The article analyzes how Alibaba Cloud’s CPFS, KVCacheStore, AgenticFS, OSS Agent and CDE Agent together address exploding AI data scales, accelerate model training and inference, and support massive AI Agent workloads with concrete performance numbers and architectural details.
FreeToken, an open‑source MoE inference engine from UC Berkeley, enables a 35B model to run on an 8 GB GPU at 39.3 tokens per second by offloading weights to system memory and dynamically scheduling between CPU and GPU, offering a detailed performance analysis and community feedback.
This guide explains how to evaluate quantized large language models by measuring VRAM reduction, inference performance (latency, throughput), and model quality (perplexity, accuracy, BLEU/ROUGE), providing step‑by‑step procedures, code examples, and best‑practice recommendations for deployment.
The article provides a detailed technical comparison of SRAM, DRAM and HBM, explains their roles in modern AI GPUs such as the H100, discusses training and inference memory placement, and evaluates emerging ideas like High‑Bandwidth SRAM and High‑Bandwidth Flash, concluding that all three memory types will coexist in future AI systems.
The article reviews Coleman Hooper's 2026 Berkeley PhD thesis, which shows that LLM inference is increasingly limited by memory bandwidth and capacity, and proposes a four‑pronged approach—weight quantization, KV‑cache quantization, selective context loading, and multipole attention—to dramatically improve memory efficiency and throughput.
In a candid YC Startup School interview, Jeff Dean explains how AI is moving from ever‑larger models to long‑running, self‑evolving agents, emphasizing inference hardware, data‑movement costs, context engineering, and the strategic 1% rule that guides startups to focus on problems where generic models still fail.
AirLLM introduces layer‑wise inference and per‑expert streaming to decouple VRAM usage from model size, enabling the 2.8‑trillion‑parameter Kimi K3 LLM to run on a single consumer‑grade GPU while preserving full‑precision accuracy and offering security‑focused insights.
The article analyzes why AI companies such as DeepSeek, Zhipu, OpenAI and Anthropic are moving toward custom inference ASICs, citing shifting compute costs, agent-driven inference demand, economic incentives, supply‑chain control, and export‑control challenges that together reshape the AI hardware landscape.
DeepSeek is quietly developing a custom AI inference chip—started a year ago and recruited for without public postings—to cut reliance on Nvidia, a move reflected in a broader industry shift toward self‑designed chips and backed by a 51‑billion‑RMB funding round for compute infrastructure and talent expansion.
Wafer demonstrates that the open‑source GLM‑5.2 model can run on AMD’s MI355X GPU at 2,626 tokens per second with single‑stream throughput of 213 tokens per second, achieving comparable performance to Nvidia’s B200 at less than half the cost, by applying MXFP4 quantization, selecting the sglang engine, and fixing naming and kernel issues.
VibeThinker-3B, a newly open‑sourced 3‑billion‑parameter model, achieves near‑state‑of‑the‑art scores on math competitions (AIME, IMO‑AnswerBench), coding (LiveCodeBench), and verification benchmarks, rivaling trillion‑parameter closed models, thanks to a Spectrum‑to‑Signal training pipeline, multi‑stage SFT, RL, and offline distillation, supporting a new parametric compression‑coverage hypothesis.
The article explains how prediction, reasoning, and inference differ, why a pure prediction model leaves the decision chain broken, and how a dynamic ontology‑driven feature framework—temporal, functional, and relational features—creates explainable, verifiable, and iterative decision loops.
The author details the deployment of an office AI system on an internal network using eight NVIDIA A800 GPUs, explaining model choices, inference engines, GPU allocations, compatibility issues, and presenting the overall architecture diagram.
Cerebras Systems raised $5.55 billion in the largest 2026 IPO, debuting the wafer‑scale WSE‑3 chip that promises unprecedented inference bandwidth and could erode Nvidia’s dominance, while navigating CFIUS scrutiny, a dramatic financial turnaround, and a shifting AI‑chip market landscape.
