Version Mapping

Product Version Information

Product	MindSDK
Version	26.0.0
Version Type	Release version

Related Product Version Mapping

Product	Version
Ascend HDK	Atlas 350: 1.0.RC1 Other products: 26.0.RC1
CANN	9.0.0

Virus Scan Results

Virus scan passed.

Version Compatibility

• Rec SDK Torch (torch_rec_v1): After the upgrade, recompile torchrec_npu and the packages related to custom operators.

Table 1 Software version compatibility

MindSDK Version	MindSDK Version to Upgrade To	CANN Version Compatibility	Ascend HDK Version Compatibility
Rec SDK 26.0.0	MindSDK 7.3.0 and 7.3.0.x	CANN 8.3.RC1 and patch versions CANN 8.5.0 and patch versions CANN 9.0.0 and patch versions	Ascend HDK 25.5.0 and patch versions Ascend HDK 25.1.RC1 and patch versions Ascend HDK 26.0.RC1 and patch versions

Usage Precautions

None

Change Description

New Features

Feature	Description	Compatible Product Models
Rec SDK TensorFlow(tf_rec_v1)	Adapted for the Atlas 350 PCIe card.	Atlas 800T A2 training server Atlas 200T A2 Box16 heterogeneous subrack Atlas 800T A3 SuperPoD server Atlas 350 PCIe card
Rec SDK Torch(torch_rec_v1)	Multi-level cache saving and loading: Incremental saving and loading and differential card loading are supported. Multi-level cache admission and eviction: The showclick admission and eviction policy is supported. Recommendation models now support inference on computing power-partitioned devices. Adapted for the Atlas 350 PCIe card.	Atlas 800T A2 training server Atlas 200T A2 Box16 heterogeneous subrack Atlas 800T A3 SuperPoD server Atlas 350 PCIe card
Rec SDK TensorFlow(tf_rec_v2)	The basic sparse table functions are implemented, including table creation, table lookup, saving, loading, feature admission, and feature eviction.	Atlas 350 PCIe card
Rec SDK Torch(torch_rec_v2)	The basic sparse table functions are implemented, including table creation, table lookup, saving, loading, feature admission, and feature eviction.	Atlas 350 PCIe card
Rec SDK operators	Fused operators for generative recommendation models, including in_linear_silu and reverse_sequence, are completed and adapted for the Atlas 350 PCIe card. Fused operators for generative recommendation models, including norm_multiply_dropout and concat_2d_jagged, are completed but not adapted for the Atlas 350 PCIe card. Enhanced HSTU operators: The performance of the backward operator is optimized. The forward and backward operators now support the int32 type. Refactored operators: The table lookup backward operator and the HSTU forward and backward operators are refactored.	Atlas 800T A2 training server Atlas 200T A2 Box16 heterogeneous subrack Atlas 800T A3 SuperPoD server Atlas 350 PCIe card
fbgemm-npu	Adapted for the Atlas 350 PCIe card.	Atlas 800T A2 training server Atlas 200T A2 Box16 heterogeneous subrack Atlas 800T A3 SuperPoD server Atlas 350 PCIe card
HKV	Adapted for the Atlas 350 PCIe card.	Atlas 350 PCIe card

Interface Changes

Rec SDK

• Rec SDK TensorFlow (tf_rec_v1): No interface changes.
• Rec SDK Torch (torch_rec_v1): No interface changes.
• Rec SDK TensorFlow (tf_rec_v2): No interface changes.
• Rec SDK Torch (torch_rec_v2): No interface changes.

Key Feature Changes

Rec SDK

• Rec SDK TensorFlow (tf_rec_v1): No key feature changes.
• Rec SDK Torch (torch_rec_v1): No key feature changes.
• Rec SDK TensorFlow (tf_rec_v2): No key feature changes.
• Rec SDK Torch (torch_rec_v2): No key feature changes.

Resolved Issues

None

Known Issues

None

Upgrade Impact

Impact on the System During the Upgrade

None

Impact on the System After the Upgrade

None

26.0.0 Documentation

Document	Description	Update Description
Rec SDK 26.0.0 User Guide	Introduces Rec SDK, its installation and deployment process, functions, model adaptation, and API reference.	For details, see Rec SDK 26.0.0 User Guide.

Vulnerability Patch List

None