Helios

Diagnose Gemma 4 MTP acceptance and speculative decoding behavior in Helios using real-context prompts, trace records, and exactness gates.

Profile and optimize Helios native MLX/C++ execution paths with A/B evidence, memory gates, and feature-flagged experimental kernels.

Build and run Helios real-context A/B benchmarks with deterministic workload manifests, correctness gates, and decision artifacts.

Use for tiny16 MacBook profiling: memory pressure, context-length limits, TTFT/decode metrics, cache effectiveness, MTP acceptance, and release-gate benchmark reports.

Use for dynamic LoRA/QLoRA adapter import, validation, loading, routing, adapter-aware KV cache keys, aLoRA planning, and adapter memory residency.

Use for Gemma4D KV cache, prefix cache, SSD cold tier, cache manifests, cache-key correctness, quantized/compressed KV experiments, and restore parity.

Use to execute one Gemma4D milestone end-to-end with scoped tasks, verification commands, evidence capture, and blocker reporting. Trigger on milestone implementation, Codex Goals, task planning, or release gates.

Use for Gemma 4 12B config, tokenizer, chat-template, prompt hashing, greedy parity, fixture creation, and reference comparison work.

Use for Gemma 4 MTP speculative decoding: drafter loading, draft/verify/accept/rollback, greedy exactness, acceptance metrics, and MTP auto-disable.

Use for Rust-to-MLX native C/C++ FFI work: opaque handles, C ABI, build.rs/CMake, MLX lifetime safety, native smoke tests, and unsafe Rust review.

Use when implementing or reviewing the Gemma4D Ratatui operator console, including page architecture, async event loops, terminal lifecycle, provider boundaries, config/benchmark UX, snapshots, and safety constraints.