Only 40% of California high schools offer computer science classes. A new law aims to make it easier to certify computer science teachers.

A shortage of teachers for computer science classes puts California near the bottom of a national instruction ranking

But a lack of qualified teachers has stalled these efforts, and left California -- a global hub for the technological industry -- ranked near the bottom of states nationally in the percentage of high schools offering computer science classes.

A Shortage Of Teachers For Computer Science Classes Puts California Near The Bottom Of A National Instruction Ranking

We've tested more than 100 anti-malware apps to help you find the top malware protection and removal software for all your devices. When the IBM PC was new, I served as the president of the San ...

The Best Malware Removal and Protection Software for 2023

Personal computer (PC) makers are embracing sustainability as part of their growth strategies by using recyclable materials and packaging.Apart from design and packaging material, US PC manufacturer HP Inc wants to promote circularity to refurbish PCs,

PC makers go green with recyclable content

Coding knowledge can also help those in nontech roles communicate and work better with software developers, architects, and programmers. Revelo analyzed what computer languages are most popular in 2023 using data from Stack Overflow, a website where ...

The top 10 programming languages used worldwide, and why employers look for them on resumes

Hewlett Packard Enterprise (HPE) today announced that it built a new AI supercomputer, called “Austral”, based on HPE Cray

Hewlett Packard Enterprise Builds AI Supercomputer for CRIANN to Accelerate Scientific Research and Innovation

computer security, software security and more. This cohort-based program trains computational specialists who can construct models, develop algorithms, and write software using state-of-the-art graduate-level knowledge and techniques. The program has been ...

Master's in Professional Computer Science

If you’re on a computer and you get a text on your phone, wouldn’t it be great if you could reply without picking it up? Yes, it would, and there are great apps and services that enable just ...

How to send a text message from a computer

You can have multiple cursors on your computer. However, you will need to use specialized software or hardware configurations. You can use the MouseMax application on Windows to get more than one ...

Can you have 2 cursors on 1 computer?

House Republicans were on the verge of open revolt Wednesday after the ideologically fractious conference failed to coalesce around a speaker nominee, leaving the chamber rudderless and leaderless ...

