ColPali reproduction with Triton MaxSim
This release completes a scoped, reproducible ColPali implementation and validates a fused Triton MaxSim kernel on one NVIDIA L4 (g6.xlarge).
Highlights
- 2.27× geometric-mean speedup over the PyTorch reference across the 12-case benchmark matrix
- 12/12 benchmark wins, ranging from 1.57× to 4.31×
- Five mandatory CUDA parity cases passed with zero skips
- 276 GPU tests passed on NVIDIA L4; 664 CPU tests passed in the release validation suite
- Largest measured incremental allocation: 41,798,144 bytes for PyTorch versus 1,536 bytes for Triton
- Winning Triton configuration:
tq16_td128_w4_s2 - Valid Nsight Compute report captured
The complete methodology, limitations, exact commands, environment, and result tables are in the final report. Machine-readable GPU results are in phase_11_gpu_results.json.
Reproducibility
- Release commit:
22c7bc2b337be523ece1b2b8cdd803b254b11744 - GPU benchmark source commit:
0f9543e30e49fc30de8c476b2767e1c29073602b - Hardware: NVIDIA L4 on AWS
g6.xlarge - The attached wheel and source archive were built directly from the release commit.
This project is a focused reproduction and engineering study, not an official ColPali implementation.