Create a boxed parser for this format.

Marconi prefix-cache region availability.

Module + function name 4-tuple consumed by `Qwen3AttentionLayer::new_with_gating`: `(reshape_mod, reshape_fn, decode_mod, decode_fn)`. The reshape pai

Weight key prefix for layer-level weights. Returns `"model.layers"` for flat models (qwen3_next), or `"model.language_model.layers"` for conditional g

A mask with `n` accepted indices — `memory_size()` is `4*n` bytes, so tests can dial the cache's memory accounting precisely.

Number of full attention layers.

Pre-dequant dense (non-expert) NVFP4 weights to FP8 for zero-overhead prefill. Only affects gate GEMM and shared expert GEMMs. Expert weights stay N

RoPE: apply rotary position embeddings to Q and K in-place. Kernel: `rope_forward(Q, K, positions, seq_len, num_q_heads, num_kv_heads, head_dim, rota

Convenience wrapper: sample without token history (no repetition penalty).

Set the start state.

Convert a structural-tag JSON document into a grammar. Port of `StructuralTagToGrammar` / `Grammar::FromStructuralTag(json)`. The document must be a

Wait for all work on the secondary stream to complete.

Token IDs `[M]` u32 — stable across the layer loop (DeepSeek-V4 hash-MoE reads `tid2eid[token_id]`). Upload the pass's token IDs here before the layer

Drop every entry whose `token_position` is strictly greater than `keep_len`. Called after a successful rollback: snapshots taken in the now-discarded

Add a rule under a name allocated via [`Self::allocate_rule_name`]. Returns the rule name. Port of `AddRuleWithAllocatedName`.

Advance past `n` *bytes* of escape input. Because every escape sequence `parse_next_escaped` consumes is pure ASCII, one byte equals one codepoint her

Reset (m, l, o) for `num_seqs` sequences. Call once at the start of each decode step before the first `step_tile`.

Build a leaf FSM from a regex fragment that is either a plain literal (`"abx"`, possibly with escapes / `.`) or a character class `[...]`. `regex` is

Random Gaussian projection matrix `P` of shape `[head_dim, r]`, BF16, in row-major layout. Variance 1/head_dim so that ⟨k, p⟩ has unit variance for un

Build a trie-based FSM. `patterns` are inserted as a character trie. `excluded_patterns` are only honored when `add_back_edges` is true (the trie bec

Claim a slot and wrap it in a [`SlotGuard`] that returns the slot to the free list when dropped. This is the leak-safe claim API: the guard is stored

F66 (2026-04-29): MiniMax M2.7 XML tool-call grammar. Native MiniMax format: ```xml <minimax:tool_call> <invoke name="tool_name"> <parameter name="ke

Compile a structural tag. Port of `GrammarCompiler::CompileStructuralTag`. Delegates the structural-tag-JSON -> `GrammarData` conversion to the W5 `s

Run the conversion — port of `RegexConverter::Convert`.

Convert Python Jinja2 syntax to minijinja-compatible syntax. Handles: slice reversal, string methods, etc.

The WGRAMMAR static/dynamic decomposition (Tier 3c) — the compile-time split of every rule body into fixed scaffolding spans (with precomputed literal

Remove an entry. Returns `true` when the id existed and matched `kind`, `false` otherwise. Also removes the entry from the persistent backend when one

The accepting-state bitmap.

Fill the next-token bitmask into `bitmask` (`index` selects the matcher's slice when several share one buffer; pass `0` for a single matcher). Returns

Divide o_state by l_state and store as BF16 in `output`.

Batched forward: GEMM gate for N tokens, per-token expert dispatch. Gate projection reads weights once for N tokens (GEMM M=N). Expert dispatch remai

Finish building and return the [`GrammarData`], setting the root rule to `root_rule_id`. Returns `Err` if `root_rule_id` is out of bounds.

Rule id of the rule named `name`, or `-1` if absent.

Replace `lhs` with the sorted union of `lhs` and `rhs`. Both inputs must already be sorted ascending; duplicates within an input are collapsed in the

True if the cache holds no entries.

True if this is an epsilon transition.

Read 'a.b.c' from nested dict. Returns None on missing.

K-side block stride in bytes for a specific attention layer. Same as `block_stride_bytes_for_layer`; named for clarity in asym call sites.

The scanable states reached at the current input position.

Number of signatures currently stored.

A digit-range pattern for one fixed-length slot: either a literal digit or a `[a-b]` class, optionally followed by `\d{n}`.

Re-open measurement when the live depth has left the regime the current decision was measured in (factor-[`REMEASURE_DEPTH_FACTOR`] crossing in either

Generic 501 "not supported" response used by the auto-probe stubs below. OpenAI-SDK auto-detection and observability wrappers expect a 501 + `error.ty

Join `items` with ` | `, wrapped in parentheses. Port of static `Or`.

Auto-detect and parse inner content of a `<tool_call>` block. Tries Gemma-4 native, JSON (hermes), qwen3_coder XML, then tag-style XML fallback.

Decode every codepoint in a UTF-8 byte slice. When `preserve_invalid_bytes` is `false` (the C++ default), an invalid byte aborts decoding and the res

Token at `cur + delta`. The token stream is always `EndOfFile`- terminated, so callers stay in bounds when reading near the end.

Pop `count` advanced states (rollback).

Render a single codepoint as a printable, escaped string. `additional_escape_map` is checked first (e.g. `{'-': "\\-"}`), then the C-escape table, th

Append one element to the most recently inserted row. # Panics Panics if there is no row to append to.

Apply an EBNF repetition suffix to `item`. Port of static `Repeat`. `max == -1` means unbounded.

Decode the repeat-edge info stored at aux index `idx`.

Resolve a model specifier to a local directory path. Resolution order: 1. If the specifier is an existing directory with `config.json`, use it direct

Sample the FIRST generated token after prefill, applying the grammar constraint when one is active. `#131`: the very first decode token is produced f

Extract a scalar f32 from a single-element weight tensor via D2H copy.

Shard an NVFP4-quantized weight. The packed weight is `[N, K/2]` u8; the per-group scale is `[N, K/group_size]` u8 (FP8); `weight_scale_2` is a per-te

SSM beta+alpha projection output size.

Parallel union of several FSMs — accepts any of their languages.

Set the lookahead-assertion expr id of rule `rule_id`. An id of `-1` means no assertion.