AISQL AI_PARSE_DOCUMENT¶

{
  "metadata": {
    "pageCount": 19
  },
  "pages": [
    {
      "content": "# SwiftKV: Fast Prefill-Optimized Inference with Knowledge-Preserving Model Transformation \n\nAurick Qiao
      Zhewei Yao Samyam Rajbhandari Yuxiong He<br>Snowflake AI Research<br>San Mateo, CA, United States<br>Correspondence:
      aurick.qiao@ snowflake.com\n\n\n#### Abstract\n\nLLM inference for enterprise applications, such as summarization, RAG,
      and code-generation, typically observe much longer prompt than generations, leading to high prefill cost and response
      latency. We present SwiftKV, a novel model transformation and distillation procedure targeted at reducing the prefill
      compute (in FLOPs) of prompt tokens while preserving high generation quality. First, SwiftKV prefills later layers' KV
      cache using an earlier layer's output, allowing prompt tokens to skip those later layers. Second, SwiftKV employs a
      lightweight knowledge-preserving distillation procedure that can adapt existing LLMs with minimal accuracy impact. Third,
      SwiftKV can naturally incorporate KV cache compression to improve inference performance in low-memory scenarios. Our
      comprehensive experiments show that SwiftKV can effectively reduce prefill computation by $25-50 \\%$ across several LLM
      families while incurring minimum quality degradation. In the end-to-end inference serving, SwiftKV realizes up to $2
      \\times$ higher aggregate throughput and $60 \\%$ lower time per output token. It can achieve a staggering 560 TFlops/GPU
      of normalized inference throughput, which translates to 16 K tokens/s for Llama-3.1-70B. SwiftKV is open-sourced at
      https://github . com/snowflakedb/arctictraining.\n\n\n## 1 Introduction\n\nLarge Language Models (LLMs) are now an
      integral enabler of enterprise applications and offerings, including code and data co-pilots (Chen et al., 2021; Pourreza
      and Rafiei, 2024), retrieval augmented generation (RAG) (Lewis et al., 2020; Lin et al., 2024), summarization (Pu et al.,
      2023; Zhang et al., 2024), and agentic workflows (Wang et al., 2024; Schick et al., 2023). However, the cost and speed of
      inference determine their practicality, and improving the throughput and latency of LLM inference has become increasingly
      important.\n\nWhile prior works, such as model pruning (Ma et al., 2023; Sreenivas et al., 2024), KV cache compression
      (Hooper et al., 2024; Shazeer, 2019; Ainslie et al., 2023b; Chang et al., 2024), and sparse attention (Zhao et al., 2024;
      Jiang et al., 2024), have been developed to accelerate LLM inference, they typically significantly degrade the model
      quality or work best in niche scenarios, such as lowmemory environments or extremely long contexts requests (e.g. $>100
      \\mathrm{~K}$ tokens). On the other hand, production deployments are often compute-bound rather than memory-bound, and
      such long-context requests are rare amongst diverse enterprise use cases (e.g. those observed at Snowflake).\n\nIn this
      paper, we take a different approach to improving LLM inference based on the key observation that typical enterprise
      workloads process more input tokens than output tokens. For example, tasks like code completion, text-to-SQL,
      summarization, and RAG each submit long prompts but produce fewer output tokens (a 10:1 ratio with average prompt length
      between 500 and 1000 is observed in our production). In these scenarios, inference throughput and latency are often
      dominated by the cost of prompt processing (i.e. prefill), and reducing this cost is key to improving their performance.
      \n\nBased on this observation, we designed SwiftKV, which improves throughput and latency by reducing the prefill
      computation for prompt tokens. SwiftKV (Fig. 1) consists of three key components:\n\nModel transformation. SwiftKV rewires
      an existing LLM so that the prefill stage during inference can skip a number of later transformer layers, and their KV
      cache are computed by the last unskipped layer. This is motivated by the observation that the hidden states of later
      layers do not change significantly (see Sec. 3.2 and (Liu et al., 2024b)). With SwiftKV, prefill compute is reduced by
      approximately the number of layers skipped.\n\nOptionally, for low-memory scenarios, we",
      "index": 0
    },
    ...
    {
      "content": "Efficient Distillation. Since only a few $\\mathbf{W}_{Q K V}$ parameters need training, we can keep just a
      single copy of the original model weights in memory that are frozen during training, and add an extra trainable copy of
      the $\\mathbf{W}_{Q K V}$ parameters for layers $>l$ initialized using the original model (See Fig. 1).\n\nDuring
