Swayam Singh @ ICML'25
@swayaminsync
Followers
1K
Following
4K
Media
590
Statuses
3K
ML Research @MSFTResearch | Core Maintainer @numpy_team (QuadDType)
living in the moment
Joined April 2021
Strong version of you is dealing with all the inner demons silently, keeping all the chaos contained within you, hidden from the outside world. It'll get exhaustive sometimes and I am proud of you. Don't give up.
0
0
13
So I sat down, attached all the codebase I had, kept prompting and prompting and prompting (in between switched to the Linus's persona from his prime) 100x the productivity.AND Closed the machine with."git restore ."
1
0
3
Feel free to reach out for any feedback and discussions. 📎.📎.📎
0
0
3
Early benchmarks (x86-64 & ARM): .✅ ~21× faster dot .✅ ~77× faster GEMV .✅ ~3× faster GEMM .QBLAS will integrate with numpy_quaddtype soon, providing a user-friendly API to leverage quad-precision with speed.
1
0
3
So I'm building QuadBLAS, a library for optimized LA ops for quad-precision, utilizing good cache blocking, SIMD, and multithreading to speed things up. GitHub:
1
0
4
Quad-precision floats are 16 bytes, 2x the size of float64 and 4x the size of float32 .That means: .1️⃣More memory = more bandwidth pressure 2️⃣Slower cache access .3️⃣When __float128 is missing or no LD 128 bit support, the implementation breaks to a struct of 2 int64 types.
1
0
4
A thread 🧵 .TL;DR: We’re working on making NumPy’s cross-platform 128-bit float operations go brrr. 🔥. So why are quad-precision (128-bit) linear algebra ops so slow and how we’re fixing it?
1
0
4
RT @LocalAI_API: 🔥 New model alert! 🔥. Microsoft NextCoder-32B is now available in LocalAI! 🚀 This code-editing LLM boasts impressive perfo….
0
1
0
RT @HiSohan: 📄 Paper 20/42: "NextCoder: Robust Adaptation of Code LMs to Diverse Code Edits".🇮🇳 Tushar Aggarwal (Microsoft). LinkedIn: h….
0
2
0
I planned a very cool thing for this PyCon India, but unfortunately some unplanned conflicts are happening. I'm sorry for if you were waiting for it, Next time then :)
0
0
12
This is really nice but no technical reports of either models and SWE-Bench as the only benchmark looks a bit suspicious. Don't get me wrong, Following the Mistral work a long way but it would be more acceptable if you guys open-up on some details and development of Devstral.
Introducing Devstral Small and Medium 2507! This latest update offers improved performance and cost efficiency, perfectly suited for coding agents and software engineering tasks.
0
0
8
We’ll be presenting this at ICML 2025 at the Vancouver Convention Centre, Canada. Drop by if you’d like to chat or dive deeper, we’d love to connect!.
0
0
4
Our recent work, "NextCoder" is now public. Dropping:.1️⃣ Models with strong code-editing capabilities (7B, 14B, 32B) .2️⃣ Complete training dataset .3️⃣ A clever training algorithm: Selective Knowledge Transfer (SeleKT). ✅ This is just the first phase, with more to come soon.
Excited to share the NextCoder family of SLMs with strong code-editing abilities. Finetuned with Selective Knowledge Transfer (SeleKT) and GitHub/synthetic data. #ICML. GitHub: Azure AI Foundry: HF:
3
3
29
This was literally my reply when they asked me to put the theorem's proof somewhere in the Appendix.
0
0
7
Sir Aditya Gopalan is a humble, intelligent person. We met at the MSR AI Summit and discussed off-policy RL scenarios, methods that work but lack mathematical backing, and efficient sparse model training regimes. He is a genuinely insightful, welcoming individual.
"DPO can give you a policy that is worse than what you started with".@today_itself reveals how the theoretical backing behind one of the most ubiquitous alignment methods breaks down for real-life LLMs, causing unpredictable alignment failures. He then shows how to fix it.
0
0
6
