deterministic-int8-llm-inference
Bitwise determinism across GPU INT8 kernels: pre-registered measurements of CUTLASS vs Triton disagreement in vLLM, its localization to the epilogue, fault injection showing tolerance-based conformance is blind to it, and power-of-two quantization scales as a deployable fix. Artifacts for arXiv:2608.13756.
- 0
- Python
- MIT
- 2 wk ago