      training, we create two modes for the later layers $>l$, one with original frozen parameters using original architecture,
      and another with the SwiftKV re-wiring using new QKV projections i.e.,\n\n$$\n\\begin{aligned}\n& \\mathbf{y}_{\\text
      {teacher }}=\\mathbf{M}(\\mathbf{x}, \\text { SwiftKV }=\\text { False }) \\\\\n& \\mathbf{y}_{\\text {student }}=\\mathbf
      {M}(\\mathbf{x}, \\text { SwiftKV }=\\text { True })\n\\end{aligned}\n$$\n\nwhere $\\mathbf{y}$ is the final logits,
      $\\mathbf{M}$ is the model, and $\\mathbf{x}$ is the input. Afterwards, we apply the standard distillation loss (Hinton et
      al., 2015) on the outputs. After the distillation, the original KV projection layers $>l$ are discarded during inference.
      \n\nThis method allows us to distill Llama-3.1-8BInstruct on 680 M tokens of data in 3 hours using 8 H100 GPUs, and
      Llama-3.1-70B-Instruct in 5 hours using 32 H100 GPUs across 4 nodes. In contrast, many prune-and-distill (Sreenivas et
      al., 2024) and layer-skipping (Elhoushi et al., 2024) methods require much larger datasets (e.g. 10-100B tokens) and incur
      greater accuracy gaps than SwiftKV.\n\n### 3.5 Optimized Implementation for Inference\n\nLLM serving systems can be
      complex and incorporate many simultaneous optimizations at multiple layers of the stack, such as PagedAttention (Kwon et
      al., 2023), Speculative Decoding (Leviathan et al., 2023), SplitFuse (Holmes et al., 2024; Agrawal et al., 2024), and
      more. A benefit of SwiftKV is that it makes minimal changes to the model architecture, so it can be integrated into
      existing serving systems without implementing new kernels (e.g. for custom attention operations or sparse computation) or
      novel inference procedures.\n\nImplementation in vLLM and SGLang. To show that the theoretical compute reductions of
      SwiftKV translates to real-world savings, we integrated it with vLLM (Kwon et al., 2023) and SGLang (Zheng et al., 2024).
      Our implementation is compatible with chunked prefill (Holmes et al., 2024; Agrawal et al., 2024), which mixes chunks of
      prefill tokens and decode tokens in each minibatch. During each forward pass, after completing layer $l$, the KV-cache for
      the remaining layers ( $>l$ ) are immediately computed, and only the decode tokens are propagated through the rest of the
      model layers.\n\n## 4 Main Results\n\nWe evaluated SwiftKV in terms of model accuracy (Sec. 4.1) compared to the original
      model and several baselines, and end-to-end inference performance (Sec. 4.2) in a real serving system.\n\nDistillation
      datasets. Our dataset is a mixture of Ultrachat (Ding et al., 2023), SlimOrca (Lian et al., 2023), and OpenHermes-2.5
      (Teknium, 2023), totaling roughly 680M Llama-3.1 tokens. For more details, please see Appendix A.1.\n\nSwiftKV Notation.
      For prefill computation, we report the approximate reduction as $(L-l) / L$ due to SwiftKV, and for KV cache, we report
      the exact memory reduction due to AcrossKV. For example, SwiftKV $(l=L / 2)$ and 4-way AcrossKV is reported as $50 \\%$
      prefill compute reduction and $37.5 \\% \\mathrm{KV}$ cache memory reduction.\n\n### 4.1 Model Quality Impact of
      SwiftKV\n\nTable 2 shows the quality results of all models we evaluated, including Llama-3.1-Instruct, Qwen2.
      5-14B-Instruct, Mistral-Small, and Deepseek-V2. Of these models, we note that the Llama models span two orders of
      magnitude in size (3B to 405B), Llama-3.1-405B-Instruct uses FP8 (W8A16) quantization, and Deepseek-V2-Lite-Chat is a
      mixture-of-experts model that implements a novel latent attention mechanism (DeepSeek-AI et al., 2024).\n\nWe also compare
      with three baselines: (1) FFN-SkipLLM (Jaiswal et al., 2024), a training-free method for skipping FFN layers (no attention
      layers are skipped) based on hidden state similarity, (2) Llama-3.1-Nemotron-51B-Instruct (Sreenivas et al., 2024), which
      is pruned and distilled from Llama-3.1-70B-Instruct using neural architecture search on 40B tokens, and (3) DarwinLM-8.4B
      (Tang et al., 2025), which is pruned and distilled from Qwen2.5-14B-Instruct using 10B tokens.\n\nSwiftKV. For Llama,
      Mistral, and Deepseek, we find the accuracy degradation for $25 \\%$ SwiftKV is less than $0.5 \\%$ from the original
      models (averaged across tasks). Additionally, the accuracy gap is within $1-2 \\%$ even at $40-50 \\%$ SwiftKV. Beyond $50
      \\%$ SwiftKV, model quality drops quickly. For example, Llama-3.1-8B-Instruct incurs a 7\\% accuracy gap at $62.5 \\%$
      SwiftKV. We find that Qwen suffers larger degradations, at $1.1 \\%$ for $25 \\%$ SwiftKV and $7.4 \\%$ for $50 \\%$
      SwiftKV, which may be due to Qwen models having lower simularity between layer at 50-75\\% depth (Fig. 2). Even still,
      SwiftKV",
      "index": 4
    },
    ...
  ]
}

