Markdown source

Google Photos Slides: LLM Inference at Scale Workshop

Source

Relationship To World's Fair 2026

These photos are matched to the following scheduled session(s):

Match Confidence

Photo Slides

Photo 002

photo-002.jpg

OCR text:

Parallelism techniques

Tensor parallelism Split matmul across GPUs,good for GEMMs

Data parallelism Split batch across GPus,good for attention

Pipeline parallelism Split model forward pipeline steps across GPus Prefill/decode disaggregation is, IMO,a special case

Expertparallelism Split MoE experts across GPUs,good for group GEMMs

Context parallelism Split sequence across GPUs, stillnew (to Oss)

Modal

Photo 003

photo-003.jpg

OCR text:

Using the NSight Systems/Torch Profilers (CuPTI)

13 1 37853734

16716 DP3L 100%

CUDAHWDOC3 300020-X

NCCI

00030-N

NCCL

NCOL Threads(187)

CUDAHW(000250010-N

AISTN

NCCL

00010-N

AR

NCCL

..

Modal Threds(91) 2ste Stra

Photo 004

photo-004.jpg

OCR text:

Helpful resources

Good for humans,good for agents

Aleksa Gordic's walkthrough of vllm,Modal Notebooks [1l,[2]

DeepWiki from Cognition Organized,diagrams,queryable s

Modal