@fugood/llama.node 0.3.12 → 0.3.14

package/src/llama.cpp/src/llama-vocab.cpp

@@ -16,6 +16,7 @@
 #include <queue>
 #include <set>
 #include <unordered_map>
+#include <cctype>
 
 //
 // helpers
@@ -392,6 +393,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_GPT4O:
+                regex_exprs = {
+                    // original regex from tokenizer.json
+                    // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                    "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1592,6 +1600,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             } else if (
                 tokenizer_pre == "megrez") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
+            } else if (
+                tokenizer_pre == "gpt-4o") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
+                clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }

package/src/llama.cpp/src/llama.cpp

@@ -4978,6 +4978,149 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_gemma3() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
+
+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
+
+        // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+        if (ubatch.token) {
+            inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+            cb(inpL, "inp_scaled", -1);
+        }
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        // gemma3 requires different mask for layers using sliding window (SWA)
+        struct ggml_tensor * KQ_mask     = build_inp_KQ_mask(true);
+        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(true);
+
+        // "5-to-1 interleaved attention"
+        // 5 layers of local attention followed by 1 layer of global attention
+        static const int sliding_window_pattern = 6;
+
+        for (int il = 0; il < n_layer; ++il) {
+            const bool is_sliding          = (il + 1) % sliding_window_pattern;
+            const float freq_base_l        = is_sliding ? 10000.0f    : freq_base;
+            const float freq_scale_l       = is_sliding ? 1.0f        : freq_scale;
+            struct ggml_tensor * KQ_mask_l = is_sliding ? KQ_mask_swa : KQ_mask;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens);
+                Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                    model.layers[il].attn_q_norm,
+                    NULL,
+                    LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens);
+                Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                    model.layers[il].attn_k_norm,
+                    NULL,
+                    LLM_NORM_RMS, cb, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask_l, n_tokens, kv_head, n_kv, hparams.f_attention_scale, cb, il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_post_norm", il);
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+            cb(sa_out, "sa_out", il);
+
+            cur = llm_build_norm(ctx0, sa_out, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, sa_out);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
 
     struct ggml_cgraph * build_starcoder2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
@@ -8298,6 +8441,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_gemma2();
             } break;
+        case LLM_ARCH_GEMMA3:
+            {
+                result = llm.build_gemma3();
+            } break;
         case LLM_ARCH_STARCODER2:
             {
                 result = llm.build_starcoder2();
@@ -8801,12 +8948,14 @@ static int llama_decode_impl(
     //llama_synchronize(&lctx);
 
     // decide if we need to defrag the kv cache
-    if (cparams.causal_attn && cparams.defrag_thold >= 0.0f) {
-        const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used)/float(kv_self.n) : 0.0f;
+    if (cparams.causal_attn && cparams.defrag_thold > 0.0f) {
+        // - do not defrag small contexts (i.e. < 2048 tokens)
+        // - count the padding towards the number of used tokens
+        const float fragmentation = kv_self.n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self.used + llama_kv_cache_get_padding(cparams))/float(kv_self.n)) : 0.0f;
 
         // queue defragmentation for next llama_kv_cache_update
         if (fragmentation > cparams.defrag_thold) {
-            //LLAMA_LOG_INFO("fragmentation: %.2f\n", fragmentation);
+            LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
 
             llama_kv_cache_defrag(kv_self);
         }
@@ -9428,8 +9577,6 @@ static struct llama_model * llama_model_load_from_file_impl(
         struct llama_model_params params) {
     ggml_time_init();
 
-    llama_model * model = new llama_model(params);
-
     unsigned cur_percentage = 0;
     if (params.progress_callback == NULL) {
         params.progress_callback_user_data = &cur_percentage;
@@ -9447,6 +9594,8 @@ static struct llama_model * llama_model_load_from_file_impl(
         };
     }
 
+    llama_model * model = new llama_model(params);
+
     // create list of devices to use with this model
     if (params.devices) {
         for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {

package/src/llama.cpp/src/unicode.cpp

@@ -618,7 +618,14 @@ std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
     result.reserve(utf8.size());
     size_t offset = 0;
     while (offset < utf8.size()) {
-        result.push_back(unicode_cpt_from_utf8(utf8, offset));
+        try {
+            result.push_back(unicode_cpt_from_utf8(utf8, offset));
+        }
+        catch (const std::invalid_argument & /*ex*/) {
+            // Silently ignore invalid UTF-8 input to avoid leaking the exception beyond llama_tokenize
+            ++offset;
+            result.emplace_back(0xFFFD); // replacement character
+        }
     }
     return result;
 }
@@ -701,7 +708,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
     const auto cpts = unicode_cpts_from_utf8(text);
 
     // generate a "collapsed" representation of the text, where all codepoints are replaced by a single byte
-    // ref: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2081479935
+    // ref: https://github.com/ggml-org/llama.cpp/pull/6920#issuecomment-2081479935
     std::string text_collapsed;
     if (need_collapse) {
         // collapse all unicode categories