Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
chore(glm53): standardize user-facing diagnostics in English
trigger2k2058498af5b67c
diff --git a/c/glm53.c b/c/glm53.cindex 3af0f2cd..ca89372f 100644--- a/c/glm53.c+++ b/c/glm53.c@@ -291,8 +291,8 @@ static void load_cfg(Cfg *c, const char *snap) {} else {jval *kinds = json_get(tc, "mlp_layer_types");if (!kinds || kinds->t != J_ARR) {- fprintf(stderr, "config.json: serve first_k_dense_replace "- "oppure mlp_layer_types\n");+ fprintf(stderr, "config.json: requires first_k_dense_replace "+ "or mlp_layer_types\n");exit(1);}c->first_dense = kinds->len;@@ -724,9 +724,9 @@ static const float *load_f32(GModel *m, const char *fmt, ...) {char name[512];va_list args; va_start(args, fmt); vsnprintf(name, sizeof(name), fmt, args); va_end(args);st_tensor *t = st_find(&m->S, name);- if (!t) { fprintf(stderr, "manca il tensore %s\n", name); exit(1); }+ if (!t) { fprintf(stderr, "missing tensor %s\n", name); exit(1); }float *buffer = malloc((size_t)t->numel * sizeof(float));- if (!buffer) { fprintf(stderr, "OOM su %s\n", name); exit(1); }+ if (!buffer) { fprintf(stderr, "OOM on %s\n", name); exit(1); }st_read_f32_cap(&m->S, name, buffer, t->numel, 0);return buffer;}@@ -741,7 +741,7 @@ static int glm53_dense_bits(void) {const char *setting = getenv("GLM53_BITS");cached = setting ? atoi(setting) : 4;if (cached != 4 && cached != 8 && cached != 32) {- fprintf(stderr, "GLM53_BITS=%s: valori ammessi 4, 8, 32\n", setting);+ fprintf(stderr, "GLM53_BITS=%s: allowed values are 4, 8, 32\n", setting);exit(1);}return cached;@@ -792,7 +792,7 @@ static Mat quantize_loaded(float *buffer, int rows, int columns) {const int groups = columns / 64;uint8_t *packed = malloc((size_t)rows * ((columns + 1) / 2));float *step = malloc((size_t)rows * groups * sizeof(float));- if (!packed || !step) { fprintf(stderr, "OOM quantizzando %dx%d\n", rows, columns); exit(1); }+ if (!packed || !step) { fprintf(stderr, "OOM quantizing %dx%d\n", rows, columns); exit(1); }quantize_i4_grouped(buffer, packed, step, rows, columns, 64);free(buffer);mat.fmt = 4; mat.q4 = packed; mat.s = step; mat.gs = 64;@@ -802,7 +802,7 @@ static Mat quantize_loaded(float *buffer, int rows, int columns) {* di 64, che capita sulle proiezioni piccole dell'indexer. */int8_t *level = malloc((size_t)rows * columns);float *step = malloc((size_t)rows * sizeof(float));- if (!level || !step) { fprintf(stderr, "OOM quantizzando %dx%d\n", rows, columns); exit(1); }+ if (!level || !step) { fprintf(stderr, "OOM quantizing %dx%d\n", rows, columns); exit(1); }quantize_rows(buffer, level, step, rows, columns, 8);free(buffer);mat.fmt = 1; mat.q8 = level; mat.s = step;@@ -833,16 +833,16 @@ static void absorb_kvb(GModel *m, GLayer *l, const char *name) {const Cfg *c = &m->c;const int H = c->n_heads, QK = c->qk_nope, V = c->v_head, L = c->kv_lora;st_tensor *t = st_find(&m->S, name);- if (!t) { fprintf(stderr, "manca %s\n", name); exit(1); }+ if (!t) { fprintf(stderr, "missing %s\n", name); exit(1); }if (t->numel != (int64_t)H * (QK + V) * L) {- fprintf(stderr, "%s: %lld valori, attesi %lld per %d teste\n", name,+ fprintf(stderr, "%s: %lld values, expected %lld for %d heads\n", name,(long long)t->numel, (long long)H * (QK + V) * L, H);exit(1);}float *whole = malloc((size_t)t->numel * sizeof(float));float *kt = malloc((size_t)H * L * QK * sizeof(float));float *vv = malloc((size_t)H * V * L * sizeof(float));- if (!whole || !kt || !vv) { fprintf(stderr, "OOM su %s\n", name); exit(1); }+ if (!whole || !kt || !vv) { fprintf(stderr, "OOM on %s\n", name); exit(1); }st_read_f32_cap(&m->S, name, whole, t->numel, 1);for (int h = 0; h < H; h++) {@@ -864,7 +864,7 @@ static Mat load_mat(GModel *m, const char *fmt, ...) {char name[512];va_list args; va_start(args, fmt); vsnprintf(name, sizeof(name), fmt, args); va_end(args);st_tensor *t = st_find(&m->S, name);- if (!t) { fprintf(stderr, "manca la matrice %s\n", name); exit(1); }+ if (!t) { fprintf(stderr, "missing matrix %s\n", name); exit(1); }Mat mat; memset(&mat, 0, sizeof(mat));/* Gia' quantizzato nel checkpoint: si prende com'e', senza passare per@@ -875,7 +875,7 @@ static Mat load_mat(GModel *m, const char *fmt, ...) {snprintf(scales, sizeof(scales), "%s.qs", name);st_tensor *qs = st_find(&m->S, scales);if (!qs) {- fprintf(stderr, "%s e' int4 ma manca %s\n", name, scales);+ fprintf(stderr, "%s e' int4 ma missing %s\n", name, scales);exit(1);}/* Il contenitore e' piatto: 4.194.304 byte di nibble e 131.072 scale,@@ -891,33 +891,33 @@ static Mat load_mat(GModel *m, const char *fmt, ...) {* la strada e' quella di kimi_k3: la forma la passa il chiamante. */const int64_t values = qs->numel * 64;if (t->nbytes * 2 != values) {- fprintf(stderr, "%s: %lld byte e %lld scale non sono un int4 gs64\n",+ fprintf(stderr, "%s: %lld bytes and %lld scales do not form int4 gs64\n",name, (long long)t->nbytes, (long long)qs->numel);exit(1);}if (t->rank != 2) {- fprintf(stderr, "%s: contenitore int4 piatto fuori dagli esperti; "- "la forma non e' nel file e non si indovina\n", name);+ fprintf(stderr, "%s: flat int4 container outside routed experts; "+ "the shape is not stored in the file and cannot be inferred\n", name);exit(1);}mat.rows = (int)t->shape[0];mat.columns = (int)(values / t->shape[0]);if (mat.columns % 64) {- fprintf(stderr, "%s: %d colonne non sono multiple di 64\n", name, mat.columns);+ fprintf(stderr, "%s: %d columns are not multiples of 64\n", name, mat.columns);exit(1);}uint8_t *packed = malloc((size_t)t->nbytes);float *step = malloc((size_t)qs->numel * sizeof(float));- if (!packed || !step) { fprintf(stderr, "OOM su %s\n", name); exit(1); }+ if (!packed || !step) { fprintf(stderr, "OOM on %s\n", name); exit(1); }st_read_raw(&m->S, name, packed, 1);st_read_f32_cap(&m->S, scales, step, qs->numel, 1);mat.fmt = 4; mat.q4 = packed; mat.s = step; mat.gs = 64; mat.resident = 1;return mat;}- if (t->rank != 2) { fprintf(stderr, "%s: rank %d, attesa 2\n", name, t->rank); exit(1); }+ if (t->rank != 2) { fprintf(stderr, "%s: rank %d, expected 2\n", name, t->rank); exit(1); }float *buffer = malloc((size_t)t->numel * sizeof(float));- if (!buffer) { fprintf(stderr, "OOM su %s\n", name); exit(1); }+ if (!buffer) { fprintf(stderr, "OOM on %s\n", name); exit(1); }st_read_f32_cap(&m->S, name, buffer, t->numel, 1);mat.rows = (int)t->shape[0];mat.columns = (int)t->shape[1];@@ -1120,7 +1120,7 @@ static void mla_layer(const Cfg *c, const GLayer *l, const float *x, int tokens,if (coli_sparse_index_select_range(selected, iq, ik, gates, head_w, l->ikpa, valid,seen, IH, ID, c->index_kpool, c->index_topk,c->index_kpool_tail, base, seen)) {- fprintf(stderr, "selezione indexer fallita\n"); exit(1);+ fprintf(stderr, "indexer selection failed\n"); exit(1);}/* GLM53_DUMP_INDEX=1 stampa le righe scelte dall'indexer: e' il primo* posto da guardare quando il motore diverge solo su certe lunghezze. */@@ -1240,7 +1240,7 @@ static void expert_table_init(GModel *m) {"down_proj.weight", "down_proj.weight.qs",};m->eref = calloc((size_t)c->n_layers * c->n_experts, sizeof(*m->eref));- if (!m->eref) { fprintf(stderr, "OOM sulla tabella degli esperti\n"); exit(1); }+ if (!m->eref) { fprintf(stderr, "OOM allocating expert table\n"); exit(1); }const int from = c->first_dense > m->layer_begin ? c->first_dense : m->layer_begin;for (int i = from; i < m->layer_end; i++) {@@ -1251,9 +1251,9 @@ static void expert_table_init(GModel *m) {snprintf(name, sizeof(name), "%slayers.%d.mlp.experts.%d.%s",m->prefix, i, e, piece[p]);st_tensor *t = st_find(&m->S, name);- if (!t) { fprintf(stderr, "manca %s\n", name); exit(1); }+ if (!t) { fprintf(stderr, "missing %s\n", name); exit(1); }if (t->nbytes != m->e_len[p]) {- fprintf(stderr, "%s: %lld byte, attesi %lld\n", name,+ fprintf(stderr, "%s: %lld bytes, expected %lld\n", name,(long long)t->nbytes, (long long)m->e_len[p]);exit(1);}@@ -1298,7 +1298,7 @@ static void expert_cache_init(GModel *m) {budget = free_now - 3.0;if (budget < 1.0) budget = 1.0;if (getenv("GLM53_VERBOSE"))- fprintf(stderr, "budget esperti: %.1f GB (%.1f disponibili, 3 di margine)\n",+ fprintf(stderr, "expert budget: %.1f GB (%.1f available, 3 GB reserved)\n",budget, free_now);}const int from = c->first_dense > m->layer_begin ? c->first_dense : m->layer_begin;@@ -1310,17 +1310,17 @@ static void expert_cache_init(GModel *m) {if (cap > c->n_experts) cap = c->n_experts;m->ecache = calloc((size_t)c->n_layers, sizeof(*m->ecache));- if (!m->ecache) { fprintf(stderr, "OOM sulla cache degli esperti\n"); exit(1); }+ if (!m->ecache) { fprintf(stderr, "OOM allocating expert cache\n"); exit(1); }for (int i = from; i < m->layer_end; i++) {LCache *cache = &m->ecache[i];cache->cap = cap;cache->s = calloc((size_t)cap, sizeof(*cache->s));- if (!cache->s) { fprintf(stderr, "OOM sugli slot del layer %d\n", i); exit(1); }+ if (!cache->s) { fprintf(stderr, "OOM allocating slots for layer %d\n", i); exit(1); }for (int j = 0; j < cap; j++) cache->s[j].eid = -1;}if (getenv("GLM53_VERBOSE"))- fprintf(stderr, "esperti: slot da %.1f MB, %d per layer su %d layer sparsi "- "(%.1f GB residenti)\n",+ fprintf(stderr, "experts: %.1f MB slots, %d per layer across %d sparse layers "+ "(%.1f GB resident)\n",m->e_slot / 1e6, cap, sparse, (double)cap * sparse * m->e_slot / 1e9);}@@ -1369,8 +1369,8 @@ static void expert_read(GModel *m, int layer, int eid, Slot *slot) {slot->own = (uint8_t *)p;slot->own_len = need;if (!slot->own) {- fprintf(stderr, "OOM su uno slot esperto (%.1f MB): la cache esperti non ci sta "- "in memoria; riduci con --ram N o GLM53_EXPERT_GB=N (#1375)\n",+ fprintf(stderr, "OOM allocating expert slot (%.1f MB): the expert cache does not fit "+ "in memory; reduce it with --ram N or GLM53_EXPERT_GB=N (#1375)\n",m->e_slot / 1e6);exit(1);}@@ -1471,7 +1471,7 @@ static void ffn_layer(GModel *m, const GLayer *l, int index, const float *x,int *chosen = malloc((size_t)tokens * topk * sizeof(int));float *weight = malloc((size_t)tokens * topk * sizeof(float));float *score = malloc((size_t)c->n_experts * sizeof(float));- if (!chosen || !weight || !score) { fprintf(stderr, "OOM nel router\n"); exit(1); }+ if (!chosen || !weight || !score) { fprintf(stderr, "OOM in router\n"); exit(1); }/* --- primo tempo: il router, per ogni token --- */for (int t = 0; t < tokens; t++) {@@ -1522,7 +1522,7 @@ static void ffn_layer(GModel *m, const GLayer *l, int index, const float *x,float *sg = malloc((size_t)wide * sizeof(float));float *su = malloc((size_t)wide * sizeof(float));float *tmp = malloc((size_t)c->hidden * sizeof(float));- if (!sg || !su || !tmp) { fprintf(stderr, "OOM nel MoE\n"); exit(1); }+ if (!sg || !su || !tmp) { fprintf(stderr, "OOM in MoE\n"); exit(1); }/* l'esperto condiviso e' sempre attivo e non passa dalla cache */for (int t = 0; t < tokens; t++)@@ -1546,7 +1546,7 @@ static void ffn_layer(GModel *m, const GLayer *l, int index, const float *x,/* unione dei distinti, nell'ordine in cui compaiono */int *union_ids = malloc((size_t)tokens * topk * sizeof(int));int n_union = 0;- if (!union_ids) { fprintf(stderr, "OOM sull'unione\n"); exit(1); }+ if (!union_ids) { fprintf(stderr, "OOM building expert union\n"); exit(1); }for (int i = 0; i < tokens * topk; i++) {int seen = 0;for (int j = 0; j < n_union; j++) if (union_ids[j] == chosen[i]) { seen = 1; break; }@@ -1557,7 +1557,7 @@ static void ffn_layer(GModel *m, const GLayer *l, int index, const float *x,const int block = cache->cap;int *slot_of = malloc((size_t)block * sizeof(int));int *to_read = malloc((size_t)block * sizeof(int));- if (!slot_of || !to_read) { fprintf(stderr, "OOM sugli slot\n"); exit(1); }+ if (!slot_of || !to_read) { fprintf(stderr, "OOM allocating slots\n"); exit(1); }for (int base = 0; base < n_union; base += block) {const int here = base + block <= n_union ? block : n_union - base;@@ -1734,7 +1734,7 @@ static void model_load_range(GModel *m, const char *dir, int layer_begin,snprintf(first, sizeof(first), "%slayers.%d.mlp.experts.0.gate_proj.weight",P, probe_layer);st_tensor *probe_expert = st_find(&m->S, first);- if (!probe_expert) { fprintf(stderr, "manca %s\n", first); exit(1); }+ if (!probe_expert) { fprintf(stderr, "missing %s\n", first); exit(1); }m->streaming = probe_expert->dtype == 3;}if (m->streaming) {@@ -1837,8 +1837,8 @@ static void model_load_range(GModel *m, const char *dir, int layer_begin,if (getenv("COLI_METAL") && atoi(getenv("COLI_METAL"))) {g_metal_ready = coli_metal_init() && coli_metal_available();fprintf(stderr, g_metal_ready- ? "Metal: attivo sulle matrici residenti\n"- : "Metal: nessun device utilizzabile, resto su CPU\n");+ ? "Metal: active for resident matrices\n"+ : "Metal: no usable device, falling back to CPU\n");}#endif#ifdef COLI_VULKAN@@ -1855,8 +1855,8 @@ static void model_load_range(GModel *m, const char *dir, int layer_begin,else snprintf(spv, sizeof(spv), "%s/qmatmul.spv", given ? given : "shaders");g_vk_ready = coli_vk_init(spv) && coli_vk_available();fprintf(stderr, g_vk_ready- ? "Vulkan: attivo sulle matrici residenti\n"- : "Vulkan: nessun device utilizzabile (%s), resto su CPU\n", spv);+ ? "Vulkan: active for resident matrices\n"+ : "Vulkan: no usable device (%s), falling back to CPU\n", spv);}#endif/* La cache si dimensiona qui, non prima: quanto si puo' spendere dipende@@ -1898,7 +1898,7 @@ static void vision_load(GModel *m) {m->vision.merger_down = load_f32(m, "%smerger.down_proj.weight", V);m->vblocks = calloc((size_t)c->vis_layers, sizeof(*m->vblocks));- if (!m->vblocks) { fprintf(stderr, "OOM sui blocchi vision\n"); exit(1); }+ if (!m->vblocks) { fprintf(stderr, "OOM allocating vision blocks\n"); exit(1); }for (int b = 0; b < c->vis_layers; b++) {ColiVisionBlock *vb = &m->vblocks[b];vb->norm1 = load_f32(m, "%sblocks.