wmantly/turing-multi-gpu-llm-server
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

tuning: add stable-fast profile for 24/7 FlashAttention inference

Browse Source
This commit is contained in:
wmantly
2026-09-02 21:01:06 +00:00
parent 9e543f1291
commit e037f7a122
1 changed files with 14 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tuning/cmp-tune.conf
+14 -4
View File
@@ -1,13 +1,23 @@
# /etc/cmp-tune.conf - Profile-driven tuning for CMP 50HX / CMP 90HX # /etc/cmp-tune.conf - Profile-driven tuning for CMP 50HX / CMP 90HX
[efficient] [stable-fast]
power_w = 170 # Recommended for 24/7 continuous LLM FlashAttention load without Xid 13 errors
core_offset = 200 power_w = 210
mem_offset = 1000 core_offset = 125
mem_offset = 600
clock_max = 1950 clock_max = 1950
[efficient]
# Low heat, energy-efficient inference
power_w = 170
core_offset = 150
mem_offset = 600
clock_max = 1850
[insane] [insane]
# Max silicon benchmark profile (Not recommended for sustained multi-minute 100% prefill)
power_w = 220 power_w = 220
core_offset = 250 core_offset = 250
mem_offset = 1200 mem_offset = 1200
clock_max = 2050 clock_max = 2050