Pythia releases 16 identically trained LLMs with full checkpoints and data tools to study training dynamics, scaling, memorization, and bias in language models.
hub Canonical reference
Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model
Canonical reference. 100% of citing Pith papers cite this work as background.
44
Pith papers citing it
Background
100% of classified citations
abstract
Pretrained general-purpose language models can achieve state-of-the-art accuracies in various natural language processing domains by adapting to downstream tasks via zero-shot, few-shot and fine-tuning techniques. Because of their success, the size of these models has increased rapidly, requiring high-performance hardware, software, and algorithmic techniques to enable training such large models. As the result of a joint effort between Microsoft and NVIDIA, we present details on the training of the largest monolithic transformer based language model, Megatron-Turing NLG 530B (MT-NLG), with 530 billion parameters. In this paper, we first focus on the infrastructure as well as the 3D parallelism methodology used to train this model using DeepSpeed and Megatron. Next, we detail the training process, the design of our training corpus, and our data curation techniques, which we believe is a key ingredient to the success of the model. Finally, we discuss various evaluation results, as well as other interesting observations and new properties exhibited by MT-NLG. We demonstrate that MT-NLG achieves superior zero-, one-, and few-shot learning accuracies on several NLP benchmarks and establishes new state-of-the-art results. We believe that our contributions will help further the development of large-scale training infrastructures, large-scale language models, and natural language generations.
hub tools
citation-role summary
background 9
citation-polarity summary
roles
background 9polarities
background 9representative citing papers
Heterogeneous parallelism decouples module layouts in multimodal LLM training via boundary communicators, yielding up to 49.3% TFLOPS/GPU gains in colocated mode and 13% throughput in non-colocated mode with convergence parity.
SELECT-LLM is the first active model selection framework for LLMs that uses expected information gain from pairwise output similarities to minimize required annotations, reporting up to 84.78% cost reduction across 23 datasets and 156 models.
Simulation study shows cold TLB misses in reverse address translation dominate latency for small collectives in multi-GPU pods, causing up to 1.4x degradation, while larger ones see diminishing returns.
o1-like models overthink easy tasks; self-training reduces compute use without accuracy loss on GSM8K, MATH500, GPQA, and AIME.
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
PoT prompting improves numerical reasoning by having language models write programs executed by a computer instead of performing calculations in natural language chains of thought, with an average 12% gain over CoT.
Adding the fixed prompt 'Let's think step by step' enables large language models to achieve substantial zero-shot gains on arithmetic, symbolic, and logical reasoning benchmarks without any task-specific examples.
OPT releases open decoder-only transformers up to 175B parameters that match GPT-3 performance at one-seventh the carbon cost, along with code and training logs.
M²RNN achieves perfect state tracking at unseen lengths and outperforms Gated DeltaNet hybrids by 0.4-0.5 perplexity on 7B models with 3x smaller recurrent states.
veScale-FSDP uses RaggedShard and structure-aware planning to support block-wise quantization and non-element-wise optimizers while delivering 5-66% higher throughput and 16-30% lower memory than prior FSDP systems at massive scale.
Chameleon provides adaptive fault tolerance for distributed training by real-time selection of optimal recovery policies via a unified performance model, demonstrated with low overhead on a 32-card cluster.
Math reasoning gains in LLMs rarely transfer to general domains; RL tuning generalizes while SFT causes forgetting and representation drift.
MiniMax-01 models match GPT-4o and Claude-3.5-Sonnet performance while providing 20-32 times longer context windows through lightning attention and MoE scaling.
Falcon-180B is a 180B-parameter open decoder-only model trained on 3.5 trillion tokens that approaches PaLM-2-Large performance at lower cost and is released with dataset extracts.
Repeating training data up to 4 epochs yields negligible loss increase versus unique data for fixed compute, and a new scaling law accounts for the decaying value of repeated tokens and excess parameters.
Distilling step-by-step uses LLM-generated rationales as additional supervision in a multi-task framework so that 770M-parameter models outperform 540B-parameter models on NLP benchmarks with only 80% of the data.
Pre-trained LLMs using recursive criticism and improvement prompting achieve state-of-the-art results on the MiniWoB++ computer task benchmark with only a handful of demonstrations and no task-specific reward function.
FP8 formats E4M3 and E5M2 match 16-bit training accuracy on CNNs, RNNs, and Transformers up to 175B parameters without hyperparameter changes.
Atlas reaches over 42% accuracy on Natural Questions with only 64 examples, outperforming a 540B-parameter model by 3% with 50x fewer parameters.
Autoregressive language models trained on data with middle spans relocated to the end learn infilling without degrading left-to-right perplexity or sampling quality.
MRKL is a modular neuro-symbolic architecture that integrates LLMs with external knowledge and discrete reasoning to overcome limitations of pure neural language models.
GPT-NeoX-20B is a publicly released 20B parameter autoregressive language model trained on the Pile that shows strong gains in five-shot reasoning over similarly sized prior models.