%d.norm1.weight", V, b);@@ -1929,19 +1929,19 @@ static void vision_load(GModel *m) {static float *vision_encode(GModel *m, const float *patches,int grid_h, int grid_w, int *out_tokens) {if (!m->has_vision) {- fprintf(stderr, "questo checkpoint non porta la torre vision\n");+ fprintf(stderr, "this checkpoint does not include the vision tower\n");exit(1);}const int tokens = coli_vision_output_tokens(&m->vision.config, grid_h, grid_w);if (tokens <= 0) {- fprintf(stderr, "griglia %dx%d non divisibile per merge %d\n",+ fprintf(stderr, "grid %dx%d is not divisible by merge size %d\n",grid_h, grid_w, m->vision.config.merge);exit(1);}float *out = malloc((size_t)tokens * m->vision.config.out_hidden * sizeof(float));- if (!out) { fprintf(stderr, "OOM sugli embedding vision\n"); exit(1); }+ if (!out) { fprintf(stderr, "OOM allocating vision embeddings\n"); exit(1); }if (coli_vision_forward(out, &m->vision, patches, grid_h, grid_w) != 0) {- fprintf(stderr, "la torre vision ha rifiutato l'ingresso\n");+ fprintf(stderr, "vision tower rejected the input\n");exit(1);}/* La torre esce a out_hidden; il flusso testuale vuole hidden. Il@@ -1961,10 +1961,10 @@ static float *vision_encode(GModel *m, const float *patches,static GSession *session_open(const GModel *m, int cap) {const Cfg *c = &m->c;GSession *s = calloc(1, sizeof(*s));- if (!s) { fprintf(stderr, "OOM sulla sessione\n"); exit(1); }+ if (!s) { fprintf(stderr, "OOM allocating session\n"); exit(1); }s->cap = cap;s->layer = calloc((size_t)c->n_layers, sizeof(*s->layer));- if (!s->layer) { fprintf(stderr, "OOM sugli stati di layer\n"); exit(1); }+ if (!s->layer) { fprintf(stderr, "OOM allocating layer states\n"); exit(1); }if (c->kda_proj)s->kda_scratch = malloc((size_t)coli_kda_scratch_floats(c->kda_heads, c->kda_hd,c->kda_hd) * sizeof(float));@@ -1975,14 +1975,14 @@ static GSession *session_open(const GModel *m, int cap) {st->ikeys = malloc((size_t)cap * c->index_hd * sizeof(float));st->igates = malloc((size_t)cap * c->index_hd * sizeof(float));if (!st->latent || !st->ikeys || !st->igates) {- fprintf(stderr, "OOM sulla cache del layer %d\n", i); exit(1);+ fprintf(stderr, "OOM allocating cache for layer %d\n", i); exit(1);}} else if (c->kda_proj) {st->kda_state = calloc((size_t)c->kda_heads * c->kda_hd * c->kda_hd,sizeof(float));st->kda_window = calloc((size_t)3 * c->kda_proj * c->conv_k, sizeof(float));if (!st->kda_state || !st->kda_window) {- fprintf(stderr, "OOM sullo stato KDA del layer %d\n", i); exit(1);+ fprintf(stderr, "OOM allocating KDA state for layer %d\n", i); exit(1);}}}@@ -1990,8 +1990,8 @@ static GSession *session_open(const GModel *m, int cap) {int full = 0;for (int i = 0; i < c->n_layers; i++) if (c->is_full[i]) full++;const double per_token = (double)full * (c->kv_lora + 2 * c->index_hd) * sizeof(float);- fprintf(stderr, "cache: %.1f KB per token su %d layer DSA "- "(%.2f GB a %d posizioni)\n",+ fprintf(stderr, "cache: %.1f KB per token across %d DSA layers "+ "(%.2f GB at %d positions)\n",per_token / 1024.0, full, per_token * cap / 1e9, cap);}return s;@@ -2026,7 +2026,7 @@ static float *run_layers(GModel *m, GSession *s, float *streams, float *next,float *post = malloc((size_t)n * H * sizeof(float));float *comb = malloc((size_t)n * H * H * sizeof(float));if (!collapsed || !normed || !branch || !post || !comb) {- fprintf(stderr, "OOM nei temporanei del passaggio\n"); exit(1);+ fprintf(stderr, "OOM allocating forward-pass temporaries\n"); exit(1);}for (int i = begin; i < end; i++) {@@ -2171,7 +2171,7 @@ static float *forward_span(GModel *m, GSession *s, const int *tokens, int n,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);+ fprintf(stderr, "context exhausted: %d positions of %d\n", start + n, s->cap);exit(1);}const int D = c->hidden;@@ -2185,14 +2185,14 @@ static float *forward_span(GModel *m, GSession *s, const int *tokens, int n,/* Un id fuori dal vocabolario legge oltre la tabella degli embedding.* Da CLI sarebbe un errore di battitura; da server e' input di rete. */if (tokens[t] < 0 || tokens[t] >= c->vocab) {- fprintf(stderr, "token %d fuori dal vocabolario (0..%d)\n",+ fprintf(stderr, "token %d is outside the vocabulary (0..%d)\n",tokens[t], c->vocab - 1);exit(1);}if (vision && c->image_token >= 0 && tokens[t] == c->image_token) {if (consumed >= n_vision) {- fprintf(stderr, "il prompt ha piu' token immagine (%d+) degli "- "embedding forniti (%d)\n", consumed + 1, n_vision);+ fprintf(stderr, "prompt has more image tokens (%d+) than "+ "provided embeddings (%d)\n", consumed + 1, n_vision);exit(1);}row = vision + (size_t)consumed++ * D;@@ -2206,8 +2206,8 @@ static float *forward_span(GModel *m, GSession *s, const int *tokens, int n,* modello un'immagine diversa da quella che il chiamante crede di aver* passato: e' un errore, non un caso limite. */if (consumed != n_vision) {- fprintf(stderr, "%d embedding vision forniti ma solo %d token immagine "- "nel prompt\n", n_vision, consumed);+ fprintf(stderr, "%d vision embeddings provided but only %d image tokens "+ "are present in the prompt\n", n_vision, consumed);exit(1);}@@ -2265,7 +2265,7 @@ static float *forward_prefill(GModel *m, GSession *s, const int *tokens, int n,if (chunk > n) chunk = n;float *all = keep_all ? malloc((size_t)n * c->vocab * sizeof(float)) : NULL;- if (keep_all && !all) { fprintf(stderr, "OOM sui logit del prefill\n"); exit(1); }+ if (keep_all && !all) { fprintf(stderr, "OOM allocating prefill logits\n"); exit(1); }float *last = NULL;int used_vision = 0;@@ -2292,7 +2292,7 @@ static float *forward_prefill(GModel *m, GSession *s, const int *tokens, int n,if (here > 1) {/* si tiene solo l'ultima riga */float *tail = malloc((size_t)c->vocab * sizeof(float));- if (!tail) { fprintf(stderr, "OOM sui logit\n"); exit(1); }+ if (!tail) { fprintf(stderr, "OOM allocating logits\n"); exit(1); }memcpy(tail, last + (size_t)(here - 1) * c->vocab,(size_t)c->vocab * sizeof(float));free(last);@@ -2301,7 +2301,7 @@ static float *forward_prefill(GModel *m, GSession *s, const int *tokens, int n,}}if (vision && used_vision != n_vision) {- fprintf(stderr, "%d embedding vision forniti ma %d segnaposto nel prompt\n",+ fprintf(stderr, "%d vision embeddings provided but %d placeholders are present in the prompt\n",n_vision, used_vision);exit(1);}@@ -2406,7 +2406,7 @@ static int sample_token(const float *logits, int