The engineering notebook / 03
Follow the work upstream.
From a failing workload to a smaller reproduction, a fix, and a contribution other people can use.
Featured investigation
A faster kernel is only half the story.
The Q8_0 work connected layout, decode, repeat-prompt correctness, and allocation pressure. One optimization exposed several contracts worth testing.
Read the engineering story →Six contributions / retained upstream history
Q8_0 weight reorder
A faster layout for quantized token generation. 0988accf8
Correctness after the first prompt
Reorder-aware dequantization and full-VRAM failure handling. b1be68e8c
Vocabulary alignment
An assertion fix for unaligned vocabulary sizes. 788fcbc5d
When VRAM allocations fill host RAM
Direct Level Zero allocation for the multi-GPU memory path. 9ed6e19b9
BF16 on the DMMV path
Decode-kernel coverage for BF16 weights. 8cc67efcd
Use the native subgroup
K-quant DMMV work matched to the device subgroup. e3bb1add8
Dates identify commits in retained upstream ancestry at 9d77fa172. They are not new test dates. Current installed behavior and later upstream changes require separate checks. These fixes do not imply that users should manually patch a current build.