minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 10 Packages Projects Releases Wiki Activity

[CI][RFC] Replace black-jupyter with ruff-format (#37210)

Browse Source 
Co-authored-by: Alison Shao <a.shao@wustl.edu>
This commit is contained in:
Alex Nails
2026-09-02 19:46:08 -07:00
committed by GitHub
co-authored by Alison Shao
parent 2641e427be
commit 28262c20df
1411 changed files with 7766 additions and 8176 deletions
Download Patch File Download Diff File Expand all files Collapse all files
benchmark/lean_kernel_sweep.py
+5 -3
View File
@@ -137,19 +137,21 @@ def main():
# b32 x 128K on the 8-KV-head config exceeds the microbench's single
# contiguous KV tensor (faults the GPU); real serving uses a paged pool.
if H_KV == 8 and B == 32 and S == 131072:
print(f"{B:>5} {S//1024:>5}K {'skipped (contiguous-KV limit)':>30}")
print(
f"{B:>5} {S // 1024:>5}K {'skipped (contiguous-KV limit)':>30}"
)
rows.append(f"{name},{H_Q},{H_KV},{B},{S},,,,skip,")
continue
try:
std, lean, cos, gate = run(H_Q, H_KV, B, S)
except torch.cuda.OutOfMemoryError:
torch.cuda.empty_cache()
print(f"{B:>5} {S//1024:>5}K {'OOM':>9}")
print(f"{B:>5} {S // 1024:>5}K {'OOM':>9}")
rows.append(f"{name},{H_Q},{H_KV},{B},{S},,,,OOM,")
continue
sp = std / lean
print(
f"{B:>5} {S//1024:>5}K {std:>9.3f} {lean:>9.3f} {sp:>7.2f}x {('ON' if gate else 'OFF'):>5} {cos:>7.4f}"
f"{B:>5} {S // 1024:>5}K {std:>9.3f} {lean:>9.3f} {sp:>7.2f}x {('ON' if gate else 'OFF'):>5} {cos:>7.4f}"
)
rows.append(
f"{name},{H_Q},{H_KV},{B},{S},{std:.4f},{lean:.4f},{sp:.4f},{cos:.4f},{int(gate)}"