# 推理 · 日跟踪 · 2026-09-16

> 日期：2026-09-16（Asia/Shanghai）。专题正文：`intel/topics/inference.md`。原料：`/workspace/intel/archive/2026-09-16/`；对照 2026-09-15 日跟踪。

## 摘要

- **新稿 · OpenShift AI agentic 全链路剖析（非裸 vLLM）：** 生产 agentic 请求经 harness（上下文/会话/工具 schema）+ 可选 OpenShell 沙箱后才到模型；GPU 在 C=10 仍有余量（tensor core 约 **35%**、KV cache 约 **14%**——**厂商基准**）；瓶颈常在 harness/response processing，而非「再加一块 GPU」。属 **推理平台容量规划方法**，见 agent-hosting/openshift。
- **续档 · AutoRAG TP（OpenShift AI 3.5）；Speculator training 托管路线未 GA；llm-d / neocloud；Nutanix Private Inference / Agent Gateway；d-Matrix NVLink Fusion；Qualcomm Hexagon 端侧（见 edge-inference）。**
- **交叉轻 · Astera Leo X：** 宣称 TTFT 最高约 **62%**、token 吞吐约 **22%** 改善（相对 PCIe H200 测例，**厂商/媒体转述，非 Blackwell 大规模**）——KV cache 近加速器存储旁路，见 asic。

## 原文摘录

### Agentic pipeline profiling · OpenShift AI

- 归档：`redhat/opening-the-black-box-profiling-a-secured-agentic-pipeline-on-red-hat--2f370b3e.txt`
- URL：https://www.redhat.com/en/blog/opening-black-box-profiling-secured-agentic-pipeline-red-hat-openshift-ai

> Most inference benchmarks test the model in isolation—a raw request to vLLM… But production agentic systems don’t work that way…  
> Even at C=10 with 10 concurrent agents, tensor core activity reaches only 35%, KV cache occupancy sits at 14%, and the vLLM request queue is empty 67% of the time…  
> To optimize an agentic deployment, benchmark the harness, not just the model.

### AutoRAG / Speculator · 续档摘录

- `redhat/autorag-advances-in-red-hat-openshift-ai-3-5-99df12dc.txt`  
  > Red Hat OpenShift AI 3.5 is already available, and AutoRAG 3.5 is currently available as a Technical Preview.
- `redhat/from-fine-tuned-model-to-cheaper-and-faster-inference-speculator-train-eb9403d6.txt`  
  > Red Hat is building managed speculator training into Red Hat OpenShift AI… We will share more details as we approach release.