arXiv subjects

Marco Chiesa

Publications and source records attributed to Marco Chiesa.

At least 19 recordsLinked to original sources

arXiv2026

From Training Exposure to Membership Evidence: A Controlled Study of OLMo 2

Whether a language model was trained on a document matters for copyright, privacy, and benchmark contamination. Yet most membership-inference (MI) evaluations on language models collect members and non-members after the fact, so success may reflect distribution shift and failure is inconclusive. We re-ran OLMo 2 1B's official 50-billion-token mid-training stage, replacing selected training chunks with documents from 7,814 pairs matched on domain and difficulty. For each pair, a coin flip determined which document was inserted and which was held out. The run released by the Allen Institute for AI (AI2) from the same seed saw identical data everywhere else and serves as the counterfactual. One exposure raises a previously unseen document's likelihood on average but rarely identifies it. Using the counterfactual as reference, the reference ratio has an area under the ROC curve (AUC) of 0.767 but, with one threshold across domains, catches 3.8% of members at a 1% false-positive rate (FPR). In contrast, 13 matched pairs give a paired test 80% power at the 5% level to detect training on a collection, given same-domain non-members and a close reference. A reference trained with another seed recovers 92% of the counterfactual's AUC advantage over chance, whereas reference-free attacks stay within 0.07 of chance when pooled over domains. The full official post-training recipe, re-run on both models, leaves 89% of the average effect. Even with a closely matched reference, substantial document-level variability limits individual membership inference. We will release the checkpoints, per-token log-probabilities, and the evaluation documents; the code is available at https://github.com/ali-bana/OLMo-Twin.

cs.CR↗

arXiv2026

Taking Tokenization off the Host for Agentic LLM Serving

Agentic workloads resend long contexts that barely change between calls. Prefix caching now skips most of their prefill, yet every request must still be tokenized in full before the cache can even be consulted. Tokenization, a CPU stage long treated as negligible, becomes the largest part of time-to-first-token (TTFT), and a frontend that tokenizes for a whole GPU fleet limits how many GPUs it can feed. We present GraphTok, a GPU tokenizer that runs the entire encoding pipeline, from pre-tokenization through byte-pair encoding (BPE) merging to token-ID assembly, as one replayable CUDA graph. GraphTok reproduces the reference token IDs bit-for-bit on GPT-2, Llama 3, Qwen 2.5, DeepSeek-V3, and Gemma 3, hands short inputs to a CPU engine where the GPU cannot win, and plugs into unmodified vLLM and NVIDIA Dynamo. GraphTok encodes one million GPT-2 tokens in 1.7-3.2 ms including data transfers, and on Llama 3 it is 4.7-6.0x faster than gigatoken, the strongest CPU tokenizer. On TokTier's own single-request benchmark, it is 2.6-3.9x faster than gigatoken and up to 18.1x faster than TokTier. In vLLM with a cached prefix, it lowers mean TTFT by 2.1-2.4x. The largest gain is capacity. Within a 2 s P99 target, one Dynamo frontend with GraphTok sustains 518 requests/s of 121k-token traffic, the busiest-hour load of 56 GPUs, against 296 with gigatoken and 18.5 with the default tokenizer. Across 121k to one million tokens, this is 1.75-2.5x the load gigatoken sustains, with a P99 up to 64x lower where gigatoken's tail breaks.

cs.DC↗

arXiv2026

Dissecting GPU Utilization for LLM Inference on Nvidia Hopper

A single SM utilization percentage can make an LLM inference workload look compute-saturated while hiding how much useful work is being done. The problem is not that the counter is wrong, but that it collapses several different mechanisms into one number. This is most severe during decode, where each request contributes only one new token and dense projection GEMMs become small-row matrix multiplications. On Hopper, the bfloat16 GMMA path executes these operations in fixed 64-row matrix fragments, so small-batch decode can fill only a small fraction of each fragment with real token rows. In this paper, we profile vLLM with FlashAttention-3 and cuBLASLt on an H100 NVL across cold prefill, warm prefill, and decode, sweeping sequence length and batch size. We replace the usual single utilization number with eight counter-validated views derived from raw Nsight Compute reports, each pinned to an NCU counter or explicit formula. Together, these views map utilization gaps to concrete mechanisms - fragment fill, occupancy limits, stall signatures, wave quantization, and kernel selection - across four production models and six per-layer kernel roles.

