TRL

Compression at the Edge — NVIDIA, Unsloth, HuggingFace, Ollama

Aug 7, 2026 · 46:01

NVIDIA's Chris Alexiuk, Unsloth's Daniel Han, NVIDIA's Asma Beevi, Hugging Face's Merve Noyan, and Ollama's Parth Sareen argue compression democratizes AI: GLM 5.2 shrinks from 1.5 terabytes to 250 GB, 86% smaller without being 86% dumber. Han says layers are unequal—first/last critical, middle near-useless, and one 'super weight' can make a model 20% dumber—so layer choice is a combinatorial search. Asma details NVFP4, 4-bit floats sharing an FP8 scale per 16 values, targeting under 1% accuracy loss and working out of the box above ~20B parameters. Benchmarks only verify tasks—Han uses KL divergence of BF16 vs quantized logits, Ollama tests quants in real harnesses—and linear-attention models can break heuristics before KV-cache compression pushes models to phones.

Everything you need to know about Fine-tuning and Merging LLMs: Maxime Labonne

Sep 25, 2024 · 17:52

Maxime Labonne from Liquid AI explains the LLM training lifecycle—pre-training, supervised fine-tuning (SFT), and preference alignment—and when to use fine-tuning over prompt engineering. He details SFT dataset creation (accuracy, diversity, complexity) and techniques like full fine-tuning, LoRA, and QLoRA, with key hyperparameters. The core of the talk is model merging: combining weights of fine-tuned models without GPU, using methods like SLERP (spherical linear interpolation for two models), TIES (pruning redundant parameters to merge many models), pass-through (concatenating layers, e.g., Meta LLaMA 3 120B Instruct by repeating layers gives strong creative writing), and Franken-MoE (extracting FFN layers from domain-specific models with a router). Labonne demonstrates these with his NeuralBeagle and Beyonder models, noting merged models dominate the OpenLLM leaderboard and that TIES merging often outperforms more experimental Mixture of Experts approaches.