Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
test(vulkan): partial-dsk's prompts-only device-only Kimi case reads each device MLA layer's kv_b back once
JustVugg8d319deaaf2b
diff --git a/c/tests/vulkan_partial_dsk.sh b/c/tests/vulkan_partial_dsk.shindex 3537ab90..e8472224 100644--- a/c/tests/vulkan_partial_dsk.sh+++ b/c/tests/vulkan_partial_dsk.sh@@ -375,6 +375,10 @@ ptl_k3_dho() {ptl_k3 "$tag" "$c" "$k" COLI_VK_DENSE_HOST=0 "$@"local n want L rn=$(dho_dropped vk.log); want=$(ptl_k3_droppable "$k"); L=$(ptl_L kimi_k3 vk.log); r=$(dho_reloaded vk.log)+ # RELOAD_MLA=1 (prompts only: decode runs on the CPU, whose MLA attention reads kv_b by+ # row): each MLA layer on the device reads its kv_b back once, as in the dense-only family+ local reloads=0+ [ "${RELOAD_MLA:-0}" = 1 ] && reloads=$(sed -n 's/^\[VK\] kimi_k3 chain: [0-9]* layers on the device ([0-9]* KDA, \([0-9]*\) MLA.*/\1/p' vk.log | tail -1)[ "$k" = "$L" ] && want=$((want + 1)) # the head, with every layer[ "$n" = "$want" ] || { grep '^\[VK\]' vk.log; fail "$tag: $n host copies dropped, the $k layers have $want"; }if [ "$k" -lt "$L" ]; then@@ -383,7 +387,7 @@ ptl_k3_dho() {fiif [ -n "${FAULT_BACK:-}" ]; then[ "$r" -gt 0 ] && [ "$r" -le "$n" ] || { grep '^\[VK\]' vk.log; fail "$tag: $r matrices read back after the loss ($n on the device only)"; }- else [ "$r" = 0 ] || { grep '^\[VK\]' vk.log; fail "$tag: $r matrices read back from disk on a healthy run"; }; fi+ else [ "$r" = "${reloads:-0}" ] || { grep '^\[VK\]' vk.log; fail "$tag: $r matrices read back from disk on a healthy run, expected ${reloads:-0}"; }; fiecho " $n host copies dropped (the $k layers' $want), $r read back"}# ptl_k3_same <tag> <case> <a> <b> <env...>: the chain's logits under settings a and b the@@ -432,7 +436,7 @@ ptl_k3_cases() {# ---- the dense weights on the device onlyfor k in 1 3 6; do ptl_k3_dho "partial kimi_k3 device only, $k of 6 layers" long $k $O; doneptl_k3_dho "partial kimi_k3 device only, staged" long 3 $O COLI_VK_STAGED=1- CHAINMODE=2 ptl_k3_dho "partial kimi_k3 device only, prompts only" long 3 $O+ CHAINMODE=2 RELOAD_MLA=1 ptl_k3_dho "partial kimi_k3 device only, prompts only" long 3 $OFAULT_BACK=3 REBUILD=1 ptl_k3_dho "partial kimi_k3 device only, device lost mid-decode" long 3 $Optl_k3_fault "partial kimi_k3 device only, an upload failing in layer 3" long 3 $O COLI_VK_DENSE_HOST=0[ "$(dho_reloaded vk.log)" = 0 ] || { grep '^\[VK\]' vk.log; fail "partial kimi_k3 device only, an upload failing: read back from disk"; }