The vLLM 0.21.0 release brings five major updates—including Transformers v4 deprecation, a C++20 build requirement, KV offload with hybrid memory, speculative decoding that respects thinking budgets, and a Blackwell token‑speed backend—while offering detailed upgrade guidance for different user groups.
In a deep interview, former Google TPU architect Reiner Pope explains that low‑concurrency fast‑mode services trade higher fees for faster streaming but are limited by memory‑bandwidth bottlenecks, that optimal concurrency balances compute and memory costs, and that pipeline‑parallel sparse expert models and reinforcement‑learning fine‑tuning introduce new inefficiencies and overtraining risks.
This article compares two recent GitHub LLM inference engines—ds4.c, a Metal‑optimized engine for DeepSeek V4 Flash on Apple Silicon Macs, and TokenSpeed, a Python/C++‑based, data‑center‑grade engine for GPU clusters—detailing their design choices, performance numbers, usage instructions, and suitable scenarios.
As large‑model inference demand outpaces training, the decode stage hits a memory‑wall that GPUs cannot efficiently cross; SRAM’s on‑chip bandwidth and low‑energy access open a path forward, though capacity and process limits still pose challenges.
Google’s Cloud Next 2026 reveal splits the 8th‑generation TPU into training‑focused Sunfish and inference‑focused Zebrafish, highlighting Ironwood’s record‑breaking performance, a multi‑vendor supply chain, Anthropic’s multi‑gigawatt order, and a broader industry shift toward custom AI chips that promise far higher profit margins than generic GPUs.
The article examines the recent Claude performance controversy, showing that engineering adjustments to inference parameters, cache handling, and system prompts rewrote the model’s behavior, making it answer faster but think shallower, leading users to perceive a degradation despite no official model downgrade.
Google unveiled its eighth‑generation TPU line at Cloud Next 2026, introducing the training‑focused TPU 8t with a 2.7× performance boost and massive scaling, and the inference‑optimized TPU 8i featuring three‑times more on‑chip SRAM and an 80% performance uplift for agentic AI workloads, while positioning the chips as a complement—not a replacement—to Nvidia's offerings.
This article explains why scaling GPU inference services on Kubernetes is challenging and presents a multi‑layer control architecture, metric upgrades, and production‑ready implementations using HPA, KEDA, KServe, and Karpenter to achieve stable, cost‑effective autoscaling.
By 2026 inference will consume over 70% of AI compute, prompting a shift from GPU‑centric training to specialized, low‑latency, low‑cost inference chips, with Nvidia, Google, Amazon, Microsoft, Intel and newcomers like Groq and CoreWeave racing to capture the new battlefield.
This article walks engineers through the full LLM stack—from tokenization and positional encoding to transformer blocks, efficient fine‑tuning, quantization, and production‑grade inference techniques such as KV‑cache, FlashAttention, PagedAttention, continuous batching, and speculative decoding—highlighting trade‑offs, toolchains, and practical workflow steps.
This article details the vLLM‑Kunlun open‑source project that adapts the high‑performance vLLM inference engine to Baidu's Kunlun XPU, covering platform overview, model‑porting workflow, plugin architecture, concrete case studies with MIMO‑Flash‑V2 and Qwen 3.5, and the performance‑tuning techniques that enable seamless, GPU‑level inference on domestic hardware.
The 2025 State of AI report from Artificial Analysis outlines five core trends—intensified competition, the rise of autonomous agents, native speech models, mainstream inference models, and booming image/video generation—showing how costs have plummeted, capabilities have surged, and AI is reshaping every industry.
This article analyzes Meta's rapid evolution of four generations of MTIA AI chips, detailing how modular hardware, inference‑first design, deep software integration, and aggressive iteration cycles delivered up to 30 PFLOPs of performance and dramatically reshaped the AI compute landscape.
This article presents a comprehensive, step‑by‑step methodology for troubleshooting and optimizing large‑language‑model inference performance, covering GPU, CPU, memory, network, configuration, and application layers, with concrete benchmark scripts, diagnostic commands, and real‑world case studies.
