a network, when all slices are to be contracted, may give incorrect results. The user can work around this for now by providing the slice ids in strictly ascending order, or by passing NULL as the sliceGroup parameter, which contracts all the network slices in sequence. This will be fixed in a subsequent release.

  • Specifying the tensor ids to be conjugated in a network contraction, via setting the attribute CUTENSORNET_NETWORK_INPUT_TENSORS_CONJUGATED with cutensornetNetworkSetAttribute(), after preparing the network contraction (cutensornetNetworkPrepareContraction()) does not take effect unless the network contraction is re-prepared. The user can work around this for now by setting the conjugation attribute ahead of preparing or by re-preparing the network contraction. This will be fixed in a subsequent release.

  • A state projection MPS of an MPS-finalized state can return environments of the previous state computation after the state was computed again with configuration changes that keep its MPS bond extents, canonical center and gauge option (for example the SVD normalization, algorithm or cutoffs). Call cutensornetWorkspacePurgeCache() on the workspace descriptor of the state projection MPS before computing environments again, or provide no cache workspace.

cuTensorNet v2.14.0#

  • New functionalities:

  • This release provides functional support for the Rubin GPU family.

  • Bugs fixed:

    • cutensornetNetworkComputeGradientsBackward() with accumulateOutput=0 now zeroes the gradient buffers of the input tensors that carry duplicate mode labels, whose gradients only occupy a diagonal sub-space of their dense storage. Previously the off-diagonal elements were left untouched, and the user had to pre-set such gradient memory to zero. Input tensors with unique mode labels are unaffected and are still fully overwritten without any extra zeroing.

    • Fixed a correctness issue in cache-backed tensor-network contractions that could ignore updated output-tensor strides. When output storage was rebound after contraction preparation, the cache-to-output copy could retain the strides used during preparation, causing output modes to be permuted.

    • Consequently, cutensornetAccessorCompute() and cutensornetMarginalCompute() could produce outputs with incorrect mode ordering when explicit non-default strides were supplied and the underlying output contraction was cache-backed. This applies to Accessors for both pure and mixed states and to full and diagonal Marginals. For example, this could occur when all state tensors were immutable; a mixed-state network composed solely of quantum channels is one such case. These APIs now honor the supplied output strides. Calls using the default (NULL) strides were unaffected.

    • Fixed a correctness issue where cutensornetStateProjectionMPSComputeTensorEnv() ignored the coefficients of network operators applied to the projected state via cutensornetStateApplyNetworkOperator() (for example, the coefficient passed to cutensornetNetworkOperatorAppendMPO()): the computed environment tensor was scaled only by the per-state coefficients passed to cutensornetCreateStateProjectionMPS(). The product of the applied operator coefficients now multiplies the environment exactly once, composed with the corresponding per-state coefficient.

    • Fixed a correctness issue where cutensornetStateProjectionMPSComputeTensorEnv() could compute environments of a state with a matrix product operator (MPO) applied via cutensornetStateApplyNetworkOperator() with adjoint=1 as if the operator had been applied without the adjoint. Adjoint MPOs now contribute the documented Hermitian conjugate, including the conjugated operator coefficient.

    • Fixed a correctness issue in MPS simulation with the CUTENSORNET_STATE_MPS_GAUGE_SIMPLE gauge option: after applying a non-unitary tensor operator (including non-unitary matrix product operators and controlled tensor operators), truncating simulations could produce incorrect states. Exact (non-truncating) simulations were unaffected.

    • cutensornetStateUpdateTensorOperator() (and the deprecated cutensornetStateUpdateTensor()) now honor a change of the unitary parameter; such a change requires re-preparing the state. Previously the change was silently ignored and could lead to incorrect results. For tensors that are part of an MPO the parameter remains not applicable.

Known issues:

Compatibility notes:

  • cuTensorNet requires cuTENSOR v2.8.0 or above.

  • Distributed tensor decomposition (experimental), including cutensornetCreateDistributedTensorDescriptor(), requires cuTENSOR v2.8.x. This functionality additionally requires NCCL and cuSOLVERMp; see distributed tensor decomposition requirements.

