fix: harden Lunar Lake Gemma 4 attention and quantized CPU copies - #172

Draft

marksverdhei wants to merge 21 commits into

htfrom

fix/xe2-attention-safety

Commits on Sep 5, 2026

fix(cpu): reject unsafe quantized copy layouts and correct row sizes
marksverdhei
committed
e25e52c
fix(vulkan): bound Gemma 4 tuning to validated Lunar Lake workloads
marksverdhei
committed
5ea6e1e
test(xe2): add pinned models and end-to-end validation harness
marksverdhei
committed
cab602c
test(xe2): retain per-stage provenance and reject invalid benchmark metrics
marksverdhei
committed
926d541
feat(xe2): add bounded serving presets and soak both profiles
marksverdhei
committed
899e994
test(xe2): honor Gemma suppressed-token logits in parity checks
marksverdhei
committed
48d5498
test(xe2): read suppression metadata through portable GGUF API
marksverdhei
committed
867b8e8
test(xe2): retain full parity diagnostics without masking failures
marksverdhei
committed
7baac2a
test(xe2): account for UMA GPU memory and require MTP engagement
marksverdhei
committed
8896fc8
test(xe2): compare MTP output with target-only decoding
marksverdhei
committed
5eae094
fix(xe2): bound serving prompt cache on the 32GB UMA device
marksverdhei
committed
4651598
test(xe2): validate instruction models with their chat template
marksverdhei
committed
d194f44
test(xe2): verify speculative decoding on a longer code response
marksverdhei
committed
ede5b05
test(xe2): bound full-model token distributions and retain strict KV checks
marksverdhei
committed
e1dcac5
test(xe2): cancel streams after an observed generation token
marksverdhei
committed
1242516
fix(xe2): tolerate telemetry faults and preserve interrupted run evidence
marksverdhei
committed
507d3e8
docs(xe2): record parity results and the failing FA-on control
marksverdhei
committed
a388e16
fix(xe2): bound per-slot checkpoints and report warm UMA memory trends
marksverdhei
committed
ab22a3a
test(xe2): check known answers throughout the server soak
marksverdhei
committed
dcbff80
test(xe2): separate cache churn from warm memory growth
marksverdhei
committed
decccf6
docs(xe2): record completed hour of sustained-load validation
marksverdhei
committed
f41b69b