Topic

Quantisation & speculative decoding

Quantisation, speculative decoding and other ways to make a model use less memory and generate faster.

Follow this topic

On this site · Elsewhere
On this site · My radar
Elsewhere · RSS reader

Topic graph

16 nodes · 68 links
The graph shows the 10 of the 17 entries under “Quantisation & speculative decoding” that are most connected to each other, and the 6 of the topic's 55 direct outside neighbours most connected to them. The full lists are below.
  • Entries in this topic
  • Neighbours outside the topic
Relation types
  • Builds on · Successor · Extends
  • Prerequisite · Uses
  • Explains · Supports
View in the network →

Skills and knowledge entries are written in Chinese, as are signals from Chinese sources.

Published signals

13 signals
Hugging Face Blog2026-10-20

MPS in PyTorch 2.15

MPS in PyTorch 2.15

量子位2026-10-20

记忆张量开源 MemT 1.0:把记忆原生写进 Transformer

记忆张量开源记忆原生模型 MemT 1.0,含 36B 总参数、激活 3B 的 Nano 与 288B 总参数、激活 17B 的 Flash 两个型号,在 Transformer 模块中加入与注意力并行的记忆分支,把经验编码为可读写的内部状态。论文称这是记忆原生模型首次完成近 300B 参数规模的工程验证。

Hugging Face Blog2026-10-19

TorchNeuron: running HF models on AWS Trainium is easier than ever

TorchNeuron: running HF models on AWS Trainium is easier than ever

vLLM Releases2026-10-18

v0.32.0

v0.32.0 Highlights This release features 814 commits from 363 contributors (121 new)! Structured decisions and generic parsers: the new POST /v1/systemone endpoint answers choice, noul and score questions over a structured state (#59299), is served by default…

Hugging Face Blog2026-10-16

Welcome Beam: a 501B open model you can own, customize, and scale

Welcome Beam: a 501B open model you can own, customize, and scale

Hugging Face Blog2026-10-15

Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

量子位2026-09-26

TPU 巨核:同样是 16 块芯片,Kimi K3 快 57%

vLLM 原班人马创办的 Inferact 开源了一组 TPU 推理内核:把 Kimi K3 的 92 层计算写进同一个程序,趁当前层还在计算时预取下一层的权重。16 块 TPU v7 配合推测解码每秒输出 709 个 token,基线是 vLLM 在 16 块 GB200 上的 452 个。两种芯片的内存带宽其实是 GB200 略高,差距来自软件怎样用满它。这是面向单个用户、小批量解码的成绩。

Hugging Face Blog2026-09-24

Accelerating vision-language models with LFM2.5-VL-DSpark

Accelerating vision-language models with LFM2.5-VL-DSpark

InfoQ 中文2026-09-23

Pinterest 数百亿向量检索:量化与 SSD 的取舍明细

Pinterest 的搜索平台 Manas 在 80 个集群上服务数百亿嵌入向量。在 1 亿条向量上的离线基准里,标量量化把 HNSW 索引从 121GB 压到 50GB、召回几乎不掉;乘积量化压到 32GB,召回从 93.7% 掉到 77.3%。上线后服务成本降低 20% 到 30%。搬到 SSD 时,SPANN 加乘积量化的吞吐是 DiskANN 的 3 倍、延迟三分之一,召回低 5%。选哪种,取决于你的召回下限,不取决于压缩率。

InfoQ 中文2026-08-21

Muse Glimmer 开源:30B 压进消费级显卡,报错不停下

Meta 发布 Muse Glimmer:300 亿参数、Apache 2.0 开放权重,专为常驻本地的智能体工作流设计。4 位动态量化把 55GB 以上的显存需求压到 17 至 20GB,配合 DFlash 推测解码在 M5 Max 与 RTX 5090 上把吞吐提到最多 3.1 倍。真正值得注意的不是尺寸,而是它把「工具调用失败之后怎么办」写进了训练目标:API 或终端命令报错时,模型诊断故障并另找路径,而不是终止执行。

InfoQ 中文2026-08-12

Meta 重回开放权重:一封长文、一个独立董事会,和一组本地部署的数字

Muse Glimmer 以 Apache 2.0 开放权重的同时,Meta 宣布重回开放路线:扎克伯格发表长文反对能力集中,公司设立独立董事会逐项审核每次模型发布,并披露了让本地部署成立的量化与推测解码数字。Muse Spark 1.2 的权重据称随后开放。

SGLang Releases2026-08-08

SGLang v0.5.17

SGLang v0.5.17: 582 PRs from 194 contributors, with day-0 support for Kimi K3 and MiniMax-H3.

