recipe/llama.cpp-patches.md
1.2 KB · 21 lines · markdown Raw
1 # Runtime used for the custom-build measurement
2
3 The model does not require this patch set. It runs on clean llama.cpp build
4 10454. These are the exact commits behind the separate 55.402 tok/s runtime
5 measurement.
6
7 Base commit: `4df29be4f4c3673f428170fda944a5b19f743bb8`
8
9 | Pull request | Pinned head |
10 |---|---|
11 | [#26001](https://github.com/ggml-org/llama.cpp/pull/26001) | `1e1885f3d7d576f0374834979b0ef03fa6d6f485` |
12 | [#26048](https://github.com/ggml-org/llama.cpp/pull/26048) | `d7d780e40de3e702e5109f712cec5d8097eca5d6` |
13 | [#26705](https://github.com/ggml-org/llama.cpp/pull/26705) | `7061f25bafaf34a0cb51ee5fdbf972315c3a8102` |
14 | [#27173](https://github.com/ggml-org/llama.cpp/pull/27173) | `2c9cb8f3fffc93c7ffc5e186a6d6d4f134b9cd5a` |
15 | [#24891](https://github.com/ggml-org/llama.cpp/pull/24891) | `55e9026d6b88d5eb8ad021f5fad7d55e896b5dbb` |
16 | [#25635](https://github.com/ggml-org/llama.cpp/pull/25635) | `fbcbe95f69502c11d77e9cdd61c9e6be74b2e017` |
17
18 The build targeted CUDA architectures 89 and 120 and enabled CUDA graphs and
19 Flash Attention. Recheck every patch against current master before rebuilding;
20 open pull requests move, conflict, and sometimes become unnecessary.
21