Skip to content

PUBLIC_API.md

Latest commit

 

History

History
82 lines (62 loc) · 3.44 KB

PUBLIC_API.md

File metadata and controls

82 lines (62 loc) · 3.44 KB
Raw

MobileFineTuner Public API

MobileFineTuner exposes one supported C++ entrypoint:

#include "mobile_finetuner/mobile_finetuner.h"

Downstream projects should link the installed CMake target:

find_package(MobileFineTuner REQUIRED)
target_link_libraries(your_target PRIVATE MobileFineTuner::operators)

The finetune_ops/* headers are still installed because the current examples use them directly, but new applications should include the umbrella header first. Historical or non-buildable APIs are excluded from the supported product surface.

Model weights and datasets are external assets. The library accepts standard HuggingFace-style model directories with config/tokenizer files and either a single model.safetensors file or model.safetensors.index.json plus all referenced shard files. See docs/MODEL_ASSETS.md for the asset contract.

Stable Auto APIs

Applications should discover model and tokenizer assets through the public registry APIs instead of hard-coding per-model training directories. For common fine-tuning code, use AutoModelForCausalLM and AutoTrainer so the application does not branch manually between GPT-2, Gemma, and Qwen graph classes.

#include "mobile_finetuner/mobile_finetuner.h"

auto spec = ops::ModelRegistry::inspect_pretrained(model_dir);
auto tokenizer = ops::TokenizerFactory::from_pretrained(model_dir);

auto model = ops::AutoModelForCausalLM::from_pretrained(model_dir);
model->init_lora(ops::AutoLoraConfig::attention_qkvo());

ops::AutoTrainerConfig trainer_cfg;
trainer_cfg.learning_rate = 2e-4f;
ops::AutoTrainer trainer(*model, trainer_cfg);
auto result = trainer.train_step(input_ids, attention_mask, labels);

ModelRegistry is the C++ equivalent of a small AutoConfig layer: it reads config.json, identifies the supported model family, records tokenizer and SafeTensors asset paths, and exposes default LoRA target names. TokenizerFactory is the matching AutoTokenizer layer: it keeps GPT-2, Qwen, and Gemma tokenizers behind one ops::Tokenizer interface while still using the model-specific tokenization algorithm required by each checkpoint.

AutoModelForCausalLM is the matching model dispatcher. It constructs the concrete graph class, loads SafeTensors with family-correct transpose defaults, exposes one forward(input_ids, attention_mask) call, initializes LoRA, and returns trainable parameters. AutoTrainer provides the shared one-step LoRA training core: forward, LM cross entropy, backward, gradient clipping, Adam update, and zero-grad.

The current Auto APIs cover the shared causal-LM/LoRA training core. Specialized dataset loops, logging, checkpoint schedules, and model-specific diagnostics can still use the lower-level graph classes directly.

Android SDK API

Android applications can consume the AAR built from android-visualizer/mft-sdk. The Java entrypoint is:

import com.mobilefinetuner.sdk.MobileFineTuner;

It wraps the same C++ AutoModelForCausalLM and AutoTrainer objects through JNI. The Android SDK is a packaging layer, not a separate implementation. See docs/ANDROID_SDK.md for build, ABI, asset layout, and Java call-order details.

The Android API also exposes MobileFineTuner.selfTest(File workingDir), which runs a tiny synthetic native LoRA training step on device. This is intended for installation validation and CI/device smoke testing; real applications should still pass their own HuggingFace-style modelDir to MobileFineTuner.open.