When Precision Meets Position: BFloat16 Breaks Down RoPE in Long-Context Training

• URL: http://arxiv.org/abs/2411.13476v2
• Date: Tue, 26 Nov 2024 09:46:25 GMT
• Title: When Precision Meets Position: BFloat16 Breaks Down RoPE in Long-Context Training
• Authors: Haonan Wang, Qian Liu, Chao Du, Tongyao Zhu, Cunxiao Du, Kenji Kawaguchi, Tianyu Pang,
• Abstract summary: Extending context window sizes allows large language models to process longer sequences and handle more complex tasks. We observe that using RoPE with BFloat16 format results in numerical issues, causing it to deviate from its intended relative positional encoding. We develop AnchorAttention, a plug-and-play attention method that alleviates numerical issues caused by BFloat16.
• Score: 51.23520027773028
• License:
• Abstract: Extending context window sizes allows large language models (LLMs) to process longer sequences and handle more complex tasks. Rotary Positional Embedding (RoPE) has become the de facto standard due to its relative positional encoding properties that benefit long-context training. However, we observe that using RoPE with BFloat16 format results in numerical issues, causing it to deviate from its intended relative positional encoding, especially in long-context scenarios. This issue arises from BFloat16's limited precision and accumulates as context length increases, with the first token contributing significantly to this problem. To address this, we develop AnchorAttention, a plug-and-play attention method that alleviates numerical issues caused by BFloat16, improves long-context capabilities, and speeds up training. AnchorAttention reduces unnecessary attention computations, maintains semantic coherence, and boosts computational efficiency by treating the first token as a shared anchor with a consistent position ID, making it visible to all documents within the training context. Experiments on three types of LLMs demonstrate that AnchorAttention significantly improves long-context performance and reduces training time by over 50\% compared to standard full attention mechanisms, while preserving the original LLM's capabilities on general tasks. Our code is available at https://github.com/haonan3/AnchorContext.

Related papers

• Squeezed Attention: Accelerating Long Context Length LLM Inference [64.11145320159126]
  We propose Squeezed Attention as a mechanism to accelerate LLM applications where a large portion of the input prompt is fixed. We use K-means clustering offline to group the keys for the fixed context based on semantic similarity and represent each cluster with a single centroid value. We then compute exact attention using only these important keys from the fixed context, thereby reducing bandwidth and computational costs.
  arXiv  Detail & Related papers  (2024-11-14T18:54:19Z)
• What is Wrong with Perplexity for Long-context Language Modeling? [71.34933096461124]
  Long-context inputs are crucial for large language models (LLMs) in tasks such as extended conversations, document summarization, and many-shot in-context learning. Perplexity (PPL) has proven unreliable for assessing long-context capabilities. We propose bfLongPPL, a novel metric that focuses on key tokens by employing a long-short context contrastive method to identify them.
  arXiv  Detail & Related papers  (2024-10-31T09:39:28Z)
• Why Does the Effective Context Length of LLMs Fall Short? [68.34573617977013]
  In this work, we introduce ShifTed Rotray position embeddING (STRING) STRING shifts well-trained positions to overwrite the original ineffective positions during inference, enhancing performance within their existing training lengths. Experimental results show that STRING dramatically improves the performance of the latest large-scale models.
  arXiv  Detail & Related papers  (2024-10-24T13:51:50Z)
• FocusLLM: Scaling LLM's Context by Parallel Decoding [16.642675785000176]
  FocusLLM is a framework designed to extend the context length of any decoder-only LLM. FocusLLM processes long text inputs by dividing them into chunks based on the model's original context length. It appends the local context to each chunk as a prompt to extract essential information from each chunk based on a novel parallel decoding mechanism.
  arXiv  Detail & Related papers  (2024-08-21T16:11:59Z)
• Long Context Alignment with Short Instructions and Synthesized Positions [56.1267385315404]
  This paper introduces Step-Skipping Alignment (SkipAlign) It is a new technique designed to enhance the long-context capabilities of Large Language Models (LLMs) With a careful selection of the base model and alignment datasets, SkipAlign with only 6B parameters achieves it's best performance and comparable with strong baselines like GPT-3.5-Turbo-16K on LongBench.
  arXiv  Detail & Related papers  (2024-05-07T01:56:22Z)
• Extending LLMs' Context Window with 100 Samples [42.52554295241792]
  Large Language Models (LLMs) are known to have limited extrapolation ability beyond their pre-trained context window. Recent studies have sought to extend the context window by modifying rotary position embedding (RoPE) We introduce a novel extension to RoPE which combines adjusting RoPE's base frequency and scaling the attention logits to help LLMs efficiently adapt to a larger context window.
  arXiv  Detail & Related papers  (2024-01-13T07:57:01Z)
• PoSE: Efficient Context Window Extension of LLMs via Positional Skip-wise Training [91.99700930388998]
  We propose Positional Skip-wisE training that simulates long inputs using a fixed context window. PoSE greatly reduces memory and time overhead compared with Full-length fine-tuning. We have successfully extended the LLaMA model to 128k tokens using a 2k training context window.
  arXiv  Detail & Related papers  (2023-09-19T08:03:38Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.