FP16xINT4 LLM inference kernel that can achieve near-ideal ~4x speedups up to medium batchsizes of 16-32 tokens.
Primary LanguagePythonApache License 2.0Apache-2.0
No issues in this repository yet.