diff --git a/.github/workflows/exact-llm.yml b/.github/workflows/exact-llm.yml index 75accda..88acb98 100644 --- a/.github/workflows/exact-llm.yml +++ b/.github/workflows/exact-llm.yml @@ -21,7 +21,8 @@ jobs: # SmolLM2-1.7B, one route per job, each inside 120 minutes in run 43759765827: f64 on the six runners, float32 # beside it on ubuntu-24.04-arm alone (53 min; 86 to over 120 elsewhere); f32 and i8 on the non-Windows # runners, except f32 on macos-latest and macos-15-intel (each ran past 120 minutes: macos-latest in run 43759765827, - # macos-15-intel in run 43777602992) + # macos-15-intel in run 43777602992) and i8 on macos-latest (4,816 s in run 43759765827, past 120 minutes twice + # in run 43806351228) include: - {model: SmolLM2-1.7B, os: ubuntu-latest, routes: f64, args: "--routes f64 --no-float"} - {model: SmolLM2-1.7B, os: ubuntu-24.04-arm, routes: f64, args: "--routes f64"} @@ -33,7 +34,6 @@ jobs: - {model: SmolLM2-1.7B, os: ubuntu-24.04-arm, routes: f32, args: "--routes f32 --no-float"} - {model: SmolLM2-1.7B, os: ubuntu-latest, routes: i8, args: "--routes i8 --no-float"} - {model: SmolLM2-1.7B, os: ubuntu-24.04-arm, routes: i8, args: "--routes i8 --no-float"} - - {model: SmolLM2-1.7B, os: macos-latest, routes: i8, args: "--routes i8 --no-float"} - {model: SmolLM2-1.7B, os: macos-15-intel, routes: i8, args: "--routes i8 --no-float"} steps: - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 diff --git a/proofs/exact-llm/README.md b/proofs/exact-llm/README.md index 6d2a31f..5829c7c 100644 --- a/proofs/exact-llm/README.md +++ b/proofs/exact-llm/README.md @@ -160,9 +160,9 @@ Seconds are the hash run of the exact-16 route. Before it, the 16-bit weights, i The workflow runs each 1.7B job that finished here: - f64 on all six runners, float32 only beside it on ubuntu-24.04-arm (53 minutes); - f32 on ubuntu-latest and ubuntu-24.04-arm; -- i8 on the four non-Windows runners. +- i8 on ubuntu-latest, ubuntu-24.04-arm and macos-15-intel. -The f32 jobs on macos-latest and macos-15-intel are left out because they ran past 120 minutes: on macos-latest the exact-16 stage alone in run 43759765827, on macos-15-intel in run [43777602992](https://github.com/lmtlssss/speedrunpapi/actions/runs/43777602992) after 6,231 s in the first. Every printed line: `record/ci-43759765827.txt`. +The f32 jobs on macos-latest and macos-15-intel are left out because they ran past 120 minutes: on macos-latest the exact-16 stage alone in run 43759765827, on macos-15-intel in run [43777602992](https://github.com/lmtlssss/speedrunpapi/actions/runs/43777602992) after 6,231 s in the first. The i8 job on macos-latest is left out too: it took 4,816 s in run 43759765827 and ran past 120 minutes twice in run [43806351228](https://github.com/lmtlssss/speedrunpapi/actions/runs/43806351228). Every printed line: `record/ci-43759765827.txt`. The bound at 1.7B: the widest product is the down projection, 8,192 inputs. On f64, 8,192 · (2^15 − 1)^2 < 2^43, under 2^53. On f32 every product at 1.7B is wider than 1,024 inputs, so every one is chunked: 2 chunks at width 2,048 and 8 for the down projection. On i8, 8,192 · 255^2 < 2^29, under 2^31, and n ≤ 33,025.