←搜索结果

搜索“Google DeepMind”

按匹配度 · 共 63 条结果 清除搜索

星期四

  1. Ollama Blog

    Ollama 推出多模态模型新引擎，支持 Llama 4、Gemma 3、Qwen 2.5 VL 等视觉模型

    Ollama 发布新的多模态引擎，开始支持 Meta Llama 4、Google Gemma 3、Qwen 2.5 VL、Mistral Small 3.1 等视觉模型，可进行多图理解、文档扫描与字符识别。此前 Ollama 依赖 ggml-org/llama.cpp 提供模型支持，新引擎让每个模型自包含并暴露自己的投影层，同时加入图像缓存、显存估算与 KV cache 优化，例如针对 Gemma 3 的滑动窗口注意力和 Llama 4 Scout 的分块注意力。后续计划支持更长上下文、思考/推理、流式工具调用和 computer use。

    KV cache optimizations to improve how memory can be efficiently used. Ollama configures causal attention at the individual model level instead of configuring as a group. Examples: Google DeepMind’s Gemma 3 leverages sliding window attention, and Ollama can leverage that to allocate a subset or a portion of the model’s context length to improve performance, and because of the resulting memory efficiency, we can increase the context length of the model on the same system or use the remaining memor

    匹配字段正文

    #company#产品更新#多模态#部署/工程

星期一

  1. Prime Intellect

    Prime Intellect 开源分布式推理栈，面向消费级 GPU 与公网高延迟

    Prime Intellect 发布分布式推理栈预览，并开源 PRIME-IROH、PRIME-VLLM、PRIME-PIPELINE 三个代码库，目标是让消费级 GPU 通过公网连接运行推理。文章指出流水线并行通信需求最低，适合分布式环境，但同步流水线会因 100ms 级跨洲延迟形成吞吐上限，而异步微批调度在内存带宽受限的解码阶段无法带来训练那样的加速。团队称已可用私有硬件通过公网跑推理，正把组件接入自家栈，用于 SYNTHETIC-2 合成数据运行。

    ut the runtime of operations on GPUs to analyze bottlenecking factors during inference workloads on a single node. Much of the notation is borrowed from How To Scale Your Model by Google DeepMind. If you are already familiar with traditional inference math, skip to Distributed Inference. Accelerator Math Whenever we run an operation on an accelerator, there are two things that take time: Computation: How long it takes to compute all the floating-point operations (FLOPs) required for the operatio

    匹配字段正文

    #company#开源/仓库#部署/工程#推理

星期四

  1. Prime Intellect

    Prime Intellect 开源 OpenDiLoCo，实现跨洲分布式低通信训练

    Prime Intellect 开源 OpenDiLoCo，这是 DeepMind DiLoCo 低通信分布式训练方法的可复现实现，代码基于 Hivemind 与 PyTorch FSDP，已发布在 GitHub。团队在加拿大、芬兰和美国两地共四个 DiLoCo worker（各 8 张 H100）上训练 1.1B 参数模型，跨两大洲三国，计算利用率保持 90-95%，all-reduce 仅占训练时间 6.9%。相比原论文 400M 参数规模，该方法被扩展到 1.1B 参数，并验证伪梯度可用 FP16 all-reduce 而不损失性能。

    OpenDiLoCo: An Open-Source Framework for Globally Distributed Low-Communication Training

    ency Insights: Our ablation studies provide valuable insights into the algorithm's scalability and compute efficiency, paving the way for future improvements DiLoCo Recent work by Google DeepMind has introduced an approach that enables the training of language models on islands of devices that are poorly connected. This method allows for data parallel training on these different islands, requiring synchronization of pseudo gradients only every 500 steps. DiLoCo introduces an inner-outer optimiza

    匹配字段正文

    #company#开源/仓库#数据/训练#部署/工程