{
  "metadata": {
    "pageCount": 53
  },
  "pages": [
    {
      "content": ...
      "index": 0
    },
    ....
    {
      "content": "# Key Operational and Business Metrics \n\nIn addition to the measures presented in our interim condensed
      consolidated financial statements, we use the following key operational and business metrics to evaluate our business,
      measure our performance, develop financial forecasts, and make strategic decisions.\n\n|  | Three Months Ended March 31, |  |
      \n| :--: | :--: | :--: |\n|  | 2025 | 2024 |\n| Ending Paid Connected Fitness Subscriptions ${ }^{(1)}$ | 2,880,176 | $3,051,
      451$ |\n| Average Net Monthly Paid Connected Fitness Subscription Churn ${ }^{(1)}$ | $1.2 \\%$ | $1.2 \\%$ |\n| Ending Paid
      App Subscriptions ${ }^{(1)}$ | 572,775 | 675,190 |\n| Average Monthly Paid App Subscription Churn ${ }^{(1)}$ | $8.1 \\%$ |
      $9.0 \\%$ |\n| Subscription Gross Profit (in millions) | \\$ 288.8 | \\$ 298.1 |\n| Subscription Contribution (in millions) $
      { }^{(2)}$ | \\$ 304.9 | \\$ 316.4 |\n| Subscription Gross Margin | $69.0 \\%$ | $68.1 \\%$ |\n| Subscription Contribution
      Margin ${ }^{(2)}$ | $72.9 \\%$ | $72.3 \\%$ |\n| Net loss (in millions) | \\$ $(47.7)$ | \\$ $(167.3)$ |\n| Adjusted EBITDA
      (in millions) ${ }^{(3)}$ | \\$ 89.4 | \\$ 5.8 |\n| Net cash provided by operating activities (in millions) | \\$ 96.7 | \\$
      11.6 |\n| Free Cash Flow (in millions) ${ }^{(4)}$ | \\$ 94.7 | \\$ 8.6 |\n\n[^0]\n## Ending Paid Connected Fitness
      Subscriptions\n\nEnding Paid Connected Fitness Subscriptions includes all Connected Fitness Subscriptions for which we are
      currently receiving payment (a successful credit card billing or prepaid subscription credit or waiver). We do not include
      paused Connected Fitness Subscriptions in our Ending Paid Connected Fitness Subscription count.\n\n## Average Net Monthly
      Paid Connected Fitness Subscription Churn\n\nTo align with the definition of Ending Paid Connected Fitness Subscriptions
      above, our quarterly Average Net Monthly Paid Connected Fitness Subscription Churn is calculated as follows: Paid Connected
      Fitness Subscriber \"churn count\" in the quarter, divided by the average number of beginning Paid Connected Fitness
      Subscribers each month, divided by three months. \"Churn count\" is defined as quarterly Connected Fitness Subscription
      churn events minus Connected Fitness Subscription unpause events minus Connected Fitness Subscription reactivations.\n\nWe
      refer to any cancellation or pausing of a subscription for our All-Access Membership as a churn event. Because we do not
      receive payment for paused Connected Fitness Subscriptions, a paused Connected Fitness Subscription is treated as a churn
      event at the time the pause goes into effect, which is the start of the next billing cycle. An unpause event occurs when a
      pause period elapses without a cancellation and the Connected Fitness Subscription resumes, and is therefore counted as a
      reduction in our churn count in that period. Our churn count is shown net of reactivations and our new quarterly Average Net
      Monthly Paid Connected Fitness Subscription Churn metric averages the monthly Connected Fitness churn percentage across the
      three months of the reported quarter.\n\n## Ending Paid App Subscriptions\n\nEnding Paid App Subscriptions include all App
