Hugging Face's logo

HelloSun
/
qwen38-27b-ssd-streaming

Model card Files
xet
 Community
qwen38-27b-ssd-streaming

Commit History

27 days ago

  1. 整理中文 model card 與 SSD streaming 定位

    HelloSun
    f6a7b8e

  1. model card (README.md): 重點整理 + 使用方法 + 證據對照; 修正 memwatch 證據檔互相覆蓋的 bug (probe 後重啟會覆蓋 1K 驗證的 json, 改用 -restart tag); 1K 峰值改報區間 0.35-0.67 GiB (三次量測, 參考上緣)

    Auto Upload Agent
    2bd0a72

  2. 文件更新: 262144 達成結果 + 更正磁碟頻寬 (舊 4.3-5.1GB/s 是上一台機器, 本機 325MB/s) + 5 個量測陷阱

    Auto Upload Agent
    ab86acb

  3. 262144 理論最高上下文達成: KV 檔案 mmap 4.78 GiB @SSD , 峰值總 RSS 2.118 GiB < 3GB, swap 0, 實際推理成功; 修正 LLAMA_KV_SSD 決定被上一次啟動污染的 bug (start_server 覆寫變數 -> probe 誤判不放 SSD)

    Auto Upload Agent
    f7eef17

  4. 1K 驗證全數通過 (all_passed=true): 峰值總 RSS 0.668 GiB, swap 0 MiB; 新增 tools/io_probe.py 實測本機 SSD (單執行緒 325 MB/s, 4 threads 1213 MB/s) — 舊 STATUS.md 的 4.3-5.1GB/s 是上一台機器, 導致舊 0.33 tok/s 證據虛高; V3 max_tokens 96->8 (每 token 約 44s, 96 token 會逾時)

    Auto Upload Agent
    f7de554

  5. swap 違規已修正 (VmSwap 612MiB->0), 峰值總 RSS 0.67 GiB @1K

    Auto Upload Agent
    ee501ad

  6. swap 違規根因修復驗證成功: 加上 --sweep 後 VmSwap 612 MiB -> 0, 峰值總 RSS 0.67 GiB; 修正 0001/0002 互相影響導致的 patch 誤判 (改用符號確認); sweep 間隔 0.25->1.0s (實測 0.25s 只是重做 4 倍功); 推論逾時 900->2400s; 修掉 -Wmissing-declarations

    Auto Upload Agent
    bd473b5

  7. checkpoint: 1K baseline + swap root-cause fix + SSD KV patch (0002)

    Auto Upload Agent
    1cca54c

  8. 1K 驗證重現 (峰值總 RSS 0.565 GiB); 找出並修掉 swap 違規根因: installpi.sh 從未傳 --sweep, GGUF page cache 灌滿 16GB cgroup -> 核心把本行程匿名頁換出 (VmSwap 612 MiB); 新增 --sweep + patches/0002-kv-ssd-pager.patch (SSD-backed KV cache) 與 262144 實際推理探測

    Auto Upload Agent
    f656e67

  9. 禁止 swap: KV cache 用 mlock 釘在 RAM (llama.cpp LLAMA_KV_MLOCK), 移除 setup_swap/swapfile 路徑, V6 換出量>1MiB 即失敗; 修正 memwatch /proc/limits 解析 (單位已是 bytes, 之前誤乘 1024 使驗證失效)

    Auto Upload Agent
    c5f4753

  10. checkpoint: 1K 驗證證據已保存 (峰值總 RSS 0.586 GiB)

    Auto Upload Agent
    aa92ffb

  11. 1K 端到端驗證通過: 峰值總 RSS 0.586 GiB (<3GB), 0.33 tok/s; pi_config 改為 skip 而非 fail; swap 使用量誠實記錄

    Auto Upload Agent
    30a3291

  12. installpi.sh: 1K 驗證 + 套用 SSD 權重釋放 patch, 更正 KV 容量計算 (17層/q4_0=0.5625), 記憶體判準改為總 RSS; 加 AGENTS.md 崩潰續作指引

    Auto Upload Agent
    7b1da14

  13. v1: 真 3GB 總 RSS 約束成立 (789 MiB @1K ctx)⁠

    SDQwen Agent
    ccd0314

  1. pin llama.cpp to validated commit 0504396 for reproducible SSD-streaming patch + build

    Auto Upload Agent
    d291952

  2. limit build parallelism by available RAM to avoid OOM on 2-3GB machines

    Auto Upload Agent
    4e81dc4

  3. recovery checkpoint: merged storage history, verified 4k evidence, plan low-RAM rebuild

    Auto Upload Agent
    e5a2376

  4. merge storage initial commit with validated 4k SSD inference

    Auto Upload Agent
    43d705e

  5. update STATUS.md with push info

    Auto Upload Agent
    65eadc0

  6. merge storage repo

    Auto Upload Agent
    90b7bf4

  7. initial working 4k validation

    Auto Upload Agent
    1a20509

  8. initial working 4k validation: SSD-streamed Qwen3.8-27B with 3GB RAM limit⁠

    Auto Upload Agent
    7a65842

  9. Merge branch 'main' of https://huggingface.co/HelloSun/sddqwen38-27b

    Auto Upload Agent
    d690d5a

  10. initial commit

    HelloSun
    Verified signature
    0b6c91a

  11. Add auto_upload.py (token from env)

    Auto Upload Agent
    3a0f444

  12. Recovery: uncommitted changes from crash

    Auto Upload Agent
    75dda9c

  13. Enforce 3GB hard memory constraint: mmap/SSD streaming, 4K validation context

    HelloSun
    0650064