  • Support for Fedora 42 and KylinOS V10 has been removed. The minimum supported versions are Fedora 44 and KylinOS V11.

  • This release adds support for additional Linux distributions:

    • x86_64: AlmaLinux 8/9/10, Fedora 44, KylinOS V11, openSUSE 16, Oracle Linux 8/9/10, SLES 16, Ubuntu 26.04

    • ARM64: Debian 13, KylinOS V11, SLES 16, Ubuntu 26.04

cuTensorNet v2.13.0#

  • New functionalities:

  • Bugs fixed:

    • Fixed a correctness bug where cutensornetStateApplyTensorOperator() with adjoint=1 applied only the ket/bra mode transposition of the gate instead of the documented Hermitian conjugate when cutensornetStateFinalizeMPS() was called.

    • Fixed a bug where cutensornetCreateStateProjectionMPS() would reject a captured MPS state that previously contained unitary channels due to a stale internal counter not being reset by cutensornetStateCaptureMPS().

    • Fixed a bug in cutensornetSamplerSample() where the sampler’s internal random number generator did not advance between successive calls, causing repeated draws to return identical samples.

    • Fixed a bug where setting the CUTENSORNET_SAMPLER_CONFIG_DETERMINISTIC attribute to 0 after a non-zero value seeded the sampler’s random number generator with seed 0 instead of restoring non-deterministic behavior. The attribute now correctly re-randomizes the generator when set to 0, matching the documented contract that a positive seed enables deterministic sampling.

    • Fixed a bug where cutensornetSamplerSample() could return inconsistent samples across processes in a distributed run when the requested number of samples was smaller than the number of processes; in this duplicated-work case all processes now produce identical outputs.

    • Fixed a bug where cutensornetExpectationComputeWithGradientsBackward() could fail or return incorrect gradients when cutensornetExpectationPrepare() was called more than once on the same expectation object before the backward pass.

    • cutensornetSamplerSample() now fails with CUTENSORNET_STATUS_INVALID_VALUE when a conditional marginal probability distribution has zero total mass (for example, an over-truncated or otherwise unnormalizable state) instead of silently returning a degenerate sample. This makes the default marginal sampler consistent with the MPS sampling path, which already rejected zero-norm states.

Compatibility notes:

  • cuTensorNet requires cuTENSOR v2.5.0 or above. However, cuTENSOR v2.7.0 or above is recommended for improved performance.

  • Starting with the next release, Fedora 42 and KylinOS V10 will no longer be supported. The minimum supported versions will be Fedora 44 and KylinOS V11.

cuTensorNet v2.12.2#

  • Bugs fixed:

    • Fixed a correctness issue where cutensornetExpectationComputeWithGradientsBackward() would produce incorrect gradients when the tensor network operator expansion contains terms with complex-valued coefficients. This did not affect operators with purely real coefficients.

cuTensorNet v2.12.1#

  • Bugs fixed:

    • Fixed a correctness bug where cutensornetStateApplyTensorOperator() with adjoint=1 applied element-wise complex conjugation instead of the documented Hermitian conjugate (complex conjugation with ket/bra mode transposition).

cuTensorNet v2.12.0#

Compatibility notes:

  • The strides of MPS output state tensors returned by cutensornetGetOutputStateDetails() may differ from those returned in previous releases. The library does not guarantee any particular stride ordering (e.g., row-major or column-major) for output state tensors, and the returned strides may vary across releases, configurations, or successive calls. Callers should always use the strides returned by cutensornetGetOutputStateDetails() when accessing output tensor data rather than assuming a specific memory layout.

  • cuTensorNet requires cuTENSOR v2.5.0 or above.

cuTensorNet v2.11.0#

  • New functionalities:

  • Bugs fixed:

    • Fixed a bug where cutensornetStateCompute() may return incorrect results for exact MPS simulation on qudit systems (qudits with different extents on different modes) with operators acting on non-adjacent qudits.

  • Performance enhancements:

cuTensorNet v2.10.1#

Known issues:

  • In certain MPS simulation with value based truncation, the workspace size returned by cutensornetStatePrepare() may be lower than the actual required workspace size.

cuTensorNet v2.10.0#

  • New functionalities:

    • Enabled network construction and path generation suites of APIs to run on machines with no GPU device, all dependencies remain the same.

