Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
fix(build): clean up Clang compiler warnings
trigger2k20e79aee823a4c
diff --git a/c/deepseek_v4.c b/c/deepseek_v4.cindex 1cd2e08f..4052ed8d 100644--- a/c/deepseek_v4.c+++ b/c/deepseek_v4.c@@ -2726,8 +2726,6 @@ int coli_v4_attention_window_batch_ref(#ifdef COLI_V4_GPU_TIERint gpu_batch = coli_v4_gpu_attn_batch_wanted() && batch > 1;-#else- int gpu_batch = 0;#endif/* Whole-chunk GPU projections for the compressor and the indexer's* compressor; the per-token state advance stays on the CPU. NULL means@@ -3759,7 +3757,9 @@ int coli_v4_indexer_select_batch(ColiDeepSeekV4Indexer *state, int *indices,const uint16_t *raw_weights = value(state->weights, "attn.indexer.weights_proj.weight", NULL);size_t qn = (size_t)heads * dimension;+#ifdef COLI_V4_GPU_TIERsize_t cols = (size_t)wq.columns;+#endif#define ALIGN32(n) (((size_t)(n) + 31) & ~(size_t)31)size_t sz_queries = ALIGN32((size_t)batch * qn * sizeof(float));@@ -3769,8 +3769,6 @@ int coli_v4_indexer_select_batch(ColiDeepSeekV4Indexer *state, int *indices,size_t sz_stoken = ALIGN32((size_t)need * sizeof(int));size_t sz_scores = ALIGN32((size_t)need * max_count * sizeof(float));size_t sz_ranked = ALIGN32((size_t)max_count * sizeof(IndexScore));- size_t sz_scales = ALIGN32((size_t)dimension / 32);- size_t sz_qdq = ALIGN32((size_t)dimension * sizeof(float));#ifdef COLI_V4_GPU_TIERsize_t sz_xq = (need <= 1024) ? ALIGN32((size_t)need * cols * sizeof(float)) : 0;size_t sz_yq = (need <= 1024) ? ALIGN32((size_t)need * qn * sizeof(float)) : 0;@@ -3780,7 +3778,7 @@ int coli_v4_indexer_select_batch(ColiDeepSeekV4Indexer *state, int *indices,#endifsize_t total_scratch = sz_queries + sz_sq + sz_head_weights + sz_scounts + sz_stoken +- sz_scores + sz_ranked + sz_scales + sz_qdq + sz_xq + sz_yq + sz_xs + 256;+ sz_scores + sz_ranked + sz_xq + sz_yq + sz_xs + 256;char *scratch_ptr = (char *)indexer_scratch_alloc(state, total_scratch);if (!scratch_ptr || !raw_weights) {@@ -3797,8 +3795,6 @@ int coli_v4_indexer_select_batch(ColiDeepSeekV4Indexer *state, int *indices,int *stoken = (int *)scratch_ptr; scratch_ptr += sz_stoken;float *scores = (float *)scratch_ptr; scratch_ptr += sz_scores;IndexScore *ranked = (IndexScore *)scratch_ptr; scratch_ptr += sz_ranked;- uint8_t *scales = (uint8_t *)scratch_ptr; scratch_ptr += sz_scales;- float *qdq = (float *)scratch_ptr; scratch_ptr += sz_qdq;int result = 0;/* Query projection. Preferred: host fp8 activation quantization (the@@ -4977,11 +4973,15 @@ static int moe_token_pipeline(float *output,#ifdef COLI_V4_EXPERIMENTAL_DUAL_EXPERT_LOADERColiExpertView *views = malloc((size_t)selected * sizeof(*views));+#ifdef COLI_V4_GPU_TIERint gpu_compute = 0;+#endifif (!views) result = -1;if (!result) {memset(views, 0, (size_t)selected * sizeof(*views));+#ifdef COLI_V4_GPU_TIERgpu_compute = 1;+#endiffor (int current = 0; !result && current < selected; current++) {int slot = current % dual_loader_lanes();if (!loader_active[slot] ||@@ -5001,9 +5001,11 @@ static int moe_token_pipeline(float *output,coli_v4_gpu_expert_attach(store, &views[current]);}#endif+#ifdef COLI_V4_GPU_TIERif (!views[current].gate.gpu || !views[current].up.gpu ||!views[current].down.gpu)gpu_compute = 0;+#endifint next = current + dual_loader_lanes();if (next < selected) {diff --git a/c/expert_store_registry.c b/c/expert_store_registry.cindex 526d8cd9..b646802b 100644--- a/c/expert_store_registry.c+++ b/c/expert_store_registry.c@@ -17,10 +17,22 @@* and call register() from their own constructors, so these symbols must* survive LTO inlining. Without this, gcc drops them when the only caller in* the current link is inside the same LTO set. */-#if defined(__GNUC__)-#define COLI_ESR_EXPORT __attribute__((externally_visible, used))+#if defined(__has_attribute)+# if __has_attribute(externally_visible)+# define COLI_ESR_EXTERNALLY_VISIBLE __attribute__((externally_visible))+# else+# define COLI_ESR_EXTERNALLY_VISIBLE+# endif+#elif