cs.PF↗

arXiv2026

Avoiding Cross-Datacenter Collective Congestion via Disaggregated Buffering

LLM training at the scale of tens of thousands of GPUs now spans multiple datacenters (DC), making cross-DC collectives over long-haul links unavoidable. A critical and overlooked bottleneck arises when these collectives collide with intra-DC traffic at the destination - a common pattern in real workloads. The multi-millisecond congestion control loop is too slow to react, triggering severe packet loss and congestion collapse. We present Spillway, a transparent in-network mechanism that buffers dropped packets in switch-disaggregated buffers in a destination data center and drains them once congestion subsides. Through large-scale end-to-end simulations and a hardware prototype, we show that Spillway eliminates performance degradation from collective collisions, reducing iteration time by up to 14 %, without changes to end hosts or training frameworks.

cs.NI↗

arXiv2026

Blink: CPU-Free LLM Inference by Delegating the Serving Stack to GPU and SmartNIC

Large Language Model (LLM) inference is rapidly becoming a core datacenter service, yet current serving stacks keep the host CPU on the critical path for orchestration and token-level control. This makes LLM performance sensitive to CPU interference, undermining application colocation and forcing operators to reserve CPU headroom, leaving substantial capacity unutilized. We introduce Blink, an end-to-end serving architecture that removes the host CPU from the steady-state inference path by redistributing responsibilities across a SmartNIC and a GPU. Blink offloads request handling to the SmartNIC, which delivers inputs directly into GPU memory via RDMA, and replaces host-driven scheduling with a persistent GPU kernel that performs batching, scheduling, and KV-cache management without CPU involvement. Evaluated against TensorRT-LLM, vLLM, and SGLang, Blink outperforms all baselines even in isolation, reducing pre-saturation P99 TTFT by up to 8.47$\times$ and P99 TPOT by up to 3.40$\times$, improving decode throughput by up to 2.1$\times$, and reducing energy per token by up to 48.6$\%$. Under CPU interference, Blink maintains stable performance, while existing systems degrade by up to two orders of magnitude.

cs.DC↗

arXiv2026

KVComm: Enabling Efficient LLM Communication through Selective KV Sharing

Large Language Models (LLMs) are increasingly deployed in multi-agent systems, where effective inter-model communication is crucial. Existing communication protocols either rely on natural language, incurring high inference costs and information loss, or on hidden states, which suffer from information concentration bias and inefficiency. To address these limitations, we propose KVComm, a novel communication framework that enables efficient communication between LLMs through selective sharing of KV pairs. KVComm leverages the rich information encoded in the KV pairs while avoiding the pitfalls of hidden states. We introduce a KV layer-wise selection strategy based on attention importance scores with a Gaussian prior to identify the most informative KV pairs for communication. Extensive experiments across diverse tasks and model pairs demonstrate that KVComm achieves comparable performance to the upper-bound method, which directly merges inputs to one model without any communication, while transmitting as few as 30\% of layers' KV pairs. Our study highlights the potential of KV pairs as an effective medium for inter-LLM communication, paving the way for scalable and efficient multi-agent systems.

