Skip to content

basal-rs v0.1.8

Choose a tag to compare

View all tags
@github-actions github-actions released this 10 Oct 14:55
· 4 commits to main since this release
v0.1.8
24ae958
This commit was signed with the committer’s verified signature.
nixuuu Damian Plewa
SSH Key Fingerprint: 2E0J6Rcr4ENJ7BzFmbIm+BXOUwlXDgj8yxBCerlPg6A
Verified
Learn about vigilant mode.

What's Changed

  • CUDA packages now embed 27 batch-invariant f16 GEMM tables for 10 GPU names, including 24 newly generated tables. Matching GPU/model/cuBLASLt 12.9.1 configurations can use them without searching for algorithms at startup.
  • Mini, 4.5B and max are covered on H100 PCIe, RTX 6000 Ada Generation, L40S, RTX A6000, A100-SXM4-80GB, L40 and RTX PRO 6000 Blackwell Server Edition.
  • Mini and 4.5B are covered on L4, A10 and RTX A5000.
  • Added a cloud runner and reports documenting table generation, batch invariance, FP32 comparisons and benchmarks.

Validation

All 24 new tables produced byte-identical exports across single/tree/budget batching on the checked cases. Each was compared with upstream FP32 on 44 examples and 900 questions. FP32 decision agreement is 44/44 for all models; on the larger set it is 899/900 for mini and 4.5B, and 900/900 for max. This does not imply bitwise equality with FP32.

Reports: initial campaign, additional GPUs.

Full Changelog: v0.1.7...v0.1.8

Assets 6
Loading