aq changelog
- aq v0.0.4
Fitting every machine
New
- Machine-aware compute: probes VRAM; VLMs no longer hardcode float32. Auto bf16/fp16, batch, gradient checkpointing; OOM auto-retry shrink.
- Windows install: the installer now works cleanly on Windows (Git Bash and
aq.cmd). -
aq update: reinstall the latest release (same entrypoint ascurl …/install.sh | bash).
Fixes
- VLM vision towers:
vision:accepts full ViT sizes (vit-h/14,ViT-H-14, …) and optional OpenCLIP weights viavision_pretrained/pretrained. aq servefor every method: generation on VLM (LLaVA/Flamingo/CLIP), vision CNN/ViT, and tabular;--image/serve.imagefor multimodal. LLM methods unchanged.
- aq v0.0.3
Teaching models to see
New
- Vision CNNs:
family: vision/method: cnnwith aq-owned LeNet, AlexNet, VGG, ResNet, Inception, EfficientNet, ConvNeXt. ImageFolder or path+label table. Train, eval, inspect. - Vision transformers:
method: vitwith aq-owned ViT, Swin, DeiT (CNN teacher), BEiT (VQ + block MIM → classify). - Vision-language:
family: vlm: CLIP / SigLIP (aq dual encoders, recall@1); LLaVA / GPT-4V-style (projector + causal LM); Flamingo (perceiver + gated cross-attn). - Train TUI: live dashboard on a real terminal (bars, stats, step table). Pipes/CI stay plain tables.
AQ_TUI=0turns it off. - Eval TUI: scoreboard of its own (metric / score / verdict / probes), not the train monitor.
Fixes
speculative: true: assisted decode at serve withdraft_model:.formats: true: real weight export after train (HF + GGUF; GPTQ/AWQ on CUDA; EXL2 whenconvert.pyis set). Preferformats: [gguf]on Mac.
- Vision CNNs:
- aq v0.0.2
Refining and accelerating
New
-
aq plot: charts for loss, jobs, and runs. Ask the agent for a graph, or runaq plotyourself. -
aq spawn --kill: optional critic agent that tries to break your idea in the background.
Cleaner train folder
- Fewer folders on
aq init. Usetools/for scripts and custom training code,jobs/plans/for cron and sweeps. - Agent memory moved to
~/.aq/memory/(not inside the train). Oldmemory/folders migrate automatically. instructions.mdis nowexperiment.md. Old files rename themselves the first time you runaq statusoraq train.- Dropped unused slots:
connections/,methods/,sandbox/,train.ts. aq schedulestill works: it runsaq job plan.
Agent
- After
aq init, tell the agent the new folder name: it won't lose track. - The agent waits for you to say go before training or eval.
Install
- Re-running the install script cleanly replaces the old copy (fixes upgrade from 0.0.1).
Training fixes
- CatBoost and LightGBM work out of the box.
- Llama/GPT transformer trains no longer fail on missing pad tokens.
- MLM works on causal models like Llama, not just BERT.
- Multi-token prediction (
mtp) actually trains extra heads now. paged_kvandspeculativeno longer block training.
-
- aq v0.0.1
Building a developer environment for ML
Concept
aqis the developer environment for foundation models: frameworks, libraries, tools, and autonomous agents that help you design, test, and refine models. Built for precision and intent. Think Next.js for JS, but for ML: architecture, experiments, and building AI the way it should be built, deliberately, not by guessing.Trains live on disk. A train is a folder with your data, recipe, and results.
aqis the CLI that runs them: train, eval, queue jobs, and chat with an agent that works inside that folder, with the same precision as writing code.Install
curl -fsSL https://aq.aquin.app/framework/install.sh | bash
Needs Node 18+, Python 3, and npm. After install:
aq helpandaq doctor.New in 0.0.1
Start a project
aq init: new experiment folder (aq-experiment, or-new1,-new2, … if that name exists)aq init my-project: same, with your name-
aq fork: copy an experiment (skips old job logs and checkpoints)
Train & evaluate
aq train: fit fromrecipe.yamlaq eval: score against files inevals/aq serve: try generation from a checkpoint- Tabular ML: linear/logistic regression, ridge, lasso, elastic net, trees, forests, boosting, Gaussian processes
- LLM work: pretrain, SFT, full fine-tune, LoRA, QLoRA (NVIDIA + 4-bit), masked LM, span corruption, fill-in-the-middle, continued training from a checkpoint
- Transformers: encoder, decoder, encoder-decoder (set
model:in the recipe)
Work in the background
-
aq job run: run a command as a queued job aq job list | log | cancel | resume | tree
Agent & chat
-
aq: interactive chat in the terminal aq ask: one-shot question, no UIaq chat list: pick up old conversations (stored under~/.aq/chats/)-
aq spawn: start a background agent on a task - Plug in your own scripts (
tools/) and skill packs (skills/)
Other
aq provider: OpenAI, Anthropic, Grok, or Ollamaaq schedule: chains like train → eval, sweeps, cron-style rerunsaq stage: sub-experiments insidestages/aq status,aq diff,aq checkout: see what ran and go back to an older stateaq version,aq doctor
Fixes in 0.0.1
- Install:
aqcommand works after curl install (no broken path / missing binary) - Install: Python deps on Ubuntu/Debian (venv + pip setup)
- Training: faster LM runs on typical GPUs; less likely to OOM on a T4-class card
- Chat: conversations live in one global place, not scattered per folder
This is only the first cut. We want aq to feel like a calm desk for teaching models: yours to keep, easy to share, honest about what it can do. If you are curious about how machines learn, you already belong here. Build something small. See what it becomes.
