Decompose `[min, max]` of packed UTF-8 values of the *same byte length* into byte-range edges from state `from` to state `to`.

Return the default CUDA stream handle.

Detach `idx` from the reference graph.

EP broadcast: send a command (u32) to all worker ranks. Called by rank 0 before each model operation to synchronize workers. Only valid when EP is en

EP broadcast: send a `(seq_id, cmd)` pair to all worker ranks. Use this at the *first* broadcast of a logical command sequence (e.g. the K=2 verify m

Look up `key` in an ordered object slice.

Optimize a structural-tag grammar so the Earley parser can run.

Load a quantized weight, dispatching by variant. Handles all three on-disk formats including FP8 block-scaled (requires dimensions for FP8→BF16→NVFP4

Sanitize a JSON schema to prevent xgrammar's C++ EBNF parser from crashing. XGrammar's `json_schema_converter.cc` can produce empty EBNF rule bodies

Append a raw `(type, data[])` expression to the CSR store and return its expr id.

── events ────────────────────────────────────────────────────────────

Mix `value` into `seed` in place. Faithful to C++ `HashCombineBinary` — the boost `hash_combine` 64-bit mixing step. Uses wrapping arithmetic to matc

Whether the grammar has been fully matched (all required structure generated).

`(id, token)` pairs sorted lexicographically by token bytes, excluding stop and special tokens.

Add a rule with a unique name derived from `hint`.

Allocate a name from `hint` and add the rule in one step. Returns the actual rule name. Port of `AddRule`.

Returns the refusal sentence when `content` opens with one of the known patterns, else `None`. The returned sentence is the first sentence (truncated

Detect [`HfMetadata`] from the raw text of a `tokenizer.json`. Returns `Err` only when `backend_str` is not parseable JSON or is not a JSON object —

Fill `bitmask` with the set of tokens acceptable for the next decode position. `index` selects the matcher's slice when the caller batches several ma

Synchronize the stream.

True when there are no rows.

Lowercase and collapse runs of `_` / `-` to a single `_`, trimming leading and trailing underscores. Used by fuzzy tool-name repair to match across se

Shard a BF16 dense weight `[out_dim, in_dim]` according to `kind`. Returns `(sharded_ptr, sharded_out, sharded_in)`. When `tp_size == 1` or `kind ==

Fused SiLU activation: output = SiLU(gate) * up. Kernel: `silu_mul_separate(gate, up, output, n)` Grid: (ceil(n/256), 1, 1) Block: (256, 1, 1)

Per-vision-pad-token-id helper. Vision prompts splice ViT embeddings into placeholder `<|image_pad|>` token positions; the hashed token-ID stream ther

Add an epsilon transition.

True if no weights are loaded.

Construct an [`StructuralTagError::InvalidJson`].

Launch a kernel with raw parameters. # Safety `params` must contain valid pointers matching the kernel signature.

Parse `schema` into a spec. `rule_name_hint` seeds the EBNF rule name; `default_type` is applied when no `type` keyword is present (used by `propertyN

Get the raw CUstream handle for Atlas's own stream.

Check text for common degeneration signals. Returns (bool, detail).

Measure kernel execution time in milliseconds using CUDA events. Runs `warmup` warmup iterations, then `iters` timed iterations, returning the minimum

Block stride in bytes for a specific attention layer. For symmetric dtypes returns the K stride (which equals V). For asymmetric dtypes returns the K-

Look up a kernel function handle with caching.

True if this is a character-range edge.

True if this is a rule reference.

Convenience: render `grammar` as EBNF text.

A tiny RAW-vocab tokenizer covering the bytes used in the test grammars: single ASCII chars plus a few multi-char tokens.

Sort edges and pack into the CSR [`CompactFsm`] form.

POST /tokenize — tokenize text or chat messages, return token IDs and count.

Async all-reduce using GPU-side event synchronization. Replaces `gpu.synchronize(stream) + all_reduce(ptr, bytes)`. Uses a dedicated comm stream + CU

Build the boxed parser for a [`ReasoningFormat`].

Convert and assert exact equality with `expected`.

Load a BF16 norm weight and subtract 1.0 from every element. Atlas's `rms_norm` kernel uses the Qwen3-Next "offset-from-1" convention (`out = x * (1

Run `sudo docker ARGS` on the given host.

Find the representative of the set containing `element`, applying path compression. # Panics Panics if `element` is not present — mirrors the C++ `XG

Dispatch GEMV to the appropriate kernel based on weight precision.

True if this is a repeat reference.

Generate a globally unique tool call ID.

Total number of elements.

Build the parser for a format.

Rotary embedding dimension. Priority: 1. Explicit `rotary_dim` field (MiniMax M2 — integer in config.json). 2. `partial_rotary_factor * head_dim` (Qw

Save SSM state from active pool slot into a snapshot slot. Returns `None` if no free snapshot slots are available. Tags the snapshot with `session_has

FP32 conv1d output for SSM recurrent path (prevents BF16 precision drift).

Turbo3 data section bytes (3-bit packed: 8 values in 3 bytes).

Set the body of rule `rule_id`. Returns `Err` if `rule_id` is out of range.

V-side block stride in bytes for a specific attention layer. Differs from K-side only for asymmetric KV cache dtypes.

Add a `ByteString` expr from a Rust string (UTF-8 bytes).

Mark `state` accepting.

Create a rule for `spec` (or return a cached rule name) and return its name. Port of `CreateRule`.

Ensure the AtlasRegistry is initialized (idempotent).

Evict up to `num_blocks` LRU zero-ref leaf nodes. Returns physical block indices that were freed plus parallel disk-block IDs (Phase 6.1.e). When HSS

Returns the (K_dtype, V_dtype) pair. For symmetric variants both elements are identical. For asymmetric variants the pair differs.

Load FP8 KV cache quantization scales from a checkpoint. Searches for `{attn_prefix}.k_proj.k_scale` and `{attn_prefix}.v_proj.v_scale` tensors in th