Known issues:

  • Applying two-qudit gates to non-adjacent, non-homogeneous qudits in MPS simulations with exact SVD enabled may cause representational errors or undefined behavior.

cuTensorNet v2.9.1#

cuTensorNet v2.9.0#

  • New functionalities:

    • Introduced the network-centric API for tensor network definition and execution with lighter signatures and simpler semantics:

      • New flow: define a network with cutensornetCreateNetwork, append inputs via cutensornetNetworkAppendTensor, set output with cutensornetNetworkSetOutputTensor, then prepare and run using cutensornetNetworkPrepareContraction and cutensornetNetworkContract.

      • Gradient workflow is explicitly separated: set adjoint and per-input gradient buffers via cutensornetNetworkSetAdjointTensorMemory and cutensornetNetworkSetGradientTensorMemory, prepare with cutensornetNetworkPrepareGradientsBackward, and execute with cutensornetNetworkComputeGradientsBackward.

      • Workspace management for the network flow is unified via cutensornetWorkspaceSetMemory with CUTENSORNET_WORKSPACE_SCRATCH and CUTENSORNET_WORKSPACE_CACHE kinds.

      • The optimizer round-trip is maintained: create config/info, call cutensornetContractionOptimize, and attach to the network via cutensornetNetworkSetOptimizerInfo.

    • The legacy descriptor/plan-based API remains available but is deprecated. Prefer the network-centric API moving forward.

  • Bugs fixed:

  • Known issues:

  • Other changes:

Compatibility notes:

  • cuTensorNet requires cuTENSOR v2.3.1 or above.

cuTensorNet v2.8.0#

cuTensorNet v2.7.0#

Known issues:

Compatibility notes:

  • cuTensorNet requires cuTENSOR v2.2.0 or above.

cuTensorNet v2.6.0#

  • New functionalities:

  • Bugs fixed:

    • Fixed a bug in large network when workspace size exceed int32 limit. In the previous versions, when such situation occurs, the code returns with an error “no path can be found”. The fix will allow the pathfinder to continue to try to find a path with slicing.

    • Fixed a potential deadlock in the pathfinder ThreadPool.

  • Other changes:

cuTensorNet v2.5.0#

  • New functionalities:

  • Bugs fixed:

    • Fixed a bug in large network planning leading to a segfault.

    • Fixed a bug where gradient computation may fail due to insufficient workspace.

    • Fixed a corner case bug where contraction results may be incorrect.

    • Fixed failed contractions with compute type CUTENSORNET_COMPUTE_3XTF32.

    • Fixed the false GPU workspace shortage failure that could show up for certain quantum circuits.

    • Fixed the ability to call non-compute API with a different active GPU device than the one associated with the library handle.

    • Fixed the check on tensor operator mutability during the tensor operator data update operation.

    • Fixed a corner case bug where illegal memory access was encountered.

    • Fixed an issue where the cutensornetState_t object in MPS simulations would throw an error when cutensornetStateCompute() was called more than once after cutensornetStatePrepare() step, even if there were no structural changes. Users can now call cutensornetStateCompute() multiple times on the same cutensornetState_t object with just one preceding cutensornetStatePrepare() step without encountering errors.

  • Performance enhancements:

    • Improved performance of contraction and gradient computations (cutensornetContractSlices() and cutensornetComputeGradientsBackward respectively) when the user allocates workspace buffers on the host memory.

  • Other changes:

    • Lowered CPU overhead of several cuTensorNet functions.

    • Enabled support of tensor operators using real data types (CUDA_R_32F and CUDA_R_64F) in the State API.

Compatibility notes:

  • cuTensorNet requires cuTENSOR v2.0.2 or above.

cuTensorNet v2.4.0#

Compatibility notes:

  • cuTensorNet requires cuTENSOR v2.0.1 or above.

  • cuQuantum will drop support for RHEL 7 in the following cuQuantum release. Please plan ahead with this in mind. Thank you.