      One, App+, and Strength+ subscriptions for which we are currently receiving payment.\n\n## Average Monthly Paid App
      Subscription Churn\n\nWhen a Subscriber to App One, App+, or Strength+ cancels their membership (a churn event) and
      resubscribes in a subsequent period, the resubscription is considered a new subscription (rather than a reactivation that is
      counted as a reduction in our churn count). Average Paid App Subscription Churn is calculated as follows: Paid App
      Subscription cancellations in the quarter, divided by the average number of beginning Paid App Subscriptions each month,
      divided by three months.\n\n\n[^0]:    (1) Beginning January 1, 2025, the Company migrated its subscription data model for
      reporting Ending Paid Connected Fitness Subscriptions, Average Net Monthly Paid Connected Fitness Subscription Churn, Ending
      Paid App Subscriptions, and Average Monthly Paid App Subscription Churn to a new data model that provides greater visibility
      to changes to a subscription's payment status when they occur. The new model gives the Company more precise and timely data
      on subscription pause and churn behavior. Prior period information has been revised to conform with current period
      presentation. The impact of this change in the model on Ending Paid Connected Fitness Subscriptions, Average Net Monthly
      Paid Connected Fitness Subscription Churn, Ending Paid App Subscriptions and Average Monthly Paid App Subscription Churn for
      the three months ended March 31, 2025 and 2024 is immaterial.\n    (2) Please see the section titled \"Non-GAAP Financial
      Measures—Subscription Contribution and Subscription Contribution Margin\" for a reconciliation of Subscription Gross Profit
      to Subscription Contribution and an explanation of why we consider Subscription Contribution and Subscription Contribution
      Margin to be helpful measures for investors.\n    (3) Please see the section titled \"Non-GAAP Financial Measures—Adjusted
      EBITDA\" for a reconciliation of Net loss to Adjusted EBITDA and an explanation of why we consider Adjusted EBITDA to be a
      helpful measure for investors.\n    (4) Please see the section titled \"Non-GAAP Financial Measures-Free Cash Flow\" for a
      reconciliation of net cash provided by (used in) operating activities to Free Cash Flow and an explanation of why we
      consider Free Cash Flow to be a helpful measure for investors.",
      "index": 28
    },
    ...
    {
      "content": "# CERTIFICATION OF PRINCIPAL FINANCIAL OFFICER PURSUANT TO 18 U.S.C. SECTION 1350, AS ADOPTED PURSUANT TO
      SECTION 906 OF THE SARBANES-OXLEY ACT OF 2002 \n\nI, Elizabeth F Coddington, Chief Financial Officer of Peloton Interactive,
      Inc. (the \"Company\"), do hereby certify, pursuant to 18 U.S.C. Section 1350, as adopted pursuant to Section 906 of the
      Sarbanes-Oxley Act of 2002, that to the best of my knowledge:\n\n1. the Quarterly Report on Form 10-Q of the Company for the
      fiscal quarter ended March 31, 2025 (the \"Report\") fully complies with the requirements of Section 13(a) or 15(d) of the
      Securities Exchange Act of 1934, as amended; and\n2. the information contained in the Report fairly presents, in all
      material respects, the financial condition, and results of operations of the Company.\n\nDate: May 8, 2025\n\nBy: /s/
      Elizabeth F Coddington\nElizabeth F Coddington\nChief Financial Officer\n(Principal Financial Officer)",
      "index": 52
    }
  ]
}

