Version Compatibility Notes

Product Version Information

Item Content
Product Name MindIE SD
Product Version 3.0.0
Version Type Official Release
Maintenance Cycle Three Months
Product Name Version
CANN 9.0.0
Ascend Extension for PyTorch 7.3.0
Ascend HDK See CANN Version Compatibility Notes for version compatibility

Version Compatibility

MindIE SD components must be used with matching versions. Do not mix components across different versions.

Table 1 Software Version Compatibility

CANN Ascend Extension for PyTorch
9.0.0 7.3.0

Version Usage Notes

None

3.0.0 Release Notes

New Features

No. Details
1 Enhanced quantization capabilities. Support for FA dynamic FP8, new W4A4_DYNAMIC quantization format, completion of W4A4 quantization algorithm common logic, and new W4A4MXFP4DualQuantLinear capability, improving adaptability and deployment flexibility across quantization scenarios.
2 Enhanced operator and plugin capabilities. New aclnn LayerNorm plugin and external API, new adaLayerNormV2 plugin and layer implementation, along with additional aclnn capabilities including sparse_block_estimate, block_sparse_attention, laser_attention, and la_preprocess, increasing operator coverage.
3 Enhanced runtime capabilities. New multi-instance shared memory for sharing weight memory across instances to reduce duplication; new block-level CPU offload for fine-grained dynamic transfer of modules between CPU and NPU, alleviating memory pressure.
4 Enhanced attention and layout adaptation. attention_forward and rf_v2 now support BNSD layout input; FA can be enabled via environment variable, reducing integration cost for upper-layer model integration.
5 Enhanced scheduling and serving capabilities. New Dynamic EPLB scheduling, new service-based examples, and wan2.2 service-side synchronization and accuracy fixes, improving availability for serving and inference scenarios.
6 Optimized quantization operator backend. Migrated existing torch-atb-based quantization operators to aclnn native operators, improving compatibility and stability, with full support for compilation optimization features such as torch.compile, enhancing framework adaptability.

Modified Features

No. Details
1 Removed _mindie_sd suffix from custom plugin operator names and unified namespace to mindiesd for consistent naming.
2 In FA quantization scenarios, FIA operator output format now matches input query format to reduce format incompatibility issues.
3 Adapted npu_quant_matmul operator with added constraints to mitigate integration risks from constraint changes.
4 Systematic adaptation of aclnn compilation project with enhanced build chains, directory management, and error handling to improve build efficiency and stability.

Removed Features

None

API Change Notes

API changes include additions, modifications, deprecations, and removals. API changes only reflect code-level modifications and do not include documentation improvements in language, formatting, or links.

  • New: APIs introduced in this version.
  • Modified: APIs changed compared to the previous version.
  • Deprecated: APIs that cease evolution from the declared version and may be removed one year after declaration.
  • Removed: APIs removed in this version.
Class Name / API Signature Change Type Description
• def mindiesd.layernorm_scale_shift
• def mindiesd.fast_layernorm
• def mindiesd.sparse_attention
New New APIs
• class mindiesd.Linear
• class mindiesd.QuantFA
Removed Removed APIs

Resolved Issues

No. Category Issue Description
1 Installation & Compatibility libopapi.so missing when running tests after installing compiled MindIE-SD, affecting post-installation testing and basic functionality verification.
2 Installation & Compatibility Insufficient compatibility with newer torch versions, affecting integration with newer inference images.
3 Installation & Compatibility Build package missing plugin, affecting package completeness and plugin capability loading.
4 Operator & Compilation Flux.1-dev failing aclnnAdaLayerNorm call with compile enabled in new environments, blocking compilation acceleration path.
5 Operator & Compilation Incorrect API usage in test_rainfusionattention.py causing test execution failures.
6 Cache & Test Quality Block_end validation in DiT Cache Agent inconsistent with left-closed right-open interval semantics, affecting cache usage.
7 Cache & Test Quality Single-metric accuracy comparison (only cosine similarity), incomplete accuracy evaluation dimensions.

Known Issues

No. Category Issue Description
1 Operators Lack of matrix multiplication operators based on CUTLASS and Triton
2 Usability Need to support more extensions, such as cache-dit
3 Performance Need to support more parallel computation masking and fusion schemes

Upgrade Impact

Impact on Running Systems During Upgrade

  • Business impact

    Software version upgrade will cause service interruption.

  • Network communication impact

    No impact on network communication.

Impact on Running Systems After Upgrade

None

Vulnerability Patch List

None