Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
qwen36 tier: the re-plan orders are total, so ties pick the same victim on every libc
kreuzzelge4a2e3c3b077
diff --git a/c/qwen36_tier.c b/c/qwen36_tier.cindex bb7aca65..e85abc6c 100644--- a/c/qwen36_tier.c+++ b/c/qwen36_tier.c@@ -1263,16 +1263,22 @@ static void replan_drain_locked(int max){* cudaFree + cudaMalloc, which synchronise the device under the async groups */static const uint32_t *g_rp_cnt; static int g_rp_base; /* counts indexed by slot - base */-static int cmp_rp_cand(const void *a,const void *b){ /* count desc */- uint32_t fa=g_rp_cnt[*(const int*)a-g_rp_base], fb=g_rp_cnt[*(const int*)b-g_rp_base];- return fa<fb ? 1 : fa>fb ? -1 : 0;+/* Both orders are total (the slot index breaks the last tie): qsort is not+ * stable, and glibc's and msvcrt's disagree on ties, so without it the same+ * counts would pick different victims on Linux and Windows. */+static int cmp_rp_cand(const void *a,const void *b){ /* count desc, index asc */+ int ia=*(const int*)a, ib=*(const int*)b;+ uint32_t fa=g_rp_cnt[ia-g_rp_base], fb=g_rp_cnt[ib-g_rp_base];+ if(fa!=fb) return fa<fb ? 1 : -1;+ return ia<ib ? -1 : ia>ib ? 1 : 0;}-static int cmp_rp_vict(const void *a,const void *b){ /* count asc, heat asc */+static int cmp_rp_vict(const void *a,const void *b){ /* count asc, heat asc, index asc */int ia=*(const int*)a, ib=*(const int*)b;uint32_t fa=g_rp_cnt[ia-g_rp_base], fb=g_rp_cnt[ib-g_rp_base];if(fa!=fb) return fa<fb ? -1 : 1;uint32_t ha=G.slot[ia].heat, hb=G.slot[ib].heat;- return ha<hb ? -1 : ha>hb ? 1 : 0;+ if(ha!=hb) return ha<hb ? -1 : 1;+ return ia<ib ? -1 : ia>ib ? 1 : 0;}int qt_replan(int layer,const uint32_t *counts,int max_swaps){if(!G.on||!counts||G_fp8_stream||max_swaps<=0||layer>=G.nl) return 0;