-
Book Overview & Buying
-
Table Of Contents
DeepSeek in Practice
By :
This chapter turned deployment from theory into a pragmatic playbook for running full-parameter DeepSeek-V3/R1 in production. You mapped the landscape (API vs. managed vs. DIY) with a control/cost/complexity lens, learned how to size hardware (VRAM, KV cache, TTFT/throughput), and chose inference engines deliberately (vLLM/TGI/TensorRT-LLM) with quantization options (FP8, GPTQ, and AWQ) to fit performance and budget. Three concrete paths anchored the guidance: local dev proxies (Ollama) for iteration, managed services (Bedrock/Vertex/Azure/Hugging Face) for enterprise reliability, and DIY (vLLM on GPUs/Kubernetes) for maximum control and economics at scale.
You then operationalized the service: monitor tokens/latency/GPU and KV pressure, scale on active generation with continuous batching and sticky routing, cache shared prefixes, and enforce token-based budgets with scale-to-zero and budget alerts. Security moved beyond network basics to prompt injection defenses and...