Daniel Wong

Quantum software firm in S’pore gets boost in real-world testing with its own quantum computer | The Straits Times https://lnkd.in/gHV45jRb

1

Continuing my work on Quantinuum Guppy: I added an efficient array swap function for quantum computing. When quantum developers needed to swap two elements in an array (like swapping qubits at different positions), they had to use a workaround function ("mem_swap") that was less efficient than it could be. In PR #1459, I added a new "array_swap" function that compiles directly to HUGR's native swap operation. This means swapping array elements now translates to optimized low-level quantum operations instead of going through multiple steps. By providing a direct path from high-level Python code to optimized quantum operations, developers can write cleaner code that runs more efficiently on quantum hardware. As always, I appreciate the Guppy contributors at Quantinuum for their help and collaboration. Edit: I used the phrase "quantum operations" loosely, but to be precise: array_swap compiles to an optimized real-time classical operation on the quantum computer's control system — not a gate acting on qubits. #QuantumComputing #OpenSource #Guppy #Quantinuum #QuantumProgramming For more technical details, refer to PR #1459 here: https://lnkd.in/geQ53YHW

8

2 Comments

Another good collaboration with DeepGlint (and of course Jiankang Deng and Roy Miles): ViCToR: Improving Visual Comprehension via Token Reconstruction for Pretraining LMMs has now being accepted to AAAI 2026. A really cool method on pre-training of VLMs with quite good results in LLaVA1.5 and LLaVA-Next architectures. The arxiv link: https://lnkd.in/ev_cqx5T?

44

🚀 Just published a companion repository for the FuriosaAI Furiosa-LLM Quick Start guide! 📦 https://lnkd.in/gyqd6gat Furiosa-LLM is a high-performance LLM serving framework built on FuriosaAI's RNGD NPU — and this repo goes beyond the official docs with: ✅ Ready-to-run Python examples (offline batch, streaming, multi-turn chat) ✅ Tool calling with a real weather API (Open-Meteo, no key required) ✅ OpenAI-compatible server test scripts ✅ Troubleshooting guide & NPU PE vs. TP model fit matrix If you're exploring NPU-based LLM inference or evaluating FuriosaAI's RNGD for your HPC workloads, this might be a useful starting point. Feedback and contributions are welcome! 🙌 #FuriosaAI #LLM #NPU #HPC #GenerativeAI #OpenSource #Supercomputing

3

Happy to be part of this CVPR 2026 paper led by Junyi Wu! 🎉 Object identity should be independent of visibility. How can we segment and understand deformable objects in a 4D scene, beyond just reconstructing it? 🔗 https://lnkd.in/gE4gXX_x 📝 https://lnkd.in/gMpjuVcW A key observation: in 4D Gaussian Splatting, opacity (visibility) and occupancy (physical existence) are often conflated. When objects deform or get occluded, their identities drift and fragment across views and time. CIF (Consistent Instance Field) addresses this by modeling each point in space-time with two decoupled quantities: whether something physically exists there, and which instance it belongs to. Three mechanisms bring this to life: ✅ Field-Aware Splatting: renders identity as a proper probability distribution, not a byproduct of RGB opacity ✅ Instance Identity Estimation: calibrates per-Gaussian identities to reduce visibility bias across views and time ✅ Instance-Guided Resampling: redistributes Gaussian capacity from redundant backgrounds toward semantically meaningful regions Grounding identity in physical existence rather than appearance yields more coherent instance understanding across motion, deformation, and occlusion. A personal thought🤔: CIF is an optimization-based approach, but as feed-forward 3D models and world models continue to advance, I wonder if the same principle of decoupling physical existence from appearance could bring similar benefits to those settings. Always curious about what's next in 3D vision and beyond🤗. For more interesting research and ideas, check out my homepage 👉 https://lnkd.in/gjmtSzt7 #CVPR2026 #ComputerVision #GaussianSplatting #4DReconstruction #DynamicScenes #SceneUnderstanding

58