# 推理 · 日跟踪 · 2026-09-15

> 日期：2026-09-15（Asia/Shanghai）。专题正文：`intel/topics/inference.md`。原料：`/workspace/intel/archive/2026-09-15/`；对照 2026-09-14 日跟踪。

## 摘要

- **新稿 · AutoRAG（OpenShift AI 3.5 Technical Preview）：** Kubeflow Pipelines + IBM ai4rag；自动评测/超参；胜出模式一键生成 ingestion KFP + Responses API/`pattern.json`；多语言（德/西/日起步）；**pgvector** 并列 Milvus——标 **TP**，OpenShift AI 3.5 已可用、AutoRAG 3.5 为 TP。
- **新稿 · Speculator training on OpenShift AI（路线）：** EAGLE3 / vLLM Speculators；定制 draft 对齐微调 verifier；同族交叉蒸馏降训练成本；Managed workflow「Coming to」OpenShift AI——**未宣称 GA**；serving 侧投机解码 vLLM 已可用。
- **新稿交叉 · neocloud：** Red Hat AI Inference / **llm-d**（CNCF sandbox；prefix-cache-aware routing、KV offload、多后端）给已有 K8s 抬利用率；完整路径 RHEL→OpenShift→OpenShift AI。
- **续档：** ZStack AIOS×910C；d-Matrix NVLink Fusion；RH AI 3.5 priority-aware / KV offload；Nutanix Private Inference。

## 原文摘录

### AutoRAG TP

- 归档：`redhat/autorag-advances-in-red-hat-openshift-ai-3-5-99df12dc.txt`
- URL：https://www.redhat.com/en/blog/autorag-advances-red-hat-openshift-ai-35

> … the upcoming advanced technical preview of AutoRAG in Red Hat OpenShift AI 3.5…  
> … automatically generates production-ready artifacts for both halves of your RAG pipeline (the ingestion pipeline and the query endpoint)…  
> … adding pgvector support alongside the existing Milvus option…  
> Red Hat OpenShift AI 3.5 is already available, and AutoRAG 3.5 is currently available as a Technical Preview.

### Speculator training

- `redhat/from-fine-tuned-model-to-cheaper-and-faster-inference-speculator-train-eb9403d6.txt`  
  > … real world vLLM production deployments consistently achieving 2.5x to 3.5x reductions in inter-token latency…  
  > Red Hat is building managed speculator training into Red Hat OpenShift AI, powered by the open source Speculators library and Kubeflow. … We will share more details as we approach release.

### llm-d / neocloud（交叉）

- `redhat/from-bare-metal-to-diverse-ai-revenue-streams-navigating-the-gpu-cloud-4975c2ea.txt`  
  > Red Hat AI Inference, powered by llm-d, provides disaggregated inference with prefix-cache-aware routing, KV-cache offloading, and multi-backend support…

## 检索范围

- `archive/2026-09-15/redhat/`、`zstack/`、`industry/`（d-Matrix 再抓）；对照 2026-09-14 inference