2 days ago
-
Sync latest vllm-fast, enable position pooling by default, document prefix/head cache usage
-
Document default acceleration flags and single-concurrency guidance
-
Pin the tested vLLM/PyTorch/Transformers stack used for the README run
-
Record local evaluation metadata and align model card with the release
-
Fix release metadata and evaluation wrapper: add SHA256SUMS/calibration.json, remove unsupported --fail-fast, correct run_metadata claims
7 days ago
-
Publish model and typed-decisions evaluation