The article outlines a comprehensive, repeatable regression testing framework for TensorRT inference pipelines, covering engine build validation, functional correctness against golden outputs, performance monitoring, common pitfalls, and CI/CD integration to ensure model updates remain both fast and reliable.
This guide reviews the Qwen3.5 model lineup, explains mixed‑inference and MoE architecture, presents benchmark comparisons with GPT‑5.2, Claude 4.5 and Gemini‑3 Pro, evaluates 4‑bit and 3‑bit quantization loss, outlines hardware requirements, and provides step‑by‑step deployment options using llama.cpp or llama‑server.
Google’s recent study shows that the length of a model’s token chain is negatively correlated with inference accuracy, and introduces the Deep Thinking Ratio (DTR) metric to identify truly reasoning tokens, enabling the Think@n strategy to halve compute cost without sacrificing performance.
This roadmap outlines a step‑by‑step learning path for building, optimizing, and safely deploying large language model systems, covering fundamentals, vector stores, RAG, advanced techniques, fine‑tuning, inference speed, deployment, observability, agents, and production safeguards.
This article reviews practical techniques for accelerating large language model inference—including reduced‑precision formats, post‑training quantization, adapter‑based fine‑tuning, pruning, continuous batch processing, and multi‑GPU deployment—while providing concrete code examples, benchmark results, and guidance on selecting the right approach for production workloads.
This guide walks through setting up a Kunlun P800 XPU host, preparing Docker containers, deploying Qwen3‑8B/‑32B/‑VL models with vLLM‑Kunlun, benchmarking performance, and running full‑parameter DPO training using LLaMA‑Factory, providing scripts, configuration files, and troubleshooting tips for AI engineers.
This article explains the NVFP4 4‑bit floating‑point quantization technique, shows how to deploy Qwen3‑30B‑A3B models with TensorRT‑LLM and vLLM, compares performance across NVFP4, AWQ and INT8 quantizations, and provides practical profiling commands for NVIDIA DGX systems.
This article details how to deploy the 235‑billion‑parameter Qwen3‑235B model using PD‑separation and MoE techniques, explains the associated challenges, and demonstrates a production‑grade solution built on the high‑performance SGLang inference engine and the RoleBasedGroup (RBG) orchestration framework, complete with benchmark results and best‑practice YAML examples.
This guide walks through building a production‑grade Kubernetes GPU cluster for large language model inference, covering hardware sizing, GPU resource scheduling, model storage options, automated scaling with HPA, health checks, monitoring, troubleshooting, and multi‑model deployment strategies.
The article details how the small Qwen‑0.6B model was adapted to generate and run WebAssembly code during inference, achieving deterministic calculations and revealing both the promise and current limitations of integrating world‑model reasoning into tiny LLMs.
This article walks through the complete workflow of loading and running the open‑source Qwen3‑8B model, explaining each core file (weights, config, generation config, tokenizer), how the model tokenizes input, applies chat templates, generates responses, and decodes output, all illustrated with code and diagrams.
This guide walks you through deploying large language models such as ChatGLM and Llama in production, covering environment setup, model quantization, dynamic batching, service configuration, Nginx load balancing, monitoring, troubleshooting, and best‑practice recommendations for high‑performance, cost‑effective AI inference.
This article explains why KV‑Cache hit rate is critical for large‑model inference, describes vLLM's automatic prefix caching, outlines the distributed cache challenges, and provides a step‑by‑step guide to deploying Alibaba Cloud ACK Gateway with Inference Extension's precise‑mode prefix‑cache‑aware routing, backed by benchmark results.
The vLLM‑Kunlun Plugin, built on the vLLM hardware‑plugin RFC, lets developers deploy any major large language model on Baidu's Kunlun XPU instantly without modifying vLLM core code, dramatically shortening migration time, providing high‑performance fusion operators, and offering open‑source tools for precision verification and profiling.