DeepSpeed-Ulysses keeps communication volume constant for sequence-parallel attention when sequence length and device count scale together, delivering 2.5x faster training on 4x longer sequences than prior SOTA.
citing papers explorer
-
Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
Pythia releases 16 identically trained LLMs with full checkpoints and data tools to study training dynamics, scaling, memorization, and bias in language models.
-
Heterogeneous Parallelism for Multimodal Large Language Model Training
Heterogeneous parallelism decouples module layouts in multimodal LLM training via boundary communicators, yielding up to 49.3% TFLOPS/GPU gains in colocated mode and 13% throughput in non-colocated mode with convergence parity.
-
Large Language Model Selection with Limited Annotations
SELECT-LLM is the first active model selection framework for LLMs that uses expected information gain from pairwise output similarities to minimize required annotations, reporting up to 84.78% cost reduction across 23 datasets and 156 models.
-
Analyzing Reverse Address Translation Overheads in Multi-GPU Scale-Up Pods
Simulation study shows cold TLB misses in reverse address translation dominate latency for small collectives in multi-GPU pods, causing up to 1.4x degradation, while larger ones see diminishing returns.
-
Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs
o1-like models overthink easy tasks; self-training reduces compute use without accuracy loss on GSM8K, MATH500, GPQA, and AIME.
-
DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
-
Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks
PoT prompting improves numerical reasoning by having language models write programs executed by a computer instead of performing calculations in natural language chains of thought, with an average 12% gain over CoT.
-
Large Language Models are Zero-Shot Reasoners
Adding the fixed prompt 'Let's think step by step' enables large language models to achieve substantial zero-shot gains on arithmetic, symbolic, and logical reasoning benchmarks without any task-specific examples.
-
OPT: Open Pre-trained Transformer Language Models
OPT releases open decoder-only transformers up to 175B parameters that match GPT-3 performance at one-seventh the carbon cost, along with code and training logs.
-
M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling
M²RNN achieves perfect state tracking at unseen lengths and outperforms Gated DeltaNet hybrids by 0.4-0.5 perplexity on 7B models with 3x smaller recurrent states.
-
veScale-FSDP: Flexible and High-Performance FSDP at Scale
veScale-FSDP uses RaggedShard and structure-aware planning to support block-wise quantization and non-element-wise optimizers while delivering 5-66% higher throughput and 16-30% lower memory than prior FSDP systems at massive scale.
-
Chameleon: Adaptive Fault Tolerance for Distributed Training via Real-time Policy Selection
Chameleon provides adaptive fault tolerance for distributed training by real-time selection of optimal recovery policies via a unified performance model, demonstrated with low overhead on a 32-card cluster.
-
Does Math Reasoning Improve General LLM Capabilities? Understanding Transferability of LLM Reasoning
Math reasoning gains in LLMs rarely transfer to general domains; RL tuning generalizes while SFT causes forgetting and representation drift.
-
MiniMax-01: Scaling Foundation Models with Lightning Attention
MiniMax-01 models match GPT-4o and Claude-3.5-Sonnet performance while providing 20-32 times longer context windows through lightning attention and MoE scaling.
-
The Falcon Series of Open Language Models
Falcon-180B is a 180B-parameter open decoder-only model trained on 3.5 trillion tokens that approaches PaLM-2-Large performance at lower cost and is released with dataset extracts.
-
Scaling Data-Constrained Language Models
Repeating training data up to 4 epochs yields negligible loss increase versus unique data for fixed compute, and a new scaling law accounts for the decaying value of repeated tokens and excess parameters.
-
Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes
Distilling step-by-step uses LLM-generated rationales as additional supervision in a multi-task framework so that 770M-parameter models outperform 540B-parameter models on NLP benchmarks with only 80% of the data.
-
Language Models can Solve Computer Tasks
Pre-trained LLMs using recursive criticism and improvement prompting achieve state-of-the-art results on the MiniWoB++ computer task benchmark with only a handful of demonstrations and no task-specific reward function.
-
FP8 Formats for Deep Learning
FP8 formats E4M3 and E5M2 match 16-bit training accuracy on CNNs, RNNs, and Transformers up to 175B parameters without hyperparameter changes.
-
Atlas: Few-shot Learning with Retrieval Augmented Language Models
Atlas reaches over 42% accuracy on Natural Questions with only 64 examples, outperforming a 540B-parameter model by 3% with 50x fewer parameters.
-
Efficient Training of Language Models to Fill in the Middle
Autoregressive language models trained on data with middle spans relocated to the end learn infilling without degrading left-to-right perplexity or sampling quality.
-
MRKL Systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning
MRKL is a modular neuro-symbolic architecture that integrates LLMs with external knowledge and discrete reasoning to overcome limitations of pure neural language models.
-
GPT-NeoX-20B: An Open-Source Autoregressive Language Model
GPT-NeoX-20B is a publicly released 20B parameter autoregressive language model trained on the Pile that shows strong gains in five-shot reasoning over similarly sized prior models.
