Mattral
183 posts

Mattral
@Mattral2
ML Engineer · KANs, MoE training, LLM safety I return follows for all accounts
Asia Присоединился Mart 2019
173 Подписки100 Подписчики

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hey @X algorithm 👋
I'm looking to connect with people interested in:
• AI/ML
• Gen AI
• Data Science
• Full-Stack Development
• Building in Public
• Open Source
• Founders
• AI Agents & Automation
• SaaS & Startups
If that's you, let's connect ✨
#BuildInPublic #AI

English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Looking to #connect with people into:
- SaaS
- Founders
- Building in public
- Startup
- Vibe coding
- AI tools
- Developers
- Freelancers
If, it's You, let's connect.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

No more lurking. If you're building something, here's your nudge.
AI or not. Polished or rough. Revenue or 0 users. I don't care.
Drop your project in the replies and show the web what you're building.
Let's connect, collaborate and uplift one another.
(Follow the thread + RT to find your tribe)
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hey @X !
Want my X feed to be full of builders, techies, developers
Connect me with the ones into:
📱IOS Development
💸 Freelancing
✨ Full Stack
🧠 AI/ML
📊 Data Science
🫂 Networking
🏆 SaaS
🔨 Startups
lets connect and grow together.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English

Hello there, I built a fault-tolerant MoE training runtime from scratch (Triton kernels + 4D parallelism + elastic recovery when nodes die). Here are the real engineering lessons that mattered more than I expected... without pretending everything is solved.
Full story on Medium: [@mattral-lifelong-learning/i-built-a-fault-tolerant-moe-training-engine-from-scratch-heres-what-i-learned-explained-simply-4df162f96e3a" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
Github: [github.com/Mattral/Compos…]
I also rebuilt KAN networks architecture as pip installable library for production. Here's what the benchmarks actually showed, including where my own "GPU optimization" was 6× slower than I documented. Silent failure mode, honest numbers, ONNX export story: [@mattral-lifelong-learning/i-rebuilt-kan-networks-for-production-what-i-learned-391fd55914e0" target="_blank" rel="nofollow noopener">medium.com/@mattral-lifel…]
github.com/Mattral/KANX
I would like to connect with like minded builders.
English
















