AI News AI资讯 1d ago Updated 1d ago 更新于 1天前 48

Liquid AI Releases LFM2.5-DSpark Draft Models That Deliver Up to 3.18x Faster Decoding Without Changing Model Outputs Liquid AI 发布 LFM2.5-DSpark 草稿模型,解码速度提升高达 3.18 倍且不改变模型输出

Liquid AI released DSpark draft model checkpoints for three LFM2.5 models (1.2B-Instruct, 2.6B, 8B-A1B), adding speculative decoding paths with ~300M-parameter drafters Achieves up to 3.18x faster decoding on H100 GPUs and 2.87x on M4 Max MacBook Pro with zero accuracy loss under greedy decoding DSpark combines a DFlash-style parallel backbone, a lightweight sequential head modeled as a Markov chain, and a confidence-scheduled verifier for token pruning Speedup is workload-dependent, ranging fro Liquid AI发布DSpark draft模型检查点,为LFM2.5系列(1.2B-Instruct、2.6B、8B-A1B)提供推测解码加速方案 在H100上最高实现3.18倍解码速度提升(MATH500基准),M4 Max MacBook Pro上最高2.87倍(HumanEval基准) 推测解码采用约300M参数的draft模型提出9个候选token块,目标模型单次前向传播验证,greedy解码输出与基线完全一致 支持llama.cpp和SGLang第一天集成,权重以Safetensors和GGUF格式提供,需自托管部署 多工具函数调用场景延迟平均降低57%,特别适合本地编码助手、设

68
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Liquid AI released DSpark draft model checkpoints for three LFM2.5 models (1.2B-Instruct, 2.6B, 8B-A1B), adding speculative decoding paths with ~300M-parameter drafters
  • Achieves up to 3.18x faster decoding on H100 GPUs and 2.87x on M4 Max MacBook Pro with zero accuracy loss under greedy decoding
  • DSpark combines a DFlash-style parallel backbone, a lightweight sequential head modeled as a Markov chain, and a confidence-scheduled verifier for token pruning
  • Speedup is workload-dependent, ranging from 1.04x to 3.18x, with multi-tool function-calling scenarios seeing 57% latency reduction on LFM2.5-2.6B
  • Licensed under LFM Open License v1.0 for commercial use under $10M annual revenue; larger enterprises require direct licensing from Liquid AI

Why It Matters

This release makes speculative decoding significantly more accessible for self-hosted deployments, offering dramatic latency improvements without any compromise to output quality. For AI practitioners building agentic systems, local coding assistants, or on-device applications, DSpark provides a practical path to faster inference that scales across both cloud GPUs and consumer hardware.

Technical Details

  • Architecture: Each drafter is approximately 300M parameters (295.7M for 1.2B-Instruct, 327.7M for 2.6B and 8B-A1B targets) with 5 full-attention layers, hidden_size=2048, intermediate_size=6144, GQA at 32 heads over 8 KV heads, and a block size of 9 tokens
  • DSpark Method: Combines three components: (1) a DFlash-style parallel backbone conditioned on target context features for single-pass hidden state generation, (2) a lightweight sequential head using a Markov chain at rank 256 to restore inter-token dependency, and (3) a confidence-scheduled verifier that prunes low-confidence suffixes when verification cost exceeds savings
  • Implementation: Weights ship as Safetensors and GGUF formats; both llama.cpp and SGLang provide day-one support with dedicated speculative decoding flags; embedding and LM head are tied from the target model at load time
  • Benchmarks: Tested on MATH500, HumanEval, MBPP, GSM8K, and MT-Bench with block size 9, batch size 1, and temperature 0; H100 results in BF16 via SGLang, M4 Max results in FP16 GGUF via llama.cpp with Metal
  • Performance: LFM2.5-1.2B-Instruct achieves 2.10x mean speedup (656→1384 tok/s) on H100 and 2.54x (138→350 tok/s) on M4 Max; LFM2.5-2.6B achieves 2.67x (323→864 tok/s) on H100; LFM2.5-8B-A1B achieves 2.54x (418→1074 tok/s) on H100 but only 1.18x on M4 Max due to MoE implementation limitations