Alibaba’s 1688 platform leverages inference‑based large language models to enhance recommendation discovery, addressing the filter‑bubble problem by analyzing long‑term buyer behavior, compressing extensive activity streams, generating nuanced demand queries, and integrating multimodal data and market trend agents to deliver more diverse, explainable product suggestions for B‑type buyers.
This guide walks you through setting up an Ubuntu environment, training a LightGBM stock‑movement model in Python, exporting it to ONNX, and deploying fast, cross‑platform inference in Go using ONNX Runtime, complete with code snippets and project structure.
Ling-1T, a trillion‑parameter flagship non‑thinking model, combines 50 billion active parameters per token, 128 K context, Evo‑CoT reasoning, and FP8 mixed‑precision training to achieve state‑of‑the‑art performance on complex reasoning, code generation, and multimodal tasks while outlining its architecture, benchmarks, limitations, and future roadmap.
LongCat-Flash-Thinking, the latest open‑source model from Meituan, introduces domain‑parallel RL training, a high‑throughput DORA infra, and a dual‑path inference framework that together achieve state‑of‑the‑art performance on logical, mathematical, coding, and agentic tasks while maintaining top‑tier speed.
The article explains how to choose suitable large AI models for product use by focusing on three key aspects—model inference, post‑training optimization techniques such as quantization, pruning and distillation, and the role of GPUs, including their kernels, software stack, and computing platforms.
Baidu’s Baige 5.0 AI Computing Platform introduces FP8 mixed‑precision training, MoE‑aware distributed strategies, adaptive parallelism, and a three‑tier KV‑Cache, delivering over 30% training speedup and 50% inference throughput gains while keeping token latency under half a second for large‑scale models.
LangChain4j 1.2.0 introduces a suite of stable modules, advanced inference and thinking capabilities, streaming tool calls, and extensive AI service enhancements, offering developers finer control, lower latency, and richer debugging for LLM‑driven applications.
Burn, a native Rust deep learning framework by Tracel AI, combines extreme flexibility, high computational efficiency, and cross‑platform portability through a modular backend abstraction, type‑safe tensor operations, asynchronous execution, and extensive tooling, offering performance‑competitive alternatives to Python‑based frameworks for both training and inference.
Large language models are booming, and this article highlights six cutting‑edge research directions—LLM‑plus synthetic data, reward modeling, inference techniques, LLM‑as‑a‑Judge, safety alignment, and long‑context handling—each illustrated with recent papers, experimental results, and links to code repositories.
The 2025 Q1 AI report from Artificial Analysis highlights six major trends—including a thousand‑fold drop in inference cost, the rise of MoE models, the growing parity of Chinese open‑source labs, the emergence of autonomous AI agents, native multimodal capabilities, and the trade‑off between performance, cost, and context windows—painting a picture of a rapidly evolving, increasingly competitive AI ecosystem.
The article describes how a large‑scale dispatch system was re‑engineered with NVIDIA TritonServer to unify GPU‑accelerated operations‑research kernels and deep‑learning models, detailing a three‑stage architecture (in‑process, cross‑process, cross‑node), the performance, stability and memory challenges addressed, and future plans for heterogeneous GPU scaling.
This article provides a comprehensive comparison of seven popular large‑language‑model inference engines—Transformers, vLLM, Llama.cpp, SGLang, MLX, Ollama and others—detailing their core features, performance characteristics, hardware compatibility, concurrency support, and ideal use‑cases, plus practical installation guidance for Xinference.
Cursor, the AI‑powered code editor, grew to handle billions of document queries and over a hundred‑million model calls daily, prompting a multi‑stage infrastructure overhaul that moved from a failing YugaByte setup to PostgreSQL RDS, then to object‑storage‑backed databases, while tackling indexing, inference scaling, and cold‑start challenges.
