[GH-ISSUE #747] ai frameworks feedback (round 2) #5525

New Issue

GiteaMirror · 2026-04-21T21:28:41-05:00

GiteaMirror commented

2026-04-21 21:28:41 -05:00

Originally created by @18jeffreyma on GitHub (Mar 13, 2025).
Original GitHub issue: https://github.com/harvard-edge/cs249r_book/issues/747

https://mlsysbook.ai/contents/core/frameworks/frameworks.html#system-level-implications, important to call out here in the 4th paragraph: neural net diagram shows an architecture, but this might in practice map to two computational graphs (forward graph + backward graph)
really good static + dynamic graph diagrams
https://mlsysbook.ai/contents/core/frameworks/frameworks.html#device-placement maybe a shout out here as to why JAX/TF can more easily determine sharding versus PyTorch might require more manual parallelism definition
this forward mode and reverse mode is written amazingly, huge shout out
For Memory implications (https://mlsysbook.ai/contents/core/frameworks/frameworks.html#memory-implications) section I ran into PipeMare paper (https://arxiv.org/pdf/1910.05124) which has a nice Figure 1 that shows the memory hungriness of autodiff
https://mlsysbook.ai/contents/core/frameworks/frameworks.html#memory-implications, i wonder if a shout to https://pytorch.org/docs/stable/distributed.tensor.html or the XLA equivalent is worth here (i.e. how ml frameworks might show tensors globally, but their actual layout is managed by the framework). additionally some shout out to sparsity here might be good (how the framework tensor API may differ from its under the hood representation)
https://mlsysbook.ai/contents/core/frameworks/frameworks.html#development-tools and https://mlsysbook.ai/contents/core/frameworks/frameworks.html#extensions-and-plugins could use some links to pytorch profling or tensorboard or example development tools.
Sections in 7.5 i think feel a bit general for AI frameworks (either some diagrams or some urls for specific system integration problems)?
7.8.4 maybe more specificity and an example on why youd choose between a framework based on some constraint? like scalbility for TFLite if you need to continuously train versus something else?

Originally created by @18jeffreyma on GitHub (Mar 13, 2025). Original GitHub issue: https://github.com/harvard-edge/cs249r_book/issues/747 - [ ] https://mlsysbook.ai/contents/core/frameworks/frameworks.html#system-level-implications, important to call out here in the 4th paragraph: neural net diagram shows an architecture, but this might in practice map to two computational graphs (forward graph + backward graph) - [ ] really good static + dynamic graph diagrams - [ ] https://mlsysbook.ai/contents/core/frameworks/frameworks.html#device-placement maybe a shout out here as to why JAX/TF can more easily determine sharding versus PyTorch might require more manual parallelism definition - [ ] this forward mode and reverse mode is written amazingly, huge shout out - [ ] For Memory implications (https://mlsysbook.ai/contents/core/frameworks/frameworks.html#memory-implications) section I ran into PipeMare paper (https://arxiv.org/pdf/1910.05124) which has a nice Figure 1 that shows the memory hungriness of autodiff - [ ] https://mlsysbook.ai/contents/core/frameworks/frameworks.html#memory-implications, i wonder if a shout to https://pytorch.org/docs/stable/distributed.tensor.html or the XLA equivalent is worth here (i.e. how ml frameworks might show tensors globally, but their actual layout is managed by the framework). additionally some shout out to sparsity here might be good (how the framework tensor API may differ from its under the hood representation) - [ ] https://mlsysbook.ai/contents/core/frameworks/frameworks.html#development-tools and https://mlsysbook.ai/contents/core/frameworks/frameworks.html#extensions-and-plugins could use some links to pytorch profling or tensorboard or example development tools. - [ ] Sections in 7.5 i think feel a bit general for AI frameworks (either some diagrams or some urls for specific system integration problems)? - [ ] 7.8.4 maybe more specificity and an example on why youd choose between a framework based on some constraint? like scalbility for TFLite if you need to continuously train versus something else?

GiteaMirror added the area: book type: improvement labels 2026-04-21 21:28:41 -05:00

GiteaMirror closed this issue

2026-04-21 21:28:42 -05:00

GiteaMirror commented

2026-04-21 21:28:43 -05:00

@profvjreddi commented on GitHub (Aug 23, 2025):

Resolved - AI frameworks feedback incorporated

@profvjreddi commented on GitHub (Aug 23, 2025): Resolved - AI frameworks feedback incorporated

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: github-starred/cs249r_book#5525