Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.devops/tools.sh
#	.devops/vulkan.Dockerfile
#	.github/workflows/build.yml
#	.github/workflows/docker.yml
#	.github/workflows/server.yml
#	Makefile
#	README.md
#	cmake/llama-config.cmake.in
#	common/CMakeLists.txt
#	examples/gbnf-validator/gbnf-validator.cpp
#	examples/run/run.cpp
#	examples/server/README.md
#	examples/server/tests/README.md
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-hip/CMakeLists.txt
#	scripts/sync-ggml.last
#	tests/CMakeLists.txt
#	tests/test-backend-ops.cpp
#	tests/test-chat-template.cpp
#	tests/test-grammar-integration.cpp

src/llama-grammar.cpp

@@ -560,7 +560,7 @@ bool llama_grammar_parser::parse(const char * src) {
             }
         }
     } catch (const std::exception & err) {
-        fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+        fprintf(stderr, "%s: error parsing grammar: %s\n\n%s\n", __func__, err.what(), src);
         rules.clear();
         return false;
     }
@@ -960,10 +960,28 @@ struct llama_grammar * llama_grammar_init_impl(
     // Important: vec_rules has to be moved here, not copied, because stacks contains
     // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar
     // then the pointers would be invalidated when the local vec_rules goes out of scope.
-    return new llama_grammar { vocab, std::move(vec_rules), std::move(stacks), {}, };
+    return new llama_grammar {
+        vocab,
+        std::move(vec_rules),
+        std::move(stacks),
+        /* .partial_utf8 = */     {},
+        /* .lazy =*/              false,
+        /* .awaiting_trigger = */ false,
+        /* .trigger_buffer = */   "",
+        /* .trigger_tokens   = */ {},
+        /* .trigger_words    = */ {},
+    };
 }
 
-struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root) {
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
     llama_grammar_parser parser;
 
     // if there is a grammar, parse it
@@ -1035,10 +1053,31 @@ struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab,
         }
     } while (true);
 
+    std::vector<llama_token>    vec_trigger_tokens;
+    std::vector<std::string> vec_trigger_words;
+    for (size_t i = 0; i < num_trigger_tokens; i++) {
+        GGML_ASSERT(trigger_tokens != nullptr);
+        vec_trigger_tokens.push_back(trigger_tokens[i]);
+    }
+    for (size_t i = 0; i < num_trigger_words; i++) {
+        GGML_ASSERT(trigger_words != nullptr);
+        vec_trigger_words.push_back(trigger_words[i]);
+    }
+
     // Important: vec_rules has to be moved here, not copied, because stacks contains
     // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar
     // then the pointers would be invalidated when the local vec_rules goes out of scope.
-    return new llama_grammar { vocab, std::move(vec_rules), std::move(stacks), {}, };
+    return new llama_grammar {
+        vocab,
+        std::move(vec_rules),
+        std::move(stacks),
+        /* .partial_utf8 = */     {},
+        /* .lazy = */             lazy,
+        /* .awaiting_trigger = */ lazy,
+        /* .trigger_buffer = */   "",
+        std::move(vec_trigger_tokens),
+        std::move(vec_trigger_words),
+    };
 }
 
 void llama_grammar_free_impl(struct llama_grammar * grammar) {
@@ -1055,6 +1094,11 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
         grammar.rules,
         grammar.stacks,
         grammar.partial_utf8,
+        grammar.lazy,
+        grammar.awaiting_trigger,
+        grammar.trigger_buffer,
+        grammar.trigger_tokens,
+        grammar.trigger_words,
     };
 
     // redirect elements in stacks to point to new rules
@@ -1076,6 +1120,10 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
 void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_data_array * cur_p) {
     GGML_ASSERT(grammar.vocab != nullptr);
 
+    if (grammar.awaiting_trigger) {
+        return;
+    }
+
     bool allow_eog = false;
     for (const auto & stack : grammar.stacks) {
         if (stack.empty()) {
@@ -1115,6 +1163,34 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
 void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token) {
     GGML_ASSERT(grammar.vocab != nullptr);
 
+    const auto & piece = grammar.vocab->token_to_piece(token);
+
+    if (grammar.awaiting_trigger) {
+        if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
+            grammar.awaiting_trigger = false;
+            grammar.trigger_buffer.clear();
+            llama_grammar_accept_str(grammar, piece);
+            LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
+            return;
+        } else {
+            // TODO: consider a smarter incremental substring search algorithm (store last position to search from).
+            grammar.trigger_buffer += piece;
+            for (const auto & word : grammar.trigger_words) {
+                auto pos = grammar.trigger_buffer.find(word);
+                if (pos != std::string::npos) {
+                    grammar.awaiting_trigger = false;
+                    auto constrained_str = grammar.trigger_buffer.substr(pos);
+                    grammar.trigger_buffer.clear();
+                    llama_grammar_accept_str(grammar, constrained_str);
+                    LLAMA_LOG_DEBUG("Grammar triggered on word `%s`", word.c_str());
+                    return;
+                }
+            }
+            LLAMA_LOG_DEBUG("Grammar still awaiting trigger after token %d (`%s`) (buffer: `%s`)\n", token, piece.c_str(), grammar.trigger_buffer.c_str());
+            return;
+        }
+    }
+
     if (grammar.vocab->is_eog(token)) {
         for (const auto & stack : grammar.stacks) {
             if (stack.empty()) {
@@ -1124,8 +1200,10 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
         GGML_ABORT("fatal error");
     }
 
