Motivation
A reference implementation for Deepseek-r1. The implementation should include the dataset (and sampled versions if needed for lower concurrency runs), as well as accuracy checks and allowed optimizations.
Proposed Solution
Possibly using the MLPerf inference setup and tweaking for different concurrency points. Concurrency=1 would require more thought.
Alternatives Considered
No response
Additional Context
No response
Motivation
A reference implementation for Deepseek-r1. The implementation should include the dataset (and sampled versions if needed for lower concurrency runs), as well as accuracy checks and allowed optimizations.
Proposed Solution
Possibly using the MLPerf inference setup and tweaking for different concurrency points. Concurrency=1 would require more thought.
Alternatives Considered
No response
Additional Context
No response