ROCm/FBGEMM
Dec 2024 – Dec 2024
1 Month active
Languages Used
C++Python
Technical Skills
Asynchronous ProgrammingDatabase IntegrationMultithreadingPerformance OptimizationPyTorch
Marko Radmilac
PROFILE
Marko Radmilac
Developed and delivered a performance-focused feature for the ROCm/FBGEMM repository, enabling asynchronous initialization of RockDB SSD tensors to reduce training startup time for large-scale jobs. Leveraging C++ and Python, the implementation moved tensor initialization to a separate thread, allowing concurrent execution of other tasks and improving overall throughput. The work included adding synchronized getter and setter methods for SSD database access, ensuring thread safety and correctness when accessing in-memory weights. By integrating asynchronous programming, multithreading, and database integration, this contribution established a scalable foundation for future asynchronous initialization patterns within ROCm/FBGEMM’s training infrastructure.
Overall Statistics
Feature vs Bugs
100%Features
Repository Contributions
1Total
Bugs
0
Commits
1
Features
1
Lines of code
69
Activity Months1
Your Network
3175 peopleSame Organization
@meta.com3078
Peter RongMember
Zain RizviMember
Aaditya NairMember
Aahan AggarwalMember
Aliaksei AndreyeuMember
Arjun ChaturvediMember
Aaron FarberMember
Aaron PollackMember
Aaryaman SagarMember
Work History
December 2024
1 Commits • 1 Features
Dec 1, 2024December 2024 monthly summary for ROCm/FBGEMM focusing on key feature delivery and technical accomplishments. Overall: Delivered a performance-oriented feature that reduces training startup time by enabling asynchronous RockDB SSD tensor initialization, with proper synchronization to maintain correctness. This work improves TTFB for larger training jobs and establishes a foundation for more asynchronous initialization patterns in ROCm/FBGEMM.
1 Commits • 1 Features
Dec 1, 2024December 2024 monthly summary for ROCm/FBGEMM focusing on key feature delivery and technical accomplishments. Overall: Delivered a performance-oriented feature that reduces training startup time by enabling asynchronous RockDB SSD tensor initialization, with proper synchronization to maintain correctness. This work improves TTFB for larger training jobs and establishes a foundation for more asynchronous initialization patterns in ROCm/FBGEMM.
December 2024
Activity
Loading activity data...
Quality Metrics
Correctness90.0%
Maintainability80.0%
Architecture90.0%
Performance90.0%
AI Usage20.0%
Skills & Technologies
Programming Languages
C++Python
Technical Skills
Asynchronous ProgrammingDatabase IntegrationMultithreadingPerformance OptimizationPyTorch
Repositories Contributed To
Overview of all repositories you've contributed to across your timeline