Posts

A Survey of Resource-efficient LLM and Multimodal Foundation Models

Extensive Reading Goal The goal of this survey is to provide an overarching understanding of how current approaches are tackling the resource challenges posed by large foundation models and to potentially inspire future breakthroughs in this field. ...

H2O Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models

Skimming Author Info Zhenyu “Allen” Zhang: A final-year Ph.D. student at the Electrical and Computer Engineering Department of UT Austin. Ying Sheng Insights Inherent Sparsity of Attention 推理过程中，其注意力矩阵表现出极高的稀疏性，超过95%的注意力值都非常小。这意味着在生成下一个 token 时，模型实际上只关注了过去所有词元中的一小部分。这为减少 KV Cache 的大小提供了可能性，因为大部分缓存的键值对实际上很少被用到 Existence of “Heavy Hitters” 通过分析词元在注意力计算中的累积得分，作者发现这些得分遵循 Power-law distribution, 这意味着只有一小部分词元 (Heavy Hitters) 贡献了绝大部分的注意力价值。这些 H₂ 词元对于维持模型的性能至关重要，如果将它们从缓存中移除，模型的准确率会急剧下降 Effectiveness of Local Statistics 理论上，要识别出真正的 Heavy Hitters 需要知道未来所有词元的注意力信息，这在自回归生成中是不现实的。论文通过实验发现，仅使用局部信息——即在每个解码步骤中，根据已经生成的词元来计算和累积注意力分数——来动态确定 H₂，其效果与使用全局信息几乎一样好。 Note 既然不是所有的历史信息都同等重要，那么就可以设计一种智能的缓存管理策略，只保留那些最关键的信息，从而在有限的显存中实现高效推理。 ...

LLM as a System Service on Mobile Devices

Intensive Reading Author Info ‪Wangsong Yin‬ - ‪Google Scholar‬ Mengwei Xu Background 论文首先提出了 LLMaaS: LLM as a system service on mobile devices (LLMaaS): The mobile OS exposes an LLM and its inference infrastructure as a system feature to mobile apps, akin to the location or notification services. LLMaaS 的提出主要有以下原因： LLMaaS needs only one copy of LLM weights in memory. 不同应用程序应该去调用由系统维护的同一个大模型，而不是自己单独去加载一个 A system-level LLM can be better customized for on-device accelerator and enjoy the performance gain over commodity hardware. 在系统层面去做大模型的管理和推理更接近底层，能够更好地利用底层的硬件资源这篇文章要解决的核心问题是 How to efficiently manage the LLM contexts ...

KV-Runahead Scalable Causal LLM Inference by Parallel Key-Value Cache Generation

Skimming Author Info Background Challenges Insights Approaches 看了好几遍都没看懂，我大概的理解是利用了 casual mask 的特性以链式的方式在不同设备之间传递 KV，避免了传统 TSP 的大量重复计算和冗余传输为了平衡整个流水线采用了 context-level load balancing，靠前的设备多算一些 KV, 靠后的设备少算一些，因为靠后的设备注意力计算会更长这里的关键点是：每个设备不仅传递 KV 缓存，也要利用收到的缓存，完成自己那部分词元的注意力计算。在 D1 上: 计算 T1-T4 的Q_0, K_0, V_0。立刻进行自己部分的注意力计算：用 Q_0 与 K_0 计算一个 4x4 的注意力矩阵，得到输出A_0。然后，它将 K_0, V_0（尺寸为 4 的缓存）发送给D2。在 D2 上: 在等待 D1 数据的同时，它可以并行计算 T5-T7 的本地Q_1, K_1, V_1。当它收到 D1 发来的 K_0, V_0 后，它将自己本地的 K_1, V_1 追加上去，形成一个包含 T1-T7 信息的、尺寸为 7 的 KV 缓存。立刻进行自己部分的注意力计算：用自己的 Q_1（来自 T5-T7）与这个尺寸为 7 的完整缓存进行计算（一个 3x7 的注意力计算），得到输出 A_1。然后，它将这个尺寸为 7 的 KV 缓存发送给 D3。在 D3 上: 并行计算 T8-T9 的本地Q_2, K_2, V_2。收到 D2 发来的尺寸为 7 的缓存后，追加自己的 K_2, V_2，形成包含全部 9 个词元信息的最终KV缓存。它进行自己部分的注意力计算：用 Q_2 与这个尺寸为 9 的完整缓存进行计算（一个 2x9 的注意力计算），得到输出 A_2。作为最后一个设备，它最终生成第一个令牌。 TSP ...

