30 lines
2.4 KiB
Markdown
30 lines
2.4 KiB
Markdown
# Calculet NPU 第八步完整模型查看索引
|
|
|
|
状态：完整模型归档已完成 1,942 个分片、整包 SHA-256、zstd/tar 结构、解包及 68 个文件逐一校验；独立复核再次全部通过。
|
|
|
|
## 查看入口
|
|
|
|
- [完整模型原始归档](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/archives/calculet-models-complete-20260801.tar.zst>)
|
|
- [完整解包目录](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/models-complete/models>)
|
|
- [Qwen3-30B-A3B 模型目录](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/models-complete/models/Qwen3-30B-A3B-dynamic-W8A8-W4AF16-full_layers_merged_2_chips_40960_fa_2026-05-22>)
|
|
- [Tokenizer](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/extracted/models-complete/models/Qwen3-30B-A3B-dynamic-W8A8-W4AF16-full_layers_merged_2_chips_40960_fa_2026-05-22/__Tokenizer>)
|
|
- [模型文件 SHA-256 权威清单](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/manifests/MODEL-FILES.sha256>)
|
|
- [模型文件尺寸清单](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/manifests/MODEL-FILES.tsv>)
|
|
- [内容摘要](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/STEP-8-COMPLETE-MODEL-CONTENT-SUMMARY.txt>)
|
|
- [本地整包 SHA-256](</Users/cogito/Documents/Codex/2026-08-01/wo-m/outputs/Calculet-NPU-analysis-bundle-20260801/remote-export/LOCAL-STEP-8-COMPLETE-MODEL.sha256>)
|
|
|
|
## 校验结果
|
|
|
|
| 项目 | 结果 |
|
|
| --- | --- |
|
|
| 压缩归档大小 | `16,284,930,970` 字节 |
|
|
| 整包 SHA-256 | `ebb143415bc44456599265d07e0b0057497d8708ebafb23f53435d0c189622d8` |
|
|
| 解包文件数 | 68 |
|
|
| 解包文件逻辑大小 | `18,749,687,350` 字节 |
|
|
| `param_blk0.bin` | `9,404,557,312` 字节，SHA-256 通过 |
|
|
| `param_blk1.bin` | `8,782,227,456` 字节，SHA-256 通过 |
|
|
|
|
模型目录包含完整参数、prefill/decode 命令流、CCU ELF、CPU/PLD ping-pong 文件、运行库、YAML、profparts、I/O buffer layout 和 Tokenizer。该资料可用于端到端性能、精度、回归测试和新模型算子优化。
|
|
|
|
完整权重和编译产物属于高敏感度技术资产，不应未经检查对外分发。
|