Republicans fail to coalesce around speaker choice, leaving House in limbo

Coalesce

coalesce

Authors:
(1) Sara Seager, Departments of Earth, Atmospheric and Planetary Sciences, Physics, Aeronautics and Astronautics, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA;
(2) Janusz J. Petkowski, Department of Earth;
(3) Peter Gao, Department of Astronomy, University of California at Berkeley, California, USA;
(4) William Bains, Department of Earth;
(5) Noelle C. Bryan, Department of Earth;
(6) Sukrit Ranjan, Department of Earth;
(7) Jane Greaves, School of Physics and Astronomy, Cardiff University, Cardiff, United Kingdom and Institute of Astronomy, Cambridge University, Cambridge, United Kingdom. Table of Links Abstract and 1. Introduction and Overview Challenges and Assumptions for Life in the Venusian Clouds
A Proposed Cycle for Venusian Aerial Microbial Life
Discussion
Summary, Acknowledgments, Author Disclosure Statement, Funding Information, and References Abstract We revisit the hypothesis that there is life in the Venusian clouds to propose a life cycle that resolves the conundrum of how life can persist aloft for hundreds of millions to billions of years. Most discussions of an aerial biosphere in the Venus atmosphere temperate layers never address whether the life—small microbial-type particles—is free floating or confined to the liquid environment inside cloud droplets. We argue that life must reside inside liquid droplets such that it will be protected from a fatal net loss of liquid to the atmosphere, an unavoidable problem for any free-floating microbial life forms. However, the droplet habitat poses a lifetime limitation: Droplets inexorably grow (over a few months) to large enough sizes that are forced by gravity to settle downward to hotter, uninhabitable layers of the Venusian atmosphere. (Droplet fragmentation—which would reduce particle size—does not occur in Venusian atmosphere conditions.) We propose for the first time that the only way life can survive indefinitely is with a life cycle that involves microbial life drying out as liquid droplets evaporate during settling, with the small desiccated ‘‘spores’’ halting at, and partially populating, the Venus atmosphere stagnant lower haze layer (33–48 km altitude). We, thus, call the Venusian lower haze layer a ‘‘depot’’ for desiccated microbial life. The spores eventually return to the cloud layer by upward diffusion caused by mixing induced by gravity waves, act as cloud condensation nuclei, and rehydrate for a continued life cycle. We also review the challenges for life in the extremely harsh conditions of the Venusian atmosphere, refuting the notion that the ‘‘habitable’’ cloud layer has an analogy in any terrestrial environment. Key Words: Venus— Clouds—Life—Habitability—Sulfuric Acid—Life Cycle—Aerial Biosphere. Astrobiology 21, xxx–xxx. 1. Introduction and Overview Life on Venus has been a topic of speculation for more than half a century, with published papers ranging from science-fiction-like to invalid conjecture to legitimate hypothesis (Morowitz and Sagan, 1967; Grinspoon, 1997; Cockell, 1999; Schulze-Makuch and Irwin, 2002, 2006; Schulze-Makuch et al., 2004; Grinspoon and Bullock, 2007; Limaye et al., 2018). Today, only Venus’ atmospheric cloud layers (a large region spanning from 48 to 60 km altitude) have seemingly habitable conditions—the surface (at 735 K) is too hot for any plausible solvent and for most organic covalent chemistry. How the clouds could become inhabited is not known. In principle, life could arise in the clouds independent from the ground (Woese, 1979; Dobson et al., 2000) with material from meteoritic input (Sleep, 2018a, 2018b) from the asteroid belt (including Ceres, and even from Mars). Life may even have been directly seeded by impacts from Earth ejecta (Melosh, 1988; Reyes-Ruiz et al., 2012; Beech et al., 2018). A more commonly agreed on, and perhaps more conceivable scenario, is that life originated on the surface, as it most likely did on Earth, and migrated into the clouds. Recent modeling by Way et al. (2016) suggests the existence of habitable surface and oceans as late as *700 Mya. Consideration of life on Venus is extensively summarized in a recent paper (Limaye et al., 2018). Almost all previous work on life in Venusian clouds does not specify what exactly ‘‘life in the clouds’’ means (for one exception see Schulze-Makuch et al., 2004). Does microscopic life reside inside cloud liquid droplets? Or is microscopic life free floating in the air between cloud droplets? Earth has an aerial biosphere created by microbial life that regularly migrates to clouds from the ground (Vaı¨tilingom et al., 2012; Amato et al., 2017). Studies of Earth’s aerial biosphere shows that microbes mostly reside inside cloud droplets but some are free-floating in the atmosphere (Section 4.1). Some microbes (both inside and outside droplets) are found to be metabolically active, even though there is no evidence, as of yet, of cell division (Amato et al., 2019). Microbial life cannot reside in Earth’s atmosphere indefinitely, mostly because of lack of continuous cloud cover, and after days to weeks microbes are deposited back down on Earth’s surface (Burrows et al., 2009; Bryan et al., 2019). Earth’s aerial biosphere is intimately connected to the habitable surface of the planet. We argue that life, if it exists in Venus’ atmosphere, must reside inside cloud liquid droplets for the majority of its life cycle (Section 2.1). Life engulfed by cloud droplets will be protected from a fatal net loss of liquid to the atmosphere, an unavoidable problem for any free-floating microbial lifeforms. But a droplet habitat implies a lifetime limitation. As liquid droplets coalesce and grow, they eventually reach a size that, due to gravity, settles out of the temperate layers of the atmosphere at an appreciable rate. Over time, the population of inhabited droplets should therefore decline to zero. (Droplet fragmentation—which would reduce particle size—does not occur in Venusian atmosphere conditions [Section 3.4].) The conundrum is that there is no way for liquid droplets—and hence life inside of them—to persist indefinitely in Venus’ temperate atmosphere layers. In this article, we describe this problem, and a solution to it. We hypothesize how life escapes being ‘‘rained out’’ down to inhospitably hot atmosphere layers by cycling between small, desiccated spores and larger, metabolically active, droplet-inhabiting cells. Venusian life escapes settling to the surface by forming a resistant, spore-like form that survives the evaporation of the inevitable downward droplet flow to atmosphere layers of high temperature. The desiccated spores become suspended in the Venus atmosphere lower haze layer, which we thus call a ‘‘depot’’ for desiccated microbial life. Because the dynamics of the relatively stagnant lower haze layer are not well known, the main uncertainty in our life cycle hypothesis is how the spores are transported back up into the clouds again. The spores most likely travel upward by vertical mixing induced by gravity waves, and once in the cloud layer they form the nucleus of a new droplet. The depot is ‘‘leaky,’’ that is spores will also vertically mix downward to atmosphere layers with fatally high temperatures. Our proposed life cycle includes cell division that occurs in the larger droplets, and sporulation for individual cells, enabling cell numbers to be replenished against loss. In this work, we use the terms ‘‘microbial life’’ or ‘‘microbes’’ for microscopic life, without intending to imply that hypothetical Venusian microbes might in any way be taxonomically related to microbial life on Earth. We use the term ‘‘spore’’ to denote a cell in a dormant state of longterm metabolic inactivity, which is further resistant to (and protected from) environmental stresses. We begin with a highlighted review on the very harsh and inhospitable conditions in the Venusian atmosphere and related, required assumptions for life to exist (Section 2). We next present our Venusian life cycle hypothesis (Section 3), which optimistically assumes that the challenges described in Section 2 can be met. We put the hypothesis in the context with Earth’s aerial biosphere and other characteristics of the Venusian hypothesized aerial biosphere in Section 4. We conclude with a summary in Section 5. