minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD] Increase gfx950 DSA model indexer topk_transform kernel occupancy (#35628)

Browse Source
This commit is contained in:
wufann
2026-08-27 22:20:52 -07:00
committed by GitHub
parent 43b5a57dbb
commit cce0b53466
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/kernels/aot/setup_rocm.py
+3 -2
View File
@@ -87,8 +87,9 @@ fp8_macro = (
# Dynamic shared-memory budget for the TopK kernels.
# - gfx942 (MI300/MI325): LDS is typically 64KB per workgroup -> keep dynamic smem <= ~48KB
# (leaves room for static shared allocations in the kernel).
# - gfx95x (MI350) and gfx1250: LDS is larger -> allow the original 128KB dynamic smem.
topk_dynamic_smem_bytes = 48 * 1024 if amdgpu_target == "gfx942" else 32 * 1024 * 4
# - gfx95x (MI350) and gfx1250: LDS is larger. Large dynamic budget wastes LDS
# and pins occupancy to 1 block/CU. Keep it small (40KB) for better occupancy.
topk_dynamic_smem_bytes = 48 * 1024 if amdgpu_target == "gfx942" else 40 * 1024
hipcc_flags = [
"-DNDEBUG",