{ "name": "LuminaV", "version": "1.6.0", "package": "luminamoon", "package_version": "1.0.9", "description": "Master-free, low-precision BF16/FP16 deep learning optimizer with Dual-Buffer Convex Lookahead Momentum (CLM), Single-Buffer Triangle-Defect Convex Lookahead Momentum (TD-CLM) via the Triangle Inequality Envelope, rational Pade [13/6] bounding, stateless 32-bit MurmurHash3 stochastic rounding, and native Rust SIMD / Vulkan / Triton execution.", "repository": "https://github.com/cloverx-id/luminamoon", "documentation": "https://cloverx-id.github.io/luminamoon/", "default_params": { "lr": 0.0003, "betas": [ 0.9, 0.999 ], "eps": 1e-08, "weight_decay": 0.08, "tau": 0.8, "lam": 0.3, "cautious": true, "cautious_clamp_min": 0.5, "buffer": 2, "stochastic_rounding": true, "bound": true, "bound_type": "coordinate", "bound_ratio": 0.05, "master_weights": "none", "execution": "auto", "fused_single_pass": true, "fuse_tensors": false, "bucket_cap_mb": 500.0, "prewarm_jit": true, "ema_decay": "auto", "fsdp_sync": false, "seed": 1337 }, "presets": { "dual_buffer_clm": { "buffer": 2, "tau": 0.8, "cautious": true, "cautious_clamp_min": 0.5, "bound": true, "bound_type": "coordinate", "bound_ratio": 0.05, "stochastic_rounding": true, "master_weights": "none", "fused_single_pass": true, "ema_decay": "auto", "fsdp_sync": false, "description": "Default Dual-Buffer Convex Lookahead Momentum (CLM) tracking m_t and innovation variance v_t with coordinate trust-region bounding." }, "single_buffer_td_clm": { "buffer": 1, "lam": 0.3, "cautious": true, "cautious_clamp_min": 0.5, "bound": true, "bound_type": "coordinate", "bound_ratio": 0.05, "stochastic_rounding": true, "master_weights": "none", "fused_single_pass": true, "ema_decay": "auto", "fsdp_sync": false, "description": "Single-Buffer Triangle-Defect Convex Lookahead Momentum (TD-CLM) using the Triangle Inequality Envelope (cuts optimizer state RAM by 50% with zero global reductions in coordinate mode)." }, "radial_trust_region_bound": { "buffer": 2, "tau": 0.8, "cautious": true, "cautious_clamp_min": 0.5, "bound": true, "bound_type": "radial", "bound_ratio": 0.03, "stochastic_rounding": true, "master_weights": "none", "fused_single_pass": true, "ema_decay": "auto", "fsdp_sync": false, "description": "Direction-preserving radial trust-region bounding (tanh(r)/r) preserving 100% update angular geometry." }, "semi_hybrid_fp32_master": { "buffer": 2, "cautious": true, "cautious_clamp_min": 0.5, "bound": true, "bound_type": "coordinate", "bound_ratio": 0.05, "stochastic_rounding": true, "master_weights": "semi", "fused_single_pass": true, "fsdp_sync": false, "description": "Hybrid precision pairing 32-bit FP32 master weights with 16-bit BF16/FP16 optimizer moment buffers." }, "full_fp32_baseline": { "buffer": 2, "cautious": true, "cautious_clamp_min": 0.5, "bound": true, "bound_type": "coordinate", "bound_ratio": 0.05, "stochastic_rounding": true, "master_weights": "full", "fused_single_pass": true, "fsdp_sync": false, "description": "Full-precision FP32 master weights and FP32 optimizer states for controlled baseline evaluation." }, "fsdp_distributed_sync": { "buffer": 2, "cautious": true, "cautious_clamp_min": 0.5, "bound": true, "bound_type": "radial", "bound_ratio": 0.03, "stochastic_rounding": true, "master_weights": "none", "fused_single_pass": true, "fsdp_sync": true, "description": "Distributed multi-GPU training mode under PyTorch FSDP or DeepSpeed ZeRO-3 with cross-rank collective scratch synchronization." } } }