Meta 开源 Muse Spark Open 模型,816B 总参数、1M 上下文
Meta 发布 Muse Spark Open 的开放权重,采用 Muse Model License 在 Hugging Face 上线,为 816B 总参数、42B 激活参数、1M token 上下文窗口的模型,原生多模态并面向编码与智能体任务。该版本在 Muse Spark 基础上增加安全后训练,可在 SGLang 和 vLLM 上运行,并针对 NVIDIA、AMD、Intel 优化。PyTorch 侧集成 Helion 内核,在 NVIDIA GB300 上相比 CUDA、CuTe DSL 自定义内核取得 1.03–3.95x 加速,AMD Mi350 端到端吞吐提升 3%;TorchTitan 已支持其 SFT、LoRA 与 TitanRL 微调。

nt shapes and hardware. Helion kernels can be further autotuned by users for broader hardware platforms and workloads – including different accelerator architectures (NVIDIA, AMD, Google TPU, etc.), quantization formats, tensor shapes, and parallelism strategies – without maintaining separate hand-tuned implementations for every configuration. Together, these capabilities make it easier to achieve portable performance while reducing the cost of developing and maintaining specialized kernels. Pos
匹配字段正文


