Files
calculet-npu-research-archive/reports/Calculet-NPU-前四步解包内容查看索引-20260801.md

35 lines
4.2 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Calculet NPU 前四步解包内容查看索引
状态:四个归档已经完成分片校验、整包 SHA-256、解包和内容级检查。解包目录合计约 1.4 GiB,可直接查看,不需要再次解压。
## 内容入口
| 步骤 | 解包目录 | 规模 | 校验状态 |
| ---: | --- | ---: | --- |
| 1 | [关键上下文](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/priority-critical-context>) | 约 20 MiB,115 个文件 | 整包及逐文件通过 |
| 2 | [基线分析资料](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/baseline>) | 约 416 MiB68 个文件 | 53 个分片及整包通过 |
| 3 | [无权重模型运行时](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/model-runtime-no-weights>) | 约 537 MiB,66 个文件 | 整包及逐文件通过;权重文件为 0 |
| 4 | [完整 llama.cpp 仓库](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/llama.cpp-full/llama.cpp>) | 约 444 MiB4,443 个文件 | 整包、Git HEAD 和 `git fsck --full` 通过 |
完整仓库 HEAD 为 `fd9bd632f346085920d523dd307a3e4c11cc0a05`
## 推荐查看顺序
1. 先查看 [环境清单](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/baseline/environment>)。这里包含 CPU、内存、PCIe、内核模块、驱动参数、软件包、Podman、服务版本、动态库依赖和 500 条 Git 历史。
2. 再查看 [模型元数据](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/baseline/model-metadata>)。重点包括 prefill/decode `profparts`、I/O buffer YAML、模型 YAML、`hparam.yaml` 和内存预留信息。
3. 然后进入 [CalRT GGML 后端](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/llama.cpp-full/llama.cpp/ggml/src/ggml-calrt>),重点看 `ggml-calrt.cpp``calrt-op.cpp``cal-runtime.h``calrt-common.h`
4. 接着查看 [llama-calrt.cpp](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/llama.cpp-full/llama.cpp/src/llama-calrt.cpp>) 与 [KV cache 适配器](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/llama.cpp-full/llama.cpp/src/llama-kv-cache-calrt-adapter.cpp>),分析模型图、KV cache 和 CalRT 生命周期的接入方式。
5. 最后对照 [无权重运行时模型目录](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/model-runtime-no-weights/models>)。其中保留 prefill/decode ELF、CPU 命令流、ping/pong 数据、Tokenizer、运行库和配置,但不含 `param_blk0.bin``param_blk1.bin`
## 其他重点资料
- [CalRT SDK 原始包](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/baseline/calrt-sdk-0.7.6.tar.gz>)
- [构建复现包](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/baseline/llama.cpp-build-repro-fd9bd632.tar.gz>)
- [历史性能日志](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/priority-critical-context/home/user2/llm/test_llama.cpp/client>)
- [启动与测试脚本](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/priority-critical-context/home/user2/llm>)
- [四个原始归档的本地 SHA-256](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/LOCAL-PRIORITY-1-4-ARCHIVES.sha256>)
## 敏感信息提示
关键上下文中包含远端 `.bash_history` 快照、环境清单、模型文件清单和容器信息。这些资料用于内部环境复现与性能分析,不应未经检查直接对外分发。