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Sara Seager, Departments of Earth, Atmospheric and Planetary Sciences, Physics, Aeronautics and Astronautics, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA; (2) Janusz J. Petkowski, Department of Earth; (3) Peter Gao, Department of Astronomy, University of California at Berkeley, California, USA; (4) William Bains, Department of Earth; (5) Noelle C. Bryan, Department of Earth; (6) Sukrit Ranjan, Department of Earth; (7) Jane Greaves, School of Physics and Astronomy, Cardiff University, Cardiff, United Kingdom and Institute of Astronomy, Cambridge University, Cambridge, United Kingdom. Authors: Authors: (1) Sara Seager, Departments of Earth, Atmospheric and Planetary Sciences, Physics, Aeronautics and Astronautics, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA; (2) Janusz J. Petkowski, Department of Earth; (3) Peter Gao, Department of Astronomy, University of California at Berkeley, California, USA; (4) William Bains, Department of Earth; (5) Noelle C. Bryan, Department of Earth; (6) Sukrit Ranjan, Department of Earth; (7) Jane Greaves, School of Physics and Astronomy, Cardiff University, Cardiff, United Kingdom and Institute of Astronomy, Cambridge University, Cambridge, United Kingdom. Table of Links Abstract and 1. Introduction and Overview Abstract and 1. Introduction and Overview Challenges and Assumptions for Life in the Venusian Clouds A Proposed Cycle for Venusian Aerial Microbial Life Discussion Summary, Acknowledgments, Author Disclosure Statement, Funding Information, and References Challenges and Assumptions for Life in the Venusian Clouds Challenges and Assumptions for Life in the Venusian Clouds A Proposed Cycle for Venusian Aerial Microbial Life A Proposed Cycle for Venusian Aerial Microbial Life Discussion Discussion Summary, Acknowledgments, Author Disclosure Statement, Funding Information, and References Summary, Acknowledgments, Author Disclosure Statement, Funding Information, and References Abstract We revisit the hypothesis that there is life in the Venusian clouds to propose a life cycle that resolves the conundrum of how life can persist aloft for hundreds of millions to billions of years. Most discussions of an aerial biosphere in the Venus atmosphere temperate layers never address whether the life—small microbial-type particles—is free floating or confined to the liquid environment inside cloud droplets. We argue that life must reside inside liquid droplets such that it will be protected from a fatal net loss of liquid to the atmosphere, an unavoidable problem for any free-floating microbial life forms. However, the droplet habitat poses a lifetime limitation: Droplets inexorably grow (over a few months) to large enough sizes that are forced by gravity to settle downward to hotter, uninhabitable layers of the Venusian atmosphere. (Droplet fragmentation—which would reduce particle size—does not occur in Venusian atmosphere conditions.) We propose for the first time that the only way life can survive indefinitely is with a life cycle that involves microbial life drying out as liquid droplets evaporate during settling, with the small desiccated ‘‘spores’’ halting at, and partially populating, the Venus atmosphere stagnant lower haze layer (33–48 km altitude). We, thus, call the Venusian lower haze layer a ‘‘depot’’ for desiccated microbial life. The spores eventually return to the cloud layer by upward diffusion caused by mixing induced by gravity waves, act as cloud condensation nuclei, and rehydrate for a continued life cycle. We also review the challenges for life in the extremely harsh conditions of the Venusian atmosphere, refuting the notion that the ‘‘habitable’’ cloud layer has an analogy in any terrestrial environment. Key Words: Venus— Clouds—Life—Habitability—Sulfuric Acid—Life Cycle—Aerial Biosphere. Astrobiology 21, xxx–xxx. 1. Introduction and Overview Life on Venus has been a topic of speculation for more than half a century, with published papers ranging from science-fiction-like to invalid conjecture to legitimate hypothesis (Morowitz and Sagan, 1967; Grinspoon, 1997; Cockell, 1999; Schulze-Makuch and Irwin, 2002, 2006; Schulze-Makuch et al., 2004; Grinspoon and Bullock, 2007; Limaye et al., 2018). Today, only Venus’ atmospheric cloud layers (a large region spanning from 48 to 60 km altitude) have seemingly habitable conditions—the surface (at 735 K) is too hot for any plausible solvent and for most organic covalent chemistry. How the clouds could become inhabited is not known. In principle, life could arise in the clouds independent from the ground (Woese, 1979; Dobson et al., 2000) with material from meteoritic input (Sleep, 2018a, 2018b) from the asteroid belt (including Ceres, and even from Mars). Life may even have been directly seeded by impacts from Earth ejecta (Melosh, 1988; Reyes-Ruiz et al., 2012; Beech et al., 2018). A more commonly agreed on, and perhaps more conceivable scenario, is that life originated on the surface, as it most likely did on Earth, and migrated into the clouds. Recent modeling by Way et al. (2016) suggests the existence of habitable surface and oceans as late as *700 Mya. Consideration of life on Venus is extensively summarized in a recent paper (Limaye et al., 2018). Almost all previous work on life in Venusian clouds does not specify what exactly ‘‘life in the clouds’’ means (for one exception see Schulze-Makuch et al., 2004). Does microscopic life reside inside cloud liquid droplets? Or is microscopic life free floating in the air between cloud droplets? Earth has an aerial biosphere created by microbial life that regularly migrates to clouds from the ground (Vaı¨tilingom et al., 2012; Amato et al., 2017). Studies of Earth’s aerial biosphere shows that microbes mostly reside inside cloud droplets but some are free-floating in the atmosphere (Section 4.1). Some microbes (both inside and outside droplets) are found to be metabolically active, even though there is no evidence, as of yet, of cell division (Amato et al., 2019). Microbial life cannot reside in Earth’s atmosphere indefinitely, mostly because of lack of continuous cloud cover, and after days to weeks microbes are deposited back down on Earth’s surface (Burrows et al., 2009; Bryan et al., 2019). Earth’s aerial biosphere is intimately connected to the habitable surface of the planet. We argue that life, if it exists in Venus’ atmosphere, must reside inside cloud liquid droplets for the majority of its life cycle (Section 2.1). Life engulfed by cloud droplets will be protected from a fatal net loss of liquid to the atmosphere, an unavoidable problem for any free-floating microbial lifeforms. But a droplet habitat implies a lifetime limitation. As liquid droplets coalesce and grow, they eventually reach a size that, due to gravity, settles out of the temperate layers of the atmosphere at an appreciable rate. Over time, the population of inhabited droplets should therefore decline to zero. (Droplet fragmentation—which would reduce particle size—does not occur in Venusian atmosphere conditions [Section 3.4].) The conundrum is that there is no way for liquid droplets—and hence life inside of them—to persist indefinitely in Venus’ temperate atmosphere layers. In this article, we describe this problem, and a solution to it. We hypothesize how life escapes being ‘‘rained out’’ down to inhospitably hot atmosphere layers by cycling between small, desiccated spores and larger, metabolically active, droplet-inhabiting cells. Venusian life escapes settling to the surface by forming a resistant, spore-like form that survives the evaporation of the inevitable downward droplet flow to atmosphere layers of high temperature. The desiccated spores become suspended in the Venus atmosphere lower haze layer, which we thus call a ‘‘depot’’ for desiccated microbial life. Because the dynamics of the relatively stagnant lower haze layer are not well known, the main uncertainty in our life cycle hypothesis is how the spores are transported back up into the clouds again. The spores most likely travel upward by vertical mixing induced by gravity waves, and once in the cloud layer they form the nucleus of a new droplet. The depot is ‘‘leaky,’’ that is spores will also vertically mix downward to atmosphere layers with fatally high temperatures. Our proposed life cycle includes cell division that occurs in the larger droplets, and sporulation for individual cells, enabling cell numbers to be replenished against loss. In this work, we use the terms ‘‘microbial life’’ or ‘‘microbes’’ for microscopic life, without intending to imply that hypothetical Venusian microbes might in any way be taxonomically related to microbial life on Earth. We use the term ‘‘spore’’ to denote a cell in a dormant state of longterm metabolic inactivity, which is further resistant to (and protected from) environmental stresses. We begin with a highlighted review on the very harsh and inhospitable conditions in the Venusian atmosphere and related, required assumptions for life to exist (Section 2). We next present our Venusian life cycle hypothesis (Section 3), which optimistically assumes that the challenges described in Section 2 can be met. We put the hypothesis in the context with Earth’s aerial biosphere and other characteristics of the Venusian hypothesized aerial biosphere in Section 4. We conclude with a summary in Section 5. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Life in Venusian Clouds: A Hypothesis for Microbial Survival and Life Cycle