vocab) {probability = realloc(probability, (size_t)vocab * sizeof(float));order = realloc(order, (size_t)vocab * sizeof(int));room = vocab;- if (!probability || !order) { fprintf(stderr, "OOM nel campionamento\n"); exit(1); }+ if (!probability || !order) { fprintf(stderr, "OOM in sampling\n"); exit(1); }}float top = -INFINITY;for (int i = 0; i < vocab; i++) if (logits[i] > top) top = logits[i];@@ -2471,7 +2471,7 @@ static void slots_init(const GModel *m) {g_slot_context = context ? atoi(context) : 8192;if (g_slot_context < 64) g_slot_context = 64;if (getenv("GLM53_VERBOSE"))- fprintf(stderr, "slot KV: %d da %d posizioni\n", g_n_slots, g_slot_context);+ fprintf(stderr, "KV slots: %d with %d positions each\n", g_n_slots, g_slot_context);(void)m;}@@ -2492,7 +2492,7 @@ static int slot_shared(const KVSlot *slot, const int *tokens, int n) {static void slot_remember(KVSlot *slot, const int *tokens, int n) {if (n > slot->cap) {slot->tokens = realloc(slot->tokens, (size_t)n * sizeof(int));- if (!slot->tokens) { fprintf(stderr, "OOM sulla storia dello slot\n"); exit(1); }+ if (!slot->tokens) { fprintf(stderr, "OOM allocating slot history\n"); exit(1); }slot->cap = n;}memcpy(slot->tokens, tokens, (size_t)n * sizeof(int));@@ -2547,9 +2547,9 @@ static void arm_stops(const char *dir, Tok *tokenizer, int batched) {for (int i = 0; i < g_nstop; i++) if (g_stop[i] == id) seen = 1;if (!seen) g_stop[g_nstop++] = id;}- fprintf(stderr, "[stop] %d token di stop:", g_nstop);+ fprintf(stderr, "[stop] %d stop tokens:", g_nstop);for (int i = 0; i < g_nstop; i++) fprintf(stderr, " %d", g_stop[i]);- fprintf(stderr, "%s\n", batched ? " (modalita' batch: solo fine testo)" : "");+ fprintf(stderr, "%s\n", batched ? " (batch mode: end-of-text only)" : "");}static int is_stop(int token) {@@ -3014,7 +3014,7 @@ int main(int argc, char **argv) {* openai_server.py. */if (getenv("SERVE")) {const char *snap = getenv("SNAP");- if (!snap) { fprintf(stderr, "SERVE senza SNAP: non so quale modello aprire\n"); return 2; }+ if (!snap) { fprintf(stderr, "SERVE requires SNAP to select a model\n"); return 2; }GModel served;memset(&served, 0, sizeof(served));model_load(&served, snap);@@ -3033,17 +3033,17 @@ int main(int argc, char **argv) {}if (!dir || (!ids && !prompt_text)) {- fprintf(stderr, "uso: %s --model DIR (--prompt TESTO | --ids a,b,c)\n"+ fprintf(stderr, "usage: %s --model DIR (--prompt TEXT | --ids a,b,c)\n"" [--greedy N] [--logits] [--patches FILE.f32 --grid HxW]\n",argv[0]);return 2;}if (ids && prompt_text) {- fprintf(stderr, "--prompt e --ids sono due modi di dire la stessa cosa\n");+ fprintf(stderr, "--prompt and --ids are mutually exclusive\n");return 2;}if (!!patch_file != (grid_h > 0 && grid_w > 0)) {- fprintf(stderr, "--patches e --grid vanno insieme\n");+ fprintf(stderr, "--patches and --grid must be used together\n");return 2;}int capacity = 1024, count = 0;@@ -3069,7 +3069,7 @@ int main(int argc, char **argv) {p = (*end == ',') ? end + 1 : end;}}- if (!count) { fprintf(stderr, "nessun token nel prompt\n"); return 2; }+ if (!count) { fprintf(stderr, "no tokens in prompt\n"); return 2; }GModel model;memset(&model, 0, sizeof(model)); /* contatori e puntatori opzionali */@@ -3084,20 +3084,20 @@ int main(int argc, char **argv) {if (patch_file) {const ColiVisionConfig *vc = &model.vision.config;if (!model.has_vision) {- fprintf(stderr, "--patches ma il checkpoint non porta la torre vision\n");+ fprintf(stderr, "--patches requested but this checkpoint does not include the vision tower\n");return 2;}const size_t per_patch = (size_t)vc->in_channels * vc->temporal * vc->patch * vc->patch;const size_t wanted = (size_t)grid_h * grid_w * per_patch;FILE *f = fopen(patch_file, "rb");- if (!f) { fprintf(stderr, "non apro %s\n", patch_file); return 2; }+ if (!f) { fprintf(stderr, "cannot open %s\n", patch_file); return 2; }float *patches = malloc(wanted * sizeof(float));- if (!patches) { fprintf(stderr, "OOM sulle patch\n"); return 2; }+ if (!patches) { fprintf(stderr, "OOM allocating patches\n"); return 2; }size_t got = fread(patches, sizeof(float), wanted, f);/* Una patch corta darebbe comunque un'uscita, con la coda letta da* memoria non inizializzata: meglio fermarsi e dire di quanto. */if (got != wanted) {- fprintf(stderr, "%s: %zu float su %zu attesi per una griglia %dx%d\n",+ fprintf(stderr, "%s: got %zu floats, expected %zu for a %dx%d grid\n",patch_file, got, wanted, grid_h, grid_w);return 2;}@@ -3113,7 +3113,7 @@ int main(int argc, char **argv) {const double prefill_start = now_s();float *logits = forward_prefill(&model, session, tokens, count, vision, n_vision, 1);if (getenv("GLM53_VERBOSE"))- fprintf(stderr, "caricamento %.1fs, prefill %d token in %.1fs\n",+ fprintf(stderr, "load %.1fs, prefill %d tokens in %.1fs\n",load_seconds, count, now_s() - prefill_start);printf("teacher_forcing");for (int t = 0; t < count; t++)diff --git a/c/tests/test_glm53_vulkan.py b/c/tests/test_glm53_vulkan.pyindex 1ddda245..5005ffc5 100644--- a/c/tests/test_glm53_vulkan.py+++ b/c/tests/test_glm53_vulkan.py@@ -62,7 +62,7 @@ def main() -> int:# Un traceback su un file che manca fa sembrare rotto il# motore; chi arriva per la prima volta non puo' distinguere# le due cose. Il generatore vuole transformers 5.16.1.- print(f"SKIP: manca {arguments.fixture}; generalo con\n"+ print(f"SKIP: missing {arguments.fixture}; generate it with\n"f" python3 tools/make_glm53_multimodal_tiny.py --output <dir>")return 0binary = os.path.abspath(arguments.binary)@@ -71,10 +71,10 @@ def main() -> int:cpu, _ = run(binary, arguments.fixture, reference, False, arguments.shaders)gpu, notes = run(binary, arguments.fixture, reference, True, arguments.shaders)- if "attivo" not in notes:- reason = ("il binario non e' costruito con VK=1"- if "Vulkan:" not in notes else "nessun device Vulkan utilizzabile")- print(f"SKIP: {reason}; il percorso Vulkan non e' stato verificato")+ if "Vulkan: active for resident matrices" not in notes:+ reason = ("binary was not built with VK=1"+ if "Vulkan:" not in notes else "no usable Vulkan device")+ print(f"SKIP: {reason}; Vulkan path was not verified")return 0device = next((line for line in notes.splitlines() if "[VK] ready:" in line), "")@@ -83,8 +83,8 @@ def main() -> int:print(f"FAIL {field}\n CPU: {cpu.get(field)}\n Vulkan: {gpu.get(field)}")return 1- print(f"PASS GLM-5.3 Vulkan: stessi token della CPU su "- f"{len(cpu['teacher_forcing'])} posizioni e {len(cpu['greedy'])} passi greedy"+ print(f"PASS GLM-5.3 Vulkan: same tokens as CPU across "+ f"{len(cpu['teacher_forcing'])} positions and {len(cpu['greedy'])} greedy steps"f"{' — ' + device.split('ready:')[1].split(',')[0].strip() if device else ''}")return 0