mirror of
https://github.com/hanzoai/trolley-bench.git
synced 2026-11-02 13:22:03 +00:00
zachlandesandGitHub
462aaf404e
* Kept the original raw answers while making later runs reproducible. * Registered a variance-sized sample before the independent rerun. * Removed article drafts and private paths from the current tree.
15 lines
1014 B
Bash
Executable File
15 lines
1014 B
Bash
Executable File
#!/bin/sh
|
|
# Winnow-12B on the v2 grid as v1 ran it: winnow-server at 77d1458 (its build pins and checks the
|
|
# patched llama.cpp commit and patches), Metal, apple-silicon profile, text-only, BF16 GGUF.
|
|
# Usage: run/v2/winnow-12b.sh <weights dir> <scratch dir>
|
|
set -eu
|
|
. "$(dirname "$0")/lib.sh"
|
|
weights winnow-12b
|
|
GGUF=$(snapshot EldanRing/Winnow-12B)/gguf/Winnow-12B-BF16.gguf
|
|
checkout winnow-inference https://github.com/EldanRing/winnow-inference 77d14580c6732ca2f3745750c1dc1fd446d8bcee
|
|
(cd "$S/code/winnow-inference" && python3 scripts/build.py --jobs 8)
|
|
drive winnow-12b --cells v2 --port 8091 --cwd "$S/code/winnow-inference" --code-repo https://github.com/EldanRing/winnow-inference --code-commit 77d14580c6732ca2f3745750c1dc1fd446d8bcee \
|
|
--runtime "winnow-server (patched llama.cpp 911f6cd), Metal, apple-silicon profile, text-only, BF16 GGUF" -- \
|
|
python3 scripts/serve.py --profile apple-silicon --text-only --model "$GGUF" --host 127.0.0.1 --port 8091
|
|
cleanup winnow-12b "$S/code/winnow-inference"
|