arXiv subjects

Athos Georgiou

Publications and source records attributed to Athos Georgiou.

4 recordsLinked to original sources

arXiv2026

The Price of Anarchy in Disaggregated Inference

Disaggregated inference architectures physically separate prefill and decode phases onto distinct GPU pools, creating competing "agents" that share a fixed hardware budget. We provide, to our knowledge, the first formal game-theoretic analysis of this architecture, using NVIDIA Dynamo as a concrete case study. We model disaggregated serving as three coupled games: a two-player resource game between prefill and decode pools, a selfish caching game over the hierarchical KV cache, and a congestion game with positive externalities for request routing. We empirically validate the latter two; the P/D resource game is treated analytically (Section 9.2). We characterize how GPU saturation induces regime transitions that shift the game's payoff structure: below saturation, selfish behavior has bounded Price of Anarchy (PoA); at saturation, superlinear latency and cache externalities drive our empirical estimator PoA-hat (defined in Section 6.4) upward. Based on this analysis, we design an adaptive controller that detects saturation transitions in real time and adjusts routing parameters accordingly, shifting from cache-affinity exploitation to load-balanced congestion avoidance. We instantiate our framework on a 3-node NVIDIA B200 cluster running Dynamo with two models, Nemotron-4-340B (TP=8, full-node workers with cross-InfiniBand KV transfers) and Llama-3.1-70B (TP=4), and find the same three-regime PoA-hat structure with the same first post-knee grid point (C=128) on both models. Adaptive routing shifts each model to a better operating point. Our strongest result is on the 70B 1P/5D topology, where PoA-hat drops 3.1x (66.4 to 21.5) in the saturated phase at a 13% throughput cost. On the 70B 1P/2D, PoA-hat drops 2.2x and TTFT P99 drops 7.6x (see Section 8.5).

cs.AR↗

arXiv2026

Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model

Visual document understanding typically requires separate retrieval and generation models, doubling memory and system complexity. We present Hydra, a dual-head approach that provides both ColBERT-style late-interaction retrieval and autoregressive generation from a single vision-language model. A single LoRA adapter, trained only for retrieval, is toggled at inference: enabling it produces multi-vector embeddings; disabling it recovers the base model's generation quality, with 426 of 426 language-model weight tensors byte-for-byte identical to a freshly-loaded Qwen3.5-4B. We identify two failure modes that can silently break generation in retrieval-fine-tuned VLMs (attention-mode restoration and lm_head preservation) plus an efficiency requirement (KV-cache-aware decoding); Hydra sidesteps the first two structurally and addresses the third in the decode loop. We release two scales, Hydra-4B and Hydra-0.8B, sharing LoRA hyperparameters (r=32, alpha=32) and optimisation recipe; data mix and projection dim differ across scales. The single-model design cuts peak GPU memory from 28.85 GB to 10.77 GB at 4B (62.7% reduction) and from 5.79 GB to 2.37 GB at 0.8B (59.1%) relative to a co-resident two-model deployment. A controlled ablation finds GritLM-style joint training matches Hydra's retrieval-only training on the evaluated modes while its LoRA-on generation mode collapses. A proof-of-concept on Qwen2.5-Omni-3B preserves generation equivalence on a non-Qwen3.5 backbone and transfers image retrieval within 2-8 pp of Hydra-4B, with zero-shot audio retrieval emerging through the frozen Whisper encoder.

cs.CV↗

arXiv2026

Architecture-Aware LLM Inference Optimization on AMD Instinct GPUs: A Comprehensive Benchmark and Deployment Study

We present a cross-architecture evaluation of production LLM inference on AMD Instinct MI325X GPUs, benchmarking four models spanning 235B to 1 trillion parameters across three architectural families (MoE+MLA, Dense+GQA, MoE+GQA) on an 8-GPU cluster with 2TB aggregate HBM3e using vLLM v0.14.1. Our results demonstrate that architecture-aware optimization is essential: MLA models require block size 1 and cannot use KV cache offloading, while GQA models benefit from both. The AMD AITER runtime is required for competitive MLA inference throughput and must be selectively disabled for architectures with incompatible attention head configurations. A controlled AITER ablation on Llama-3.1-405B (n=5 per condition) reveals a modest 3-5% throughput benefit at high concurrency but 2-16x higher measurement variability, confirming that AITER's large speedups target MoE/MLA kernels specifically. Under text-only workloads, Llama-405B and DeepSeek V3.2 achieve comparable peak throughput (15,944 and 15,343 tok/s) despite an order-of-magnitude difference in active parameters. Under vision workloads, Qwen3-VL-235B reaches 47,873 tok/s, 6.5x higher than Kimi-K2.5 (7,327 tok/s). Active parameter count per token is associated with inference throughput, though confounded by differences in quantization, AITER acceleration, and tensor parallelism. All four models exhibit a common throughput saturation point consistent with a memory-bandwidth bottleneck (~500 concurrent for short sequences, ~100-200 for longer sequences). All models maintain 100% HTTP-level success rates through 1,000 concurrent users, processing 18.9 million tokens across 17,406 requests without failures.

cs.AR↗

arXiv2026

Spatially-Grounded Document Retrieval via Patch-to-Region Relevance Propagation

Late-interaction multimodal retrieval models like ColPali achieve state-of-the-art document retrieval by embedding pages as images and computing fine-grained similarity between query tokens and visual patches. However, they operate at page-level granularity, limiting utility for retrieval-augmented generation (RAG) where precise context is paramount. Conversely, OCR-based systems extract structured text with bounding box coordinates but lack semantic grounding for relevance assessment. We propose a hybrid architecture that unifies these paradigms: using ColPali's patch-level similarity scores as spatial relevance filters over OCR-extracted regions. We formalize the coordinate mapping between vision transformer patch grids and OCR bounding boxes, introduce intersection metrics for relevance propagation, and establish theoretical bounds on area efficiency. We evaluate on BBox-DocVQA with ground-truth bounding boxes. For within-page localization (given correct page retrieval), ColQwen3-4B with percentile-50 thresholding achieves 59.7% hit rate at IoU@0.5 (84.4% at IoU@0.25, 35.8% at IoU@0.7), with mean IoU of 0.569, compared to ~6.7% for random region selection. Our approach reduces context tokens by 28.8% compared to returning all OCR regions and by 52.3% compared to full-page image tokens. Our approach operates at inference time without additional training. We release Snappy, an open-source implementation at https://github.com/athrael-soju/Snappy.

cs.CV↗