Table of Links Abstract and 1 Introduction 2 Background and 2.1 Transformer-Based Large Language Models 2.2 LLM Service & Autoregressive Generation 2.3 Batching Techniques for LLMs 3 Memory Challenges in LLM Serving 3.1 Memory Management in Existing Systems 4 Method and 4.1 PagedAttention 4.2 KV Cache Manager 4.3 Decoding with PagedAttention and vLLM 4.4 Application to Other Decoding Scenarios 4.5 Scheduling and Preemption 4.6 Distributed Execution 5 Implementation 6 Evaluation and 6.1 Experimental Setup 6.2 Basic Sampling 6.3 Parallel Sampling and Beam Search 6.4 Shared prefix 6.5 Chatbot 7 Ablation Studies 8 Discussion 9 Related Work 10 Conclusion, Acknowledgement and References 5 Implementation vLLM is an end-to-end serving system with a FastAPI [15] frontend and a GPU-based inference engine. The frontend extends the OpenAI API [34] interface, allowing users to customize sampling parameters for each request, such as the maximum sequence length and the beam width 𝑘. The vLLM engine is written in 8.5K lines of Python and 2K lines of C++/CUDA code. We develop control-related components including the scheduler and the block manager in Python while developing custom CUDA kernels for key operations such as PagedAttention. For the model executor, we implement popular LLMs such as GPT [5], OPT [62], and LLaMA [52] using PyTorch [39] and Transformers [58]. We use NCCL [32] for tensor communication across the distributed GPU workers. 5.1 Kernel-level Optimization Since PagedAttention introduces memory access patterns that are not efficiently supported by existing systems, we develop several GPU kernels for optimizing it. (1) Fused reshape and block write. In every Transformer layer, the new KV cache are split into blocks, reshaped to a memory layout optimized for block read, then saved at positions specified by the block table. To minimize kernel launch overheads, we fuse them into a single kernel. (2) Fusing block read and attention. We adapt the attention kernel in FasterTransformer [31] to read KV cache according to the block table and perform attention operations on the fly. To ensure coalesced memory access, we assign a GPU warp to read each block. Moreover, we add support for variable sequence lengths within a request batch. (3) Fused block copy. Block copy operations, issued by the copy-on-write mechanism, may operate on discontinuous blocks. This can lead to numerous invocations of small data movements if we use the cudaMemcpyAsync API. To mitigate the overhead, we implement a kernel that batches the copy operations for different blocks into a single kernel launch. 5.2 Supporting Various Decoding Algorithms vLLM implements various decoding algorithms using three key methods: fork, append, and free. The fork method creates a new sequence from an existing one. The append method appends a new token to the sequence. Finally, the free method deletes the sequence. For instance, in parallel sampling, vLLM creates multiple output sequences from the single input sequence using the fork method. It then adds new tokens to these sequences in every iteration with append, and deletes sequences that meet a stopping condition using free. The same strategy is also applied in beam search and prefix sharing by vLLM. We believe future decoding algorithms can also be supported by combining these methods. This paper is available on arxiv under CC BY 4.0 DEED license. Authors:
(1) Woosuk Kwon, UC Berkeley with Equal contribution;
(2) Zhuohan Li, UC Berkeley with Equal contribution;
(3) Siyuan Zhuang, UC Berkeley;
(4) Ying Sheng, UC Berkeley and Stanford University;
(5) Lianmin Zheng, UC Berkeley;
(6) Cody Hao Yu, Independent Researcher;
(7) Cody Hao Yu, Independent Researcher;
(8) Joseph E. Gonzalez, UC Berkeley;
(9) Hao Zhang, UC San Diego;
(10) Ion Stoica, UC Berkeley. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Background and 2.1 Transformer-Based Large Language Models 2 Background and 2.1 Transformer-Based Large Language Models 2.2 LLM Service & Autoregressive Generation 2.2 LLM Service & Autoregressive Generation 2.3 Batching Techniques for LLMs 2.3 Batching Techniques for LLMs 3 Memory Challenges in LLM Serving 3 Memory Challenges in LLM Serving 3.1 Memory Management in Existing Systems 3.1 Memory Management in Existing Systems 4 Method and 4.1 PagedAttention 4 Method and 4.1 PagedAttention 4.2 KV Cache Manager 4.2 KV Cache Manager 4.3 Decoding with PagedAttention and vLLM 4.3 Decoding with PagedAttention and vLLM 4.4 Application to Other Decoding Scenarios 4.4 Application to Other Decoding Scenarios 4.5 Scheduling and Preemption 4.5 Scheduling and Preemption 4.6 Distributed Execution 4.6 Distributed Execution 5 Implementation 5 Implementation 6 Evaluation and 6.1 Experimental Setup 6 Evaluation and 6.1 Experimental Setup 6.2 Basic Sampling 6.2 Basic Sampling 6.3 Parallel Sampling and Beam Search 6.3 Parallel Sampling and Beam Search 6.4 Shared prefix 6.4 Shared prefix 6.5 Chatbot 6.5 Chatbot 7 Ablation Studies 7 Ablation Studies 8 Discussion 8 Discussion 9 Related Work 9 Related Work 10 Conclusion, Acknowledgement and References 10 Conclusion, Acknowledgement and References 5 Implementation vLLM is an end-to-end serving system with a FastAPI [15] frontend and a GPU-based inference engine. The frontend extends the OpenAI API [34] interface, allowing users to customize sampling parameters for each request, such as the maximum sequence length and the beam width 𝑘. The vLLM engine is written in 8.5K lines of Python and 2K lines of C++/CUDA code. We develop control-related components including the scheduler and the block manager in Python while developing custom CUDA kernels for key operations such as PagedAttention. For the model executor, we implement popular LLMs such as GPT [5], OPT [62], and LLaMA [52] using PyTorch [39] and Transformers [58]. We use NCCL [32] for tensor communication across the distributed GPU workers. 5.1 Kernel-level Optimization Since PagedAttention introduces memory access patterns that are not efficiently supported by existing systems, we develop several GPU kernels for optimizing it. (1) Fused reshape and block write. In every Transformer layer, the new KV cache are split into blocks, reshaped to a memory layout optimized for block read, then saved at positions specified by the block table. To minimize kernel launch overheads, we fuse them into a single kernel. (2) Fusing block read and attention. We adapt the attention kernel in FasterTransformer [31] to read KV cache according to the block table and perform attention operations on the fly. To ensure coalesced memory access, we assign a GPU warp to read each block. Moreover, we add support for variable sequence lengths within a request batch. (3) Fused block copy. Block copy operations, issued by the copy-on-write mechanism, may operate on discontinuous blocks. This can lead to numerous invocations of small data movements if we use the cudaMemcpyAsync API. To mitigate the overhead, we implement a kernel that batches the copy operations for different blocks into a single kernel launch. 5.2 Supporting Various Decoding Algorithms vLLM implements various decoding algorithms using three key methods: fork, append, and free. The fork method creates a new sequence from an existing one. The append method appends a new token to the sequence. Finally, the free method deletes the sequence. For instance, in parallel sampling, vLLM creates multiple output sequences from the single input sequence using the fork method. It then adds new tokens to these sequences in every iteration with append, and deletes sequences that meet a stopping condition using free. The same strategy is also applied in beam search and prefix sharing by vLLM. We believe future decoding algorithms can also be supported by combining these methods. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv Authors: (1) Woosuk Kwon, UC Berkeley with Equal contribution; (2) Zhuohan Li, UC Berkeley with Equal contribution; (3) Siyuan Zhuang, UC Berkeley; (4) Ying Sheng, UC Berkeley and Stanford University; (5) Lianmin Zheng, UC Berkeley; (6) Cody Hao Yu, Independent Researcher; (7) Cody Hao Yu, Independent Researcher; (8) Joseph E. Gonzalez, UC Berkeley; (9) Hao Zhang, UC San Diego; (10) Ion Stoica, UC Berkeley. Authors: Authors: (1) Woosuk Kwon, UC Berkeley with Equal contribution; (2) Zhuohan Li, UC Berkeley with Equal contribution; (3) Siyuan Zhuang, UC Berkeley; (4) Ying Sheng, UC Berkeley and Stanford University; (5) Lianmin Zheng, UC Berkeley; (6) Cody Hao Yu, Independent Researcher; (7) Cody Hao Yu, Independent Researcher; (8) Joseph E. Gonzalez, UC Berkeley; (9) Hao Zhang, UC San Diego; (10) Ion Stoica, UC Berkeley.