DeepSeek‑VL2 is a state‑of‑the‑art multimodal model built on a Mixture‑of‑Experts architecture that combines a SigLIP‑L vision encoder with dynamic tiling, a two‑layer VL adaptor, and a DeepSeek‑MoE language model using Multi‑head Latent Attention, trained in three stages on diverse visual‑language and text data, and achieving strong results on benchmarks such as DocVQA and TextVQA, with full implementation and inference code available in PaddleMIX.
Xinference is a powerful distributed inference framework that enables quick deployment and efficient serving of open‑source large language models via Docker or source installation, offering Web UI, CLI, and API interfaces with detailed setup, model launching, and Chatbox integration instructions.
This guide walks you through deploying the open‑source QwQ‑32B inference model using Alibaba Cloud's Serverless AI platform CAP, covering benchmark highlights, preparation steps, two deployment methods (application template and model service), verification, and project cleanup.
This guide walks you through using Alibaba Cloud Container Compute Service (ACS) to provision GPU resources, prepare the QwQ-32B model, configure persistent storage, deploy the model with vLLM, set up OpenWebUI, verify the service, and optionally benchmark its performance, all with detailed commands and YAML examples.
This guide explains how to deploy the QwQ-32B large language model using vLLM on an Alibaba Cloud ACK Kubernetes cluster, configure storage, set up OpenWebUI, enable ACK Gateway with AI Extension for intelligent routing, and benchmark the inference service performance.
JD Retail’s Nine‑Number Algorithm Platform delivers an end‑to‑end AI engine that unifies GPU and domestic NPU resources across a thousand‑card cluster, offering zero‑cost model migration, optimized training and inference pipelines, support for over 40 LLM and multimodal models, and proven business‑level performance that reduces dependence on overseas chips.
The article analyses DeepSeek‑R1’s low‑cost open‑source strategy, superior inference performance (including GPQA benchmark gains over GPT‑4o), its focus on complex reasoning, math and programming, and how these traits reshape AIGC across industries while highlighting remaining privacy and ethical challenges.
This article describes JD Retail's AI engine built on domestic NPU chips, covering challenges, heterogeneous GPU‑NPU scheduling, high‑performance training and inference engines, extensive model support, real‑world deployment cases, and future plans for large‑scale chip clusters and ecosystem development.
This article details JD.com's challenges and solutions for deploying domestic NPU chips across heterogeneous GPU‑NPU clusters, covering architecture, scheduling, high‑performance training and inference engines, real‑world case studies, and future plans to scale AI workloads securely and efficiently.
This article explains how inference‑oriented large language models such as DeepSeek‑R1 and OpenAI o1‑mini shift AI research from training‑time scaling to test‑time computation, detailing the underlying principles, new scaling laws, verification techniques, reinforcement‑learning pipelines, and practical methods for distilling reasoning capabilities into smaller models.
The author benchmarks DeepSeek's FlashMLA against FlashInfer and several Triton-based implementations, detailing setup challenges, decode‑only bandwidth results, and observations that the official DeepSeek version leads while Triton optimizations show mixed performance across different head sizes.
This article introduces Jlama, an open‑source Java LLM inference engine, outlines its key features, provides step‑by‑step CLI and Maven integration instructions, shows code examples, run logs, and special setup notes for using large language models efficiently within Java applications.
The article examines the end‑to‑end architecture for large‑model inference, detailing seven layers—from chip hardware and programming toolkits to deep‑learning frameworks, inference accelerators, model providers, compute platforms, application orchestration, and traffic management—highlighting key vendors, open‑source projects, and performance‑optimizing techniques.
This guide walks you through creating a lightweight compute instance, adding it to Baidu Baige AI heterogeneous computing platform, deploying the vLLM tool, loading and serving small‑scale dense models such as DeepSeek, Llama and Qwen, and provides recommended configuration lists to achieve low‑cost, high‑performance inference.
The article examines DeepSeek’s rapid rise, its open‑source R1 model and distilled variants, the resurgence of AI PCs, hardware support from Nvidia, AMD and others, and how this ecosystem is reshaping personal AI experiences and the broader large‑model landscape.