{
  "metadata": {
    "pageCount": 5
  },
  "pages": [
    {
      "content": "# Deloitte.\n\n![img-0.jpeg](img-0.jpeg)\n\n## IFRS 16\n\nAdvisory assistance with determining discount rates
      (the IBR method)",
      "index": 0
    },
    {
      "content": "## **Background**\n\nFor accounting periods beginning on or after 1 January 2019, IFRS/FRS101 adopters will be
      required to bring all operating leases onto the balance sheet.\n\nThe adoption of this standard will particularly
      affect:\n\n- companies with significant operating lease commitments on premises/property; and\n- companies with leased
      cars; office equipment; machinery, or other such assets.\n\nUnder IFRS 16, companies are required to calculate and
      disclose the net present value of their lease obligations.\n\nThis creates a lease liability and right of use asset on the
      balance sheet for the first time, which can result in a significant change in the profile and overall net asset value
      presented in the statutory accounts. Companies need to assess the impact, and implement the change, and this is also a
      primary focus of audit teams.\n\n![img-1.jpeg](img-1.jpeg)\n\n*The most challenging aspect that groups are typically
      facing when seeking to implement the standard is determining an appropriate discount rate.*\n\nIf a company is adopting
      the Full Retrospective Approach, the Standard sets out that the discount rate can be calculated in one of two ways (the
      incremental borrowing rate (\"IBR\") or interest rate implicit in the lease (\"IRIIL\") method).\n\nHowever, if a company
      is adopting the modified retrospective approach, (most companies we have assisted thus far have done), the standard
      mandates that the IBR method must be applied.\n\nIn some circumstances, where the information is available and where the
      company adopts a 'Full Retrospective Approach', the rate used may be the IRIIL. However, many businesses either do not
      have this information or are adopting the 'Modified Retrospective Approach', and that means they need to work out the IBR.
      \n\n![img-2.jpeg](img-2.jpeg)\n\n![img-3.jpeg](img-3.jpeg)",
      "index": 1
    },

