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

test: guard router benchmark equivalence

EricSpencer00855baac929bf
diff --git a/c/tests/bench_router_select.c b/c/tests/bench_router_select.c
index 6e79af07..44902ecf 100644
--- a/c/tests/bench_router_select.c
+++ b/c/tests/bench_router_select.c
@@ -2,8 +2,8 @@
*
* This measures only top-K expert selection. The input is prepared before each
* timed call, and both arms use the same finite score arrays and fallback path.
- * The full run uses 11 seeds and 2,000 repetitions per cell; correctness is
- * covered separately by test_router_nan.c.
+ * The full run uses 11 seeds and 2,000 repetitions per cell. Each cell also
+ * checks that both selectors return the same ordered expert ids before timing.
*/
#define main coli_glm_main_unused
#include "../colibri.c"
@@ -81,6 +81,19 @@ static double bench_new(const float *choice, int E, int K, int *idx, float *work
return samples[N_REPEAT/2];
}
+static int verify_selection(const float *choice, int E, int K, int layer){
+ int old_idx[32], new_idx[32]; float work[256];
+ select_old(choice,E,K,old_idx,layer);
+ memcpy(work,choice,(size_t)E*sizeof(float));
+ router_select_topk(work,E,K,new_idx,layer);
+ for(int kk=0;kk<K;kk++) if(old_idx[kk]!=new_idx[kk]){
+ fprintf(stderr,"selection mismatch: E=%d K=%d slot=%d old=%d new=%d\n",
+ E,K,kk,old_idx[kk],new_idx[kk]);
+ return 0;
+ }
+ return 1;
+}
+
int main(void){
const int cases[][2]={{64,8},{256,8},{256,32}};
const char *shapes[]={"random","peaked","plateau"};
@@ -95,6 +108,7 @@ int main(void){
for(int seed=0;seed<N_SEEDS;seed++){
seed_rng(0xA5A5A5A5u+(uint32_t)(seed*0x9E3779B9u));
fill_scores(choice,E,shape);
+ if(!verify_selection(choice,E,K,0)) return 2;
old_seed[seed]=bench_old(choice,E,K,idx);
new_seed[seed]=bench_new(choice,E,K,idx,work);
}
diff --git a/docs/experiments/sparse-routing-2026-09-13.json b/docs/experiments/sparse-routing-2026-09-13.json
index c94c7756..f088f83f 100644
--- a/docs/experiments/sparse-routing-2026-09-13.json
+++ b/docs/experiments/sparse-routing-2026-09-13.json
@@ -21,14 +21,14 @@
"samples": {
"tok_s": [
1000000.0,
- 250000.0,
- 14084.50704225352,
+ 200000.0,
+ 13333.333333333334,
1000000.0,
200000.0,
- 14084.50704225352,
+ 10752.68817204301,
1000000.0,
200000.0,
- 12987.012987012987
+ 10204.08163265306
]
},
"median_tok_s": 200000.0,
@@ -38,7 +38,7 @@
},
"evidence": {
"uri": "docs/experiments/sparse-routing-2026-09-13.txt",
- "sha256": "350dc67288a496f8eab469a1b4feca5619d8f169ad0cb0bd0a9d37f1e7abc7a2"
+ "sha256": "9815b4437e0beeb7a2731cf3deadd21f4bf8266cbfc7c83ac558556d7466c3bd"
}
},
"trial": {
@@ -53,10 +53,10 @@
76923.07692307692,
1000000.0,
250000.0,
- 71428.57142857143,
+ 58823.5294117647,
1000000.0,
333333.3333333333,
- 83333.33333333333
+ 62500.0
]
},
"median_tok_s": 333333.3333333333,
@@ -66,7 +66,7 @@
},
"evidence": {
"uri": "docs/experiments/sparse-routing-2026-09-13.txt",
- "sha256": "350dc67288a496f8eab469a1b4feca5619d8f169ad0cb0bd0a9d37f1e7abc7a2"
+ "sha256": "9815b4437e0beeb7a2731cf3deadd21f4bf8266cbfc7c83ac558556d7466c3bd"
}
},
"outcome": "improvement"
diff --git a/docs/experiments/sparse-routing-2026-09-13.txt b/docs/experiments/sparse-routing-2026-09-13.txt
index 81b83aa5..7ad32049 100644
--- a/docs/experiments/sparse-routing-2026-09-13.txt
+++ b/docs/experiments/sparse-routing-2026-09-13.txt
@@ -26,14 +26,14 @@ bench_router_select: old duplicate scan vs marked-score scan (median of 11 seeds
shape E K old ns/call new ns/call speedup
---------------------------------------------------------------------
random 64 8 1000 1000 1.00x
-random 256 8 4000 3000 1.33x
-random 256 32 71000 13000 5.46x
+random 256 8 5000 3000 1.67x
+random 256 32 75000 13000 5.77x
peaked 64 8 1000 1000 1.00x
peaked 256 8 5000 4000 1.25x
-peaked 256 32 71000 14000 5.07x
+peaked 256 32 93000 17000 5.47x
plateau 64 8 1000 1000 1.00x
plateau 256 8 5000 3000 1.67x
-plateau 256 32 77000 12000 6.42x
+plateau 256 32 98000 16000 6.13x
bench_router_select: done (lower ns is better; speedup = old/new)
Full DSA benchmark output (11 seeds x 2,000 repetitions per cell):