Searcharxiv⌕ Search

arXiv subjects

Shobhit Kanaujia

Publications and source records attributed to Shobhit Kanaujia.

7 recordsLinked to original sources

ForeSpark: Joint Draft and Verification Budget Scheduling for Parallel Speculative Decoding

Parallel drafters accelerate LLM decoding by proposing a whole block of draft tokens in a single forward pass, making them attractive for LLM serving. Recent systems optimize goodput (i.e., accepted tokens per second) because rejected drafts steal compute from other requests. However, these systems control only how many tokens are verified, so every request still drafts a full block. At high serving load, we find that drafting consumes up to 40% of step time, of which a majority is discarded because it is never verified. Yet, draft-length control from autoregressive drafters does not carry over, because a parallel drafter commits to its length before its single pass. Moreover, draft and verification lengths are coupled across the batch, so they must be selected jointly for every request while keeping the decision computationally cheap at every step. In this paper, we propose work-conserving speculation (WCS), a design that jointly allocates draft and verification budgets per request in LLM serving, and present ForeSpark as its prototype. ForeSpark realizes WCS with (i) a context-aware pre-draft estimator that predicts acceptance for every candidate length from the target model's context features, (ii) a lightweight joint draft-verify scheduler that selects both lengths through a global sort, and (iii) a post-draft budget distribution that spends verification on the most promising tokens. Implemented on vLLM, ForeSpark improves throughput by up to 1.77x over the state-of-the-art DSpark and by 2.49x over autoregressive decoding at a matched per-request generation rate, without changing the output distribution. It also tolerates higher load, sustaining up to 1.51x the request rate of DSpark under latency service level objectives (SLOs).

cs.DC↗

Architecting for Hardware Diversity: Meta's End-to-End Co-Design for High-Performance AI Training on AMD GPUs

Meta deploys a heterogeneous fleet comprising NVIDIA GPUs, AMD GPUs, and proprietary MTIA accelerators to reduce costs and mitigate vendor lock-in. However, the extensive history of NVIDIA-specific optimizations hinders performance portability to other hardware. This paper presents our methods for optimizing alternative accelerators, using AMD Instinct MI350X versus NVIDIA B200 for recommendation model training as a case study. We propose three reusable principles. First, hardware-model co-design is essential: rather than directly porting models from NVIDIA GPUs, revising model structures and software recipes allows full exploitation of accelerator-specific strengths (e.g., MI350X's 1.6x HBM capacity relative to B200). Second, our analytical performance model enables zero-warmup optimal kernel selection across diverse datatypes and matrix shapes in production models, drastically simplifying tuning. Third, increasingly mature agentic workflows have lowered the barrier to porting from the CUDA ecosystem; we extend these workflows for microarchitecture-aware kernel optimizations and custom shim generation. Compared to the directly ported baseline, these methods deliver up to 1.5x improvement in end-to-end training throughput. To our knowledge, this is the first hyperscaler report on productionizing AMD GPUs for large-scale training. Our contributions have been upstreamed to the open-source PyTorch, Triton, and ROCm ecosystems.

cs.DC↗

Provisioning to Runtime Optimization of a 100 MW-Scale AI Cluster

The electric power supply for AI data centers is now the most significant bottleneck in the race toward Artificial General Intelligence, surpassing even the constraint of AI accelerator availability. To our knowledge, this paper is the first to describe the end-to-end power management process for a hyper-scale AI datacenter; from early power planning to accommodate next-generation accelerators 6--12 months before their general availability, to tuning power settings after large scale deployment, and finally to dynamic, runtime power management for evolving workloads. We present detailed power measurements for a 150 MW datacenter hosting a cluster of 83K GB200 GPUs. We share insights from building this state-of-the-art AI cluster. We hope this work encourages practitioners across the industry to share their own experiences as well.