How vLLM Implements Decoding Algorithms

Wildcat: Undercollateralized Credit Expansion for Fun and Profit 2.0

Wildcat: Crédito Subcolateralizado Jach’anchayaña Kusisiñataki ukhamaraki Ganancia 2.0

Wildcat: Əyləncə və Mənfəət üçün Təminatsız Kreditin Genişlənməsi 2.0

Wildcat: пашырэнне крэдыту з недастатковым забеспячэннем для забавы і прыбытку 2.0

ওয়াইল্ডক্যাট: মজা এবং লাভের জন্য আন্ডারকোলেটরালাইজড ক্রেডিট এক্সপানশন 2.0

Wildcat: Expansión del crédito sin garantías suficientes para divertirse y obtener ganancias 2.0

Wildcat: expansión de crédito subcolateralizada para diversión e beneficio 2.0

ワイルドキャット: 楽しみと利益のための担保不足の信用拡大 2.0

„Wildcat“: nepakankamai užtikrinto kredito išplėtimas pramogoms ir pelnui 2.0

Wildcat: Fanitarana trosa tsy voahaja ho an'ny fahafinaretana sy tombony 2.0

Wildcat: Upanuzi wa Mikopo Isiyo na Dhamana kwa Furaha na Faida 2.0

வைல்ட்கேட்: கேளிக்கை மற்றும் லாபத்திற்காக அண்டர்கலாடரலைஸ்டு கிரெடிட் விரிவாக்கம் 2.0

ዋይልድካት፡ ንመዘናግዕን መኽሰብን ዝኸውን ትሑት ውሕስነት ዘለዎ ምስፍሕፋሕ ልቓሕ 2.0

How to Accurately Measure Binomial Proportions for Reliable Conversion Metrics

Metamorphosis Commenced, vol2: Crecimiento Aleph Review & Q4 in Lisbon and Thailand

Metamorphosis Commended, vol2: Crecimiento Aleph Review & Q4 v Lisabonu a Thajsku

Metamorphosis Commenced, τόμος 2: Crecimiento Aleph Review & Q4 στη Λισαβόνα και την Ταϊλάνδη

Metamorfosis iniciada, vol. 2: Revisión y cuarto trimestre de Crecimiento Aleph en Lisboa y Tailandia

Metamorphosis Commenced, vol2: Crecimiento Aleph Review & Q4 en Lisboa e Tailandia

変身開始、第2巻：クレシミエント・アレフのレビューとリスボンとタイでの第4四半期

Метаморфоз басталды, 2-том: Crecimiento Aleph шолуы және Лиссабондағы және Таиландтағы 4-тоқсан

Metamorphosis ເລີ່ມຕົ້ນ, vol2: Crecimiento Aleph Review & Q4 ໃນ Lisbon ແລະປະເທດໄທ

میټامورفوسس پیل شوی، vol2: Crecimiento Aleph Review & Q4 په لیزبون او تایلینډ کې

Metamorfosis qallarisqa, vol2: Crecimiento Aleph Revisión & Q4 Lisboapi, Tailandiapipas

Metamorphosis Yakatangwa, vol2: Crecimiento Aleph Ongororo & Q4 muLisbon neThailand

Metamorfoza filloi, vëll 2: Crecimiento Aleph Review & Q4 në Lisbonë dhe Tajlandë

Metamorphosis bắt đầu, tập 2: Đánh giá Crecimiento Aleph & Q4 tại Lisbon và Thái Lan

Decentralized Micro-Governance Model for the Kingdom of Tonga Based on Proof-of-Work Consensus

