mirror of
https://github.com/TheTom/turboquant_plus.git
synced 2026-10-21 18:42:44 -05:00
0e235113ff
hw_replay.py — parse diagnostic outputs, compare hardware profiles, predict performance on different hardware: - HardwareProfile dataclass with full serialization (JSON round-trip) - Diagnostic output parser (tags → structured data) - Decode/prefill curve extraction and ratio computation - Profile comparison with anomaly detection - Constant cache degradation prediction model (GPU family-based) profiles/baseline-m5-max-128gb.json — ground truth baseline: - M5 Max 128GB, Apple10 (1010), has_tensor=true - 23 benchmark results (prefill + decode + mode 2) - 6 PPL results (8-chunk + 32-chunk, q8_0/turbo3/mode2) 36 tests: serialization, curve extraction, comparison, prediction, diagnostic parsing (all from embedded sample data — no external deps). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-Authored-By: tturney@psyguard.ai