Every step of the LLM workflow, from finding a model to shipping it on-device, lives inside LLMForge. No context switching. No config files.
Step 01
Search the entire model hub. See architecture, size, and RAM requirement before you download. One click pulls it into your local workspace.
Phi-3 Mini Llama 3.2 Qwen 2.5 Gemma 3 Many more
Phi-3 Mini microsoft/phi-3-mini-4k · 3.8B params
2.2 GB
Llama 3.2 3B meta-llama/llama-3.2-3b-instruct
2.0 GB
Qwen 2.5 1.5B Qwen/qwen2.5-1.5b-instruct
1.0 GB
Gemma 3 1B google/gemma-3-1b-it
0.7 GB
Step 02
Import CSV/JSONL, label manually, or have a local model generate pairs you accept or reject. Always outputs clean Alpaca or ChatML, ready for MLX.
Import Manual Label AI-Assisted JSONL
// dataset.jsonl — 847 examples { "instruction": "Explain LoRA in one line", "input": "", "output": "LoRA adds small trainable matrices to frozen layers for efficient fine-tuning." } { "instruction": "What is quantization?", "input": "", "output": "Reducing model precision..." }
Step 03
Runs natively on MLX — no CUDA, no cloud GPUs. Configure LoRA rank, learning rate, and epochs. Watch the loss curve descend in real time.
MLX Native LoRA / QLoRA Live Loss Checkpoints
Epoch 2/3 · Step 142/20067%
Step 04
Pick your quantization level. Balance file size against quality. One click converts and exports a ship-ready model you can drop into Xcode.
GGUF CoreML llama.cpp Xcode-ready
Q8_0 — Best quality7.2 GB
Q4_K_M — Recommended3.8 GB
Step 05
Same prompt, two models, simultaneous responses. Compare quality and speed. Save great outputs back to your dataset for the next training run.
A/B Compare tok/sec Feedback Loop
Base Model
LoRA is a method for adapting large language models using low-rank matrix decomposition techniques...
12.4 tok/sec
Fine-tuned ✓
LoRA injects small trainable rank-decomposition matrices alongside frozen weights, enabling efficient domain-specific adaptation.
11.8 tok/sec
Step 06
Spin up an OpenAI-compatible API from any fine-tuned model. Point your app at localhost, test with real requests, iterate in seconds. No deployment needed.
Local API Server OpenAI Compatible One-click Start cURL Ready
You
Explain what LoRA does in one sentence.
API · Fine-tuned Model
LoRA injects small trainable low-rank matrices into frozen layers, enabling efficient task-specific adaptation without full retraining.
curl http://localhost:8080/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"messages": [{"role": "user", "content": "Hello!"}], "max_tokens": 256, "temperature": 0.7}'
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。