-
DeepSpeed Ulysses: System Optimizations for Enabling Training of Extreme Long Sequence Transformer Models
DeepSpeed-Ulysses keeps communication volume constant for sequence-parallel attention when sequence length and device count scale together, delivering 2.5x faster training on 4x longer sequences than prior SOTA.
-
MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models
MiniGPT-4 shows that aligning a frozen vision encoder to Vicuna via one projection layer plus a second-stage detailed-description fine-tune produces GPT-4-like vision-language abilities including detailed captions, creative writing, and instruction following.
-
BloombergGPT: A Large Language Model for Finance
BloombergGPT is a 50B parameter LLM trained on a 708B token mixed financial and general dataset that outperforms prior models on financial benchmarks while preserving general LLM performance.
-
PaLM: Scaling Language Modeling with Pathways
PaLM 540B demonstrates continued scaling benefits by setting new few-shot SOTA results on hundreds of benchmarks and outperforming humans on BIG-bench.
-
Specific Domain Ontology Construction Using Large Language Models
LLMs produced coherent but incomplete ontologies for the Blue Amazon domain that required human refinement to be fully satisfactory.
-
Charon: A Unified and Fine-Grained Simulator for Large-Scale LLM Training and Inference
Charon is a unified modular simulator that predicts LLM training and inference performance with under 5.35% error and identifies throughput improvements over baselines in a real deployment case.
-
SEDD: Scalable and Efficient Dataset Deduplication with GPUs
SEDD delivers a distributed GPU deduplication system that reports up to 158x speedup over CPU baselines and 7.8x over NeMo Curator on 30M documents while preserving MinHash fidelity above 0.95 Jaccard.
-
MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning
MiniGPT-v2 adds unique task identifiers to a large language model so one system can perform image description, visual question answering, and visual grounding after three-stage training.
-
RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment
RAFT aligns generative models by ranking samples with a reward model and fine-tuning only on the top-ranked outputs, reporting gains on reward scores and automated metrics for LLMs and diffusion models.
-
Transforming the Use of Earth Observation Data: Exascale Training of a Generative Compression Model with Historical Priors for up to 10,000x Data Reduction
A generative compression model using historical priors for Earth observation data achieves up to 10,000x reduction after exascale training on an Armv9 supercomputer.
-
TACO: Efficient Communication Compression of Intermediate Tensors for Scalable Tensor-Parallel LLM Training
TACO compresses tensor-parallel intermediate tensors with an adaptive FP8 scheme and fused kernels, yielding up to 1.87X throughput gains on GPT and Qwen models with near-lossless accuracy.
-
SparseBalance: Load-Balanced Long Context Training with Dynamic Sparse Attention
SparseBalance dynamically adjusts sparsity and batches workloads to load-balance sparse attention training, delivering up to 1.33x speedup and 0.46% better long-context performance on LongBench.
-
StarCoder: may the source be with you!
StarCoderBase matches or beats OpenAI's code-cushman-001 on multi-language code benchmarks; the Python-fine-tuned StarCoder reaches 40% pass@1 on HumanEval while retaining other-language performance.
-
Cross-Layer Energy Analysis of Multimodal Training on Grace Hopper Superchips
On Grace Hopper superchips, energy efficiency during multimodal training is governed by data movement and overlap rather than compute utilization, and runtime-optimal configurations are not always energy-optimal.
-
DeepSeek LLM: Scaling Open-Source Language Models with Longtermism
DeepSeek LLM 67B exceeds LLaMA-2 70B on code, mathematics and reasoning benchmarks after pre-training on 2 trillion tokens and alignment via SFT and DPO.
-
Towards EnergyGPT: A Large Language Model Specialized for the Energy Sector
Fine-tuned LLaMA 3.1-8B variants for the energy sector outperform the base model on domain QA benchmarks, with LoRA delivering similar gains at lower training cost.
-
Phoenix-VL 1.5 Medium Technical Report
Phoenix-VL 1.5 Medium is a 123B-parameter natively multimodal model that reaches state-of-the-art results on Singapore multimodal, legal, and policy benchmarks after localized training on 1T+ tokens while staying competitive on global benchmarks.
-
A Scalable Recipe on SuperMUC-NG Phase 2: Efficient Large-Scale Training of Language Models
A combined parallelism recipe on SuperMUC-NG Phase 2 delivers 10% of theoretical peak throughput for 175B models plus 93% weak and 82% strong scaling efficiency on 128 nodes using unmodified public software.
-
Large Language Models: A Survey
The paper surveys key large language models, their training methods, datasets, evaluation benchmarks, and future research directions in the field.
-
A Survey of Large Language Models
This survey reviews the background, key techniques, and evaluation methods for large language models, emphasizing emergent abilities that appear at large scales.
-
A Comprehensive Overview of Large Language Models
A survey paper providing an overview of Large Language Models, their background, and recent advances in the field.