Introduction The Data space is almost as volatile as the AI space this year, with many players consolidating. In the summer of 2022, I wrote about The Data Table Format Wars, and the ecosystem has only gotten more interesting since then. I won’t get deep into the weeds on Data Lakes vs. Data Warehouse, etc., but suffice it to say, we are seeing everything coalesce around decoupling data storage from computing. Table formats are a way to standardize lightly opinionated storage. As I opined a couple of years ago, Apache Iceberg seems to be where they are settling. These formats tend to use Apache Parquet on object storage and provide a method for tracking new data and data state changes. In 2023, Databricks announced the open-sourcing of their Unity Catalog, and Onehouse announced OneTable (now X-Table). These projects are meant to give you a single interface to the three popular table formats: Delta, Hudi, and Iceberg. Then, June 2024 happened… What Happened? Snowflake was having its annual Snowflake Summit conference, and on June 3, 2024, they announced Apache Polaris (incubating), a vendor-neutral, open catalog for Apache Iceberg. The next day, and a week before the Databricks conference, Databricks announced their acquisition of Tabular, a company founded by the creators of Iceberg. This still seems weird to everyone from the outside because Iceberg is a spec; it isn’t a product, and Tabular isn’t even the main contributor to Iceberg. In my opinion, Databricks will somehow roll their Deltalake into Iceberg. Behind the scenes, Snowflake had been trying to acquire Tabular. Something happened at the 11th hour, and Databricks came in and laid down a ton more money, and the Snowflake bid was discarded. Snowflake was already two years into their commitment to supporting Iceberg as external storage, which made a lot of sense for them regardless. Snowflake has a reputation concerning the release speed after an announcement, so there was some skepticism, and some even wondered if any code had been written. It turns out Snowflake had been working on this for a while, and it was made available in very short order, with vendors like Dremio, Starburst, CelerData, and Upsolver showing solutions using it in less than two months: a delightful surprise, and a positive sign of things to come. All of that is a long way to set up Apache Polaris, and hopefully, it will give you a sense of the importance that a good catalog plays in all of this. How will any of your compute tools work if you don’t have a unified system to describe all those parquet files in your data lake? Polaris Overview Apache Polaris is an open-source catalog implementation for Apache Iceberg tables. Its main features include: REST API implementation: Polaris implements Iceberg's REST API, enabling interoperability across multiple query engines.
Multi-engine compatibility: It allows various query engines, such as Apache Spark, Apache Flink, Trino, and others, to read and write Iceberg tables through a single catalog.
Centralized management: Polaris provides centralized and secure access to Iceberg tables across REST-compatible query engines.
Namespace and metadata management: It supports creating namespaces for logical grouping and manages metadata for Iceberg tables.
Storage configurations: Polaris offers robust storage configurations for cloud providers like S3, Azure, and Google Cloud Storage.
Security features: It implements role-based access control (RBAC) and uses credential vending to secure query execution.
Open-source nature: Polaris is fully open-source, allowing for community contributions and vendor neutrality.
Interoperability: It enables multiple engines to be used on a single copy of data, reducing storage duplication and ETL costs.
Consistent access control: Polaris allows you to manage role-based access control and storage layer security consistently across engines from one place.
Flexibility in hosting: It can be hosted on various infrastructures, providing flexibility and reducing vendor lock-in. And a lovely little graphic from the docs to put a picture on everything. Borrowing from the Polaris docs for a picture once again, let’s dive into what is happening here in this entity hierarchy graphic: We can have multiple catalogs in a single Polaris install; in this diagram, it is only Catalog1 that has any namespaces or tables defined. A single catalog can be either internal or external. The internal tables allow for read/write functionality, meaning you have full control. The external tables are currently read-only and are intended to mirror data that is in some external catalog, such as a Glue catalog, which Polaris will see as the source of truth once connected. The catalog organizes your iceberg tables; you’ll want to configure it with your storage configuration for the applicable cloud system that you are on. This forms the first architectural layer in the Iceberg table spec. A namespace can have an arbitrary level of nesting and is used to logically group Iceberg tables within a catalog. An Iceberg table will always belong to a namespace. With this configuration in place, you can then use the Polaris catalog to describe your Iceberg tables in such a way as to make them easily accessible and manageable using a large variety of tools. That’s a quick overview of what is happening and what it looks like. The implementation inside of Snowflake provides a really handy UI to manage the implementation; other vendors have their own, and you can just do it through Spark if you like. Summary I continue to believe that Apache Iceberg is going to “win” the table format wars. Apache Polaris focusing on Iceberg is not a bad thing. I want to dig into how RBAC has been implemented in Polaris another time and see if it is as robust as what the Tabular product implemented, which I was a big fan of. I believe Polaris is a big win for the Iceberg community and will help accelerate adoption even further. I’ve seen talk about the Dremio Nessie project getting rolled in, and I’m curious to see what that will look like. I haven’t looked at Nessie in a couple of years, and I thought their git-like semantics mainly had been covered by the Iceberg tagging and branching, which I wrote about last year—yet another story for another day. I’m a big fan of what is happening here. If you are in this space, this is the right train to jump on. Check out my other What the Heck is… articles at the links below: What The Heck Is DuckDB?
What the Heck Is Malloy?
What the Heck is PRQL?
What the Heck is GlareDB?
What the Heck is SeaTunnel?
What the Heck is LanceDB?
What the heck is SDF?
What the Heck is Paimon?
What the Heck is Proton?
What the Heck is PuppyGraph?
What the Heck is GPTScript?
