Will it be possible to fine-tune a 65B parameter model with 30GB of GPU memory (average) by the end of 2023?
14
290Ṁ299resolved Mar 10
Resolved
N/A1H
6H
1D
1W
1M
ALL
QLoRA reduced the avg memory requirements from 750+ GB to < 48 GB of GPU memory (average) for a 65B model.
They checked by training 1000 models across several different instruction sets + architectures + parameter ranges [80M, 65B].
Will it be possible to reduce it further? Not just on 1 model but reliably, need consistent and compelling evidence.
This question is managed and resolved by Manifold.
Get 
1,000 to start trading!
1,000People are also trading
Will we see a public GPU compute sharing pool for LLM model training or inference before 2026 ?
86% chance
Will a model be trained using at least as much compute as GPT-3 using AMD GPUs before Jan 1 2026?
84% chance
Will there be an announcement of a model with a training compute of over 1e30 FLOPs by the end of 2025?
5% chance
An AI model with 100 trillion parameters exists by the end of 2025?
20% chance
Will a single model running on a single consumer GPU (<1.5k 2020 USD) outperform GPT-3 175B on all benchmarks in the original paper by 2025?
86% chance
Will an AI model use more than 1e28 FLOPS in training before 2026?
8% chance
By March 14, 2025, will there be an AI model with over 10 trillion parameters?
5% chance
Will a lab train a >=1e26 FLOP state space model before the end of 2025?
15% chance
Limits on AI model size by 2026?
15% chance
Will Aidan McLau's claim that very large models are "refusing instruction tuning" be validated by 2030?
42% chance
Sort by:
This market needs clarification regarding time to finetune, and finetuned model performance, required for something to count as "finetuning".
Otherwise, I can trivially finetune even a 1T parameter model with zero gpus, because finetuning a model is a computational operation and regular non-gpu computers are Turing-complete.
People are also trading
Related questions
Will we see a public GPU compute sharing pool for LLM model training or inference before 2026 ?
86% chance
Will a model be trained using at least as much compute as GPT-3 using AMD GPUs before Jan 1 2026?
84% chance
Will there be an announcement of a model with a training compute of over 1e30 FLOPs by the end of 2025?
5% chance
An AI model with 100 trillion parameters exists by the end of 2025?
20% chance
Will a single model running on a single consumer GPU (<1.5k 2020 USD) outperform GPT-3 175B on all benchmarks in the original paper by 2025?
86% chance
Will an AI model use more than 1e28 FLOPS in training before 2026?
8% chance
By March 14, 2025, will there be an AI model with over 10 trillion parameters?
5% chance
Will a lab train a >=1e26 FLOP state space model before the end of 2025?
15% chance
Limits on AI model size by 2026?
15% chance
Will Aidan McLau's claim that very large models are "refusing instruction tuning" be validated by 2030?
42% chance