cs.LG↗

arXiv2025

Dissect-and-Restore: AI-based Code Verification with Transient Refactoring

Formal verification is increasingly recognized as a critical foundation for building reliable software systems. However, the need for specialized expertise to write precise specifications, navigate complex proof obligations, and learn annotations often makes verification an order of magnitude more expensive than implementation. While modern AI systems can recognize patterns in mathematical proofs and interpret natural language, effectively integrating them into the formal verification process remains an open challenge. We present Prometheus, a novel AI-assisted system that facilitates automated code verification with current AI capabilities in conjunction with modular software engineering principles (e.g., modular refactoring). Our approach begins by decomposing complex program logic, such as nested loops, into smaller, verifiable components. Once verified, these components are recomposed to construct a proof of the original program. This decomposition-recomposition workflow is non-trivial. Prometheus addresses this by guiding the proof search through structured decomposition of complex lemmas into smaller, verifiable sub-lemmas. When automated tools are insufficient, users can provide lightweight natural language guidance to steer the proof process effectively. Our evaluation demonstrates that transiently applying modular restructuring to the code substantially improves the AI's effectiveness in verifying individual components. This approach successfully verifies 86% of tasks in our curated dataset, compared to 68% for the baseline. Gains are more pronounced with increasing specification complexity, improving from 30% to 69%, and when integrating proof outlines for complex programs, from 25% to 87%.

cs.SE↗

arXiv2025

Can LLMs Forecast Internet Traffic from Social Media?

Societal events shape the Internet's behavior. The death of a prominent public figure, a software launch, or a major sports match can trigger sudden demand surges that overwhelm peering points and content delivery networks. Although these events fall outside regular traffic patterns, forecasting systems still rely solely on those patterns and therefore miss these critical anomalies. Thus, we argue for socio-technical systems that supplement technical measurements with an active understanding of the underlying drivers, including how events and collective behavior shape digital demands. We propose traffic forecasting using signals from public discourse, such as headlines, forums, and social media, as early demand indicators. To validate our intuition, we present a proof-of-concept system that autonomously scrapes online discussions, infers real-world events, clusters and enriches them semantically, and correlates them with traffic measurements at a major Internet Exchange Point. This prototype predicted between 56-92% of society-driven traffic spikes after scraping a moderate amount of online discussions. We believe this approach opens new research opportunities in cross-domain forecasting, scheduling, demand anticipation, and society-informed decision making.

cs.NI↗

arXiv2025

Priority-Aware Preemptive Scheduling for Mixed-Priority Workloads in MoE Inference

Large Language Models have revolutionized natural language processing, yet serving them efficiently in data centers remains challenging due to mixed workloads comprising latency-sensitive (LS) and best-effort (BE) jobs. Existing inference systems employ iteration-level first-come-first-served scheduling, causing head-of-line blocking when BE jobs delay LS jobs. We introduce QLLM, a novel inference system designed for Mixture of Experts (MoE) models, featuring a fine-grained, priority-aware preemptive scheduler. QLLM enables expert-level preemption, deferring BE job execution while minimizing LS time-to-first-token (TTFT). Our approach removes iteration-level scheduling constraints, enabling the scheduler to preempt jobs at any layer based on priority. Evaluations on an Nvidia A100 GPU show that QLLM significantly improves performance. It reduces LS TTFT by an average of $65.5\times$ and meets the SLO at up to $7$ requests/sec, whereas the baseline fails to do so under the tested workload. Additionally, it cuts LS turnaround time by up to $12.8\times$ without impacting throughput. QLLM is modular, extensible, and seamlessly integrates with Hugging Face MoE models.

