Skip to content
The methodOpen data
Back to the collection

SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot

Elias Frantar, Dan Alistarh

2023 · arXiv

Language models
Read the paper
HN score
0
Mentions
6
Readers
5
Months
4
What’s in a score?

From the abstract

We show for the first time that large-scale generative pretrained transformer (GPT) family models can be pruned to at least 50% sparsity in one-shot, without any retraining, at minimal loss of accuracy. This is achieved via a new pruning method called SparseGPT, specifically designed to work efficiently and accurately on massive GPT-family models. We can execute SparseGPT on the largest available open-source models, OPT-175B and BLOOM-176B, in under 4.5 hours, and can reach 60% unstructured sparsity with negligible increase in perplexity: remarkably, more than 100 billion weights from these models can be ignored at inference time. SparseGPT generalizes to semi-structured (2:4 and 4:8) patterns, and is compatible with weight quantization approaches. The code is available at: https://github.com/IST-DASLab/sparsegpt.

Metadata: arXiv. Abstracts retain their original authors’ rights.

In the conversation

Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels

6 selected excerpts · 6 total mentions

renonce
Neutral mention · rule
Essentially the algorithm is about pruning parameters on the fly and making the weight matrix sparse by setting least significant weights to zero, where "least significant" is defined in terms of the rank of the absolute value of the pruned weight within its group? A search for model pruning turns out many results, including https://arxiv.org/abs/2305.11627 which discusses "magnitude pruning" as a baseline, and refers to https://arxiv.org/pdf/2301.00774.pdf, which asserts in the introduction: > First, as shown in Figure 1, SparseGPT can induce uniform layerwise sparsity of up to 60% in e.g. the 175-billion-parameter variant of the OPT family (Zhang et al., 2022), with…
Read the full comment
minor accuracy loss. By contrast, the only known one-shot baseline which easily extends to this scale, Magnitude Pruning (Hagiwara, 1994; Han et al., 2015), preserves accuracy only until 10% sparsity, and completely collapses beyond 30% sparsity. I don't like how these papers boast their own methods by poorly implementing a baseline and describing their own methods using lots of mathematical jargon - the OP's blog post is a breeze in making the methods accessible to people with very little background.
Original comment Show HN: Speeding up LLM inference 2x times (possibly)
yarg
Neutral mention · rule
There's no chance that we've peeked from a bang for buck sense - we still haven't adequately investigated sparse networks. Relevantish: https://arxiv.org/abs/2301.00774 The fact that we can reach those levels of sparseness with pruning also indicates that we're not doing a very good job of generating the initial network conditions. Being able to come up with trainable initial settings for sparse networks across different topologies is hard, but given that we've had a degree of success with pre-trained networks, pre-training and pre-pruning might also allow for sparse networks with minimally compromised learning capabilities. If it's possible to pre-train composable network modules, it might also be feasible to define trainable sparse networks with significantly relaxed topological constraints.
Original comment OpenAI chief goes before US Congress to propose licenses for building AI
sp332
Neutral mention · rule
Training uses gradient descent, so you want to have good precision during that process. But once you have the overall structure of the network, https://arxiv.org/abs/2210.17323 (GPTQ) showed that you can cut down the precision quite a bit without losing a lot of accuracy. It seems you can cut down further for larger models. For the 13B Llama-based ones, going below 5 bits per parameter is noticeably worse, but for 30B models you can do 4 bits. The same group did another paper https://arxiv.org/abs/2301.00774 which shows that in addition to reducing the precision of each parameter, you can also prune out a bunch of parameters entirely. It's harder to apply this optimization because models are usually loaded into RAM densely, but I hope someone figures out how to do it for popular models.
Original comment How to run Llama 13B with a 6GB graphics card
lhl
Neutral mention · rule
While I agree we probably aren't getting exponentially increasing parameter counts (GPT4 is by all accounts 1T paramaters and of course, it is significantly better than GPT3) we are still seeing lots of improvements - 3.5 is much better than 3, based "just" on InstructGPT/RLHF training. Models are getting better as well - LLaMA 30B beats/matches GPT-3 on raw eval benchmarks at 1/6 the parameter count. We're also seeing lots of optimizations with new models (RoPE/RoPER embedding, Swish/GeLU activation, Flash Attention, etc) but I think some the most interesting gains we'll be seeing soon is with inference-optimized training (-70% parameters for +100% compute) [1] combined…
Read the full comment
with sparsity pruning (-50% size w/ almost no loss in accuracy) [2] and quantization [3] which will lead to significantly smaller models performing well. [1] https://www.harmdevries.com/post/model-size-vs-compute-overh... [2] https://arxiv.org/abs/2301.00774 [3] https://openreview.net/forum?id=tcbBPnfwxS
Original comment GitHub Copilot emits GPL code
lhl
Neutral mention · rule
I think "sweet spot" is going to depend on your task, but here's a good recent paper that may give you some more context on thinking about training and model sizes: https://www.harmdevries.com/post/model-size-vs-compute-overh... There have also been quite a few developments on sparsity lately. Here's a technique SparseGPT which suggests that you can prune 50% of parameters with almost no loss in performance for example: https://arxiv.org/abs/2301.00774
Original comment Stability AI Launches the First of Its StableLM Suite of Language Models

Keep following the thread.

More in Language models

More from the same topics.