aclgraph-dev

Use NPUGraph static capture and replay through the mindiesd aclgraph backend. Set CompilationConfig.aclgraph_only or aclgraph_with_compile to True and compile the model with MindieSDBackend, which captures the operator sequence once and replays it without per-step host launches.

Set CompilationConfig.aclgraph_with_compile to True so MindieSDBackend.compile() runs pattern fusion first, then NPUGraph capture wraps the compiled graph. If aclgraph_only is set instead, compilation is skipped and capture applies to the original graph.

The mismatch error means a replay input's shape or dtype differs from what was captured, since entries are cached per input shape. Keep input shapes stable, allow a new capture for the new shape, or disable aclgraph for dynamic-shape models.

No, aclgraph requires static input shapes and dtypes. Shape changes trigger a costly new capture or a validation failure at replay, so dynamic sequence length or resolution workloads should use the default Inductor backend instead.

The backend checks npu_graph_available by verifying torch.npu.NPUGraph and torch.npu.graph both exist. If either is missing, the aclgraph switches are automatically disabled and execution falls back to the default aot_autograd plus Inductor path.

aclgraph-dev covers NPUGraph batched dispatch: capture, graph pools, replay, and input validation. pattern-dev covers pattern matcher registration and Inductor codegen in the default backend. They connect through the aclgraph_only and aclgraph_with_compile switches.