# Сопутствующие статьи по теме Speedup

Новостной центр HTX предлагает последние статьи и углубленный анализ по "Speedup", охватывающие рыночные тренды, новости проектов, развитие технологий и политику регулирования в криптоиндустрии.

AI Image Generation Without Training Speeds Up 1000%: Method: The Most Concise "Three-Stage Pipeline"

AI image generation is powerful but often feels slow for users. To address this without hardware dependencies, model fine-tuning, or dynamic runtime optimizations, researchers propose MrFlow (Multi-Resolution Flow Matching). This method uses a simple three-stage pipeline: first, generate a low-resolution structural draft; second, upscale it to high resolution in pixel space using a super-resolution model like Real-ESRGAN; third, add a small amount of noise and perform a single-step high-resolution refinement with the original model. This approach shifts most computations to the cheaper low-resolution stage, achieving up to 10.35x end-to-end speedup on models like Qwen-Image while maintaining image quality within ~1% of original outputs. Compared to other training-free acceleration methods, MrFlow offers better speed-quality trade-offs, works with various advanced models, and can be combined orthogonally with distillation techniques for further acceleration. The code and a ComfyUI plugin are openly available.

marsbit07/08 07:13

AI Image Generation Without Training Speeds Up 1000%: Method: The Most Concise "Three-Stage Pipeline"

marsbit07/08 07:13

Fable 5 Crafts First CUDA 'Megakernel' from Scratch, Achieves 18.7x Speedup in 2.5 Hours

AI model Fable 5 (a safety-limited version of Anthropic's Claude Mythos) has achieved a breakthrough in GPU kernel optimization. In the rigorous KernelBench-Mega benchmark—which requires fusing an entire model's compute block into a single kernel—Fable 5 autonomously wrote a highly optimized CUDA "megakernel." This kernel executes a complete Kimi-Linear W4A16 hybrid decoding task within a single GPU kernel launch, using 14 grid barriers to sequence operations. The result was a performance increase of 18.7x over the baseline on an RTX PRO 6000 GPU, significantly outpacing competitors like Claude Opus 4.8 (14.4x) and GPT-5.5 (4.34x). Notably, its performance advantage widened with longer context lengths. The model spent the majority of its 2.5-hour, 550k-token session analyzing benchmarks and theoretical limits before coding, leading to an exceptionally efficient final design. Anthropic co-founder Jack Clark described this as the beginning of a "recursive self-improvement (RSI) loop," where AI's ability to optimize its own underlying computational infrastructure could rapidly accelerate its own development cycle. This advance highlights AI's growing capability in complex, low-level engineering tasks that were previously a human stronghold.

marsbit07/07 07:36

Fable 5 Crafts First CUDA 'Megakernel' from Scratch, Achieves 18.7x Speedup in 2.5 Hours

marsbit07/07 07:36

NVIDIA's New Open-Source MoE: One Line of Import, Fine-Tuning Accelerated by 3.7x

NVIDIA has open-sourced NeMo AutoModel, a tool designed to significantly accelerate the fine-tuning of Mixture-of-Experts (MoE) large language models. By adding just one import line to existing code based on Hugging Face Transformers v5, users can achieve a 3.4x to 3.7x increase in training throughput and reduce GPU memory usage by 29% to 32% without altering their API. The key innovations include Expert Parallelism (EP) to distribute expert weights across GPUs, lowering memory pressure; DeepEP to fuse computation and communication; and TransformerEngine kernels for accelerated core operations. Benchmarks on models like Qwen3-30B-A3B show training throughput per GPU jumping from 3075 to 11340 tokens per second. The solution also enables the fine-tuning of very large models, such as the 550B parameter Nemotron 3 Ultra, which would exceed memory limits with the standard Transformers v5. Code and benchmarks are available on GitHub.

marsbit06/26 07:29

NVIDIA's New Open-Source MoE: One Line of Import, Fine-Tuning Accelerated by 3.7x

marsbit06/26 07:29

Running MoE on Mobile Phones? Meta Proposes MobileMoE, Speeding Up iPhone 16 Pro by 3.8x

Meta's MobileMoE, a mobile-optimized Mixture-of-Experts (MoE) language model architecture, enables efficient on-device large language model (LLM) inference for the first time on commercial smartphones. Designed for decoder-only Transformers, it replaces dense feed-forward layers with MoE layers. Key design choices include 8 experts with granularity g=8, top-4 routing, and a shared expert. The model undergoes a four-stage training process: pre-training, intermediate training, supervised fine-tuning, and quantization-aware training. Results show MobileMoE models, with similar memory footprint, achieve equal or higher average accuracy across 14 foundational benchmarks while using only 1/2 to 1/4 of the FLOPs compared to dense baselines. After INT4 quantization, they remain competitive. Notably, on an iPhone 16 Pro, MobileMoE-S demonstrates significant speedups: up to 3.8x faster in the prompt phase and 2.2-3.4x faster in per-token generation compared to a dense counterpart, with lower peak memory usage. While MobileMoE establishes a new Pareto frontier for on-device LLMs in accuracy-compute trade-offs, particularly excelling in code and math tasks, it currently lags behind models like Qwen3.5 2B in advanced instruction following and knowledge reasoning. Future work includes improving post-training techniques, exploring NPU deployment, and managing the runtime memory sensitivity of MoE models to varying inputs.

marsbit06/01 06:09

Running MoE on Mobile Phones? Meta Proposes MobileMoE, Speeding Up iPhone 16 Pro by 3.8x

marsbit06/01 06:09

活动图片

Популярные категорииright-arrow

Популярные тегиright-arrow

Technology Trendsright-arrow

Regulatory Policyright-arrow