IDfy

torch.compile, with just one decorator, outperformed naive triton kernels! I benchmarked four implementations of GeLU and softmax on an NVIDIA L4 (FP32) on synthetic random tensors: **GeLU** - Naive PyTorch: 101.6 ms - F.gelu: 9.28 ms - torch.compile: 8.93 ms - Naive Triton: 9.31 ms **Softmax** - Naive PyTorch: 35.79 ms - F.softmax: 9.13 ms - torch.compile: 9.20 ms - Naive Triton: 9.28 ms In every case, torch.compile either won or tied. I acknowledge that my Triton kernels were naive, with a fixed BLOCK_SIZE, no autotuning, no num_warps tuning, and no swizzling. A well-optimized Triton kernel could certainly outperform torch.compile. This highlights the challenge: achieving better performance with Triton requires significant effort, while torch.compile is as simple as applying one decorator. Initially, I believed that "Triton = control = faster." However, the data indicates that for patterns visible to the compiler (like chains of elementwise operations and simple reductions), the compiler is already operating at the HBM lower bound, leaving little room for hand-tuning. Triton still excels in areas such as cross-step state, irregular access, and fused multi-op reductions like attention patterns that the compiler cannot detect. Conversely, torch.compile is sufficient for most everyday elementwise and simple-reduction tasks in model code.

Some journeys are less about clarity, and more about figuring things out along the way. What started as a simple idea — “𝘸𝘩𝘺 𝘴𝘩𝘰𝘶𝘭𝘥 𝘒𝘠𝘊 𝘣𝘦 𝘳𝘦𝘱𝘦𝘢𝘵𝘦𝘥 𝘢𝘨𝘢𝘪𝘯 𝘢𝘯𝘥 𝘢𝘨𝘢𝘪𝘯?” — turned into one of the most intense, chaotic, and rewarding rides of my life. 💡 𝐊𝐚𝐯𝐚𝐜𝐡 – 𝐕𝐞𝐫𝐢𝐟𝐲 𝐎𝐧𝐜𝐞, 𝐀𝐜𝐜𝐞𝐬𝐬 𝐀𝐥𝐰𝐚𝐲𝐬 It is not just a solution we built for a hackathon. It’s something we truly believe can change how identity works in the financial ecosystem — giving users control, reducing friction, and building trust by design. 🏁 𝐅𝐫𝐨𝐦 𝟓𝟎𝟎+ 𝐚𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬 𝐚𝐜𝐫𝐨𝐬𝐬 𝟏𝟓 𝐜𝐨𝐮𝐧𝐭𝐫𝐢𝐞𝐬 (𝐏𝐡𝐚𝐬𝐞 𝟏) — 𝐡𝐞𝐫𝐞’𝐬 𝐡𝐨𝐰 𝐨𝐮𝐫 𝐣𝐨𝐮𝐫𝐧𝐞𝐲 𝐮𝐧𝐟𝐨𝐥𝐝𝐞𝐝 : ✨ 𝐓𝐨𝐩 𝟒𝟗 — 𝐏𝐡𝐚𝐬𝐞 𝟐 🚀 𝐓𝐨𝐩 𝟐𝟏 — 𝐏𝐡𝐚𝐬𝐞 𝟑 🔥 𝐓𝐨𝐩 𝟖 — 𝐓𝐨𝐤𝐞𝐧𝐢𝐬𝐞𝐝 𝐊𝐘𝐂 𝐭𝐫𝐚𝐜𝐤 🏆 𝐅𝐢𝐧𝐢𝐬𝐡𝐞𝐝 𝟐𝐧𝐝 — and proud of every step it took to get here 🧩 𝐖𝐡𝐚𝐭 𝐰𝐞 𝐰𝐞𝐫𝐞 𝐭𝐫𝐲𝐢𝐧𝐠 𝐭𝐨 𝐟𝐢𝐱 KYC today is simple… but painfully repetitive. Every new bank, wallet, or financial app asks you to start from scratch. That means: • Users deal with constant friction • Companies spend more time and money on onboarding • And the ecosystem carries unnecessary risk 💡 𝐖𝐡𝐚𝐭 𝐰𝐞 𝐛𝐮𝐢𝐥𝐭 𝐢𝐧𝐬𝐭𝐞𝐚𝐝 We asked — 𝘸𝘩𝘢𝘵 𝘪𝘧 𝘒𝘠𝘊 𝘸𝘰𝘳𝘬𝘦𝘥 𝘰𝘯𝘤𝘦, 𝘢𝘯𝘥 𝘸𝘰𝘳𝘬𝘦𝘥 𝘦𝘷𝘦𝘳𝘺𝘸𝘩𝘦𝘳𝘦? With Kavach: • Users verify themselves just once • Institutions can reuse that verification securely • And data is shared selectively — only what’s needed, nothing more The outcome? A smoother onboarding experience, stronger privacy, and lower costs across the board. 🎢 𝐖𝐡𝐚𝐭 𝐢𝐭 𝐫𝐞𝐚𝐥𝐥𝐲 𝐭𝐨𝐨𝐤 This win feels like more than just a trophy. It’s a massive vote of confidence in the future of digital identity. We didn’t just build it. We debated it. We tore it down. We rebuilt it — over and over again. 🌍 One of the most valuable parts of this journey was getting the chance to pitch our vision directly to banks, VCs, and financial institutions — the kind of exposure that truly shapes how you build. A massive shoutout to Ashish Sahni our CTO at IDfy — thank you for being in our corner through every single phase of this competition. To Rishabh Jindal , Vivek Kalyanarangan , and Siva Kiran — thank you for stepping in with the advice and expertise that helped us sharpen our vision. And to my incredible teammates — Karthik Gogisetty , Ammar B. , and Om Uskaikar , we really lived this together. From the 2 AM design fixes to the final pitch — we pulled it off. 🚀 𝐖𝐡𝐚𝐭’𝐬 𝐧𝐞𝐱𝐭 RBI Harbinger was just the beginning. We’re more excited than ever to bring Kavach closer to real-world adoption. #RBI #Harbinger2025 #Fintech #Innovation #DigitalIdentity #Kavach #Teamwork #IDfy #KYC

