Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Merge remote-tracking branch 'origin/dev' into pr270
JustVuggd34461c8d8db
diff --cc c/Makefileindex db94a9a3,3f35ca2c..b19ee3e3--- a/c/Makefile+++ b/c/Makefile@@@ -166,7 -171,7 +171,7 @@@ elsPYTHON ?= python3endifCUDA_OBJ =- TEST_BINS = tests/test_json$(EXE) tests/test_st$(EXE) tests/test_tier$(EXE) tests/test_grammar$(EXE) tests/test_schema_gbnf$(EXE) tests/test_decode_batch$(EXE) tests/test_idot$(EXE) tests/test_i4_grouped$(EXE) tests/test_stops$(EXE) tests/test_kv_alloc$(EXE) tests/test_i4_acc512$(EXE) tests/test_compat_direct$(EXE) tests/test_pipe_block$(EXE)-TEST_BINS = tests/test_json$(EXE) tests/test_st$(EXE) tests/test_st_pread$(EXE) tests/test_tier$(EXE) tests/test_grammar$(EXE) tests/test_schema_gbnf$(EXE) tests/test_decode_batch$(EXE) tests/test_idot$(EXE) tests/test_i4_grouped$(EXE) tests/test_stops$(EXE) tests/test_topp$(EXE) tests/test_sample_nan$(EXE) tests/test_tok_o200k$(EXE) tests/test_kv_alloc$(EXE) tests/test_i4_acc512$(EXE) tests/test_compat_direct$(EXE) tests/test_dsa_select$(EXE) tests/test_logit_nan$(EXE)++TEST_BINS = tests/test_json$(EXE) tests/test_st$(EXE) tests/test_st_pread$(EXE) tests/test_tier$(EXE) tests/test_grammar$(EXE) tests/test_schema_gbnf$(EXE) tests/test_decode_batch$(EXE) tests/test_idot$(EXE) tests/test_i4_grouped$(EXE) tests/test_stops$(EXE) tests/test_topp$(EXE) tests/test_sample_nan$(EXE) tests/test_tok_o200k$(EXE) tests/test_kv_alloc$(EXE) tests/test_i4_acc512$(EXE) tests/test_compat_direct$(EXE) tests/test_dsa_select$(EXE) tests/test_logit_nan$(EXE) tests/test_pipe_block$(EXE)ifneq (,$(LINUX))TEST_BINS += tests/test_uring$(EXE)endif@@@ -326,12 -352,17 +352,20 @@@ tests/test_i4_acc512$(EXE): tests/test_tests/test_compat_direct$(EXE): tests/test_compat_direct.c compat.h$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)- tests/test_uring$(EXE): tests/test_uring.c glm.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h+ tests/test_dsa_select$(EXE): tests/test_dsa_select.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h+ $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)++ # bench_dsa_select is a microbenchmark (old qsort vs new quickselect partial-select, #356),+ # NOT a test gate -- intentionally absent from TEST_BINS. Build on demand: make tests/bench_dsa_select+ tests/bench_dsa_select$(EXE): tests/bench_dsa_select.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h+ $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)++ tests/test_uring$(EXE): tests/test_uring.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)+tests/test_pipe_block$(EXE): tests/test_pipe_block.c glm.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h+ $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)+test-c: $(TEST_BINS)$(PYTHON) tools/run_tests.py $(TEST_BINS)diff --cc c/colibri.cindex 6bd66f40,9528ee94..1f4aa5b9--- a/c/colibri.c+++ b/c/colibri.c@@@ -2192,13 -1618,8 +1635,13 @@@ static void *pipe_worker(void *arg)memory_order_acq_rel,memory_order_relaxed)){int L =atomic_load_explicit(&p->layer,memory_order_relaxed);int eid=atomic_load_explicit(&p->eids[i],memory_order_relaxed); /* AFTER winning CAS */- expert_load(p->m,L,eid,&p->m->ws[i],1); /* needed-now load: fatal on I/O error (matches serial path) */+ expert_load(p->m,L,eid,&p->m->ws[i],1,1); /* needed-now load: fatal on I/O error (matches serial path); demand=1: this IS moe()'s own miss path */atomic_store_explicit(&p->ready[i],1,memory_order_release);+ if(g_pipe_block){ /* wake a main thread parked in pipe_wait */+ pthread_mutex_lock(&p->mx);+ pthread_cond_broadcast(&p->cv_done);+ pthread_mutex_unlock(&p->mx);+ }}/* CAS failed → another worker advanced index (or gen advanced): re-loop */}@@@ -6142,14 -5419,11 +5453,16 @@@ int main(int argc, char **argv)0#endif;+ if(pipe_workers_imply_pipe(getenv("PIPE"),getenv("PIPE_WORKERS"),g_pipe)){+ g_pipe=1;+ fprintf(stderr,"[PIPE] PIPE_WORKERS is set — enabling the async pipe (PIPE=1 implied; set PIPE=0 to override)\n");+ }g_pipe_nw = getenv("PIPE_WORKERS")?atoi(getenv("PIPE_WORKERS")):8; /* I/O worker threads */if(g_pipe_nw<1) g_pipe_nw=1;+ g_pipe_block = getenv("COLI_PIPE_BLOCK")?atoi(getenv("COLI_PIPE_BLOCK")):0; /* blocking pipe_wait (default: spin) */g_direct = getenv("DIRECT")?atoi(getenv("DIRECT")):0;+ { const char *dh=getenv("COLI_DISKCLASS_WINDOW"); /* DISK-CLASS recency window, see its declaration */+ if(dh){ g_direct_heat_ticks=(uint32_t)strtoul(dh,NULL,10); g_direct_heat_explicit=1; } }g_uring = getenv("URING")?atoi(getenv("URING")):0;if(g_uring){#ifdef __linux__