small100-onnx / examples /ios /Small100Translator.swift
casawolice's picture
Upload SMaLL-100 universal ONNX (int8 model, tokenizer.json, lang map, 4-platform examples)
a0d8498 verified
Raw History Blame Contribute Delete
3.64 kB
// Reference SMaLL-100 translator for iOS (onnxruntime-swift + swift-transformers).
//
// Swift Package deps:
// .package(url: "https://github.com/microsoft/onnxruntime-swift-package-manager", from: "1.20.0")
// .package(url: "https://github.com/huggingface/swift-transformers", from: "0.1.8")
//
// Bundle onnx/encoder_model.onnx, onnx/decoder_model_merged.onnx, tokenizer.json,
// lang_tokens.json. This is a reference implementation of USAGE.md; adjust to your
// onnxruntime-swift / swift-transformers versions.
import Foundation
import onnxruntime_objc
import Tokenizers // from swift-transformers
final class Small100Translator {
private let env = try! ORTEnv(loggingLevel: .warning)
private let encoder: ORTSession
private let decoder: ORTSession
private let tokenizer: Tokenizer // loaded from tokenizer.json
private let langToId: [String: Int64]
private let nLayers = 3, eos: Int64 = 2, decoderStart: Int64 = 2
init(encoderPath: String, decoderPath: String, tokenizer: Tokenizer, langToId: [String: Int64]) throws {
let opts = try ORTSessionOptions()
self.encoder = try ORTSession(env: env, modelPath: encoderPath, sessionOptions: opts)
self.decoder = try ORTSession(env: env, modelPath: decoderPath, sessionOptions: opts)
self.tokenizer = tokenizer
self.langToId = langToId
}
func translate(_ text: String, to tgt: String, maxLen: Int = 128) throws -> String {
// 1) ids = [tgt lang token] + tokenizer(text) (tokenizer appends </s>=2)
var ids: [Int64] = [langToId[tgt]!]
ids += tokenizer.encode(text: text).map { Int64($0) }
let mask = [Int64](repeating: 1, count: ids.count)
// 2) encoder
let idsT = try tensor(ids, [1, ids.count])
let maskT = try tensor(mask, [1, mask.count])
let encOut = try encoder.run(withInputs: ["input_ids": idsT, "attention_mask": maskT],
outputNames: ["last_hidden_state"], runOptions: nil)
let encHidden = encOut["last_hidden_state"]!
// 3) greedy decode over merged decoder (KV cache)
var out: [Int] = []
var cur = decoderStart, useCache = false
var past = try emptyPast()
for _ in 0..<maxLen {
var feeds: [String: ORTValue] = [
"encoder_attention_mask": maskT,
"input_ids": try tensor([cur], [1, 1]),
"encoder_hidden_states": encHidden,
"use_cache_branch": try boolTensor(useCache),
]
past.forEach { feeds[$0.key] = $0.value }
let res = try decoder.run(withInputs: feeds,
outputNames: decoderOutputNames(), runOptions: nil)
let next = try argmaxLastLogits(res["logits"]!)
if next == eos { break }
out.append(Int(next))
past = rollPast(res, wasCache: useCache, prev: past)
cur = next; useCache = true
}
return tokenizer.decode(tokens: out)
}
// --- helpers (tensor creation, KV rolling, argmax) omitted for brevity ---
// tensor(_:)/boolTensor(_:): wrap [Int64]/Bool into ORTValue(tensorData:elementType:.int64/.bool shape:).
// emptyPast(): 3 layers × {decoder,encoder} × {key,value} float tensors of shape [1,16,0,64].
// rollPast(): present.i.decoder.* every step; present.i.encoder.* only when !wasCache, else keep prev.
// argmaxLastLogits(): argmax over the last position of logits [1, seq, 128112].
// See USAGE.md and examples/android/Small100Translator.kt for the exact mapping.
}