This guide explains how to deploy a production‑ready DeepSeek‑R1 inference service on Alibaba Cloud ACK using KServe, covering model preparation, storage configuration, service deployment, observability, autoscaling, model acceleration, gray‑release and GPU‑shared inference.
This article reviews a comprehensive study that applies Chain‑of‑Thought reasoning to autoregressive text‑to‑image generation, introducing extended test‑time computation, direct preference optimization, and two custom reward models (PARM and PARM++) that together improve generation quality by up to 15% over Stable Diffusion 3.
Direct Preference OptimizationImage Generationchain-of-thought
This guide walks you through deploying the open‑source DeepSeek‑R1 inference model on Alibaba Cloud's Compute Nest platform, covering service creation, instance configuration, login procedures, and API calls with sample curl commands for text generation and chat.
This article explains how to reduce inference cost and improve performance for large language models on Alibaba Cloud ACK by using Knative's request‑based autoscaling, custom ResourcePolicy priority scheduling, and Fluid data‑caching to achieve elastic scaling, resource pre‑emption, and faster model loading.
The article evaluates the Chinese GLM‑Zero‑Preview inference model by subjecting it to a wide range of math, logic, language, coding, and multimodal questions, compares its token efficiency and reasoning style to other models, and discusses its current strengths, limitations, and public availability.
This article evaluates NVIDIA's China‑specific H20 and L20 inference chips, comparing their compute and memory‑bandwidth characteristics against A100, H100 and H200, and shows how they achieve superior throughput in large‑model inference despite reduced specifications.
This article explains how the inference component of OpenAI’s o1 model can be implemented using Monte‑Carlo Tree Search, detailing the action space, rollout process, UCT scoring, and best‑path selection, with a concrete walkthrough of Microsoft’s open‑source rStar code.
The article provides an in‑depth technical analysis of OpenAI’s multimodal o1 model, explaining its self‑play reinforcement‑learning pipeline, the novel train‑time and test‑time compute scaling laws, its long‑think reasoning abilities demonstrated through a cipher example, and speculative architectures for generator‑verifier systems.
This article provides a thorough, yet concise, overview of Llama 3’s training pipeline, data handling, model architecture, scaling laws, post‑training techniques like SFT and DPO, and inference optimizations such as KV‑Cache, GQA, PagedAttention, and FP8 quantization, highlighting practical insights and benchmark results.
The article details JD Advertising's technical challenges and solutions for large‑scale sparse recommendation models, describing GPU‑focused storage, compute and I/O optimizations for both training and low‑latency inference, including distributed pipelines, heterogeneous deployment, batch aggregation, multi‑stream execution, and compiler extensions.
This article presents a comprehensive overview of Xiaohongshu's heterogeneous training and inference engine, covering the challenges of model engineering, the design of elastic heterogeneous engines, future HPC training frameworks, AI compilation techniques, and a forward‑looking outlook on scalability and performance.
The article traces the evolution of large‑model training and inference infrastructure from the early “black‑iron” era (2019‑2021) through the “golden” boom (2022‑2023) to the emerging “silver” phase (2024‑), highlighting key research breakthroughs, open‑source frameworks, hardware trends, market dynamics, and practical challenges for engineers entering the field.
The article analyzes how generative AI models from GPT‑1 to the upcoming GPT‑5 are driving exponential growth in compute requirements, prompting massive cloud capital expenditures and intense competition among GPU vendors such as NVIDIA, AMD, Google, and emerging domestic chip makers, while also highlighting interconnect innovations and cost‑effective solutions.
This article details the background, goals, and evolution of Tencent's FinTech AI development platform, outlines the technical challenges faced in feature engineering, model training, and inference services, and presents the comprehensive solutions and future plans implemented to improve efficiency, stability, and scalability.
This article provides a comprehensive overview of Transformer models, covering their attention-based architecture, encoder-decoder structure, training procedures including teacher forcing, inference workflow, advantages over RNNs, and various applications in natural language processing such as translation, summarization, and classification.
