Crate aria_inference 

Aria inference: bundle loader, family registry, Session.

Re-exports

pub use multimodal::action_head;

pub use multimodal::asr_transcribe_pcm16le;

pub use multimodal::rag_pack_context;

pub use multimodal::vision_encode;

Modules

fixture

Synthetic tiny Aria bundles / QuantTensors for contract tests (mirrors model tiny layout).

multimodal

Stage C multimodal helpers: ASR stub + RAG pack. Vision / VLA action require real bundle tensors — Session APIs return Unsupported rather than RGB mean-pool / fake linear stubs.

Structs

Bundle

BundleTokenizer

ChatTurn

EngineProfile

Family

Runtime family handle used by Session.

GenerateOpts

GenerateProfile

Generation

LoadProfile

ModelConfig

QuantTensor

Session

SessionBuilder

Enums

ArchClass

Architecture class for graph / loader hooks.

ComputePref

CLI / config preference for local GEMM.

EngineError

Unified engine error (requirements §3.6).

FamilyPhase

Delivery phase for a registered family (requirements §1.1).

TensorData

Constants

FAMILY_REGISTRY

Full registry mirroring model/requirements.md §1.1.

Functions

apply_chat_template

Render messages with a generation prompt for the next assistant turn.

arch_class_representatives

Representative path per arch class for tiny E2E tests.

confidence_from_logits

Optional confidence heuristic for hybrid: mean max-softmax over last logits proxy.

decode_placeholders

Fallback when no tokenizer sidecar: stable <id> placeholders (legacy demos).

dequantize

Rotated-space reconstruction (matches Python dequantize).

effective_rope_theta

Qwen3 uses θ=1e6; Qwen3.5 gated attention uses θ=1e7 (rope_parameters.rope_theta). Bundles that omit rope_theta hit serde default 10000 and produce empty / garbage completions (Hello → content: "" with a full max_tokens budget).

encode_naive

Fallback encode when no sidecar: map UTF-8 bytes into [0, vocab).

family_phase

graph_hook

Graph hook id for an architecture class (stage B/C dispatch).

infer_family_path

Map a bundle directory or slug_q4 / slug_q326_channel name onto a registry path.

load_bundle

lookup_family

require_runnable

Any registered family that can run the shared text decoder (+ stage C extras).

require_stage_a

Stage A only: gemma-4-e2b-it.

require_stage_b

Stage B: text / MoE families (+ stage-A golden path).

strip_assistant_visible

Drop Qwen3 thinking and ChatML specials from decoded assistant text.