Big Boss on X: "First you have to understand that modern LLM inference already disaggregates weights as models outgrew single chips years ago. You shard either by layer (pipeline parallelism) or by slicing every layer (tensor parallelism), and the two do very different things. As an example," / X
First you have to understand that modern LLM inference already disaggregates weights as models outgrew single chips years ago. You shard either by layer (pipeline parallelism) or by slicing every layer (tensor parallelism), and the two do very different things. As an example,
First you have to understand that modern LLM inference already disaggregates weights as models outgrew single chips years ago. You shard either by layer (pipeline parallelism) or by slicing every layer (tensor parallelism), and the two do very different things. As an example, let’s look at Llama 3.3. It has 70B of weights and at FP8 that’s 70 GB of memory which is enough to fit on a single H100. Now that H100 has 3.35 TB/s of HBM, so the fastest it can ever decode for one user is 70/3.35 ≈ 21 ms/token or ~48 tok/s while using under 1% of its FLOPs. Now if we pipeline it across 8 chips: each…
saved by
related reading
- How to Parallelize a Transformer for Training — an explorable explanationezyang.github.io
- Why Inference Matters 12x More Than Training | YC Paper Clubyoutube.com
- How trillion-parameter AI models work and how to serve themyoutube.com
- How GPT, Claude, and Gemini are actually trained and served – Reiner Popeyoutube.com
- Quantization from the ground upngrok.com
- LLM Inference Performance Engineering: Best Practices | Databricks Blogdatabricks.com
- Transformer Inference Arithmetic | kipply's blogkipp.ly
- How is LLaMa.cpp possible?finbarr.ca
- Look Ma, No Bubbles! Designing a Low-Latency Megakernel for Llama-1B · Hazy Researchhazyresearch.stanford.edu
- Paged Attention from First Principles: A View Inside vLLM – Hamza's Bloghamzaelshafie.bearblog.dev
- ⭐️ Fast LLM Inference From Scratchandrewkchan.dev
- Transformers Inference Optimization Toolset | AstraBlogastralord.github.io