model optimization notes #344

New Issue

GiteaMirror · 2026-03-22T15:37:22-05:00

GiteaMirror commented

2026-03-22 15:37:22 -05:00

Originally created by @18jeffreyma on GitHub (Mar 6, 2025).

https://mlsysbook.ai/contents/core/hw_acceleration/hw_acceleration.html#mixed-precision-computing-and-hardware-evolution, really great taxonomy to introduce
https://mlsysbook.ai/contents/core/optimizations/optimizations.html#dynamic-pruning maybe a citation here to a representative paper
https://mlsysbook.ai/contents/core/optimizations/optimizations.html#tbl-pruning bold the left hand labels for readability
https://mlsysbook.ai/contents/core/optimizations/optimizations.html#applications-of-td probably just expand TD to tensor decomposition for easier readability
https://mlsysbook.ai/contents/core/optimizations/optimizations.html#fig-quantized-energy figure wip
https://mlsysbook.ai/contents/core/optimizations/optimizations.html#early-exit-architectures diagram of an example early exit
https://mlsysbook.ai/contents/core/optimizations/optimizations.html#gate-based-conditional-computation moe diagram insert
https://mlsysbook.ai/contents/core/optimizations/optimizations.html#hardware-execution-inefficiencies maybe a link here to XLA compilation and static shapes documentation
the sparsity section could use a reference to https://arxiv.org/pdf/2211.15841 (essentially showing how sparsity can still be arranged into dense ish and hardware friendly copmutation)

otherwise rlly good organization, these are only minor nits imo

Originally created by @18jeffreyma on GitHub (Mar 6, 2025). - [ ] https://mlsysbook.ai/contents/core/hw_acceleration/hw_acceleration.html#mixed-precision-computing-and-hardware-evolution, really great taxonomy to introduce - [x] https://mlsysbook.ai/contents/core/optimizations/optimizations.html#dynamic-pruning maybe a citation here to a representative paper - [x] https://mlsysbook.ai/contents/core/optimizations/optimizations.html#tbl-pruning bold the left hand labels for readability - [x] https://mlsysbook.ai/contents/core/optimizations/optimizations.html#applications-of-td probably just expand TD to tensor decomposition for easier readability - [ ] https://mlsysbook.ai/contents/core/optimizations/optimizations.html#fig-quantized-energy figure wip - [x] https://mlsysbook.ai/contents/core/optimizations/optimizations.html#early-exit-architectures diagram of an example early exit - [x] https://mlsysbook.ai/contents/core/optimizations/optimizations.html#gate-based-conditional-computation moe diagram insert - [x] https://mlsysbook.ai/contents/core/optimizations/optimizations.html#hardware-execution-inefficiencies maybe a link here to XLA compilation and static shapes documentation - [x] the sparsity section could use a reference to https://arxiv.org/pdf/2211.15841 (essentially showing how sparsity can still be arranged into dense ish and hardware friendly copmutation) otherwise rlly good organization, these are only minor nits imo

GiteaMirror added the area: book label 2026-03-22 15:37:22 -05:00

GiteaMirror closed this issue

2026-03-22 15:37:22 -05:00

GiteaMirror commented

2026-03-22 15:37:22 -05:00

@18jeffreyma commented on GitHub (Mar 6, 2025):

what compressed formats in sparsity?

@18jeffreyma commented on GitHub (Mar 6, 2025): what compressed formats in sparsity?

GiteaMirror commented

2026-03-22 15:37:23 -05:00

@18jeffreyma commented on GitHub (Mar 6, 2025):

take a look at spatial computing (and joel emer)