Ollama Releases2026-07-25

Ollama v0.32.4

Ollama v0.32.4 adds Laguna support on Apple GPUs via the MLX engine, quantizes speculative-decoding draft output heads at the requested type, and fixes Qwen3 MoE decoding for differently-quantized experts.

Related skills

1 skill
Active

验证量化与推测解码的实际提速

遇到「快 3 倍」「省 70% 显存」,判断它在自己的负载上还剩多少:分清加速的是哪一段,在同样的条件下比,按质量下限选量化档位。

Background knowledge

3 entries
Intermediate

模型量化与数值精度

把权重和激活从高精度压到低比特,是让大模型跑进有限显存的主要手段;代价是精度损失,而损失落在哪一层决定了它是无感还是灾难。

Intermediate

推测解码与推理加速技术

自回归生成的逐 token 串行是大模型推理延迟的根源;推测解码、连续批处理与 KV 缓存管理从算法与系统两层,把同样的硬件榨出数倍吞吐。

Advanced

KV 缓存与长上下文推理

生成每个新 token 都要回看前面所有 token;KV 缓存把这份回看要用的中间结果存下来,于是上下文越长,显存和带宽越紧——读懂推理引擎更新和长上下文定价的前提。

Graph neighbours

Direct neighbours from /network
Signal · SupportsAnyLanguageModel 1.0: One Year of Local Models on Apple PlatformsSkill · Uses端侧模型部署与硬件适配Skill · Supports模型与输出评估Knowledge · Explains模型选型与约束匹配Knowledge · Supports交互系统中的延迟权衡Signal · ExtendsHow Much Memory Does Your Agent Actually Need?Skill · Supports评估开放权重模型能否自托管Knowledge · Explains混合专家模型(MoE)架构Knowledge · Prerequisite开放权重与模型许可Knowledge · Extends长时程智能体的记忆与上下文管理Skill · Supports模型与推理引擎的支持路径Skill · Supports推理服务容量规划Knowledge · Supports本地与云混合推理架构Signal · Successorv0.30.0Signal · SupportsJev introduces a new shape of LLM - System One, aka Decision ModelsKnowledge · Supports工具使用与函数调用Signal · Builds onMeta is back with Muse Glimmer: local, agentic, multimodal, and open sourceSignal · ExplainsGemini 4 Argon: Built to solve your hardest problemsSignal · SupportsDeepSeek-V4 的周期性盲点:信息换个位置,检索准确率差 40 个百分点Signal · ExplainsOne Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMOSignal · ExtendsEmbeddingGemma 2: an open, lightweight multimodal embedding modelSignal · RelatedOllama v0.35.1: Clef decision modelsSignal · UsesIntroducing Olmo-core 3: Open, scalable training infrastructure for large MoEsSignal · SupportsDeepSeek 开源昇腾版 DeepGEMM 等组件:接口与 NVIDIA 版相同Signal · RelatedHolo4: powering generalist computer-use agentsSignal · UsesKimi K3, and what we can still learn from the pelican benchmarkSignal · ExtendsColibrì:25GB 内存跑 744B 模型,权重按需从硬盘读Signal · Builds onDeploy local agents everywhere with LFM2.5-2.6BSkill · Prerequisite视觉输入的组织与核验Signal · ExtendsLFM2.5-Encoders for Fast Long-Context Inference on CPUSkill · Prerequisite检索流水线调优Knowledge · Explains检索增强生成基础Knowledge · Supports系统设计的权衡Signal · Explains模型之间只传 17 比特:一座桥补上一半差距Signal · SupportsFreeToken:消费级机器不是缩水的服务器,是异构资源Signal · ExplainsGPT-6 Astra: A new generation of intelligenceSignal · SupportsQwen3.8-27B 开源:消费级显卡跑得动,尺子却是自己命名的Skill · Supports旗舰模型发布解读与换代判断Skill · Extends工具集成模式Knowledge · Supports完成判定与验收信号Knowledge · Explains强化学习后训练与可验证奖励Knowledge · Explains知识蒸馏Knowledge · Explains监督微调与参数高效微调(LoRA)Knowledge · Explains基于执行的代码评测Signal · SupportsSame Cluster, 33 Points More Utilization: What Changed Was the OrderSignal · SupportsQwen 3.8 27B is excellent, but it defaults to wildly overthinking thingsSkill · Builds on模型微调与后训练定制Knowledge · Supports人在回路的审查Signal · UsesvLLM v0.27.0Signal · SupportsvLLM v0.26.0Signal · UsesWelcome Inkling by Thinking MachinesSignal · SuccessorOllama v0.32.0Signal · ExplainsvLLM v0.25.0Signal · ExplainsGPT-5.6: Frontier intelligence that scales with your ambitionSignal · ExplainsNative-speed vLLM transformers modeling backend