Loading...
Loading...
Found 52 Skills
Validate and use CPU offloading in Megatron Bridge, including layer-level activation offloading and fractional optimizer state offloading with HybridDeviceOptimizer.
Apple Silicon skill for M-series development and profiling. Use when leveraging unified memory, Metal Performance Shaders, Instruments profiling, sysctl hardware queries, Rosetta 2 behavior, or 16KB page size considerations. Activates on queries about Apple Silicon, unified memory, AMX, MPS, Instruments, Rosetta, or M-series page size.
Process large datasets efficiently using chunk(), chunkById(), lazy(), and cursor() to reduce memory consumption and improve performance
Use when "training LLM", "finetuning", "RLHF", "distributed training", "DeepSpeed", "Accelerate", "PyTorch Lightning", "Ray Train", "TRL", "Unsloth", "LoRA training", "flash attention", "gradient checkpointing"
Optimizing .NET allocations/throughput. Span, ArrayPool, ref struct, sealed, stackalloc.
Use when a Unity 6.3 game is slow or stutters: low fps, frame spikes, hitches when entering areas or on first effects, GC spikes, too many draw calls or SetPass calls, CPU vs GPU bound questions, Profiler, Profile Analyzer, Memory Profiler, Frame Debugger or Project Auditor work, GPU Resident Drawer or SRP Batcher setup, LOD and Mesh LOD, shader warm-up and variants, memory or VRAM budgets, IL2CPP and build size, or proving a fix with numbers from a player.
Reclaim DRAM by disabling unused subsystems across MB1 BCT, MB2 BCT, kernel reserved-memory, and SWIOTLB. Use for headless or no-camera Jetson deployments; not for CPU/GPU frequency tuning.
Techniques for reducing peak GPU memory in Megatron Bridge — expandable segments, parallelism resizing, activation recompute, CPU offloading constraints, and common OOM fixes.
Conserve memory by reusing existing instances when working with large numbers of similar objects.
Quantizes LLMs to 8-bit or 4-bit for 50-75% memory reduction with minimal accuracy loss. Use when GPU memory is limited, need to fit larger models, or want faster inference. Supports INT8, NF4, FP4 formats, QLoRA training, and 8-bit optimizers. Works with HuggingFace Transformers.
Performance diagnosis and optimization - MicroProfiler, Script Profiler, Developer Console, Scene Analysis, Luau heap and memory categories, memory leaks, network/replication cost, rendering and physics cost, streaming, native codegen, Parallel Luau with Actors. Use when a game lags, stutters, drops FPS, crashes on mobile, has high ping, or before optimizing hot code.
Analyze and optimize Roblox scene and script performance. Use for mobile frame rate, memory, instance count, rendering cost, physics, particles, lighting, asset budgets, streaming, leaks, and using the rbx-scene-analysis workflow.