Measure cosine similarity between h[l] and h[l+1] (Fate §4.2). High similarity validates cross-layer prediction: if cos(h[l], h[l+1]) ≈ 1, th

Find the index one past the `}` that matches `text[open] == '{'`. Respects nested {}/[] depth and skips over "..." / '...' / `...` string literals (wi

Convert an array of bf16 values to fp16 via f32 intermediate.

Convert an array of bf16 values to f32 in place into a destination buffer.

Build a cached pre-FFN graph for a FULL-ATTENTION layer. With the step-invariant set_rows KV write (kv_write_rows input) and a 256-aligned FA span (kv

Cached batched COLD routed graph (CPU backend, no shared expert). Mirror of build_cached_hot_batched_graph for the cold expert stack; used by the mixe

Build a cached pre-FFN graph for a DeltaNet layer. DeltaNet layers have no kv_start-dependent views — the graph structure is identical across tokens.

Build Fate-style cross-layer training pairs: (h[l], experts[l+1]). Records are token-major: layers 0..num_layers-1 per token in sequence. For

Attention block for a single layer (handles both full and SWA).

Attention block. Handles BOTH full and SWA layers; the only differences are - n_head[il] (per-layer) - rope_freq_base + n_rot (per layer-type) - YaRN

MoE block: sigmoid router with score-correction bias, sum-normalize selected weights, scale routed combine by expert_weights_scale (=2.5 for Laguna),

Phase 2.1: full MoE dispatch (sigmoid + score-correction bias + sum-norm + scale 2.5 + always-on shared expert). Mirrors llama.cpp's build_moe_ffn for

Hybrid MoE block: same router as build_laguna_moe_block_full, but the routed experts are served from the bounded hot stack via a global->local LUT, an

Needle-In-A-Haystack-shaped prompt for the cache test. Uses a coarse `target_tokens * 4.0` sizing — this script only needs a long, determinis

Returns true when this build should keep draft projection weights as BF16. CUDA builds keep BF16 only when built for native BF16 tensor-core support;

Build a chat-templated prompt that hits ``target_tokens`` ±20%. Returns (text, n_tokens).

Build a fresh single-token forward graph. We rebuild per step so that `kv_start` updates drive the correct KV cache slot. The graph is cheap to rebuil

Compare greedy spec-decode (df) against greedy autoregressive (ar). A raw bit-exact compare is too strict on its own: the target sees draft t

Claude Code message.content (string or array of typed blocks) -> text.

Try strict json::parse first; on failure rewrite single- and backtick-quoted strings to double-quoted, wrap bare identifier keys in double quotes, and

[(session_file, parser_fn, source_tag), ...] from the enabled sources.

C++ drafter score+compress via daemon. Returns compressed token ids. Daemon command: compress <bin> <keep_x1000> <drafter_gguf> <drafter_arch

Batched variant used by prefill. The default keeps implementations simple and lets remote backends override it to amortize IPC overhead.

─── Compute per-layer compression ratios (matches ds4.c logic) ─────────

Compute the effort tiers from max_tokens + complex_problem_max_tokens using the spec §3.3 formula. `think_max` and `complex_think_max` are derived in

─── Disk-cache identity salt ─────────────────────────────────────────── Compute a 16-byte salt from inputs that affect KV cache validity: model path

Get the continued-interval setting.