Topic hub
AI Development Tools
Tools, frameworks, and best practices for building production AI applications.
All posts · 52
AI DEVELOPMENT TOOLS01
Failed to Initialize NVML: Unknown Error on GPU NodesToolkit v1.18.0 moved the NVIDIA Container Runtime off legacy mode on 2025-10-21, but the hook behind docker run --gpus still resolves to legacy at v1.20.0.
AI DEVELOPMENT TOOLS02
Mutation Testing for AI-Generated Tests: The CI GateOne four-line function, three test suites, all at 100 percent line and branch coverage, with mutation scores of 0, 50 and 100 percent. Here is the gate.
AI DEVELOPMENT TOOLS03
Claude Code Local Model: What Actually Breaks in 2026Read at pinned versions, three local Anthropic-format servers agree on only 3 of 12 request rows. On vLLM the losses come back as a 200, not an error.
AI DEVELOPMENT TOOLS04
MIG vs MPS vs Time-Slicing: Sharing One GPU for vLLMTime-slicing caps no memory, MPS shares a fault domain, and vLLM sizes itself against the whole card. The decision path across all three mechanisms.
AI DEVELOPMENT TOOLS05
Transformers Does Not Recognize This Architecture in vLLMvLLM 0.27.1 consults a 64-entry config registry of its own before Transformers runs. Seven startup errors look alike and only one is fixed by a pip upgrade.
AI DEVELOPMENT TOOLS06
OpenTelemetry GenAI Semantic Conventions: 0 of 63 Are StableAll 63 gen_ai attribute keys still read stability: development. Here is the deprecation map, the package flip that blanks dashboards, and what to pin.
AI DEVELOPMENT TOOLS07
Firecracker vs libkrun: Boot Time, Memory, and SizingThe 125ms boot number is the VMM's, not yours. Where the other 200ms goes, what the 5 MiB overhead excludes, and how to size a host that runs 200 microVMs.
AI DEVELOPMENT TOOLS08
LangChain vs Mastra: TypeScript Agent Frameworks in 2026A public head-to-head put LangChain 25-45% faster on the same pipeline with Mastra burning 1.5-2.5x the tokens. The reason is config, and it changes the answer.
AI DEVELOPMENT TOOLS09
Claude Code vs OpenCode: Why One Burns 4.7x More TokensClaude Code sends about 33k tokens before your prompt arrives; OpenCode sends about 7k. Where the other 26k goes, and what it costs at scale.