Pith. sign in

REVIEW 6 cited by

Scaling FP8 training to trillion-token LLMs

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2409.12517 v2 pith:NUGXK6Y4 submitted 2024-09-19 cs.LG cs.AI

classification cs.LGcs.AI
keywords trainingamplificationfirstfunctioninstabilitiesprecisionswiglutime
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
abstract

We train, for the first time, large language models using FP8 precision on datasets up to 2 trillion tokens -- a 20-fold increase over previous limits. Through these extended training runs, we uncover critical instabilities in FP8 training that were not observable in earlier works with shorter durations. We trace these instabilities to outlier amplification by the SwiGLU activation function. Interestingly, we show, both analytically and empirically, that this amplification happens only over prolonged training periods, and link it to a SwiGLU weight alignment process. To address this newly identified issue, we introduce Smooth-SwiGLU, a novel modification that ensures stable FP8 training without altering function behavior. We also demonstrate, for the first time, FP8 quantization of both Adam optimizer moments. Combining these innovations, we successfully train a 7B parameter model using FP8 precision on 256 Intel Gaudi2 accelerators, achieving on-par results with the BF16 baseline while delivering up to a $\sim 34 \%$ throughput improvement. A reference implementation is supplied in https://github.com/Anonymous1252022/Megatron-DeepSpeed.

Discussion (0). Sign in to comment.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. One QK Channel, Many Sources: Guarding Low-Precision Attention Collapse

    cs.LG 2026-08 conditional novelty 7.0 of 10

    Different low-precision errors converge on the same query-key spectral runaway, entry is gated by temporal sign-coherence, and a dormant query-key normalization guard contains it.

  2. Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention

    cs.LG 2026-07 conditional novelty 7.0 of 10

    Full-stack NVFP4 LLM pretraining is stable with a 1.47% loss gap to BF16 via LoRA-SVD projections, transformed AdamW/Root optimizers, and mixed-precision attention.

  3. Characterization and Mitigation of Training Instabilities in Microscaling Formats

    cs.LG 2025-06 conditional novelty 7.0 of 10

    Block-scaled MX low-precision training is unstable because quantization of tightly clustered layer-norm weights and some activations injects multiplicative gradient bias, and this can be fixed by keeping activations i...

  4. Thunder-LLM: Efficiently Adapting LLMs to Korean with Minimal Resources

    cs.CL 2025-06 conditional novelty 6.0 of 10

    A cost-effective recipe consisting of tokenizer extension, continual pretraining, FP8 training, and SFT/DPO post-training yields Korean-English bilingual 8B models with top Korean benchmark scores.

  5. A Comprehensive FP8 Training Recipe for Reasoning-Enhanced Language Models

    cs.CL 2025-09 conditional novelty 4.0 of 10

    A hybrid-granularity FP8 training recipe achieves near-lossless accuracy versus BF16 on reasoning benchmarks with up to 22% faster training and 14% lower memory.

  6. DeepSeek: Paradigm Shifts and Technical Evolution in Large AI Models

    cs.AI 2025-07 unverdicted

    A survey of DeepSeek's V3 and R1 models covering MLA, MoE, MTP, GRPO, and training engineering, with no new experimental results.

Pith tools