jgrusewski
ebc4e66578
spec: add Phase 3 performance optimizations (8 items, prioritized)
3.1 Fuse bias+ReLU into cuBLAS epilogue (highest impact)
3.2 IQN quantiles 64→32 (4.3GB VRAM + 15-20% aux speedup)
3.3 Double-buffered experience collection
3.4 cuGraphExecUpdate for hyperparameter changes
3.5 Tensor core dimension padding (adv_h 32→128)
3.6 Batch size 16384→8192 (unlocks multi-stream benefit)
3.7 Persistent Adam kernel
3.8 Memory pool (cuMemAllocAsync)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>