What the Heck is WarpStream? Introduction Introduction The Data space is almost as volatile as the AI space this year, with many players consolidating. In the summer of 2022, I wrote about The Data Table Format Wars , and the ecosystem has only gotten more interesting since then. I won’t get deep into the weeds on Data Lakes vs. Data Warehouse, etc., but suffice it to say, we are seeing everything coalesce around decoupling data storage from computing. The Data Table Format Wars Table formats are a way to standardize lightly opinionated storage. As I opined a couple of years ago, Apache Iceberg seems to be where they are settling. These formats tend to use Apache Parquet on object storage and provide a method for tracking new data and data state changes. In 2023, Databricks announced the open-sourcing of their Unity Catalog, and Onehouse announced OneTable (now X-Table). These projects are meant to give you a single interface to the three popular table formats: Delta, Hudi, and Iceberg. Then, June 2024 happened… announced announced What Happened? Snowflake was having its annual Snowflake Summit conference, and on June 3, 2024, they announced Apache Polaris (incubating), a vendor-neutral, open catalog for Apache Iceberg. The next day, and a week before the Databricks conference, Databricks announced their acquisition of Tabular, a company founded by the creators of Iceberg. This still seems weird to everyone from the outside because Iceberg is a spec; it isn’t a product, and Tabular isn’t even the main contributor to Iceberg. In my opinion, Databricks will somehow roll their Deltalake into Iceberg. announced announced Behind the scenes, Snowflake had been trying to acquire Tabular. Something happened at the 11th hour, and Databricks came in and laid down a ton more money, and the Snowflake bid was discarded. Snowflake was already two years into their commitment to supporting Iceberg as external storage, which made a lot of sense for them regardless. Snowflake has a reputation concerning the release speed after an announcement, so there was some skepticism, and some even wondered if any code had been written. It turns out Snowflake had been working on this for a while, and it was made available in very short order, with vendors like Dremio, Starburst, CelerData, and Upsolver showing solutions using it in less than two months: a delightful surprise, and a positive sign of things to come. All of that is a long way to set up Apache Polaris , and hopefully, it will give you a sense of the importance that a good catalog plays in all of this. How will any of your compute tools work if you don’t have a unified system to describe all those parquet files in your data lake? Apache Polaris Polaris Overview Apache Polaris is an open-source catalog implementation for Apache Iceberg tables. Its main features include: REST API implementation: Polaris implements Iceberg's REST API, enabling interoperability across multiple query engines. Multi-engine compatibility: It allows various query engines, such as Apache Spark, Apache Flink, Trino, and others, to read and write Iceberg tables through a single catalog. Centralized management: Polaris provides centralized and secure access to Iceberg tables across REST-compatible query engines. Namespace and metadata management: It supports creating namespaces for logical grouping and manages metadata for Iceberg tables. Storage configurations: Polaris offers robust storage configurations for cloud providers like S3, Azure, and Google Cloud Storage. Security features: It implements role-based access control (RBAC) and uses credential vending to secure query execution. Open-source nature: Polaris is fully open-source, allowing for community contributions and vendor neutrality. Interoperability: It enables multiple engines to be used on a single copy of data, reducing storage duplication and ETL costs. Consistent access control: Polaris allows you to manage role-based access control and storage layer security consistently across engines from one place. Flexibility in hosting: It can be hosted on various infrastructures, providing flexibility and reducing vendor lock-in. REST API implementation: Polaris implements Iceberg's REST API, enabling interoperability across multiple query engines. Multi-engine compatibility: It allows various query engines, such as Apache Spark, Apache Flink, Trino, and others, to read and write Iceberg tables through a single catalog. Centralized management: Polaris provides centralized and secure access to Iceberg tables across REST-compatible query engines. Namespace and metadata management: It supports creating namespaces for logical grouping and manages metadata for Iceberg tables. Storage configurations: Polaris offers robust storage configurations for cloud providers like S3, Azure, and Google Cloud Storage. Security features: It implements role-based access control (RBAC) and uses credential vending to secure query execution. Open-source nature: Polaris is fully open-source, allowing for community contributions and vendor neutrality. Interoperability: It enables multiple engines to be used on a single copy of data, reducing storage duplication and ETL costs. Consistent access control: Polaris allows you to manage role-based access control and storage layer security consistently across engines from one place. Flexibility in hosting: It can be hosted on various infrastructures, providing flexibility and reducing vendor lock-in. And a lovely little graphic from the docs to put a picture on everything. Borrowing from the Polaris docs for a picture once again, let’s dive into what is happening here in this entity hierarchy graphic: We can have multiple catalogs in a single Polaris install; in this diagram, it is only Catalog1 that has any namespaces or tables defined. A single catalog can be either internal or external. The internal tables allow for read/write functionality, meaning you have full control. The external tables are currently read-only and are intended to mirror data that is in some external catalog, such as a Glue catalog, which Polaris will see as the source of truth once connected. The catalog organizes your iceberg tables; you’ll want to configure it with your storage configuration for the applicable cloud system that you are on. This forms the first architectural layer in the Iceberg table spec. A namespace can have an arbitrary level of nesting and is used to logically group Iceberg tables within a catalog. An Iceberg table will always belong to a namespace. With this configuration in place, you can then use the Polaris catalog to describe your Iceberg tables in such a way as to make them easily accessible and manageable using a large variety of tools. That’s a quick overview of what is happening and what it looks like. The implementation inside of Snowflake provides a really handy UI to manage the implementation; other vendors have their own, and you can just do it through Spark if you like. Summary I continue to believe that Apache Iceberg is going to “win” the table format wars. Apache Polaris focusing on Iceberg is not a bad thing. I want to dig into how RBAC has been implemented in Polaris another time and see if it is as robust as what the Tabular product implemented, which I was a big fan of. I believe Polaris is a big win for the Iceberg community and will help accelerate adoption even further. I’ve seen talk about the Dremio Nessie project getting rolled in, and I’m curious to see what that will look like. I haven’t looked at Nessie in a couple of years, and I thought their git-like semantics mainly had been covered by the Iceberg tagging and branching, which I wrote about last year—yet another story for another day. wrote about I’m a big fan of what is happening here. If you are in this space, this is the right train to jump on. Check out my other What the Heck is… articles at the links below: What The Heck Is DuckDB? What the Heck Is Malloy? What the Heck is PRQL? What the Heck is GlareDB? What the Heck is SeaTunnel? What the Heck is LanceDB? What the heck is SDF? What the Heck is Paimon? What the Heck is Proton? What the Heck is PuppyGraph? What the Heck is GPTScript? What the Heck is WarpStream? What The Heck Is DuckDB? What The Heck Is DuckDB? What the Heck Is Malloy? What the Heck Is Malloy? What the Heck is PRQL? What the Heck is PRQL? What the Heck is GlareDB? What the Heck is GlareDB? What the Heck is SeaTunnel? What the Heck is SeaTunnel? What the Heck is LanceDB? What the Heck is LanceDB? What the heck is SDF? What the heck is SDF? What the Heck is Paimon? What the Heck is Paimon? What the Heck is Proton? What the Heck is Proton? What the Heck is PuppyGraph? What the Heck is PuppyGraph? What the Heck is GPTScript? What the Heck is GPTScript? What the Heck is WarpStream? What the Heck is WarpStream?

