# T4 cold-start cache Chip-scoped cache for one NVIDIA Tesla T4 (`sm_75`, CUDA 13.0) on Colab. It contains the **one** prebuilt `llama.cpp` server used by every T4 model home, plus a per-model manifest. It contains **no model weights**. Do not use this repo on L4 (`sm_89`), A100 (`sm_80`), or G4 (`sm_120`): a binary built for `sm_75` will not load there. Each chip gets its own `compile-cache-` repo. ## Layout ``` manifest.json llama.cpp//sm75/cuda13.0/llama-server-t4-sm75-cuda13.0.tar.gz models/NVIDIA/Dory1.0-4B/manifest.json models/XHToken/Spark-X2.6-4B/manifest.json ``` ## Artifact | Field | Value | | :--- | :--- | | File | `llama.cpp/98a74c97ea09ac72f6ceda1559808fc0a2f447ce/sm75/cuda13.0/llama-server-t4-sm75-cuda13.0.tar.gz` | | Size | 48,476,283 B | | SHA-256 | `d58c4433aabe5a500b800c1949498591e94f3e887c5f21e6729308633524e5d5` | | llama.cpp source | `https://github.com/brianchmielNV/llama.cpp` | | Revision | `98a74c97ea09ac72f6ceda1559808fc0a2f447ce` (PR #30306) | | Arch | `sm_75` only | | CUDA | 13.0 (Colab T4 image) | | Built on | Colab Tesla T4 `--high-mem` (8 vCPU), 2026-11-06 | | Reuse | `NVIDIA/Dory1.0-4B`, `XHToken/Spark-X2.6-4B` | The archive contains `llama-server`, the `libllama*` / `libggml*` shared objects, and `BUILD-INFO.txt`. It serves both models because the pinned revision contains both `src/models/dory.cpp` and `src/models/spark2-5.cpp`. CMake flags: ``` -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=75 -DLLAMA_BUILD_SERVER=ON -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_APP=OFF -DLLAMA_BUILD_UI=OFF -DLLAMA_USE_PREBUILT_UI=OFF -DGGML_NATIVE=OFF ``` ## Consuming it A model config pins the repo, path, revision, and SHA-256: ```toml binary_repo = "avgrocketman/compile-cache-t4" binary_file = "llama.cpp/98a74c97ea09ac72f6ceda1559808fc0a2f447ce/sm75/cuda13.0/llama-server-t4-sm75-cuda13.0.tar.gz" binary_revision = "a1a394aa84d03bd2c6779724655aa9aae4712cb4" # commit that added this file binary_sha256 = "d58c4433aabe5a500b800c1949498591e94f3e887c5f21e6729308633524e5d5" ``` The launcher verifies the SHA-256 before unpacking. A mismatch stops the run. It unpacks to `/content/llama`, not `/dev/shm` (tmpfs is `noexec` on Colab T4). ## Rebuild trigger Rebuild only when the llama.cpp revision, CUDA arch, or CMake feature set changes. Build on a CPU host with the CUDA 13 toolkit; never compile on a GPU runtime. Verify `--version`, `ldd`, the embedded SASS arch, and the SHA-256, then upload one new file and update `manifest.json` and the model manifests. ## History Moved from `avgrocketman/llamacpp-dory-t4` on 2026-11-06. The old repo is kept read-only as a redirect; new runs use this repo. License: llama.cpp is MIT-licensed; see the upstream repository.