cs.LG↗

arXiv2025

Deriving Coding-Specific Sub-Models from LLMs using Resource-Efficient Pruning

Large Language Models (LLMs) have demonstrated their exceptional performance in various complex code generation tasks. However, their broader adoption is limited by significant computational demands and high resource requirements, particularly memory and processing power. To mitigate such requirements, model pruning techniques are used to create more compact models with significantly fewer parameters. However, current approaches do not focus on the efficient extraction of programming-language-specific sub-models. In this work, we explore the idea of efficiently deriving coding-specific sub-models through unstructured pruning (i.e., Wanda). We investigate the impact of different domain-specific calibration datasets on pruning outcomes across three distinct domains and extend our analysis to extracting four language-specific sub-models: Python, Java, C++, and JavaScript. We are the first to efficiently extract programming-language-specific sub-models using appropriate calibration datasets while maintaining acceptable accuracy w.r.t. full models. We are also the first to provide analytical evidence that domain-specific tasks activate distinct regions within LLMs, supporting the creation of specialized sub-models through unstructured pruning. We believe that this work has significant potential to enhance LLM accessibility for coding by reducing computational requirements to enable local execution on consumer-grade hardware, and supporting faster inference times critical for real-time development feedback.

cs.LG↗

arXiv2025

From Scientific Texts to Verifiable Code: Automating the Process with Transformers

Despite the vast body of research literature proposing algorithms with formal guarantees, the amount of verifiable code in today's systems remains minimal. This discrepancy stems from the inherent difficulty of verifying code, particularly due to the time-consuming nature and strict formalism of proof details that formal verification tools require. However, the emergence of transformers in Large Language Models presents a promising solution to this challenge. In this position paper, we believe that transformers have the potential to read research papers that propose algorithms with formal proofs and translate these proofs into verifiable code. We leverage transformers to first build a formal structure of the proof using the original text from the paper, and then to handle the tedious, low-level aspects of proofs that are often omitted by humans. We argue that this approach can significantly reduce the barrier to formal verification. The above idea of reading papers to write verifiable code opens new avenues for automating the verification of complex systems, enabling a future where formally verified algorithms from academic research can more seamlessly transition into real-world software systems, thereby improving code reliability and security.

cs.SE↗

arXiv2025

Automating the Detection of Code Vulnerabilities by Analyzing GitHub Issues

In today's digital landscape, the importance of timely and accurate vulnerability detection has significantly increased. This paper presents a novel approach that leverages transformer-based models and machine learning techniques to automate the identification of software vulnerabilities by analyzing GitHub issues. We introduce a new dataset specifically designed for classifying GitHub issues relevant to vulnerability detection. We then examine various classification techniques to determine their effectiveness. The results demonstrate the potential of this approach for real-world application in early vulnerability detection, which could substantially reduce the window of exploitation for software vulnerabilities. This research makes a key contribution to the field by providing a scalable and computationally efficient framework for automated detection, enabling the prevention of compromised software usage before official notifications. This work has the potential to enhance the security of open-source software ecosystems.

cs.SE↗

arXiv2024

High-speed Source-Device-Independent Quantum Random Number Generator on a Chip

A wide range of applications require, by hypothesis, to have access to a high-speed, private, and genuine random source. Quantum Random Number Generators (QRNGs) are currently the sole technology capable of producing true randomness. However, the bulkiness of current implementations significantly limits their adoption. In this work, we present a high-performance source-device independent QRNG leveraging a custom made integrated photonic chip. The proposed scheme exploits the properties of a heterodyne receiver to enhance security and integration to promote spatial footprint reduction while simplifying its implementation. This characteristics could represents a significant advancement toward the development of generators better suited to meet the demands of portable and space applications. The system can deliver secure random numbers at a rate greater than 20 Gbps with a reduced spatial and power footprint.

quant-ph↗

arXiv2024

Just-in-Time Packet State Prefetching

Could information about future incoming packets be used to build more efficient CPU-based packet processors? Can such information be obtained accurately? This paper studies novel packet processing architectures that receive external hints about which packets are soon to arrive, thus enabling prefetching into fast cache memories of the state needed to process them, just-in-time for the packets' arrival. We explore possible approaches to (i) obtain such hints either from network devices or the end hosts in the communication and (ii) use these hints to better utilize cache memories. We show that such information (if accurate) can improve packet processing throughput by at least 50%.