Ring Attention with Blockwise Transformers for Near-Infinite Context

Extensive Reading Author Info Hao Liu: A research scientist at Google DeepMind. Matei Zaharia: An associate professor at UC Berkeley (previously Stanford), where he works on computer systems and AI in the Sky Lab. Related Blogs Ring Attention Explained | Coconut Mode Background Transformer 的核心组件“自注意力机制”的内存消耗会随着输入序列长度的增加而呈二次方增长。这导致即便是最先进的 GPU/TPU，其有限的显存（通常小于 100GB）也无法处理超长序列，例如处理百万甚至千万级别的 token. 注意力模块的显存占用分析 $B$: Batch size ...

Striped Attention Faster Ring Attention for Causal Transformers

Skimming Author Info Implementation and Benchmark zhuzilin/ring-flash-attention: Ring attention implementation with flash attention Corresponding virtualization is here Background Challenges Insights Ring attention suffers from workload imbalance Due to the casual mask mechanism, some devices are doing meaningless computations in the iterations while other devices stays busy all the time. Stripped attention propose an another way to distribute workloads across devices to eliminate the imbalance. Approaches Striped Attention 让每个设备都持有了在原始序列中均匀分布的、不连续的词元 Example Important 理解这个例子最重要的一点：Ring Attention 和 Striped Attention 都不是采用朴素的注意力计算 ...

TPI-LLM Serving 70B-scale LLMs Efficiently on Low-resource Mobile Devices

Extensive Reading A similar paper is found: arxiv.org/pdf/2504.08791? Author Info ‪Zonghang Li‬ - ‪Google 学术搜索‬ Background LLM serving is shifting from the cloud to edge devices like smartphones and laptops. This trend is driven by growing privacy concerns, as users want to avoid sending their sensitive interaction data to cloud providers. The goal is to process user requests locally on their own devices. Preliminaries TP 场景下 KV Cache 的维护 Challenges Hardware Limitations: Mobile devices have very limited memory (typically 4-16 GiB) and computing power, often lacking GPUs. Running a 70B-scale model can require over 40 GiB of memory, which far exceeds the capacity of a single device. Inefficient Parallelism: The standard solution for distributed systems, pipeline parallelism, is inefficient for home scenarios where only one request is processed at a time. This leads to many devices being idle most of the time, wasting resources. Slow Memory Offloading: Existing on-device solutions like llama.cpp and Accelerate offload model data to disk to save RAM. However, their blocking disk I/O operations significantly slow down the inference speed. Insights 在低资源设备协同的环境下，应该选择 Tensor Parallelism 用户的请求一次性只有一条，并行的目的应该是降低延迟而不是增加吞吐量 Tensor Parallelism 依赖 allreduce 操作来同步和聚合计算结果，在低资源设备协同的环境下，通信瓶颈并非网络带宽而是链路延迟使用 star-based allreduce 来降低网络跳数进而降低延迟使用滑动窗口内存调度器来异步加载和卸载权重由一个独立的线程后台进行将权重的加载隐藏在计算和同步的过程中 Approaches ...

LLM.int8() 8-bit Matrix Multiplication for Transformers at Scale

Extensive Reading Author Info About Me — Tim Dettmers: A research scientist at the Allen Institute for Artificial Intelligence (Ai2) and an incoming Assistant Professor at Carnegie Mellon University (CMU). ‪Mike Lewis‬ - ‪Google Scholar‬ Related Blogs LLM.int8() and Emergent Features — Tim Dettmers A Gentle Introduction to 8-bit Matrix Multiplication for transformers at scale using transformers, accelerate and bitsandbytes Background 常见的 8 bit 量化有两种： Absmax quantization 在所有数据中，找到绝对值的最大值，我们称之为 abs_max，然后计算一个全局的 scaling factor 用数据中的每一个数字乘以这个 scaling factor, 再四舍五入到最近的整数完成量化 Zeropoint Quantization 找到数据中的最大值和最小值，计算 scaling factor 同时引入一个偏移量 zeropoint 来利用整个映射后的数值范围精度更高，但是开销更大 Challenges How to preserve high quantization precision at scales beyond 1B parameters? How to deal with the systematic outliers emerged in all transformer layers starting at scales of 6.7B parameters? Insights Regular quantization methods introduce larger quantization errors for outliers. The amount of outlier can be small, but contributes the majority to the LLM’s quality. Isolate the outlier feature dimensions into a 16-bit matrix multiplication while other values are multiplied in 8-bit. Approaches 主要包括两部分内容： ...