This article presents a comprehensive overview of large language model inference, describing the prefill and decoding stages, key performance metrics such as throughput, latency and QPS, and detailing a series of system-level optimizations—including pipeline parallelism, dynamic batching, KV‑cache quantization, and hardware considerations—to significantly improve inference efficiency on modern GPUs.
Early‑Stopping Self‑Consistency (ESC) dynamically halts sampling once a sliding‑window answer distribution reaches zero entropy, cutting the number of required LLM reasoning samples by 34‑84 % across arithmetic, commonsense, and symbolic benchmarks while preserving accuracy and offering a theoretically‑bounded, robust, budget‑adaptive alternative to traditional Self‑Consistency.
This article explains the Transformer model’s core architecture, self‑attention mechanism, encoder‑decoder workflow, training with teacher forcing, inference steps, and why it surpasses RNNs and CNNs, while also outlining its major NLP applications.
This article introduces the ModelScope Python library, explains its support for PyTorch and TensorFlow, and provides step‑by‑step code examples for loading models, running inference pipelines on text and images, handling batch inputs, and customizing preprocessors in both Python and PHP.
This article explains how to efficiently deploy speech AI models—including ASR and TTS—on GPUs using NVIDIA's Triton Inference Server and TensorRT, covering background challenges, GPU‑based solutions, decoding optimizations, Whisper acceleration with TensorRT‑LLM, streaming TTS improvements, voice‑cloning pipelines, future plans, and a Q&A session.
This article reviews the LLaMA large‑language‑model series, covering its background, architectural innovations such as Add&Norm, SwiGLU, and RoPE, a known reversal‑curse bug, and provides step‑by‑step MindSpore Transformers code for model configuration, inference, and pipeline usage while previewing the upcoming LLaMA‑2 session.
This article details Xiaohongshu's end‑to‑end GPU‑centric transformation for search‑related machine‑learning models, covering model characteristics, training and inference frameworks, system‑level GPU and CPU optimizations, multi‑card and compilation techniques, and future directions for scaling large sparse and dense models.
This article compares NVIDIA's high‑end AI GPUs—H100, A100, and the newer L40S—detailing their specifications, performance trade‑offs, pricing, availability, and suitability for training and inference workloads, while highlighting why L40S can be a cost‑effective alternative for many enterprises.
This article details a step‑by‑step guide to deploying the Vicuna open‑source LLM on a single server, covering model preparation, environment setup, dependency installation, GPU and CUDA configuration, inference commands, performance evaluation, and attempted fine‑tuning, while sharing practical observations and results.
This article details the step‑by‑step process of installing GPU drivers, setting up a Python environment, deploying the open‑source Alpaca‑LoRA large language model, fine‑tuning it with Chinese data on a multi‑GPU server, and running inference, while discussing practical challenges and performance observations.
The article presents the background, challenges, and objectives of Bilibili's AI services, introduces the self‑developed InferX inference framework with its quantization and sparsity optimizations, details OCR‑specific enhancements, and describes how integrating InferX with Nvidia Triton dramatically improves throughput, latency, and GPU utilization.
This article details the step‑by‑step process of preparing the environment, merging weights, installing dependencies, running inference, evaluating Vicuna‑7B against other models, and attempting fine‑tuning, while highlighting performance results, encountered issues, and future work for large language model deployment.
The article analyses the computational resources, hardware costs, and human investment needed to train and serve large language models such as GPT‑3, discusses practical cost calculations, highlights the challenges faced by Chinese AI teams, and argues for sustained, long‑term funding to achieve meaningful breakthroughs.
dl_inference is an open‑source, production‑grade deep learning inference platform that supports TensorFlow, PyTorch and Caffe models, offering GPU and CPU deployment, TensorRT and MKL acceleration, multi‑node load balancing, and extensive Q&A on model conversion, hardware requirements, INT8 quantization, and performance gains.