What The Heck is Apache Polaris?

How to Add Full Text Search to Your Django App with django-watson

How to Build a Production-Grade Text2SQL Engine

কিভাবে একটি প্রোডাকশন-গ্রেড Text2SQL ইঞ্জিন তৈরি করবেন

So erstellen Sie eine produktionsreife Text2SQL-Engine

Cómo crear un motor Text2SQL de nivel de producción

Comment créer un moteur Text2SQL de niveau production

प्रोडक्शन-ग्रेड टेक्स्ट2एसक्यूएल इंजन कैसे बनाएं

本番環境レベルの Text2SQL エンジンを構築する方法

프로덕션 등급 Text2SQL 엔진을 구축하는 방법

Como construir um mecanismo Text2SQL de nível de produção

Как создать Text2SQL-движок промышленного уровня

Üretim Seviyesinde Bir Text2SQL Motoru Nasıl Oluşturulur

Cách xây dựng một công cụ Text2SQL cấp sản xuất

如何构建生产级 Text2SQL 引擎

Authors:
(1) Keivan Alizadeh;
(2) Iman Mirzadeh, Major Contribution;
(3) Dmitry Belenko, Major Contribution;
(4) S. Karen Khatamifard;
(5) Minsik Cho;
(6) Carlo C Del Mundo;
(7) Mohammad Rastegari;
(8) Mehrdad Farajtabar. Table of Links Abstract and 1. Introduction 2. Flash Memory & LLM Inference and 2.1 Bandwidth and Energy Constraints 2.2 Read Throughput 3 Load From Flash 3.1 Reducing Data Transfer 3.2 Improving Transfer Throughput with Increased Chunk Sizes 3.3 Optimized Data Management in DRAM 4 Results 4.1 Results for OPT 6.7B Model 4.2 Results for Falcon 7B Model 5 Related Works 6 Conclusion and Discussion, Acknowledgements and References 2.2 Read Throughput Flash memory systems perform optimally with large sequential reads. For instance, benchmarks on an Apple MacBook Pro M2 with 2TB flash demonstrate speeds exceeding 6GiB/s for a 1GiB linear read of an uncached file. However, this high bandwidth is not replicated for smaller, random reads due to the inherent multi-phase nature of these reads, encompassing the operating system, drivers, interrupt handling, and the flash controller, among others. Each phase introduces latency, disproportionately affecting smaller reads. To circumvent these limitations, we advocate two primary strategies, which can be employed jointly. The first involves reading larger chunks of data. For smaller blocks, a substantial part of the overall read time is spent waiting for data transfer to begin. This is often referred to as latency to first byte. This latency reduces the overall throughput of each read operation considerably, because the overall measured throughput has to take into account not just the speed of transfer once it begins, but the latency before it begins as well, which penalizes small reads. This means that if we coalesce the reads for rows and colums of the FFN matrices, we can pay the latency cost only once for any given row/column pair in both matrices, and higher throughput can be realized. This principle is depicted in Figure 2b. Perhaps a counterintuitive yet interesting observation is that in some scenarios, it will be worthwhile to read more than needed (but in larger chunks) and then discard, than only reading strictly the necessary parts but in smaller chunks. The second strategy leverages parallelized reads, utilizing the inherent parallelism within storage stacks and flash controllers. Our results indicate that throughputs appropriate for sparse LLM inference are achievable on modern off-the-shelf hardware using 32KiB or larger random reads across multiple threads. Motivated by the challenges described in this section, in section 3, we propose methods to optimize data transfer volume and enhance read throughput to significantly enhance inference speeds. This paper is available on arxiv under CC BY-SA 4.0 DEED license. Authors: (1) Keivan Alizadeh; (2) Iman Mirzadeh, Major Contribution; (3) Dmitry Belenko, Major Contribution; (4) S. Karen Khatamifard; (5) Minsik Cho; (6) Carlo C Del Mundo; (7) Mohammad Rastegari; (8) Mehrdad Farajtabar. Authors: Authors: (1) Keivan Alizadeh; (2) Iman Mirzadeh, Major Contribution; (3) Dmitry Belenko, Major Contribution; (4) S. Karen Khatamifard; (5) Minsik Cho; (6) Carlo C Del Mundo; (7) Mohammad Rastegari; (8) Mehrdad Farajtabar. Table of Links Abstract and 1. Introduction Abstract and 1. Introduction 2. Flash Memory & LLM Inference and 2.1 Bandwidth and Energy Constraints 2. Flash Memory & LLM Inference and 2.1 Bandwidth and Energy Constraints 2.2 Read Throughput 2.2 Read Throughput 3 Load From Flash 3 Load From Flash 3.1 Reducing Data Transfer 3.1 Reducing Data Transfer 3.2 Improving Transfer Throughput with Increased Chunk Sizes 3.2 Improving Transfer Throughput with Increased Chunk Sizes 3.3 Optimized Data Management in DRAM 3.3 Optimized Data Management in DRAM 4 Results 4 Results 4.1 Results for OPT 6.7B Model 4.1 Results for OPT 6.7B Model 4.2 Results for Falcon 7B Model 4.2 Results for Falcon 7B Model 5 Related Works 5 Related Works 6 Conclusion and Discussion, Acknowledgements and References 6 Conclusion and Discussion, Acknowledgements and References 2.2 Read Throughput Flash memory systems perform optimally with large sequential reads. For instance, benchmarks on an Apple MacBook Pro M2 with 2TB flash demonstrate speeds exceeding 6GiB/s for a 1GiB linear read of an uncached file. However, this high bandwidth is not replicated for smaller, random reads due to the inherent multi-phase nature of these reads, encompassing the operating system, drivers, interrupt handling, and the flash controller, among others. Each phase introduces latency, disproportionately affecting smaller reads. To circumvent these limitations, we advocate two primary strategies, which can be employed jointly. The first involves reading larger chunks of data. For smaller blocks, a substantial part of the overall read time is spent waiting for data transfer to begin. This is often referred to as latency to first byte. This latency reduces the overall throughput of each read operation considerably, because the overall measured throughput has to take into account not just the speed of transfer once it begins, but the latency before it begins as well, which penalizes small reads. This means that if we coalesce the reads for rows and colums of the FFN matrices, we can pay the latency cost only once for any given row/column pair in both matrices, and higher throughput can be realized. This principle is depicted in Figure 2b. Perhaps a counterintuitive yet interesting observation is that in some scenarios, it will be worthwhile to read more than needed (but in larger chunks) and then discard, than only reading strictly the necessary parts but in smaller chunks. The second strategy leverages parallelized reads, utilizing the inherent parallelism within storage stacks and flash controllers. Our results indicate that throughputs appropriate for sparse LLM inference are achievable on modern off-the-shelf hardware using 32KiB or larger random reads across multiple threads. Motivated by the challenges described in this section, in section 3, we propose methods to optimize data transfer volume and enhance read throughput to significantly enhance inference speeds. This paper is available on arxiv under CC BY-SA 4.0 DEED license. This paper is available on arxiv under CC BY-SA 4.0 DEED license. available on arxiv

Large Language Models on Memory-Constrained Devices Using Flash Memory: Read Throughput

Stories, News & Mentions about COALESCE

Coalesce’s Stories on HackerNoon

Coalesce's News & Mentions