Fix glm4moe bug (#15088)

* webui: fix markdown table

* webui: fix table display with themes

src/llama-context.cpp

@@ -786,7 +786,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     const int64_t n_embd  = hparams.n_embd;
-    const int32_t n_vocab = model.vocab.n_tokens();
+    const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
     if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
@@ -959,7 +959,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const auto & vocab   = model.vocab;
     const auto & hparams = model.hparams;
 
-    const int32_t n_vocab = vocab.n_tokens();
+    const int64_t n_vocab = vocab.n_tokens();
     const int64_t n_embd  = hparams.n_embd;
 
     // when computing embeddings, all tokens are output
@@ -1328,21 +1328,21 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 }
 
 void llama_context::output_reorder() {
-    const uint32_t n_vocab = model.vocab.n_tokens();
+    const uint64_t n_vocab = model.vocab.n_tokens();
     const uint64_t n_embd  = model.hparams.n_embd;
 
-    for (uint32_t s = 0; s < output_swaps.size(); ++s) {
-        const uint32_t i0 = output_swaps[s].i0;
-        const uint32_t i1 = output_swaps[s].i1;
+    for (size_t s = 0; s < output_swaps.size(); ++s) {
+        const uint64_t i0 = output_swaps[s].i0;
+        const uint64_t i1 = output_swaps[s].i1;
 
         if (logits_size > 0) {
-            for (uint32_t k = 0; k < n_vocab; k++) {
+            for (uint64_t k = 0; k < n_vocab; k++) {
                 std::swap(logits[i0*n_vocab + k], logits[i1*n_vocab + k]);
             }
         }
 
         if (embd_size > 0) {
-            for (uint32_t k = 0; k < n_embd; k++) {
+            for (uint64_t k = 0; k < n_embd; k++) {
                 std::swap(embd[i0*n_embd + k], embd[i1*n_embd + k]);
             }
         }

src/llama-model.cpp

@@ -13800,10 +13800,6 @@ struct llm_build_glm4_moe : public llm_graph_context {
                         LLM_FFN_SILU, LLM_FFN_PAR, il);
                 cb(cur, "ffn_out", il);
             } else {
-                // MoE layer with shared experts
-                const int64_t n_expert      = hparams.n_expert;
-                const int64_t n_expert_used = hparams.n_expert_used;
-
                 // Process routed experts using existing MoE infrastructure
                 ggml_tensor * routed_out = build_moe_ffn(cur,
                         model.layers[il].ffn_gate_inp,

tools/server/webui/src/index.scss

@@ -31,7 +31,24 @@ html {
   hr {
     @apply my-4 border-base-content/20 border-1;
   }
-  /* TODO: fix markdown table */
+  table {
+    @apply w-full border-collapse text-sm font-sans my-4 text-base-content;
+  }
+  thead {
+    @apply bg-base-200 text-base-content;
+  }
+  th {
+    @apply border border-base-300 px-4 py-2 text-left font-semibold;
+  }
+  td {
+    @apply border border-base-300 px-4 py-2 align-top;
+  }
+  tbody tr:nth-child(even) {
+    @apply bg-base-100;
+  }
+  tbody tr:hover {
+    @apply bg-base-200;
+  }
 }
 
 .btn-mini {