sparse-quantization

To load W8A8SC sparse-quantized models in vLLM-MindSpore, place weights under the rank_* directory and initialize the loader to switch load_format to sparse_quant for proper inference.

No, sparse-quantization for MindSpore enforces 310P-only execution for sparse paths. Attempting to run on non-310P devices triggers explicit error handling and stops the loading process.

vLLM-MindSpore handles deq_scale conversion by validating weight and index naming, then applying rank-based weight layout loading and integration into QuantLinearSparse layers.

The correct weight directory layout for MindSpore sparse quantization models requires placing sparse-quantized weights into rank_* directories to enable proper rank-based or split loader detection.

vLLM-MindSpore fails to detect is_sparse_quantization when weight and index naming validations are not met, preventing the load_format from automatically switching to sparse_quant.

Sparse-quantization in MindSpore is used to load and adapt W8A8SC quantized models within vLLM inference workflows, ensuring proper sparse configuration detection and weight layout validation on 310P devices.