blob: 94e6bb8f37c074757eee6a894a5d5f97f409d81e (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
|
# Working /etc/sglang/gemma_4_31b_nvfp4.env on RTX 5090 (SM 12.0a)
# with sglang-git ≥ r12451 + this AUR package's patches applied.
# See SM120-NVFP4-NOTES.md in this directory for what each line resolves.
#
# Replace <snapshot-hash> with the actual HF snapshot directory name
# under /var/lib/sglang/hub/ once the model has been downloaded.
#
# NOTE: LD_LIBRARY_PATH below hardcodes python3.14 (Arch as of 2026-05).
# Update the path component when Arch moves to a new Python minor version.
LD_PRELOAD=/usr/lib/libcuda.so.1
LD_LIBRARY_PATH=/usr/lib/python3.14/site-packages/tvm_ffi/lib:/usr/lib
SGLANG_OPTS="--port 30000 --fp4-gemm-backend=cutlass --chat-template /var/lib/sglang/hub/models--RedHatAI--gemma-4-31B-it-NVFP4/snapshots/<snapshot-hash>/chat_template.jinja"
|