Skip to content

ROADMAP.md

Latest commit

 

History

History
133 lines (95 loc) · 4.1 KB

ROADMAP.md

File metadata and controls

133 lines (95 loc) · 4.1 KB
Raw

UMC Roadmap

Universal Model Converter — the ffmpeg of AI models.

This document describes the planned development for the next versions of UMC. It is a living document: priorities are reviewed with each release.

Legend:

  • ✅ done
  • 🚧 in progress
  • 📋 planned
  • 💡 idea (not yet scheduled)

Current state (v1.0.0)

Feature Status
Universal Intermediate Representation (IR) ✅
Format detection (magic bytes, extension, content) ✅
GGUF loader ✅
SafeTensors loader / saver ✅
SentencePiece loader ✅
ONNX loader / saver 🚧 (opset 13–21)
PyTorch loader / saver 🚧
TFLite loader / saver 🚧
AWQ / GPTQ / LoRA support 📋
Conversion pipeline (mmap, rayon, cancel) ✅
Structural & numeric validation, certificates ✅
CLI (convert, inspect, formats, path) ✅
Desktop app (egui/eframe) ✅
REST API (umc-api, Actix-Web) ✅

v1.1 — Hardening & format completeness

Target: reliability of the core before adding surface area.

Formats

  • 📋 Finish ONNX loader/saver: full opset 13–21 tensor support, batch dimensions, external data files (.onnx.data).
  • 📋 Finish PyTorch saver: quantized tensors, non-contiguous storages, proper stride handling on load.
  • 📋 Finish TFLite saver: quantized (int8/uint8) tensors, buffer dedup.
  • 📋 GGUF saver (currently loader only).
  • 📋 Quantization support in the pipeline: --q8, --q4 (GGUF-style block quant) as a first-class IR transformation.

Reliability

  • 📋 Fuzz harness for every loader (cargo-fuzz) — malformed files must never panic or OOM.
  • 📋 Property-based tests for round-trips across dtypes and shapes.
  • 📋 Memory budget: cap peak RSS during huge-model conversions, add streaming path for tensors > 2 GiB.
  • 📋 Error taxonomy: machine-readable error codes in CLI/API.

UX

  • 📋 Desktop: batch conversion queue (multiple files in, multiple out).
  • 📋 Desktop: per-format options panel (quantization, dtype cast).
  • 💡 Desktop: conversion profiles (presets).

v1.2 — Quantization & compression

Target: make UMC the reference tool for shrinking models losslessly.

  • 📋 Block-quantization library (Q4_0, Q4_K, Q5_K, Q6_K, Q8_0) matching llama.cpp bit layouts exactly, verified against ggml.
  • 📋 Dtype cast with round-trip certificates (F32 → F16 → F32 bit-identical checks).
  • 📋 LoRA merge (adapter + base → full weights) in the pipeline.
  • 📋 AWQ / GPTQ dequantize → requantize flows.
  • 💡 Weight clustering / pruning passes (experimental, opt-in).

v1.3 — More formats

Target: broad coverage of the AI ecosystem.

  • 📋 Keras H5 (read-only) via native HDF5 parser.
  • 📋 GGML legacy (read-only).
  • 📋 CoreML (export), TensorRT engine (via plugin), OpenVINO (export).
  • 📋 MLIR / StableHLO import (research-grade).
  • 💡 vLLM / tensor-parallel checkpoint sharding (split & merge).

v2.0 — Ecosystem & automation

Target: embeddable, scriptable, verified at scale.

  • 📋 umc-core as a stable public API (libumc) with docs.rs coverage.
  • 📋 Library API for embeddings: UMCPlugin trait, metadata schemas.
  • 📋 WASM target for the detection + core pipeline (browser demos).
  • 📋 API: webhook notifications, job priority, quotas.
  • 💡 Model cards: read/write HuggingFace model card metadata.
  • 💡 Diff/summary of two models (umc diff a.bin b.bin) — structural + numeric report.

v2.1 — Speed

Target: "ffmpeg-fast" claim.

  • 📋 Threaded loader overlap (read-ahead + parse on multiple threads).
  • 📋 SIMD kernels for dtype conversion (AVX2/NEON via std::simd).
  • 💡 GPU-accelerated conversion (WGSL/WebGPU backend for quant ops).
  • 💡 Incremental re-conversion cache keyed by content hash.

Non-goals

  • Training or fine-tuning models.
  • Runtime/inference engines (UMC converts, it does not run models).
  • Web frontend (desktop + CLI are the interfaces; API serves automation).

How to contribute

See CONTRIBUTING.md. Roadmap items are tracked as GitHub issues tagged with their target milestone.