LLMOps

Answer-first: For 80% of domain enterprise tasks, fine-tuned Small Language Models (1B–14B) match frontier performance at 1/50th inference cost and sub-40ms latency. This playbook provides the production engineering blueprint: synthetic data curation, QLoRA fine-tuning with Axolotl on 24GB GPUs, DeepSeek-R1 reasoning distillation, DPO/GRPO alignment, and vLLM continuous batching deployment. 🎯 Series Overview: Why Small Language Models in 2026? Relying exclusively on proprietary frontier API models (GPT-4.5, Claude 3.5 Sonnet) introduces three fatal enterprise vulnerabilities: ...

← Previous Chapter: Executive Summary | Series Hub | Next Chapter: Part 2: SFT Data Engineering → Prerequisite: Read Executive Summary: The Rise of Specialized Small Language Models for cost break-even formulas and hybrid AI architectural framing. Answer-first: The Hybrid AI Routing architecture evaluates incoming request complexity and token uncertainty in under 3.5ms. 80% of structured queries are served locally by fine-tuned 7B models on vLLM within 35ms TTFT, while low-confidence requests automatically cascade to Claude 3.5 Sonnet through a localized PII sanitization proxy, cutting monthly API bills by 85%. ...

Tech Radar: Disaggregated Prefill-Decode Architecture: Decoupling Compute & Memory Bandwidth via RoCEv2 KV-Transfer Answer-First: Disaggregated Prefill-Decode serving defines 2026 enterprise LLM infrastructure, resolving the tension between compute-heavy prefill and memory-bound decode. By streaming KV caches across 400Gbps RoCEv2 fabrics, it cuts P99 TTFT by 11x (420ms to 38ms) and eliminates decode latency jitter on NVIDIA H100 clusters. name: "Disaggregated Prefill-Decode Serving" ring: "Adopt" quadrant: "AI Infrastructure & Large Language Models" rationale: "Decouples compute-bound prompt prefill from memory-bandwidth-bound token decode, eliminating head-of-line blocking and slashing P99 TTFT by 11x via zero-copy RoCEv2 KV transfer." adr_link: "/radar/2026-09/disaggregated-prefill-decode/" justification: "Empirically verified across 64x NVIDIA H100 SXM5 GPUs on DeepSeek-V3 and Llama-3.1-70B; production-ready in vLLM v1 and Mooncake architectures with 2.8x higher throughput per dollar." 1. The Compute vs. Memory-Bandwidth Dichotomy in Autoregressive Serving Autoregressive large language model serving is governed by two radically divergent computational regimes, creating an insurmountable structural tension within traditional monolithic GPU deployments: ...

Tech Radar: SGLang EAGLE-2: Speculative Decoding & Tree-Attention Latency Acceleration Answer-First: SGLang’s native EAGLE-2 implementation establishes the 2026 state-of-the-art for autoregressive latency acceleration, overcoming the memory-bandwidth wall (<1 FLOP/byte) by combining lightweight multi-layer feature extrapolation with dynamic tree-attention verification. On 8x NVIDIA H100 GPU clusters, EAGLE-2 achieves an empirical 2.5x to 3.5x generation speedup and reduces Time-Per-Output-Token from 35.7ms to 11.6ms on Llama-3-70B, with mathematically zero output distribution degradation. name: "SGLang EAGLE-2 Speculative Decoding" ring: "Adopt" quadrant: "AI Infrastructure & Large Language Models" rationale: "Overcomes autoregressive memory bandwidth saturation by verifying dynamic candidate trees in a single forward pass, delivering 3x generation speedup with zero quality loss." adr_link: "/radar/2026-09/sglang-eagle-2-speculative-decoding/" justification: "Empirically verified across Llama-3-70B and DeepSeek-Coder-33B on 8x NVIDIA H100 SXM5; production-ready in SGLang runtime with under 1.5GB VRAM overhead." 1. The Autoregressive Bottleneck & Evolution of Speculative Inference Large language model inference is characterized by two distinct computational phases: the Prefill Phase and the Decode Phase. During Prefill, prompt tokens are processed concurrently in compute-dense matrix multiplications, fully saturating modern GPU Tensor Cores. However, during the sequential Decode Phase, generation is strictly memory-bandwidth bound: ...

Tech Radar: DeepSeek-V3 Multi-Head Latent Attention (MLA) Architecture & KV Cache Compression Answer-First: DeepSeek-V3’s Multi-Head Latent Attention (MLA) fundamentally addresses the memory bandwidth and capacity bottlenecks in large language model inference. By projecting Keys and Values into a low-rank latent compressed space (latent space d_c = 512) during KV cache generation, MLA achieves a 75% reduction in runtime VRAM consumption compared to traditional Multi-Head Attention (MHA) and Grouped-Query Attention (GQA), while simultaneously retaining the high expressive representational capacity of full attention matrices through Decoupled Rotary Position Embedding (RoPE). ...

Tech Radar: vLLM Context-Aware Routing & MLA KV Cache Architecture Answer-First: Multi-Head Latent Attention (MLA) combined with Context-Aware Prefix Routing in vLLM resolves the GPU VRAM memory wall in autonomous multi-turn agent execution loops. Compressing Key-Value caches into low-dimensional latent vectors ($d_{latent} = 512$) and routing shared-prefix tool invocations to the warm GPU worker reduces VRAM consumption by 75.8% and slashes Time-to-First-Token (TTFT) from 840ms to 165ms. 1. The VRAM Explosion in Autonomous Agent Multi-Turn Loops When scaling autonomous AI agent swarms (automated code refactorers, SQL analytics bots, customer support agents), inference pipelines execute iterative loops: $$ ext{User Prompt} \longrightarrow ext{Tool Call} \longrightarrow ext{Observation} \longrightarrow ext{Next Tool} \dots \longrightarrow ext{Final Answer}$$ ...