torch-npu-fusion-optimizer

To optimize PyTorch models with torch_npu fusion operators, this skill analyzes your model code to identify compute patterns in QKV, RoPE, KV cache, and FFN paths, generating a validated replacement blueprint for faster Ascend hardware inference.

Torch_npu fusion-operator replacement maps standard PyTorch compute patterns to fused NPU kernels. You need it when optimizing Transformer models on Ascend hardware to reduce inference time and resource usage across Attention and MoE modules.

Yes, you can optimize Attention and MoE modules on Ascend hardware. The skill detects fusion opportunities specific to these scenarios and applies substitutions respecting data types and layouts for deterministic automated replacement.

The best way to identify QKV and RoPE fusion opportunities is automated pattern discovery. The skill scans your PyTorch model code to detect these common paths and proposes substitutions mapping to official torch_npu interfaces.

To generate a validated replacement plan, the skill produces a blueprint with validation hooks for safety. It maps substitutions to official torch_npu interfaces and generates deterministic replacement steps for automated application.