MOLT AI Infrastructure

Local model training. Hardware-aware execution. Measurable results.

MOLT is a Windows-first training runtime for developers and researchers working on consumer NVIDIA laptops and workstations. It brings small-model pretraining, QLoRA fine-tuning, thermal pacing, checkpoint recovery, and experiment reporting into one command-line workflow.

Current release: 0.10.0a1 · Open-source research alpha. Suitable for evaluation and controlled experiments. Production use requires workload-specific validation; MOLT does not currently offer a commercial support SLA or certified reliability.

Install · Quick start · Documentation · Contribute

Install in one command

Run in Windows Command Prompt, from a directory where you want a new MOLT folder:

git clone https://github.com/PraveenNimilka/MOLT.git && cd MOLT && powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1

Prerequisites: Git, Windows 10/11 x64, a supported NVIDIA GPU with a compatible driver, internet access, and several GB of free disk space. Repository access is required. Stop existing training before installation or updates.

The installer provisions a project-local .venv, obtains uv if needed, and installs the locked CUDA PyTorch, QLoRA, and Windows Triton dependencies. It then checks dependency imports, CUDA backward computation, and a small compiled backward pass against eager results. A failed check stops setup with an error.

No administrator rights are required by the MOLT script. It does not install GPU drivers, change Defender settings, suspend applications, or adjust power limits. Review install.ps1 before running it; the execution-policy override applies only to that PowerShell invocation.

Already downloaded the repository? Open its folder and run:

powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1

Use -EagerOnly to omit Triton and compilation checks, or -Plan to preview setup without installing anything. Full prerequisites and troubleshooting are in the installation guide.

For an isolated Python environment, the tagged source can also be installed with pip. Install CUDA PyTorch from its official index first; otherwise pip can resolve the CPU-only wheel on Windows:

py -m pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cu128; if ($?) { py -m pip install "moltengine[qlora,data,windows-fusion] @ git+https://github.com/PraveenNimilka/MOLT.git@v0.10.0-alpha.1" }

After the first PyPI release is published, the second command becomes:

py -m pip install "moltengine[qlora,data,windows-fusion]==0.10.0a1"

Maintainer publishing and supply-chain instructions are in docs/PUBLISHING.md.

Quick start

From the repository folder, launch the guided interface:

.venv\Scripts\molt.exe

Or inspect the environment directly:

.venv\Scripts\molt.exe --ui inspect

The explicit executable path works without activating a virtual environment or adding MOLT to your global PATH. If uv is available on PATH, you can also use uv run molt.

Recommended workflow

Install → doctor → prepare → fit test → train → evaluate → export.

molt doctor identifies exactly which Python environment and source checkout you are running. molt config --init creates a workspace manifest. info and inspect remain supported aliases for their older diagnostic views.

Text, JSONL, and Parquet preparation and text generation are available. Use molt prepare --help and molt generate --help; use molt research --help for experimental benchmarks. Legacy top-level research commands remain compatible.

Train with your own data

MOLT does not download model weights or datasets during setup. A common local fine-tuning flow is now three commands:

.venv\Scripts\molt.exe prepare data\examples.jsonl --model models\Qwen --output molt-workspace\datasets\examples
.venv\Scripts\molt.exe fit-test --config molt-workspace\datasets\examples\training.json
.venv\Scripts\molt.exe --ui train --config molt-workspace\datasets\examples\training.json

Preparation accepts UTF-8 .txt, line-delimited JSON objects (.jsonl), and .parquet. Common text, chat messages, and prompt/completion records are recognized. JSONL and Parquet are streamed and split at record boundaries; they are not advertised as supporting every possible third-party schema.

1. Choose a profile from configs/ and update its data, model, and artifact paths. Example paths are not bundled datasets.
2. Validate the configuration before allocating training resources.
3. Start training, then inspect the resulting run artifacts.

For example, after adapting configs/molt-stream-production.json:

.venv\Scripts\molt.exe train --config configs\molt-stream-production.json --dry-run
.venv\Scripts\molt.exe --ui train --config configs\molt-stream-production.json

A dry run validates the specification and referenced data paths. It does not prove that the model fits in VRAM or that a full run will complete.

