P1-EP-AMD-NPU-002: AMD NPU (Ryzen AI) — Subgraph Pattern Testing

[DingmaomaoBJTU]

Summary

Develop and validate subgraph pattern tests for the AMD Ryzen AI NPU, enabling the static analyzer to identify and validate supported computational patterns.

Context

AMD NPU (Ryzen AI) requires subgraph-level pattern testing — the same requirement that QNN HTP backend has (#366). Subgraph patterns define which fused operator sequences the NPU can execute natively. Without this, the static analyzer cannot accurately predict AMD NPU compatibility.

From plans/release/0501_release_plan/P0_CHECKLIST.md (P1-EP-002). Companion to P1-EP-AMD-NPU-001 (op coverage).

Current State

• No subgraph pattern data for AMD Ryzen AI
• QNN subgraph infrastructure exists and can be reused (Fix unhashable cache key error. #366)
• AMD NPU hardware required for validation

Desired State

• AMD NPU subgraph pattern library built and integrated into the static analyzer
• Automated subgraph tests run on AMD Ryzen AI hardware
• Pattern matching accuracy validated against runtime results

Acceptance Criteria

• AMD NPU subgraph patterns collected from SDK docs + empirical testing
• Subgraph test framework extended to support AMD NPU backend
• Pattern matching results validated on AMD Ryzen AI hardware
• Static analyzer correctly identifies AMD NPU-compatible subgraph patterns in P0 models
• All existing tests pass

Technical Notes

• Reuse QNN subgraph test framework (Fix unhashable cache key error. #366, fix: allow more EP in static analyzer, infer device automatically #392) as the foundation
• AMD Ryzen AI SDK docs for supported operator fusions
• Risk: Very high effort — similar to QNN subgraph discovery; hardware dependency
• Requires AMD Ryzen AI hardware (Ryzen AI 300 series preferred)

Related Files

• plans/release/0501_release_plan/ep-scale.md — AMD NPU detailed spec
• plans/release/0501_release_plan/P0_CHECKLIST.md — P1-EP-002
• Fix unhashable cache key error. #366 — QNN subgraph patterns (reference implementation)
• fix: allow more EP in static analyzer, infer device automatically #392 — Subgraph test framework