348 kB

Ctrl+K

2 contributors

History: 21 commits

Auto Upload Agent

model card: 連到 REPRODUCE.md

c293672 3 days ago

patches
v0: 4B 專案骨架 (installpi.sh/memwatch.py/patches/STATUS/AGENTS/sync.sh) 3 days ago
validate
上傳解除限制實驗的完整證據: validate/bench-logs-nolimit/ 六個 case 的 memwatch+chat 原始記錄 (swap 全 0, over_limit_samples 全 0) + bench-nolimit.jsonl; 另含 cmake 編譯記錄 3 days ago
.gitattributes

1.52 kB
initial commit 3 days ago
.gitignore

121 Bytes
v0: 4B 專案骨架 (installpi.sh/memwatch.py/patches/STATUS/AGENTS/sync.sh) 3 days ago
.gitignore.sync

142 Bytes
v0: 4B 專案骨架 3 days ago
AGENTS.md

4.1 kB
v0: 4B 專案骨架 (installpi.sh/memwatch.py/patches/STATUS/AGENTS/sync.sh) 3 days ago
README.md

11.5 kB
model card: 連到 REPRODUCE.md 3 days ago
REPRODUCE.md

11.1 kB
REPRODUCE.md: 可復現整個專案的 self-contained prompt (變數集中/驗收門檻/換模型指引/已知實測值); installpi.sh export MODEL_PATH 讓 bench.sh 能繼承 3 days ago
STATUS.md

8.97 kB
STATUS: 記錄解除記憶體限制的實測 — 權重常駐 RAM 只快 6%, I/O 不是瓶頸; KV 放 SSD 比放 RAM 快; 本機僅 2vCPU 故 8 threads 慢 9x 3 days ago
app.py

27.6 kB
v0: 4B 專案骨架 (installpi.sh/memwatch.py/patches/STATUS/AGENTS/sync.sh) 3 days ago
auto_upload.py

4.9 kB
把 GGUF/架構參數化: (1) auto_upload.py 寫死的 27B repo + 不存在的 Space 會讓成果推錯地方 → 改用 STORAGE_REPO/SPACE_REPO 變數 (2) KV elems/token 三處寫死 qwen35 的 8x4x256 → 改 KV_FULL_ATTN_LAYERS/KV_HEAD_COUNT/KV_DIM 變數, 換模型不會算錯 3 days ago
bench.sh

7.08 kB
解除記憶體限制的量測: 權重常駐 RAM 只快 6% (2.82 vs 2.67 tok/s) 但記憶體 7.2x (2.77 vs 0.384 GiB); 本機僅 2 vCPU, 8 threads 反而慢 9x 3 days ago
installpi.sh

54.6 kB
REPRODUCE.md: 可復現整個專案的 self-contained prompt (變數集中/驗收門檻/換模型指引/已知實測值); installpi.sh export MODEL_PATH 讓 bench.sh 能繼承 3 days ago
memwatch.py

16.4 kB
v0: 4B 專案骨架 (installpi.sh/memwatch.py/patches/STATUS/AGENTS/sync.sh) 3 days ago
requirements.txt

36 Bytes
v0: 4B 專案骨架 (installpi.sh/memwatch.py/patches/STATUS/AGENTS/sync.sh) 3 days ago
sync.sh

3.31 kB
v0: 4B 專案骨架 (installpi.sh/memwatch.py/patches 0001+0002/STATUS/AGENTS/sync.sh) 3 days ago