Extract the last \\boxed{...} from a string, handling nested braces.

Build a diagnostic error for a tensor that failed gguf_tensor_in_file(). Reports every operand so a false rejection on a genuinely valid file (#318) c

Deterministic logits: a smooth but non-monotonic spread over the vocab so the argmax is unambiguous and the softmax has spread-out mass.

Check if a token is an end-of-sequence marker for the given target weights.

Wires the FULL + SWA causal masks, runs the backend graph, and returns last-token logits and/or a GPU-computed argmax on the host. Updates cache.cur_p

Parse SSE stream into list of (event_type, data_dict) tuples.

Next stdout line (str), or None on EOF. Raises DaemonTimeout past `timeout` s.

verify_derived_scalars wq_ne1 : weight_q->ne[1] (output dim of Q projection) wk_ne1 : weight_k->ne[1] (output dim of K projection) w

Zero every currently-free block. reset() drops mappings but leaves the previous request's bytes in place; maskless consumers (the qwen35moe pipelined

Decide the KV reservation for VRAM-budget expert placement. kvf_pool: resident KVFlash pool in tokens (0 = KVFlash inactive). all_hot_full_kv reports

Const lookup (no alloc / LRU touch): physical slot currently holding logical `pos`, or -1 if its chunk is not resident. Callers that may need an alloc

Build the shared-expert FFN subgraph onto an existing ggml_context. Returns the output tensor (or nullptr if no shared expert is present).

Post-compress gate: check whether the effective context overflows max_ctx. effective_tokens — raw effective prompt size (post FlowKV / pFlas

─── Special tokens ──────────────────────────────────────────────

Do an iteration step towards the next element. At the end of the step the * iterator key will represent the (new) current key. If it is not possible

Returns false only when dual-residency is unsafe (VMM VA-fragmentation risk).

Snapshot/restore SSM+conv state for speculative rollback. Uses device-side tensor copy (ggml_backend_tensor_copy). Called outside of any compute graph