Use the guided resume command to select a saved run:

.venv\Scripts\molt.exe --ui resume

See the CLI reference for profiles, run reports, and automation.

Interchange formats

rem Auto selects PEFT safetensors for QLoRA and a MOLT bundle for scratch training
.venv\Scripts\molt.exe export --run RUN_DIRECTORY --output-dir exported-adapter

rem Optional consumer adapter conversion; llama.cpp is deliberately not bundled
.venv\Scripts\molt.exe export --run RUN_DIRECTORY --format gguf --llama-cpp C:\src\llama.cpp --output-dir exported-gguf

What MOLT provides

Experimental layer streaming and optimization components remain research paths. They should not be interpreted as universal support for streaming arbitrary Hugging Face models or as validated improvements over tuned baselines.

Verified single-machine endurance result

One preregistered engineering run used Windows 11 and an RTX 4060 Laptop GPU to train all 28 Qwen2.5-1.5B decoder layers with 9,232,384 active LoRA parameters, context 512, batch 1, accumulation 4, and fused FP32 AdamW. Graphics clocks were temporarily constrained to 1,500-1,650 MHz and restored afterward.

This is a single-machine, single-seed result, not an official comparison with Unsloth, a universal throughput claim, or proof of a novel algorithm. Independent reproduction and the registered multi-seed AB/BA comparison remain open evidence gates.

Safe endurance profile

On supported NVIDIA Windows systems, an Administrator can explicitly authorize MOLT's measured endurance clock range. MOLT restores automatic clocks in a finally block on completion, error, Ctrl+C, or thermal stop:

.venv\Scripts\molt.exe optimize-gpu --profile endurance --config YOUR_CONFIG.json

The command never changes clocks without interactive confirmation (or an explicit -y for automation), refuses non-elevated execution, and verifies the clock range recorded by run telemetry. It does not change firmware, fan curves, Defender, other applications, or unsupported laptop power limits.

See the all-layer thermal frontier and negative results register for full methodology, limitations, and rejected experiments.

Readiness, performance, and safety

MOLT distinguishes installed dependencies, successful runtime checks, and validated training workloads. They are not interchangeable.

• Model fit is workload-dependent. An 8B dependency-readiness flag does not guarantee an 8B model fits in 8 GB of VRAM.
• Memory mapping is not zero-RAM. Mapped pages, OS caches, batches, model weights, and optimizer state still consume memory.
• Pacing is software control, not hardware protection. It cannot guarantee flat temperatures or prevent every throttle, OOM, or power shutdown.
• Checkpoints reduce recovery risk, not all data loss. Unsaved steps can be lost; checksums and atomic publication do not guarantee survival of every filesystem or hardware failure.
• Optional tools may remain unavailable. Liger and native MSVC/nvcc development tooling are not required for every training path and are not included in the default installer.
• Prioritized mode affects only MOLT. Its temporary process priority is restored after normal completion or a training exception. Batch geometry, thermal settings, other applications, and Defender are not silently changed.

Historical measurements and methodology are retained in the benchmarking guide. Results with different validation quality are not proof of an equal-quality speed or energy advantage. This release does not claim a universal throughput target or a new training-algorithm breakthrough.

See the release verification notes for automated tests and physical runtime checks. Fresh-machine bootstrap and sustained workload behavior still require broader reproduction. The live CI badge represents the latest GitHub test status.

Update an existing installation

From your existing checkout, with training stopped:

git status
git pull --ff-only origin main
powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1

If Git reports local changes or diverged history, resolve them before updating; do not force-reset your work. Setup does not delete model weights, datasets, or run artifacts. Back up important checkpoints before changing environments.

Older MOLT versions could add Defender exclusions. These are not removed automatically because ownership cannot be inferred safely. Review unwanted entries manually in Windows Security; see the 0.9.1 release notes.

Documentation

• Installation and troubleshooting
• Command-line reference
• Architecture
• Benchmark methodology
• Development guide
• Release notes
• Contributing
• Security policy

License and feedback

MOLT is distributed under the MIT License, which permits commercial use subject to its terms. Model weights, datasets, and dependencies retain their own licenses.

Report reproducible bugs through GitHub Issues. Include the commit, Python/PyTorch versions, relevant configuration, and error traceback. Remove credentials, private data, and sensitive paths before sharing.