defined(__GNUC__)+# define COLI_ESR_EXTERNALLY_VISIBLE __attribute__((externally_visible))#else-#define COLI_ESR_EXPORT+# define COLI_ESR_EXTERNALLY_VISIBLE+#endif++#if defined(__GNUC__) || defined(__clang__)+# define COLI_ESR_EXPORT COLI_ESR_EXTERNALLY_VISIBLE __attribute__((used))+#else+# define COLI_ESR_EXPORT#endif/* The built-in on-disk/mmap backend, defined in the COLI_V4_UNIT_EXPERT_STORE_AUTOdiff --git a/c/glm53.c b/c/glm53.cindex 13dc21cd..048be622 100644--- a/c/glm53.c+++ b/c/glm53.c@@ -1340,11 +1340,11 @@ static void expert_mats(const GModel *m, const Slot *slot, Mat *gate, Mat *up, Mconst int hidden = m->c.hidden, inter = m->c.moe_inter;const Mat shape[3] = {{ 4, NULL, NULL, slot->piece[0], (const float *)slot->piece[1],- inter, hidden, 64 },+ inter, hidden, 64, NULL },{ 4, NULL, NULL, slot->piece[2], (const float *)slot->piece[3],- inter, hidden, 64 },+ inter, hidden, 64, NULL },{ 4, NULL, NULL, slot->piece[4], (const float *)slot->piece[5],- hidden, inter, 64 },+ hidden, inter, 64, NULL },};*gate = shape[0]; *up = shape[1]; *down = shape[2];}@@ -1987,12 +1987,13 @@ static void model_load(GModel *m, const char *dir) {static float *forward_span(GModel *m, GSession *s, const int *tokens, int n,const float *vision, int n_vision) {const Cfg *c = &m->c;+ const int H = c->hc_mult;const int start = s->filled; /* NON 'base': nel ciclo dei layer e' gia' preso */if (start + n > s->cap) {fprintf(stderr, "contesto esaurito: %d posizioni su %d\n", start + n, s->cap);exit(1);}- const int H = c->hc_mult, D = c->hidden;+ const int D = c->hidden;float *streams = malloc((size_t)n * H * D * sizeof(float));float *next = malloc((size_t)n * H * D * sizeof(float));/* l'embedding entra replicato in ognuno degli H flussi residui; sui token@@ -3229,7 +3230,8 @@ static int glm53_edge_embed(void *engine_impl, const ColiEdgeEmbedRequest *requestatic int glm53_edge_final(const Glm53EdgeEngine *engine, const float *streams,float *logits, float *collapsed, float *normed) {const Cfg *c = &engine->model.c;- const int H = c->hc_mult, D = c->hidden;+ const int H = c->hc_mult;+ const int D = c->hidden;for (int d = 0; d < D; d++) {float sum = 0.0f;for (int h = 0; h < H; h++) sum += streams[(size_t)h * D + d];@@ -3284,7 +3286,7 @@ static int glm53_edge_select(void *engine_impl, const ColiEdgeSelectRequest *reqreturn coli_edge_adapter_error(error, error_size,"GLM-5.3 Edge select got bad arguments");const Cfg *c = &engine->model.c;- const int H = c->hc_mult, D = c->hidden;+ const int D = c->hidden;const size_t width = engine->state_width;if (request->input_bytes < (size_t)request->rows * width * sizeof(float) ||request->token_capacity < request->rows)diff --git a/c/qwen38.c b/c/qwen38.cindex d7cdd67a..3dab1075 100644--- a/c/qwen38.c+++ b/c/qwen38.c@@ -2606,10 +2606,16 @@ static int qwen38_edge_select(void *engine_impl,}static const ColiEdgeAdapter qwen38_edge_adapter={- sizeof(ColiEdgeAdapter),COLI_EDGE_ABI_VERSION,"qwen38",- qwen38_edge_engine_open,qwen38_edge_engine_destroy,- qwen38_edge_tokenize,qwen38_edge_detokenize,- qwen38_edge_embed,qwen38_edge_select,{0}+ .struct_size = sizeof(ColiEdgeAdapter),+ .abi_version = COLI_EDGE_ABI_VERSION,+ .engine_id = "qwen38",+ .engine_open = qwen38_edge_engine_open,+ .engine_destroy = qwen38_edge_engine_destroy,+ .tokenize = qwen38_edge_tokenize,+ .detokenize = qwen38_edge_detokenize,+ .embed = qwen38_edge_embed,+ .select = qwen38_edge_select,+ .reserved_fn = {0}};int coli_qwen38_edge_adapter_register(void) {diff --git a/c/tests/test_expert_store_ops.c b/c/tests/test_expert_store_ops.cindex 23c25e05..04f3c4d6 100644--- a/c/tests/test_expert_store_ops.c+++ b/c/tests/test_expert_store_ops.c@@ -67,7 +67,7 @@ int main(void) {mock_lookup, mock_release, mock_prefetch, mock_stats, mock_destroy};MockState state = {0};- ColiExpertStore store = {&ops, &state};+ ColiExpertStore store = {&ops, &state, NULL};ColiExpertView view;ColiExpertView other;ColiExpertKey key = {7, 19};