cs.AR↗

Crossflow: Prefill-Decode Elasticity for Agentic LLM Serving

As serving capacity demand surpasses that of training, serving efficiency becomes increasingly important. Prefill-decode (P/D) disaggregation improves serving efficiency through specialization and isolation of the two phases. These benefits rest on a static partitioning. Phase demand, however, is not static. We observe that in a large LLM fleet the ratio of uncached input to output tokens has peak-to-mean ratios up to 4.7x at minute timescales, and that in a public agentic trace the hourly ratio spans a median 24.5x within a single day, while reassigning a replica takes tens of minutes. Agentic traffic sharpens the mismatch. Sizing each pool at its ninety-fifth percentile leaves up to 17% of cluster capacity unused; sizing below it converts the same imbalance into queueing and unrealized throughput. We present Crossflow, which makes this boundary elastic without changing node roles. Each decode node publishes a short-lived, revocable lease that bounds local-prefill compute, KV capacity, transfer work, and projected output. Across public and internal traces, Crossflow improves token throughput by 16.2-17.4% on geometric mean over static P/D, and by up to 43.4% at high load, while reducing mean TTFT at every evaluated point.

cs.DC↗

TPP: Transparent Page Placement for CXL-Enabled Tiered-Memory

The increasing demand for memory in hyperscale applications has led to memory becoming a large portion of the overall datacenter spend. The emergence of coherent interfaces like CXL enables main memory expansion and offers an efficient solution to this problem. In such systems, the main memory can constitute different memory technologies with varied characteristics. In this paper, we characterize memory usage patterns of a wide range of datacenter applications across the server fleet of Meta. We, therefore, demonstrate the opportunities to offload colder pages to slower memory tiers for these applications. Without efficient memory management, however, such systems can significantly degrade performance. We propose a novel OS-level application-transparent page placement mechanism (TPP) for CXL-enabled memory. TPP employs a lightweight mechanism to identify and place hot/cold pages to appropriate memory tiers. It enables a proactive page demotion from local memory to CXL-Memory. This technique ensures a memory headroom for new page allocations that are often related to request processing and tend to be short-lived and hot. At the same time, TPP can promptly promote performance-critical hot pages trapped in the slow CXL-Memory to the fast local memory, while minimizing both sampling overhead and unnecessary migrations. TPP works transparently without any application-specific knowledge and can be deployed globally as a kernel release. We evaluate TPP in the production server fleet with early samples of new x86 CPUs with CXL 1.1 support. TPP makes a tiered memory system performant as an ideal baseline (<1% gap) that has all the memory in the local tier. It is 18% better than today's Linux, and 5-17% better than existing solutions including NUMA Balancing and AutoTiering. Most of the TPP patches have been merged in the Linux v5.18 release.

cs.DC↗

First-Generation Inference Accelerator Deployment at Facebook

In this paper, we provide a deep dive into the deployment of inference accelerators at Facebook. Many of our ML workloads have unique characteristics, such as sparse memory accesses, large model sizes, as well as high compute, memory and network bandwidth requirements. We co-designed a high-performance, energy-efficient inference accelerator platform based on these requirements. We describe the inference accelerator platform ecosystem we developed and deployed at Facebook: both hardware, through Open Compute Platform (OCP), and software framework and tooling, through Pytorch/Caffe2/Glow. A characteristic of this ecosystem from the start is its openness to enable a variety of AI accelerators from different vendors. This platform, with six low-power accelerator cards alongside a single-socket host CPU, allows us to serve models of high complexity that cannot be easily or efficiently run on CPUs. We describe various performance optimizations, at both platform and accelerator level, which enables this platform to serve production traffic at Facebook. We also share deployment challenges, lessons learned during performance optimization, as well as provide guidance for future inference hardware co-design.

cs.AR↗

Live Recovery of Bit Corruptions in Datacenter Storage Systems

Due to its high performance and decreasing cost per bit, flash is becoming the main storage medium in datacenters for hot data. However, flash endurance is a perpetual problem, and due to technology trends, subsequent generations of flash devices exhibit progressively shorter lifetimes before they experience uncorrectable bit errors. In this paper we propose extending flash lifetime by allowing devices to expose higher bit error rates. To do so, we present DIRECT, a novel set of policies that leverages latent redundancy in distributed storage systems to recover from bit corruption errors with minimal performance and recovery overhead. In doing so, DIRECT can significantly extend the lifetime of flash devices by effectively utilizing these devices even after they begin exposing bit errors. We implemented DIRECT on two real-world storage systems: ZippyDB, a distributed key-value store backed by RocksDB, and HDFS, a distributed file system. When tested on production traces at Facebook, DIRECT reduces application-visible error rates in ZippyDB by more than 10^2 and recovery time by more than 10^4. DIRECT also allows HDFS to tolerate a 10^4--10^5 higher bit error rate without experiencing application-visible errors.

cs.DC↗