來源NVIDIA Developer Blog•較早收集於 15m
在 NVIDIA 平台上運行 DiffusionGemma 以實現高吞吐量文字生成
#gpu-optimization#inference-speed#real-time-aidiffusiongemmagoogle deepmindnvidiadiffusiongemma
💡了解如何利用 NVIDIA 上的 DiffusionGemma 優化即時 AI 代理的響應速度並降低服務成本。
⚡ 30 秒速覽
有什麼變化
針對 NVIDIA 平台進行高吞吐量文字生成優化
為什麼重要
此發布使開發者能夠構建響應更迅速、流暢的 AI 互動，同時降低與高延遲 token 生成相關的基礎設施成本。
下一步行動
在您現有的 NVIDIA GPU 堆疊上對 DiffusionGemma 進行基準測試，以評估其對您特定代理工作流的延遲改善效果。
誰應關注：Developers & AI Engineers
關鍵要點
- •針對 NVIDIA 平台進行高吞吐量文字生成優化
- •旨在降低即時 AI 代理與聊天助理的延遲
- •解決逐個 token 生成的瓶頸以降低服務成本
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 6 個來源。
🔑 增強重點摘要
- •DiffusionGemma is an experimental open model released under a permissive Apache 2.0 license, making its weights freely accessible for developers to download and integrate.
- •The model achieves up to 4x faster text generation by employing a diffusion process that denoises up to 256 tokens per step in parallel, a significant departure from the sequential, token-by-token generation of traditional autoregressive models.
- •Built on the Gemma 4 architecture, DiffusionGemma is a 26-billion-parameter Mixture-of-Experts (MoE) model that efficiently activates only 3.8 billion parameters during inference, enabling quantized deployment within 18 GB VRAM limits.
- •It incorporates bidirectional attention and a self-correction mechanism, allowing it to evaluate and refine entire text blocks simultaneously, which facilitates real-time error correction and parallel context propagation.
- •Optimized across NVIDIA's hardware stack, DiffusionGemma supports a range of GPUs from consumer GeForce RTX 5090 and 4090 to enterprise Hopper and Blackwell systems, leveraging native NVFP4 (4-bit floating-point) kernels for accelerated compute throughput.
🛠️ 技術深入
- Model Architecture: DiffusionGemma is built on the Gemma 4 backbone, a 26B Mixture of Experts (MoE) model that activates 3.8B parameters during inference, pairing a diffusion head with this architecture.
- Generation Mechanism: Unlike autoregressive models that predict tokens sequentially, DiffusionGemma utilizes a diffusion process, starting with a "canvas" of random placeholder tokens and iteratively refining them in parallel over multiple denoising passes.
- Parallel Generation: It denoises up to 256 tokens per step simultaneously, which significantly reduces memory transfers compared to the sequential nature of autoregressive generation.
- Bidirectional Context & Self-Correction: The model employs bidirectional attention, allowing it to evaluate the entire text block at once and correct errors during the generation process.
- Hardware Optimization: Native support for NVFP4 (4-bit floating-point) accelerates compute throughput with near-lossless accuracy, optimizing performance on NVIDIA GPUs.
- Deployment Efficiency: Designed for deployment within 18 GB VRAM limits, making it compatible with high-end consumer GPUs (like RTX 4090 and 5090) and enterprise systems (Hopper and Blackwell).
- Framework Support: The model can be served efficiently using inference frameworks such as vLLM (with Red Hat integration), Hugging Face Transformers, and MLX. Fine-tuning is supported via Hackable Diffusion, Unsloth, and NVIDIA NeMo.
- Variable Length Generation: For sequences longer than 256 tokens, DiffusionGemma employs a "Block Autoregressive Diffusion" approach, where a 256-token block is fully denoised, committed to the KV cache, and then the model processes the subsequent block.
🔮 前景展望基於引用來源的 AI 分析
Diffusion models like DiffusionGemma will become increasingly prevalent for real-time, interactive AI applications.
Their ability to generate text in parallel and self-correct directly addresses critical latency bottlenecks and coherence issues inherent in traditional autoregressive models, which are essential for responsive chat assistants and agentic workflows.
The adoption of diffusion models for text generation will drive further hardware optimization and specialized AI accelerators.
DiffusionGemma's performance gains are explicitly linked to exploiting high arithmetic intensity on NVIDIA GPUs and utilizing features like NVFP4, indicating a trend towards hardware-software co-design for this new generative AI paradigm.
The distinction between diffusion and autoregressive models for language generation will blur, leading to hybrid architectures.
Research suggests that Masked Diffusion Models are theoretically equivalent to random-order autoregressive models and can achieve strong performance on compression-heavy tasks while enabling parallel sampling, hinting at a convergence of these paradigms.
⏳ 時間線
2024-02
Initial release of Gemma (2B and 7B sizes).
2024-06
Initial release of Gemma 2 (9B and 27B sizes).
2025-03
Gemma 3 released, introducing multimodal support and larger variants.
2025-05
Google DeepMind unveils Gemini Diffusion (research preview), demonstrating commercial-grade performance parity with autoregressive models and high token generation speed.
2026-04
Google DeepMind releases Gemma 4, an open model family with extended multimodal capabilities and diverse architectures.
2026-06-10
Google DeepMind releases DiffusionGemma, an experimental open model optimized for NVIDIA hardware, offering up to 4x faster text generation.
📎 來源 (6)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： NVIDIA Developer Blog ↗
每週電子報
每週一封，可隨時退訂。