---
license: mit
tags:
- llama.cpp
- cuda
- colab
- t4
- sm75
---

# llama.cpp b10355 CUDA runtime for Colab T4 (SM75)

This public repository stores a reproducible `llama-server` runtime, not model
weights. It was cross-built on an actual Google Colab G4 with CUDA 12.8 and
`CMAKE_CUDA_ARCHITECTURES=75`, then downloaded without authentication and
validated on an actual Colab Tesla T4.

## Provenance

- Source: `ggml-org/llama.cpp`
- Release: `b10355`
- Commit: `dd1ea524333b1e697489067d7a4c39c60d32beee`
- Build host: NVIDIA RTX PRO 6000 Blackwell Server Edition (Colab G4)
- Target: Linux x86-64, CUDA 12.8, SM75, Colab Tesla T4
- Archive: `llama-cpp-b10355-cuda128-sm75-colab.tar.gz`
- Archive SHA-256: `4469c6f18805f2cf2917df6bb8c0dbd37937344ab6c2e77c312cccdf7b145a9a`
- `llama-server` SHA-256: `f4f421a690bc949174752111bd20ac861fcdc2ba65fca361955edf7c67b53252`

## Verified T4 result

The public archive was restored at revision
`5738166d5979c9d39d0b3cacbaa1c5c2a3d44b16` with no HF token. It loaded
`unsloth/Qwen2.5-VL-7B-Instruct-GGUF` revision
`68bb8bc4b7df5289c143aaec0ab477a7d4051aab`, using
`UD-Q4_K_XL` plus `mmproj-BF16`.

- GPU memory after load: 6,647 MiB
- Model load to healthy server: 4.01 seconds on the verified rerun
- Fixed workload: uncached 512 prompt tokens to fixed 128 output tokens,
  batch 1, concurrency 1, five repetitions
- Client end-to-end median: 34.53 tok/s
- llama.cpp decode median: 39.20 tok/s
- Five output hashes were identical
- Multimodal smoke: 1024x768 candy photo, 4.00 seconds; correctly reported five
  candies and the teal, orange, and green colors

See `t4_validation.json` and `t4_setup_receipt.json` for raw evidence.

## Restore

```python
from huggingface_hub import hf_hub_download
import hashlib, tarfile

path = hf_hub_download(
    repo_id="WJO/llama-cpp-b10355-colab-t4-sm75",
    filename="llama-cpp-b10355-cuda128-sm75-colab.tar.gz",
    revision="5738166d5979c9d39d0b3cacbaa1c5c2a3d44b16",
)
assert hashlib.sha256(open(path, "rb").read()).hexdigest() == (
    "4469c6f18805f2cf2917df6bb8c0dbd37937344ab6c2e77c312cccdf7b145a9a"
)
with tarfile.open(path, "r:gz") as archive:
    archive.extractall("/content", filter="data")
```

The runtime expects the CUDA and standard system libraries included in the
current Colab CUDA 12.8 image. Verify the archive and binary hashes before use.