IDfy is conducting a webinar - The Issuer’s Trust Paradox: Balancing AI-Led Growth in the Era of Deepfake & Synthetic Fraud AI is helping issuers grow faster than ever. Fraudsters are learning just as quickly. Join IDfy, CRED, and Unity Small Finance Bank for a conversation on how issuers are navigating the rise of deepfakes, synthetic identities, and AI-driven fraud while continuing to scale growth. Register here: https://lnkd.in/dNB5Pr3T Details below: 📅22 April ⏰ 2:00 – 3:00 PM IST 📍 Zoom Webinar

When we started building our Privacy suite of products the DPDP act was still just a bill being tabled. Privy by IDfy has been the pioneer in this space, building for the DPDP act when people still thought this law will never get passed :) today we are happy to have been at the core of some of the largest implementations in this space. I am so proud of the IDfy team which has been at it for the last 4 years in building the most compehensive suite for DPDP compliance. This is not just about consent but about how data is secured and how data is contextualised in a fiduciary's environment. DPDP has to be thought about holistically, thinking about it as a singular control center highlight risk, compliance and audit from a single place. Its also about thinking about this as a people process and technology problem. DPDP compliance can be a trust/revenue enabler for fiduciaries, people should be thinking about it as a core function rather than a cost function. The next couple of years are going to be exciting in this space. We are here to make India think deeply about privacy :)

A few days ago, I wrote about how Google’s TurboQuant paper validated what many of us in the quantization trenches have known for a while: compressing the KV cache is infrastructure-critical. I stand by the math. The approach of decomposing the problem - using rotation to redistribute information and patching residuals - is structurally solid. But when a research paper moves markets and wipes billions off memory chip stocks in a single news cycle, it’s our responsibility to look past the abstract and evaluate the claims objectively. Once you dig down and peel the layers, this serves as an excellent case study in the difference between steady engineering progress and corporate PR. Here is what you need to know to separate the signal from the marketing noise: 1. The "8x Speedup" Baseline Illusion: The headline numbers are staggering - until you look at the baseline. The 8x speedup and 6x memory reduction claims are measured against a 32-bit unquantized standard. Nobody in modern LLM production runs 32-bit inference. It's like comparing yourself in a race against a toddler. 2. The Scope of the Memory Savings: Headlines implied massive reductions in total hardware requirements. In reality, TurboQuant applies exclusively to the KV cache. While optimizing the cache is absolutely vital for scaling long-context windows, it does not shrink the model weights. The baseline VRAM required to load the model onto your GPUs remains exactly the same. 3. The CPU vs. A100 Benchmark: This is where the scientific rigor gets a bit muddy. A highly similar technique incorporating rotation, RaBitQ, was published a year earlier. In the initial evaluations, the authors compared TurboQuant running on an A100 GPU against RaBitQ ported to Python and running single-threaded on a CPU. That isn’t an apples-to-apples baseline; that’s an architectural mismatch. Does this mean the paper is bad? Not at all. The underlying mechanism remains entirely sound. Using random rotation to remove high-dimensional structure so that simple scalar quantization can shine is the right way to solve this. The field is catching up to what the math has been saying. But when we study AI efficiency - we have to protect the integrity of the math from the hype cycle. TurboQuant is a solid, incremental step forward in a crucial domain. It is an evolution, not an earthquake. Read the appendices, check the baselines, and stay skeptical. Check the comments for a follow up read by the paper by the RaBitQ author.