Search Results for "smartgpu2.h"

ToMe (Token Merging) is a PyTorch-based optimization framework designed to significantly accelerate Vision Transformer (ViT) architectures without retraining. Developed by researchers at Facebook (Meta AI), ToMe introduces an efficient technique that merges similar tokens within transformer layers, reducing redundant computation while preserving model accuracy. This approach differs from token pruning, which removes background tokens entirely; instead, ToMe merges tokens based on feature...

...Compressed from Nemotron 3.5 Super VL using NVIDIA's Star Elastic method, it contains 74B total parameters with 9B active per token. Its hybrid Mamba-2–Transformer LatentMoE architecture integrates a C-RADIO v4-H vision encoder for processing text, images, and video. The model supports a 256K-token context window, structured tool calling, JSON output, and configurable reasoning modes. It is optimized for coding, browser automation, terminal operations, and multi-step agent workflows. BF16 weights preserve flexibility for supervised fine-tuning, reinforcement learning, distillation, and creating quantized variants. ...

...Derived from Nemotron 3.5 Super VL through NVIDIA's Star Elastic compression method, it features 74B total parameters with approximately 9B activated per token. Its hybrid Mamba-2–Transformer LatentMoE architecture incorporates a C-RADIO v4-H vision encoder for processing text, images, and video. The model supports a 256K-token context window and uses IQ2 quantization with quantization-aware distillation to reduce memory requirements while preserving reasoning capabilities. Designed for local coding assistants, browser automation, terminal operations, and tool calling, it supports configurable reasoning modes for balancing speed and accuracy. ...

...Compressed from Nemotron 3.5 Super VL using NVIDIA's Star Elastic method, it contains 74B total parameters with 9B active per token. Its hybrid Mamba-2–Transformer LatentMoE architecture integrates a C-RADIO v4-H vision encoder for understanding text, images, and video. The model supports a 256K-token context window and uses native IQ2 quantization with quantization-aware distillation to preserve accuracy. It targets everyday agentic workflows involving coding, web browsing, terminal operations, and tool calling. Configurable thinking modes balance direct responses and deeper reasoning, while LiLiCorr speculative decoding improves generation speed. ...