v0.1.2 — HBFSim 优化加速版

HBFSim v0.1.2 优化加速版

本版整理了已验收的 HBF 推理优化、源码与构建指南、精确计时表和适用范围。优化配置按需启用，保留原有正常 HBF 模拟延迟；包括 PTX 输入核验缓存、分块读取、共享库摘要缓存和主机统一记账。同页合并、持久化 ELF 缓存及 provider API-activity 计时保持关闭。

主机计时部分	旧 r24	本次	旧耗时／本次耗时
初始化	1015.70 秒	363.54 秒	2.79×
生成	5695.33 秒	108.21 秒	52.63×
初始化＋生成	6711.03 秒	471.76 秒	14.23×
生成中的输出头，含 logits	2411.04 秒	15.68 秒	153.73×

这些数据来自两个不同 attempt、同一固定输入及两个输出 token；不是多次重复实验、单项优化的因果贡献或跨机器性能保证。范围验收包括 147 份注册 storage（49 份选择＋98 份既有）、代表性实际地址与服务闭合、19 项同步主机计时及约定输出。未建立全张量、逐字节或全部专家消费的证明。

模型自然退出为 0。原批后处理失败记录保持 FAILED；独立离线修复通过范围验收，未重跑模型或 collector。原 ACK 覆盖 199 份文件。

项目既有 CPU 构建、测试和仓库检查 CI已在代码提交 ed6f5559673767bd620eb8174fdf61bdc7aa3639 上通过，保留原有测试排除项。本地源码包核验及整数／布局夹具通过 10,002 项检查。这些检查尚不能证明优化后 GPU 代码可从头完整重建，也不能证明另一台机器上的 GPU 加速倍率。

• 简明使用指南
• 源码与重建指南
• 精确计时数据
• 后续优化方向

附件中的 425,433 字节压缩包包含增量源码与精简证据；需配合完整 v0.1.2 仓库使用，解压到 reproduction/。SHA256SUMS 用于核对原包。模型权重、生成或供应商 PTX、二进制缓存、大型 raw 日志与 SHM 快照不在发布附件中；原件保留并提供索引。本次文档更名未修改源码标签、压缩包字节、模拟参数或实验数据，也未新增 GPU 实验。

English

HBFSim v0.1.2 — Optimized Edition packages the accepted optional HBF inference optimizations, source/build instructions and compact timing evidence. Nominal HBF waits and simulator parameters are preserved. The measured initialization-plus-generation interval is 471.76 s versus 6711.03 s in the historical r24 attempt (14.23×); this is one fixed prompt and two output tokens across different attempts, not a general or cross-machine guarantee. Existing project CPU CI passed; a complete clean optimized-GPU rebuild and cross-machine GPU performance remain unverified. The original code tag and source archive are unchanged; the linked documentation adds the clearer edition name.

v0.1.1 — UCIe multistack interface

HBFSim v0.1.1

This release adds the UCIe multistack interface to HBFSim's public source tree, including the current host backend, configuration profiles and CPU reproduction tools.

New implementation

• UCIe request and device interfaces: typed link/device/multistack profiles, bounded admission, backing-range registration, AXI read transport, page striping and exact completion reassembly.
• Multistack MQSim service: independent stack engines and service limits, conservative event advancement and shared-link accounting. The online media simulator remains in the request path.
• Host transport and completion handling: shared-memory worker RPC, ordered admission, compact completion batches, completion waves and deferred worker startup.
• Workload integration: explicit daemon backend selection and vLLM weight-placement/binding support. Nominal and zero-injected-wait modes retain real MQSim execution.
• Portable build integration: matching public headers, worker and daemon targets, the MQSim next-event patch, configuration schemas, fixed CPU tests and reproduction scripts. UCIe is optional and defaults to off.

Documentation

The English and Chinese READMEs keep their existing structure. SASS final147 reproduction and full-model registration/binding evidence now appear in the validation section. Interface contracts and configuration details are documented in docs/ucie, with a dedicated v0.1.1 release document.

Reproduction

git clone --branch v0.1.1 https://github.com/SlugLab/HBFSim.git
cd HBFSim
git submodule update --init --recursive third_party/bpftime third_party/mqsim
HBFSIM_BUILD_JOBS=4 bash scripts/reproduce_ucie_current_cpu.sh

Build commands, source provenance and validation results are recorded in the release CPU checkpoint. The release branch and v0.1.1 tag identify the same source revision. This is a source release, with no prebuilt binaries.

v0.1.0 — HBFSim as a public project

First tagged release. It carries no change to the simulator itself: it is the point at which the repository became usable from outside.

The defect this release fixes

git clone --recurse-submodules https://github.com/SlugLab/HBFSim.git failed authentication for every visitor, because the paper submodule points at a private Overleaf repository. The submodule is now marked active = false and update = none, and scripts/bootstrap.sh initialises only the two build dependencies.

Verified from a clone holding no credentials for that private repository: bpftime and MQSim check out, paper is skipped, all 152 targets build, and ctest reports 31 of 31 passing with the three environment-bound tests excluded.

git clone https://github.com/SlugLab/HBFSim.git
cd HBFSim
HBFSIM_ENABLE_CUDA=OFF HBFSIM_ENABLE_MQSIM=ON ./scripts/bootstrap.sh
cmake --build build -j"$(nproc)"
ctest --test-dir build --output-on-failure

What else is in it

• Apache-2.0 licence, CITATION.cff, code of conduct, security policy.
• README.md rebuilt around a quick start that runs, with the architecture figure from the paper and every measured result carrying the qualifier its proof document attaches. README.zh-CN.md added.
• TODO.md: 39 open entries grouped by theme, each naming the artifact that would close it, and 10 ticked.
• CONTRIBUTING.md, stating the rule this project runs on: a number enters the repository only together with the checkpoint document that produced it.
• Continuous integration: a CPU-only build and test, plus a hygiene job that checks every in-repository Markdown link. 710 links resolve.
• 52 internal working documents withdrawn from the public repository; docs/reference/cuda-architecture-compatibility.md kept and relocated.
• The default branch is now main, renamed from eval_base.

Paper

arXiv:2609.09800 — HBFSim: Fast and Faithful Simulation of High-Bandwidth Flash Under Real GPU Execution