Send a completion request and measure prefill time.

Build draft graph at a given ctx_len into sg. Does NOT touch sg.alloc. mirror_view: if true, uses a view into mirror->target_feat at slot0. ctx_alloc:

Dense GELU-gated FFN (layer 0 / lead dense layers). Gemma4 uses GELU not SiLU: cur = down( gelu(gate(x)) * up(x) )

MoE block: shared expert (GELU-gated) + routed experts (softmax gating). Gemma4-specific routing: attn_out → rms_norm → scale by 1/sqrt(n_embd) → mul

One graph: base logits for all n_positions hidden columns (a single lm_head matmul), then for rows [first_corrected, n_positions) the low-rank Markov

Build a sigmoid top-k MoE router using the node order recognized by ggml-cuda's topk-moe fusion: sigmoid -> reshape -> optional bias add -> argsort_to

Build a single layer of the Qwen3.5-27B model. layer_idx: which of the 64 layers to build (0-based). inp: input activation [hidden, n_tokens] Ret

For KV (single head): x is [head_dim, n_tokens]

Convert expert indices to multi-label binary targets.

Forward GPT-2 byte encoding: raw byte → Unicode codepoint (UTF-8 string). This is the inverse of gpt2_unicode_to_byte (defined later, near decode). By

Draws a token from `cand`, whose .first fields are proportional probabilities (need not already sum to 1 or be sorted). `r_uniform` is a pre-drawn uni

Guards shared by the fused Markov paths: head present, usable inputs, and the target lm_head vocab matching the head's training vocab.

── helpers ─────────────────────────────────────────────────────

Compute precision, recall, and per-sample FP rate at a confidence threshold. Returns a dict with keys: precision — of experts predicted ≥

Logical position held by each pool slot, -1 for free blocks. `dst` must hold pool_tokens entries. Lets callers build masks that need POSITION semantic

Force chunk and its radius-neighborhood into `forced`.

Helper: get a 2D slice from a 3D tensor along ne[2] (same as llama.cpp ggml_view_2d_slice).

Whether geometric_sample_logits_cuda can (and should) handle this config entirely on the GPU. false for top_k>0 or top_p in (0,1) — not because geomet