LLM Inference Optimization Explained — From 8 Tokens/sec to 50+

Why does a 70B language model crawl at 8 tokens per second on one setup, then feel instant on another? The difference is inference optimization: KV cache management, PagedAttention, continuous batching, quantization, speculative decoding, model parallelism, and production serving frameworks like vLLM and TensorRT-LLM.

In this AI Deep Dive, we break down the systems engineering behind fast LLM serving — the techniques that turn expensive, slow autoregressive generation into real-time user experiences.

Timestamps:
0:00 — Hook: The Inference Optimization Gap
0:45 — KV Cache: The Bottleneck Behind LLM Serving
2:00 — PagedAttention: Virtual Memory for Attention
3:10 — Continuous Batching: Keeping GPUs Full
4:20 — Quantization: Shrinking Model Weights
5:30 — Speculative Decoding: Parallelizing Token Generation
6:25 — Model Parallelism: Splitting Giant Models Across GPUs
7:20 — Serving Frameworks: vLLM vs TensorRT-LLM
8:20 — Bottom Line: Good Engineering Applied to the Right Bottlenecks

Subscribe to AI Deep Dive for more AI infrastructure explainers: https://www.youtube.com/@AIdeepdive-x8i

#LLM #InferenceOptimization #AIInfrastructure #vLLM #TensorRTLLM #PagedAttention #Quantization #SpeculativeDecoding #LocalLLM #MachineLearning

Видео LLM Inference Optimization Explained — From 8 Tokens/sec to 50+ канала AI deepdive