@RisingSayak: Published my first kernel to go the last mile to optimize LTX-2.3 from @Lightricks! torch.compile + cuDNN attn already …
Summary
Published a custom kernel to further optimize LTX-2.3 from Lightricks, achieving 1.52x speedup on GB10, building upon previous torch.compile and cuDNN attention optimizations.
View Cached Full Text
Cached at: 06/13/26, 02:27 PM
Published my first kernel to go the last mile to optimize LTX-2.3 from @Lightricks!
torch.compile + cuDNN attn already gave a 1.42x boost. W/ the custom kernel added, I got 1.52x on a GB10 🔥
This was my systematic exploration of a simple agentic kernel dev workflow.
More 👇 https://t.co/u4iDpzSir0
Similar Articles
Multi-perspective Imbalance-Conscious 6G Beamforming Optimization and Performance
This paper presents a systematic machine learning study of 6G-IoT beamforming optimization, comparing network, environmental, device, and vision feature groups for predictive power, and applying clustering methods to enhance performance.
Momentum as Residual-Driven Multiplier Correction for Deep Learning Optimization
This paper introduces an ADMM-inspired momentum framework (AIM) and a new optimizer RADAR, offering a theoretical interpretation of momentum as a residual-driven multiplier correction, with experiments showing consistent gains over strong adaptive baselines.
Training Under Challenge: Executable Certificates and Challenge-Closed Optimality for Neural Networks
Introduces 'Training Under Challenge', an executable-certificate framework that uses architecture-valid procedures to construct alternative models and estimate the empirical global-optimality gap of neural network checkpoints, with theoretical guarantees and experiments on ResNet-18 distillation and quantized denoising.
Predicting When Random Low-Dimensional Reparameterizations Train Neural Networks
This paper analyzes when random low-dimensional reparameterizations can train neural networks, deriving an orientation-resolved master formula for the random-slice residual and introducing RaMaN, a scalable framework that predicts required latent dimensions while dramatically reducing memory costs.
$\varepsilon$-MemEvo: Adaptive Cross-Task Memory Transfer for LLM Program Evolution
The paper introduces ε-MemEvo, a framework for cross-task knowledge transfer in LLM-based program evolution, storing tactic memories as natural-language summaries and using an adaptive injection gate. It reports consistent gains on 8 optimization benchmarks with less than 1% computational overhead.