[AMD] Update search-space for ATOM DSR1 configs#699
[AMD] Update search-space for ATOM DSR1 configs#699cquil11 merged 8 commits intoSemiAnalysisAI:mainfrom
Conversation
|
The original search space was configured with limited settings and didn’t fully cover the Pareto frontier, especially for the newly added MTP cases. We’ve updated the configuration so we can now observe the full curve in the full sweep. |
|
/sweep test-config --config-files .github/configs/amd-master.yaml --runner-config .github/configs/runners.yaml --config-keys dsr1-fp4-mi355x-atom-mtp dsr1-fp4-mi355x-atom |
|
@seungrokj Kicking off a sweep. Run: https://github.com/InferenceMAX/InferenceMAX/actions/runs/21997709058 |
|
@pr-claude review plz |
revise the search space config
Updated to fullscan
Reduce the search space to depends less on the compute as most of the pareto frontier fall under TP=8.
79b1775 to
861c918
Compare
|
@claude add perf changelog.yaml to dsr1-fp4-mi355x-atom & dsr1-fp4-mi355x-atom-mtp and add an description of change and then review the PR plz. lgtm to me but want ur second eye claude please |
|
I'll analyze this and get back to you. |
|
@claude add perf changelog.yaml to dsr1-fp4-mi355x-atom & dsr1-fp4-mi355x-atom-mtp and add an description of change and then review the PR plz. lgtm to me but want ur second eye claude please |
|
I'll analyze this and get back to you. |
|
@seungrokj @ChuanLi1101 Was it your intention to comment out TP 4 scenarios?? |
|
Revise the changes to make sure we get the pareto's frontier.
update the search space
ChuanLi1101
left a comment
There was a problem hiding this comment.
Reduce the search space to keep only the pareto's frontier points.

Summary
conc-startfrom 256 → 4 to enable full concurrency sweep