Giuseppe
@giusfra.bsky.social
ML Software Engineer @ AMD. Working on open source quantization.
Hey, thanks for reaching out.
Feel free to just open an issue on the repo with all your questions and examples if necessary. We'll do our best to reply as soon as we can!
Feel free to just open an issue on the repo with all your questions and examples if necessary. We'll do our best to reply as soon as we can!
April 9, 2025 at 1:38 PM
Hey, thanks for reaching out.
Feel free to just open an issue on the repo with all your questions and examples if necessary. We'll do our best to reply as soon as we can!
Feel free to just open an issue on the repo with all your questions and examples if necessary. We'll do our best to reply as soon as we can!
Some open questions remain, like the potential over-fitting that we noticed across our experiments.
If you are interested, check out our work at:
arxiv.org/abs/2503.17513
And if you have questions, reach out on Brevitas!
If you are interested, check out our work at:
arxiv.org/abs/2503.17513
And if you have questions, reach out on Brevitas!
Improving Quantization with Post-Training Model Expansion
The size of a model has been a strong predictor of its quality, as well as its cost. As such, the trade-off between model cost and quality has been well-studied. Post-training optimizations like quant...
arxiv.org
March 27, 2025 at 9:59 AM
Some open questions remain, like the potential over-fitting that we noticed across our experiments.
If you are interested, check out our work at:
arxiv.org/abs/2503.17513
And if you have questions, reach out on Brevitas!
If you are interested, check out our work at:
arxiv.org/abs/2503.17513
And if you have questions, reach out on Brevitas!
By combining intuitions coming from recent quantization papers, we show how slightly increasing the number of parameters in the layers that are more sensitive to quantization can yield better perplexity and zero-shot performance with SOTA algorithms like QuaRot and SpinQuant.
March 27, 2025 at 9:59 AM
By combining intuitions coming from recent quantization papers, we show how slightly increasing the number of parameters in the layers that are more sensitive to quantization can yield better perplexity and zero-shot performance with SOTA algorithms like QuaRot and SpinQuant.
The talk will be December 10th, at 1 PM in East Exhibit Hall A, as part of a larger AMD workshop, running from 12-1:30 PM.
If you can't make it, you can find Ian at booth 142 :)
If you can't make it, you can find Ian at booth 142 :)
December 9, 2024 at 6:57 PM
The talk will be December 10th, at 1 PM in East Exhibit Hall A, as part of a larger AMD workshop, running from 12-1:30 PM.
If you can't make it, you can find Ian at booth 142 :)
If you can't make it, you can find Ian at booth 142 :)