Overview

Implement GPU performance benchmarks for APR format per Section Y.2 of the spec.

Requirement

• Y7: APR decode speed must be ≥200 tok/s on GPU (RTX 4090 reference)
• Must match or exceed GGUF decode speed on same hardware

Falsification Condition

APR < 200 tok/s when GGUF ≥ 200 tok/s on same GPU

Implementation Tasks

• Add CUDA feature flag to realizar
• Implement GPU kernels for APR inference
• Create benchmark harness for GPU performance
• Verify parity with GGUF on RTX 4090 or equivalent
• Add to CI with GPU runner (optional)

Blocked By

• Requires GPU hardware for development and testing

References

• Spec: docs/specifications/apr-whisper-and-cookbook-support-eoy-2025.md Section Y.2
• Related: Y6 (CPU benchmarks) - ✅ Verified at 206.4 tok/s

Priority

P2 - Deferred (no GPU hardware available currently)