Known issues:

  • For the MPS computations based on cutensornetStateFinalizeMPS() APIs, if the state has different extents on different modes and there are operators applied to two non-adjacent modes, the exact MPS factorization may not be computed.

cuTensorNet v2.3.0#

  • New functionalities:

    • New tensor network state APIs for defining tensor network operators and computing their expectation values over given tensor network states.

    • New tensor network state APIs for computing arbitrary slices of amplitudes for a given tensor network state.

    • New tensor network state APIs for computing the Matrix-Product-State (MPS) factorization of a given tensor network state.

    • New truncation option CUTENSORNET_TENSOR_SVD_CONFIG_DISCARDED_WEIGHT_CUTOFF for tensor SVD computation.

  • Bugs fixed:

  • Other changes:

    • Complex-valued gradients computed with the experimental API cutensornetComputeGradientsBackward are complex conjugated now, as compared to what would be returned in the previous release.

Compatibility notes:

  • cuTensorNet requires cuTENSOR v1.6.1 or above, but is not compatible with v2.x.y. cuTENSOR v1.7.0 is recommended, for performance improvements, bug fixes, and the CUDA Lazy Loading support.

cuTensorNet v2.2.1#

  • Bugs fixed:

    • Fix a regression leading to a “not supported” error for unary (single-operand) contractions.

cuTensorNet v2.2.0#

Compatibility notes:

  • cuTensorNet supports Ubuntu 20.04+.

Known issues:

cuTensorNet v2.1.0#

  • New functionalities:

    • 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.

  • Bugs fixed:

    • Failure of 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.

  • Other changes:

    • 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.sh script, this is taken care for you.

    • Introduce support for CUDA 12.

    • A set of new wheels with suffix -cu12 are released on PyPI.org for CUDA 12 users.

      • Example: pip install cutensornet-cu12 for installing cuTensorNet compatible with CUDA 12.

      • The existing cuquantum wheel (without the -cuXX suffix) 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.

  • Performance enhancements:

    • 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.

Compatibility notes:

  • 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.

cuTensorNet v2.0.0#

  • 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.

  • Major release:

    • 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.

      • Currently, the cuquantum meta-wheel points to the cuquantum-cu11 meta-wheel (which then points to cutensornet-cu11 and custatevec-cu11 wheels). This may change in a future release when a new CUDA version becomes available. Using wheels with the -cuXX suffix is encouraged.

  • New functionalities:

    • 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:

  • Bugs fixed:

    • 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 cutensornetGetOutputTensorDetails() API.

  • Performance enhancements:

    • 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.

  • Other changes:

Compatibility notes:

  • 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.

Known issues:

  • 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.

cuTensorNet v1.1.1#

  • Bugs fixed:

    • The version constraint cuTENSOR>=1.5,<2 as promised elsewhere in the documentation was not correctly respected. Both the code and various package sources are now fixed.

cuTensorNet v1.1.0#

  • New APIs and functionalities introduced:

  • Functionality/performance improvements:

    • 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.

  • Bugs fixed:

    • 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.

  • Other changes:

    • 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 cutensornetContractSlices().

cuTensorNet v1.0.1#

  • Bugs fixed:

  • Performance enhancements:

    • This release improved the support for generalized einsum expression to provide a better contraction path.

  • Other changes:

    • The Overview and Examples pages are significantly improved!

    • 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.

cuTensorNet v1.0.0#

Compatibility notes:

  • 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.

Limitation notes:

  • 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.

cuTensorNet v0.1.0#

  • Initial public release

  • Add support for Linux ppc64le

  • 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

  • API changes:

Compatibility notes:

  • cuTensorNet requires cuTENSOR 1.4.0 or above

  • cuTensorNet requires NVIDIA HPC SDK 21.11 or above

cuTensorNet v0.0.1#

  • Initial release

  • Support Linux x86_64 and Linux Arm64

  • Support Volta and Ampere architectures (compute capability 7.0+)

Compatibility notes:

  • cuTensorNet requires CUDA 11.4 or above

  • cuTensorNet requires cuTENSOR 1.3.3 or above

  • cuTensorNet supports NVIDIA HPC SDK 21.7 or above

Limitation notes:

  • This release is optimized for NVIDIA A100 and V100 GPUs.