Skip to content

bench: add MTP options to ds4-bench#205

Open
tnhnyc wants to merge 1 commit into
antirez:mainfrom
tnhnyc:codex/mtp-bench-pr
Open

bench: add MTP options to ds4-bench#205
tnhnyc wants to merge 1 commit into
antirez:mainfrom
tnhnyc:codex/mtp-bench-pr

Conversation

@tnhnyc
Copy link
Copy Markdown

@tnhnyc tnhnyc commented May 19, 2026

Summary

  • add --mtp, --mtp-draft, and --mtp-margin to ds4-bench
  • route the benchmark generation loop through the existing greedy MTP path when draft depth is greater than one
  • document a minimal normal-vs-MTP benchmark invocation and add an opt-in DS4_MTP_VERIFY_TIMING breakdown for verifier profiling

Overlap check

I checked the current issue/PR list before opening this. Related discussions exist around speculative decoding (#80, #63) and the large CUDA/backend PR #187 also mentions speculative-decoding plumbing, but I did not find a focused PR that makes ds4-bench directly benchmark the existing MTP path on main. This PR is intentionally small and measurement-oriented; it does not change default runtime behavior.

Testing

  • make
  • git diff --check origin/main..HEAD
  • ./ds4-bench --help shows the new MTP options

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant