I went for Triton because it was already closely integrated into PyTorch. Taichi looks really polished in comparison, but not tried it yet!
@alexjc
-
GPU Kernel Memory Optimization Challenges in AI Development
By
–
I know what you mean. For me there was nothing between 1x and 70x, it either worked or it didn't — had to redesign it each time. Now running out of shared memory for the GPU kernel, just a tad too small to be useful!
-
Triton Library Lacks 3.0-Level Maturity: API Limitations
By
–
I don't just mean the limitations of the GPU, that's part of it but done that before. Triton itself doesn't feel like a 3.0-level library. Stuff like http://
tl.dot() doesn't work when you specify the output array as a parameter (per docs), you have to use the retval. -
Triton MLP Implementation Achieves 70x Performance Improvement
By
–
Triton is exciting to work with. I built a type of MLP with it and it's 70x faster than the Torch version… Conversely, the set of features that actually works is so razor thin you have to code-golf every single decision for hours! https://
github.com/triton-lang/tr
iton
… -
ReMoE Implementation Optimization Challenges Wall-Clock Performance
By
–
Thought the same thing and was trying to reimplement ReMoE over the past days… Achieving the same level of utilization as the default codebase, and thus beating the wall-clock record, will not be trivial.
-
Banking Practices and AI Bubble: Economic Crash Risks Ahead
By
–
Crash will be for more fundamental reasons than just an AI bubble. Many investment banking practices that lead to 2008 were suspended shortly and were resumed, the problem wasn't fixed. I think it's getting close now…
-
Copyright AI-Generated Work Legal Implications and Ethics
By
–
❝ You can't Copyright electricity, therefore it's legal for me to steal your work to compete with you. ❞
-
Data legality fallacy: composition argument in IP law
By
–
It's a basic Fallacy of Composition: this data is transferred as electricity therefore any use is legal; data is made up of numbers & stats therefore our activity is legal. It's why the P2P lawsuits favored the rightholders: you have to look at the uses and its implications!
-
Microsoft Database IP Protection and Unprotected Facts Legal Implications
By
–
Not only are protected words required to express facts, what they seem to forget is that collections of "unprotected" facts are covered by database laws. If Microsoft gets its way, then none of its own databases would be protected by IP laws?
-
German Criminal Code Liability for AI Systems Omission
By
–
Criminal Code (StGB). Multiple clauses could apply here, for example: §271 Causing False Records, and §356 Betrayal of Client. Section StGB §13 "Commission by Omission" explains one can be found guilty for failing/omitting to take action — if it turns out the same a positive act.