cs.NI↗

arXiv2023

FMM-Head: Enhancing Autoencoder-based ECG anomaly detection with prior knowledge

Detecting anomalies in electrocardiogram data is crucial to identifying deviations from normal heartbeat patterns and providing timely intervention to at-risk patients. Various AutoEncoder models (AE) have been proposed to tackle the anomaly detection task with ML. However, these models do not consider the specific patterns of ECG leads and are unexplainable black boxes. In contrast, we replace the decoding part of the AE with a reconstruction head (namely, FMM-Head) based on prior knowledge of the ECG shape. Our model consistently achieves higher anomaly detection capabilities than state-of-the-art models, up to 0.31 increase in area under the ROC curve (AUROC), with as little as half the original model size and explainable extracted features. The processing time of our model is four orders of magnitude lower than solving an optimization problem to obtain the same parameters, thus making it suitable for real-time ECG parameters extraction and anomaly detection.

cs.LG↗

arXiv2023

Making Network Configuration Human Friendly

This paper explores opportunities to utilize Large Language Models (LLMs) to make network configuration human-friendly, simplifying the configuration of network devices and minimizing errors. We examine the effectiveness of these models in translating high-level policies and requirements (i.e., specified in natural language) into low-level network APIs, which requires understanding the hardware and protocols. More specifically, we propose NETBUDDY for generating network configurations from scratch and modifying them at runtime. NETBUDDY splits the generation of network configurations into fine-grained steps and relies on self-healing code-generation approaches to better take advantage of the full potential of LLMs. We first thoroughly examine the challenges of using these models to produce a fully functional & correct configuration, and then evaluate the feasibility of realizing NETBUDDY by building a proof-of-concept solution using GPT-4 to translate a set of high-level requirements into P4 and BGP configurations and run them using the Kathará network emulator.

cs.NI↗

arXiv2023

Oscilloscope: Detecting BGP Hijacks in the Data Plane

The lack of security of the Internet routing protocol (BGP) has allowed attackers to divert Internet traffic and consequently perpetrate service disruptions, monetary frauds, and even citizen surveillance for decades. State-of-the-art defenses rely on geo-distributed BGP monitors to detect rogue BGP announcements. As we show, though, attackers can easily evade detection by engineering their announcements. This paper presents Oscilloscope, an approach to accurately detect BGP hijacks by relying on real-time traffic analysis. As hijacks inevitably change the characteristics of the diverted traffic, the key idea is to track these changes in real time and flag them. The main challenge is that "normal" Internet events (e.g., network reconfigurations, link failures, load balancing) also change the underlying traffic characteristics - and they are way more frequent than hijacks. Naive traffic analyses would hence lead to too many false positives. We observe that hijacks typically target a subset of the prefixes announced by Internet service providers and only divert a subset of their traffic. In contrast, normal events lead to more uniform changes across prefixes and traffic. Oscilloscope uses this observation to filter out non-hijack events by checking whether they affect multiple related prefixes or not. Our experimental evaluation demonstrates that Oscilloscope quickly and accurately detects hijacks in realistic traffic traces containing hundreds of events.

cs.NI↗

arXiv2022

Fast Server Learning Rate Tuning for Coded Federated Dropout

In cross-device Federated Learning (FL), clients with low computational power train a common\linebreak[4] machine model by exchanging parameters via updates instead of potentially private data. Federated Dropout (FD) is a technique that improves the communication efficiency of a FL session by selecting a \emph{subset} of model parameters to be updated in each training round. However, compared to standard FL, FD produces considerably lower accuracy and faces a longer convergence time. In this paper, we leverage \textit{coding theory} to enhance FD by allowing different sub-models to be used at each client. We also show that by carefully tuning the server learning rate hyper-parameter, we can achieve higher training speed while also achieving up to the same final accuracy as the no dropout case. For the EMNIST dataset, our mechanism achieves 99.6\% of the final accuracy of the no dropout case while requiring $2.43\times$ less bandwidth to achieve this level of accuracy.

cs.LG↗