ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Enable conv-direct in the default native speed profile

Browse source 
Measured on the fresh linux x64 prebuilt (z-image Q8_0, sd-cli, 192 CPU
threads, 512x512, 9 steps, steady state): sampling 56.1s vs 51.3s (about
9 percent faster), VAE decode unchanged, peak RSS identical. The sd.cpp
engine only serves the no-GPU tier, so the default profile now matches
max: --diffusion-fa plus --diffusion-conv-direct.
This commit is contained in:
Daniel Han 2026-07-04 13:46:59 +00:00
commit 24de50062c
2 changed files with 11 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

4
studio/backend/tests/test_sd_cpp_args.py
View file

@ -54,7 +54,9 @@ def test_native_speed_flags():
assert native_speed_flags(None) == []
assert native_speed_flags("off") == []
assert native_speed_flags("") == []
assert native_speed_flags("default") == ["--diffusion-fa"]
# default now includes conv-direct: measured ~9% faster sampling on CPU
# (z-image Q8_0, 192 threads) with identical RSS and unchanged decode.
assert native_speed_flags("default") == ["--diffusion-fa", "--diffusion-conv-direct"]
assert native_speed_flags("max") == ["--diffusion-fa", "--diffusion-conv-direct"]
with pytest.raises(ValueError):
native_speed_flags("ludicrous")