Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

tests: glm53's pin-branch harness runs with the tier off

JustVugg2074688e7c38
diff --git a/c/tests/vulkan_engines.sh b/c/tests/vulkan_engines.sh
index 17ab9cde..606a4d11 100755
--- a/c/tests/vulkan_engines.sh
+++ b/c/tests/vulkan_engines.sh
@@ -2124,8 +2124,9 @@ family_glm_chain() {
CHAIN_SERVE_SLOTS=2 CHAIN_SERVE_DIALECT=numeric $PY tests/vulkan_chain_serve.py ./glm53 glm53_serve GLM53_BITS=32 KV_SLOTS=2
COLI_VK_CHAIN=2 CHAIN_SERVE_DIALECT=numeric $PY tests/vulkan_chain_serve.py ./glm53 glm53_serve GLM53_BITS=32
# a pin restored over rows another branch rewrote: the KDA state goes up from the pin,
- # the MLA rows' watermark comes down
- COLI_VULKAN=1 COLI_VK_CHAIN=1 COLI_USAGE=$PWD/chain.usage COLI_VK_TIER_SYNC=1 COLI_VK_TIER_BALANCE=0 $PY tests/glm53_pin_branch_harness.py --binary ./glm53 --fixture glm53_mm_tiny
+ # the MLA rows' watermark comes down (the tier off: two engine processes compared bit
+ # for bit, see vulkan_partial_glm.sh)
+ COLI_VULKAN=1 COLI_VK_CHAIN=1 COLI_USAGE=$PWD/chain.usage COLI_VK_TIER=0 $PY tests/glm53_pin_branch_harness.py --binary ./glm53 --fixture glm53_mm_tiny
unset OMP_NUM_THREADS CAP_RAISE
}
@@ -3716,7 +3717,7 @@ kv_split_glm53() {
CHAIN_SERVE_EXPECT="$E" CHAIN_SERVE_DIALECT=numeric $PY tests/vulkan_chain_serve.py ./glm53 glm53_serve GLM53_BITS=32 $K
CHAIN_SERVE_EXPECT="$E" CHAIN_SERVE_DIALECT=numeric $PY tests/vulkan_chain_serve.py ./glm53 glm53_serve GLM53_BITS=4 COLI_VK_CHAIN_ROWS=3 $K
CHAIN_SERVE_EXPECT="$E" CHAIN_SERVE_SLOTS=2 CHAIN_SERVE_DIALECT=numeric $PY tests/vulkan_chain_serve.py ./glm53 glm53_serve GLM53_BITS=32 KV_SLOTS=2 $K
- env $K COLI_VULKAN=1 COLI_VK_CHAIN=1 COLI_USAGE=$PWD/chain.usage COLI_VK_TIER_SYNC=1 COLI_VK_TIER_BALANCE=0 $PY tests/glm53_pin_branch_harness.py --binary ./glm53 --fixture glm53_mm_tiny
+ env $K COLI_VULKAN=1 COLI_VK_CHAIN=1 COLI_USAGE=$PWD/chain.usage COLI_VK_TIER=0 $PY tests/glm53_pin_branch_harness.py --binary ./glm53 --fixture glm53_mm_tiny
}
}
kv_split_glm53_sanitize() {
diff --git a/c/tests/vulkan_partial_glm.sh b/c/tests/vulkan_partial_glm.sh
index 77c83e41..ed57ee4f 100644
--- a/c/tests/vulkan_partial_glm.sh
+++ b/c/tests/vulkan_partial_glm.sh
@@ -376,8 +376,11 @@ ptl_family_glm() {
CHAIN_SERVE_EXPECT='glm53: [0-9]+ dense matrices on the device only.*the 3 of 6 layers on the device' COLI_VK_DENSE_HOST=0 \
CHAIN_SERVE_DIALECT=numeric $PY tests/vulkan_chain_serve.py ./glm53 glm53_l6_serve GLM53_BITS=32 COLI_VK_CHAIN_LAYERS=3
unset CHAIN_SERVE_EXPECT
- # a pin restored over rows another branch rewrote, 2 of 4 layers on the device
- COLI_VULKAN=1 COLI_VK_CHAIN=1 COLI_VK_CHAIN_LAYERS=2 COLI_USAGE=$PWD/chain.usage COLI_VK_TIER_SYNC=1 COLI_VK_TIER_BALANCE=0 $PY tests/glm53_pin_branch_harness.py --binary ./glm53 --fixture glm53_mm_tiny
+ # a pin restored over rows another branch rewrote, 2 of 4 layers on the device. The
+ # harness compares two engine processes bit for bit, so the tier is off: which experts
+ # a device computes depends on each process's history, and a device expert's f32 sum
+ # can differ from the CPU's in its last bit (CI's runners did, at one position).
+ COLI_VULKAN=1 COLI_VK_CHAIN=1 COLI_VK_CHAIN_LAYERS=2 COLI_USAGE=$PWD/chain.usage COLI_VK_TIER=0 $PY tests/glm53_pin_branch_harness.py --binary ./glm53 --fixture glm53_mm_tiny
rm -rf glm_tiny_shx glm53_l6 glm53_l6s-i4 glm53_l6_serve chain.usage
rm -f ptl-probe.err ptl-ram*.log same-*.log same-*.tok same-*.f32
unset OMP_NUM_THREADS CAP_RAISE