  .......

    {
      "content": "# Deloitte. \n\nThis publication has been written in general terms and we recommend that you obtain
      professional advice before acting or refraining from action on any of the contents of this publication. Deloitte LLP
      accepts no liability for any loss occasioned to any person acting or refraining from action as a result of any material in
      this publication.\n\nDeloitte LLP is a limited liability partnership registered in England and Wales with registered
      number OC303675 and its registered office at 1 New Street Square, London, EC4A 3HQ, United Kingdom.\n\nDeloitte LLP is the
      United Kingdom affiliate of Deloitte NSE LLP, a member firm of Deloitte Touche Tohmatsu Limited, a UK private company
      limited by guarantee (\"DTTL\"). DTTL and each of its member firms are legally separate and independent entities. DTTL and
      Deloitte NSE LLP do not provide services to clients. Please see www.deloitte.com/about to learn more about our global
      network of member firms.\n(c) 2019 Deloitte LLP. All rights reserved.",
      "index": 4
    }
  ]
}

{
  "metadata": {
    "pageCount": 1
  }
  "content": "![img-0.jpeg](img-0.jpeg)\n\nSchulen haben es verdient, gute Orte zu sein. Hier sollen wir Wissen und Fähigkeiten
  erlernen, die uns durch das Leben tragen. Hier verbringen viele einen Großteil ihres Tages, und das in einer Lebensphase, in
  der sich Zeit beinahe grenzenlos und eine Doppelstunde wie ein halbes Leben anfühlen kann.\n\nOb es die Freundin ist, ohne die
  man auf dem Schulhof verloren wäre. Der Lehrer, mit dem man nicht klarkommt, den man aber trotzdem jeden Tag aushalten muss.
  Die Klassenfahrt, auf der man zum ersten Mal das Meer sieht und knutscht. In Schulen entstehen Erfahrungen, Beziehungen und
  Erinnerungen, die uns ein ganzes Leben prägen.\n\nDie Erwartungen an Schulen sind dementsprechend hoch. Trotzdem werden sie
  von der Gesellschaft schnell vergessen und von der Politik hinten angestellt. Seit Jahrzehnten kriegt das deutsche Schulsystem
  verheerende Zeugnisse.\n\nNoch immer entscheiden Bildungsgrad und Kontostand der Eltern darüber, welchen Schulabschluss Kinder
  und Jugendliche machen. Noch immer funktioniert es vielerorts nur auf dem Papier, dass alle gut zusammen lernen. Im Alltag
  fehlen dann die Lehrkräfte und Mittel, um zum Beispiel einen geflüchteten Jugendlichen oder einen mit ADHS so zu unterstützen,
  dass alle möglichst gleichberechtigt in einem Klassenraum sitzen. Auch die gesellschaftliche Einsicht, dass alle
  Schulabschlüsse ihren Wert haben und gebraucht werden, muss erst wieder zurückgewonnen werden.\n\nJetzt aber hoch mit
  euch!\nDass Schule so irre früh anfangen muss, ist kein Gesetz. Und auch gar nicht ratsam: Jugendliche haben einen anderen
  Biorhythmus und brauchen mehr Schlaf als Erwachsene. Ein Schulbeginn gegen 9 oder 10 Uhr wäre für die meisten besser, da ist
  sich die Forschung einig\n\nAn Schulen tritt die Realität sehr schnell ein. Während sich die Gesellschaft noch fragt, wie mit
  künstlicher Intelligenz umzugehen ist, nutzen sie Lehrkräfte, Schülerinnen und Schüler längst für ihre Zwecke. Während über
  Jahre diskutiert wurde, ob Deutschland ein Einwanderungsland sei, war es das an Schulen längst. Und während andere Themen den
  Klimawandel in der Öffentlichkeit verdrängen, sind es besonders Schülerinnen und Schüler, die laut auf das drängendste Problem
  unserer Zeit hinweisen. Die Herausforderungen und Fragen, die sich an Schulen stellen, betreffen uns alle. Schule ist Zukunft.
  \n\nSchulleitungen, Lehrkräfte, pädagogisches Personal und alle, die sich sonst noch um das Gelingen des Schulalltags kümmern,
  stellen sich dem jeden Tag aufs Neue. Sie versuchen, Schule trotz vieler Probleme und fehlender Wertschätzung zu gestalten,
  sie versuchen, den Schülerinnen und Schülern zu vermitteln, dass es auf sie ankommt. Damit sie selbst an sich glauben. Sie
  haben es verdient.",
}

"Schools deserve to be good places. Here, we are supposed to learn knowledge and skills that will carry us through life. Many
spend a large part of their day here, and this is during a phase of life when time can seem almost endless and a double period
can feel like half a lifetime.

Whether it's the friend you would be lost without in the schoolyard. The teacher you can't get along with, but still have to
endure every day. The class trip where you see the sea for the first time and make out. In schools, experiences,
relationships, and memories are created that shape us for a lifetime.

The expectations for schools are correspondingly high. Nevertheless, they are quickly forgotten by society and pushed to the
back by politics. For decades, the German school system has been receiving devastating reports.

Even now, the level of education and the financial status of the parents still determine which school certificate children and
young people receive. It still only works on paper that everyone learns well together. In everyday life, the teachers and
resources are lacking to support, for example, a refugee youth or a student with ADHD so that they can sit in a classroom on
an equal footing. The societal insight that all school certificates have value and are needed also needs to be regained.

Now, let's get going!

The fact that school has to start so early is not a law. And it's not advisable either: teenagers have a different biological
rhythm and need more sleep than adults. A start time of 9 or 10 o'clock would be better for most, research agrees.

Reality sets in very quickly at schools. While society is still wondering how to deal with artificial intelligence, teachers,
students, and pupils are already being used for their purposes. While it was debated for years whether Germany is an
immigration country, it has been one in schools for a long time. And while other topics are pushing climate change out of the
public eye, it is especially students who are loudly pointing out the most pressing problem of our time. The challenges and
questions that schools face affect us all. School is the future.

School administrations, teachers, educational staff, and all those who take care of the success of everyday school life face
this every day. They try to shape school despite many problems and lack of appreciation, they try to convey to the students
that it's up to them. So that they believe in themselves. They deserve it."

{
  "content": "content of the document"
}