Industry Insight

  • Agentic AI deployment: The 57% latency reduction in multi-tool function-calling scenarios makes DSpark particularly valuable for agent systems that repeatedly decode during planning and tool execution cycles, directly improving user experience in conversational AI applications
  • Hardware-aware optimization: The significant performance gap between dense and MoE models on Apple Silicon (3.18x vs 1.18x on M4 Max) highlights that speculative decoding benefits are not uniform across architectures; practitioners should benchmark their specific model types before deployment
  • Licensing strategy: The $10M revenue threshold in the LFM Open License creates a clear bifurcation between indie/startup adoption and enterprise licensing, suggesting Liquid AI is targeting the long-tail developer market while preserving revenue from larger organizations

TL;DR

  • Liquid AI发布DSpark draft模型检查点,为LFM2.5系列(1.2B-Instruct、2.6B、8B-A1B)提供推测解码加速方案
  • 在H100上最高实现3.18倍解码速度提升(MATH500基准),M4 Max MacBook Pro上最高2.87倍(HumanEval基准)
  • 推测解码采用约300M参数的draft模型提出9个候选token块,目标模型单次前向传播验证,greedy解码输出与基线完全一致
  • 支持llama.cpp和SGLang第一天集成,权重以Safetensors和GGUF格式提供,需自托管部署
  • 多工具函数调用场景延迟平均降低57%,特别适合本地编码助手、设备端Agent和离线copilot应用

为什么值得看

这篇文章展示了推测解码技术在LFM2.5模型上的生产级实现,为AI从业者提供了在资源受限环境下加速大模型推理的可行方案。对于关注本地部署、边缘计算和Agent应用的开发者来说,这项技术可以显著改善用户体验并降低推理成本。

技术解析

  • DSpark draft模型约300M参数(1.2B-Instruct对应295.7M,2.6B和8B-A1B对应327.7M),骨干网络为5层全注意力架构,hidden_size=2048,intermediate_size=6144,GQA配置32头对8头KV头,block size=9,无独立词表权重(embedding和LM head从目标模型tie加载)
  • 技术架构由三部分组成:DFlash-style并行骨干网络(基于目标模型上下文特征单次前向传播生成所有draft token隐藏状态)、轻量级序列头(256秩马尔可夫链恢复相邻token间依赖关系,提升后续位置接受率)、置信度调度验证器(预测token存活概率,当验证成本超过收益时剪枝低置信度后缀)
  • 基准测试在1xH100(BF16,SGLang)和M4 Max MacBook Pro(FP16 GGUF,llama.cpp Metal)上进行,block size=9,batch size=1,temperature=0,覆盖MATH500、HumanEval、MBPP、GSM8K和MT-Bench五个基准
  • 权重以Safetensors和GGUF格式提供,但Hugging Face托管推理提供商目前不支持,需自托管并配置带DSpark支持的SGLang或llama.cpp构建
  • 许可证为LFM Open License v1.0,年收入低于1000万美元的企业可免费商业使用,大型企业需联系Liquid AI获取商业许可

行业启示

  • 推测解码正从研究走向生产部署,DSpark证明了小参数draft模型(~300M)可以显著加速大模型推理而不牺牲准确性,为边缘设备和本地部署提供了可行路径,特别适合batch size=1的单用户场景
  • MoE架构在Apple Silicon上的加速效果有限(8B-A1B仅1.18x),提示开发者在选择硬件和模型架构时需要权衡,当前llama.cpp Metal后端对MoE的支持仍有优化空间,验证k个token会激活更多expert导致权重传输成本增加
  • Agent应用场景(多工具函数调用)是推测解码的最大受益者,延迟降低57%意味着更流畅的交互体验,这将推动本地Agent和边缘AI的发展,建议开发者在规划agent系统时优先考虑DSpark等推测解码方案

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Open Source 开源 Inference 推理 GPU GPU Product Launch 产品发布