Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
tests: the Vulkan families run with speculation off, the cases that test it turn it on
JustVuggda974b331b01
diff --git a/c/tests/vulkan_engines.sh b/c/tests/vulkan_engines.shindex 2f302a61..8a31c2db 100755--- a/c/tests/vulkan_engines.sh+++ b/c/tests/vulkan_engines.sh@@ -44,6 +44,11 @@ cd "$(dirname "$0")/.."export VK_ICD_FILENAMES=${VK_ICD_FILENAMES:-/usr/share/vulkan/icd.d/lvp_icd.json}export COLI_NO_OMP_TUNE=1PY=${PY:-python3}+# Speculative decoding is on by default (MTP and prompt lookup, docs/speculative.md): its+# verifies change how many rows a forward dumps and when (the gate decides from measured+# times), which the families' row-by-row logits comparisons do not expect. They run with it+# off; the cases that test it turn it on themselves (Q38_MTP=1, COLI_LOOKUP=1, DRAFT=n).+export COLI_LOOKUP=${COLI_LOOKUP:-0} Q38_MTP=${Q38_MTP:-0}fail() { echo "FAIL: $*"; exit 1; }