Search Results for "gpu speed"

...Unlike many prior approaches, it does not require camera intrinsics or extra metadata, yet still outputs metric depth suitable for downstream 3D tasks. Apple highlights both accuracy and speed: the model can synthesize a ~2.25-megapixel depth map in around 0.3 seconds on a standard GPU, enabling near real-time applications. The repo and research page emphasize boundary fidelity and crisp geometry, addressing a common weakness in monocular depth where edges can blur. Community integrations (e.g., inference wrappers and UI nodes) have sprung up around the model, reflecting practical interest in video, AR, and generative pipelines. ...

FlashMLA is a high-performance decoding kernel library designed especially for Multi-Head Latent Attention (MLA) workloads, targeting NVIDIA Hopper GPU architectures. It provides optimized kernels for MLA decoding, including support for variable-length sequences, helping reduce latency and increase throughput in model inference systems using that attention style. The library supports both BF16 and FP16 data types, and includes a paged KV cache implementation with a block size of 64 to...

...Its flagship models—TAPIR, BootsTAPIR, and the latest TAPNext—use matching plus temporal refinement or next-token style propagation to achieve state-of-the-art accuracy and speed on TAP-Vid. RoboTAP demonstrates how TAPIR-style tracks can drive real-world robot manipulation via efficient imitation, and ships with a dataset of annotated robotics videos. The repo provides JAX and PyTorch checkpoints, Colab demos, and a real-time live demo that runs on a GPU to let you select and track points interactively.

GPT-OSS-120B is a powerful open-weight language model by OpenAI, optimized for high-level reasoning, tool use, and agentic tasks. With 117B total parameters and 5.1B active parameters, it’s designed to fit on a single H100 GPU using native MXFP4 quantization. The model supports fine-tuning, chain-of-thought reasoning, and structured outputs, making it ideal for complex workflows. It operates in OpenAI’s Harmony response format and can be deployed via Transformers, vLLM, Ollama, LM Studio, and PyTorch. Developers can control the reasoning level (low, medium, high) to balance speed and depth depending on the task. ...

...Built on the Gemma 4 26B A4B Mixture-of-Experts architecture, it has 25.2B total parameters and 3.8B active parameters, balancing capability with efficient inference. Its diffusion-based generation produces tokens in parallel 256-token blocks, enabling very high-speed output, with reported generation above 1,100 tokens per second on NVIDIA Hopper H100 in FP8. The model supports a 256K-token context window, configurable thinking mode, native function calling, structured JSON output, and multilingual inference across 35+ languages. The NVFP4 quantization reduces weights and activations from 16-bit to 4-bit, lowering disk size and GPU memory needs for vLLM deployment.

granite-timeseries-ttm-r2 is part of IBM’s TinyTimeMixers (TTM) series—compact, pre-trained models for multivariate time series forecasting. Unlike massive foundation models, TTM models are designed to be lightweight yet powerful, with only ~805K parameters, enabling high performance even on CPU or single-GPU machines. The r2 version is pre-trained on ~700M samples (r2.1 expands to ~1B), delivering up to 15% better accuracy than the r1 version. TTM supports both zero-shot and fine-tuned...