-    const std::string & piece = grammar.vocab->token_to_piece(token);
+    llama_grammar_accept_str(grammar, piece);
+}
 
+void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string & piece) {
     // Note terminating 0 in decoded string
     const auto   decoded     = decode_utf8(piece, grammar.partial_utf8);
     const auto & code_points = decoded.first;

src/llama-grammar.h

@@ -114,6 +114,15 @@ struct llama_grammar {
 
     // buffer for partially generated UTF-8 sequence from accepted tokens
     llama_partial_utf8 partial_utf8;
+
+    // lazy grammars wait for trigger words or tokens before constraining the sampling.
+    // we still ahve trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
+    // (useful e.g. for tool_choice=required)
+    bool                     lazy             = false;
+    bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
+    std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
+    std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
+    std::vector<std::string> trigger_words;
 };
 
 //
@@ -127,7 +136,15 @@ struct llama_grammar * llama_grammar_init_impl(
         size_t n_rules,
         size_t start_rule_index);
 
-struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root);
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
 
 void llama_grammar_free_impl(struct llama_grammar * grammar);
 
@@ -141,3 +158,7 @@ void llama_grammar_apply_impl(
 void llama_grammar_accept_impl(
               struct llama_grammar & grammar,
                        llama_token   token);
+
+void llama_grammar_accept_str(
+              struct llama_grammar & grammar,
+                 const std::string & piece);

src/llama-sampling.cpp

@@ -1433,13 +1433,30 @@ static void llama_sampler_grammar_apply(struct llama_sampler * smpl, llama_token
     }
 }
 
+// Fwd declare to break reset --> init_impl --> llama_sampler_grammar_i --> reset cycle.
+static struct llama_sampler * llama_sampler_init_grammar_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
+
 static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
     auto * ctx = (llama_sampler_grammar *) smpl->ctx;
     if (!ctx->grammar) {
         return;
     }
 
-    auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str());
+    std::vector<const char *>  trigger_words;
+    for (auto & word : ctx->grammar->trigger_words) {
+        trigger_words.push_back(word.c_str());
+    }
+    auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str(),
+                                                 ctx->grammar->lazy, trigger_words.data(), trigger_words.size(),
+                                                 ctx->grammar->trigger_tokens.data(), ctx->grammar->trigger_tokens.size());
 
     llama_grammar_free_impl(ctx->grammar);
     ctx->grammar = grammar_new;
@@ -1448,7 +1465,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
 static struct llama_sampler * llama_sampler_grammar_clone(const struct llama_sampler * smpl) {
     const auto * ctx = (const llama_sampler_grammar *) smpl->ctx;
 
-    auto * result = llama_sampler_init_grammar(ctx->vocab, nullptr, nullptr);
+    auto * result = llama_sampler_init_grammar_impl(ctx->vocab, nullptr, nullptr, false, nullptr, 0, nullptr, 0);
 
     // copy the state
     {
@@ -1484,7 +1501,15 @@ static struct llama_sampler_i llama_sampler_grammar_i = {
     /* .free   = */ llama_sampler_grammar_free,
 };
 
-struct llama_sampler * llama_sampler_init_grammar(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root) {
+static struct llama_sampler * llama_sampler_init_grammar_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
     auto * ctx = new llama_sampler_grammar;
 
     if (grammar_str != nullptr && grammar_str[0] != '\0') {
@@ -1492,7 +1517,7 @@ struct llama_sampler * llama_sampler_init_grammar(const struct llama_vocab * voc
             /* .vocab        = */ vocab,
             /* .grammar_str  = */ grammar_str,
             /* .grammar_root = */ grammar_root,
-            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root),
+            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens),
         };
     } else {
         *ctx = {
@@ -1509,6 +1534,24 @@ struct llama_sampler * llama_sampler_init_grammar(const struct llama_vocab * voc
     };
 }
 
+struct llama_sampler * llama_sampler_init_grammar(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root) {
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ false, nullptr, 0, nullptr, 0);
+}
+
+struct llama_sampler * llama_sampler_init_grammar_lazy(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens);
+}
+
 // penalties
 
 struct llama_sampler_penalties {

src/llama-vocab.cpp

@@ -1930,7 +1930,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         linefeed_id = ids[0];
     } else {
-        const std::vector<int> ids = tokenize("\xC4\x8A", false); // U+010A
+        const std::vector<int> ids = tokenize("\n", false);
 
         //GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         if (ids.empty()) {

src/llama.cpp

@@ -7736,17 +7736,13 @@ struct llm_build_context {
                 1
             );
 
+            struct ggml_tensor * last_norm_att = ggml_view_3d(ctx0, x_norm_att, n_embd, 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_att));
             ggml_build_forward_expand(
                 gf,
                 ggml_cpy(
                     ctx0,
-                    wkv_states,
-                    ggml_view_1d(
-                        ctx0,
-                        kv_self.v_l[il],
-                        hparams.n_embd_v_s() * n_seqs,
-                        hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self.v_l[il])
-                    )
+                    ggml_view_1d(ctx0, last_norm_att, n_embd * n_seqs, 0),
+                    ggml_view_1d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self.k_l[il]))
                 )
             );