Module infer 

ONNX inference for IMF v1 artifacts: the greedy single-row decode loop (the artifact contract per CONFORMANCE.md) over a mockable InferenceEngine trait, plus the ort-backed implementation.

Byte-exactness contract (mirror of pipeline.conformance.ort_single_row_decoder, the reference the fixtures were generated with):

• encode: ids = UTF-8 byte + 3, one trailing EOS; int64, one row;
• encoder: one row per call, no padding, no attention mask;
• decoder: greedy argmax until EOS, 512-step cap, prefix [PAD] then growing;
• output: max(0, t - 3) per id, lossy UTF-8.

Structs

OrtEngine

An ort-loaded IMF artifact: encoder + decoder sessions plus the ONNX Runtime environment they share.

Constants

MAX_DECODE_STEPS

Decode-step step cap. CONFORMANCE.md: “greedy argmax until EOS, 512-step cap”.

Traits

InferenceEngine

The engine abstraction: one encoder pass + one decoder step at a time. ort sessions implement it; tests drive the loop with a deterministic mock.

Functions

greedy_generate

The greedy single-row decode loop (model.generate + tokens.decode parity). Empty input (only the trailing EOS) short-circuits to no tokens, exactly like the reference loader.

ids_to_fixture_bytes

The fixture-generation byte rule: max(0, t - 3) per id (no pad/unk skipping — a PAD argmax yields byte 0x00 there), then lossy UTF-8. Used by the conformance harness so frozen vectors are compared byte-exactly under the rule that produced them.

ids_to_fixture_string

translate

translate(text): greedy ids -> text. Conformance fixtures were generated with the max(0, t - 3) byte rule (see ids_to_fixture_bytes); tokens::decode implements the SPEC §4 table (skip pad/unk, mod 256). This function uses the SPEC table.