LLM Pruning and Distillation in Practice: The Minitron Approach
read the original abstract
We present a comprehensive report on compressing the Llama 3.1 8B and Mistral NeMo 12B models to 4B and 8B parameters, respectively, using pruning and distillation. We explore two distinct pruning strategies: (1) depth pruning and (2) joint hidden/attention/MLP (width) pruning, and evaluate the results on common benchmarks from the LM Evaluation Harness. The models are then aligned with NeMo Aligner and tested in instruct-tuned versions. This approach produces a compelling 4B model from Llama 3.1 8B and a state-of-the-art Mistral-NeMo-Minitron-8B (MN-Minitron-8B for brevity) model from Mistral NeMo 12B. We found that with no access to the original data, it is beneficial to slightly fine-tune teacher models on the distillation dataset. We open-source our base model weights on Hugging Face with a permissive license.
This paper has not been read by Pith yet.
Forward citations
Cited by 20 Pith papers
-
FTerViT: Fully Ternary Vision Transformer
FTerViT introduces fully ternary Vision Transformers with TernaryBitConv2d and TernaryLayerNorm operators, achieving 82.43% ImageNet top-1 at 6.09 MB with 15x compression.
-
Computational Lesions in Multilingual Language Models Separate Shared and Language-specific Brain Alignment
Lesioning a shared core in multilingual LLMs drops whole-brain fMRI encoding correlation by 60.32%, while language-specific lesions selectively weaken predictions only for the matched native language.
-
Orak: A Foundational Benchmark for Training and Evaluating LLM Agents on Diverse Video Games
Orak is a foundational benchmark providing training data, interfaces, and evaluation tools for LLM agents across diverse video game genres.
-
Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing
Magpie synthesizes 300K high-quality alignment instructions from Llama-3-Instruct via auto-regressive prompting on partial templates, enabling fine-tuned models to match official instruct performance on AlpacaEval, Ar...
-
DiARC: Distinguishing Positive and Negative Samples Helps Improving ARC-like Reasoning Ability of Large Language Models
DiARC improves LLM performance on ARC-like benchmarks by constructing and training on preference pairs from three types of negative samples while keeping demonstrations fixed.
-
DiARC: Distinguishing Positive and Negative Samples Helps Improving ARC-like Reasoning Ability of Large Language Models
DiARC improves LLM performance on ARC-like tasks by fine-tuning on preference pairs of positive demonstrations and three classes of constructed negative samples.
-
LLM Compression with Jointly Optimizing Architectural and Quantization choices
A differentiable NAS framework jointly optimizes LLM architecture and mixed-precision quantization for linear layers, yielding up to 1.4x faster inference or 6% higher accuracy than sequential baselines on reasoning tasks.
-
Contribution Weights: A Geometrical Analysis of Self-Attention Transformers
Contribution Weights combine attention, value magnitude, and directional alignment to measure token influence more faithfully than attention alone, and show attention sinks actively suppress information via a convex s...
-
Pruning and Distilling Mixture-of-Experts into Dense Language Models
A systematic MoE-to-dense conversion via expert scoring, grouping, and distillation yields +6.3 pp average accuracy over dense-to-dense pruning at matched parameter count on tested models.
-
Resting Neurons, Active Insights: Robustifying Activation Sparsity in LLMs via Spontaneity
SPON adds learnable persistent activation anchors trained via distribution matching to restore LLM accuracy under high activation sparsity by preventing representational distribution shifts.
-
MaskPro: Linear-Space Probabilistic Learning for Strict (N:M)-Sparsity on LLMs
MaskPro learns categorical distributions over groups of M weights to generate exact (N:M) sparsity via N-way sampling without replacement and stabilizes training with a moving average tracker of loss residuals.
-
Small LLMs: Pruning vs. Training from Scratch
Controlled experiments show pruning a large LLM outperforms training small models from scratch with matched token budgets, but the advantage narrows with extra tokens or coarse pruning.
-
Small LLMs: Pruning vs. Training from Scratch
Pruned initializations from an 8B model outperform random starts with equal training tokens, but with full token budgets fine-grained pruning retains advantage while coarse structured pruning does not.
-
Joint Structural Pruning and Mixed-Precision Quantization for LLM Compression
Joint structural pruning and mixed-precision quantization framework for LLMs that optimizes global error and reports lower perplexity at 1-3 bits than prior methods.
-
Structural Pruning of Large Vision Language Models: A Comprehensive Study on Pruning Dynamics, Recovery, and Data Efficiency
Widthwise pruning of LVLM language backbones combined with supervised finetuning and hidden-state distillation recovers over 95% performance using just 5% of data across 3B-7B models.
-
On the Limits of Layer Pruning for Generative Reasoning in Large Language Models
Layer pruning preserves classification performance in LLMs but fundamentally limits recovery of generative reasoning capabilities even after extensive self-supervised finetuning.
-
NVIDIA Nemotron 3: Efficient and Open Intelligence
NVIDIA releases the Nemotron 3 model family with hybrid Mamba-Transformer architecture, LatentMoE, NVFP4 training, MTP layers, and multi-environment RL post-training for reasoning and agentic tasks.
-
Resting Neurons, Active Insights: Robustifying Activation Sparsity in LLMs via Spontaneity
SPON adds a small set of trainable input-independent activation vectors as representational anchors, trained by distribution matching, to stabilize sparse activation in LLMs and recover performance lost to hidden-stat...
-
Beyond Sunk Costs: Boosting LLM Pre-training Efficiency via Orthogonal Growth of Mixture-of-Experts
Orthogonal growth recycles pre-trained MoE checkpoints via layer copying and noisy expert duplication, delivering 10.6% higher accuracy than training from scratch with equivalent extra compute.
-
Ministral 3
Ministral 3 releases 3B/8B/14B parameter-efficient language models with base, instruction, and reasoning variants derived via iterative pruning and distillation, including image understanding capabilities.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.