@fugood/llama.node 1.4.8 → 1.4.9

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@fugood/llama.node",
   "access": "public",
-  "version": "1.4.8",
+  "version": "1.4.9",
   "description": "An another Node binding of llama.cpp",
   "main": "lib/index.js",
   "scripts": {
@@ -72,20 +72,20 @@
     "CMakeLists.txt"
   ],
   "optionalDependencies": {
-    "@fugood/node-llama-darwin-arm64": "1.4.8",
-    "@fugood/node-llama-darwin-x64": "1.4.8",
-    "@fugood/node-llama-linux-arm64": "1.4.8",
-    "@fugood/node-llama-linux-arm64-cuda": "1.4.8",
-    "@fugood/node-llama-linux-arm64-snapdragon": "1.4.8",
-    "@fugood/node-llama-linux-arm64-vulkan": "1.4.8",
-    "@fugood/node-llama-linux-x64": "1.4.8",
-    "@fugood/node-llama-linux-x64-cuda": "1.4.8",
-    "@fugood/node-llama-linux-x64-vulkan": "1.4.8",
-    "@fugood/node-llama-win32-arm64": "1.4.8",
-    "@fugood/node-llama-win32-arm64-vulkan": "1.4.8",
-    "@fugood/node-llama-win32-x64": "1.4.8",
-    "@fugood/node-llama-win32-x64-cuda": "1.4.8",
-    "@fugood/node-llama-win32-x64-vulkan": "1.4.8"
+    "@fugood/node-llama-darwin-arm64": "1.4.9",
+    "@fugood/node-llama-darwin-x64": "1.4.9",
+    "@fugood/node-llama-linux-arm64": "1.4.9",
+    "@fugood/node-llama-linux-arm64-cuda": "1.4.9",
+    "@fugood/node-llama-linux-arm64-snapdragon": "1.4.9",
+    "@fugood/node-llama-linux-arm64-vulkan": "1.4.9",
+    "@fugood/node-llama-linux-x64": "1.4.9",
+    "@fugood/node-llama-linux-x64-cuda": "1.4.9",
+    "@fugood/node-llama-linux-x64-vulkan": "1.4.9",
+    "@fugood/node-llama-win32-arm64": "1.4.9",
+    "@fugood/node-llama-win32-arm64-vulkan": "1.4.9",
+    "@fugood/node-llama-win32-x64": "1.4.9",
+    "@fugood/node-llama-win32-x64-cuda": "1.4.9",
+    "@fugood/node-llama-win32-x64-vulkan": "1.4.9"
   },
   "devDependencies": {
     "@babel/preset-env": "^7.24.4",

package/scripts/llama.cpp.patch

@@ -98,7 +98,7 @@ index 6085510a4..263076ce2 100644
  struct common_chat_tool_call {
      std::string name;
 diff --git a/src/llama.cpp/common/common.cpp b/src/llama.cpp/common/common.cpp
-index 5a8cf5248..8010a990e 100644
+index d4e8c7405..af3dec813 100644
 --- a/src/llama.cpp/common/common.cpp
 +++ b/src/llama.cpp/common/common.cpp
 @@ -1343,6 +1343,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
@@ -110,7 +110,7 @@ index 5a8cf5248..8010a990e 100644
      mparams.split_mode      = params.split_mode;
      mparams.tensor_split    = params.tensor_split;
 diff --git a/src/llama.cpp/common/common.h b/src/llama.cpp/common/common.h
-index d70744840..dea8c4546 100644
+index 3e314f4c8..5750a4057 100644
 --- a/src/llama.cpp/common/common.h
 +++ b/src/llama.cpp/common/common.h
 @@ -307,6 +307,7 @@ struct lr_opt {
@@ -122,7 +122,7 @@ index d70744840..dea8c4546 100644
      int32_t n_ctx                 =     0; // context size, 0 == context the model was trained with
      int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
 diff --git a/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt b/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt
-index fc31089f3..aa9befe4c 100644
+index 28fb7612e..63f7e1ca1 100644
 --- a/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt
 +++ b/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt
 @@ -106,7 +106,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
@@ -135,10 +135,10 @@ index fc31089f3..aa9befe4c 100644
              check_cxx_compiler_flag(-mfp16-format=ieee GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E)
              if (NOT "${GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
 diff --git a/src/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp b/src/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp
-index 514f086f6..792abaa58 100644
+index 6a00abacc..9e12459b6 100644
 --- a/src/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp
 +++ b/src/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp
-@@ -3213,11 +3213,26 @@ static const char * ggml_backend_hexagon_device_get_description(ggml_backend_dev
+@@ -3226,11 +3226,26 @@ static const char * ggml_backend_hexagon_device_get_description(ggml_backend_dev
      GGML_UNUSED(dev);
  }
  
@@ -168,7 +168,7 @@ index 514f086f6..792abaa58 100644
      GGML_UNUSED(dev);
  }
  
-@@ -3398,10 +3413,17 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
+@@ -3413,10 +3428,17 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
          }
      }
  
@@ -187,7 +187,7 @@ index 514f086f6..792abaa58 100644
  
      GGML_LOG_INFO("ggml-hex: Hexagon Arch version v%d\n", opt_arch);
  
-@@ -3414,6 +3436,8 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
+@@ -3429,6 +3451,8 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
          } catch (std::exception const &exc) {
              GGML_LOG_ERROR("ggml-hex: failed to create device/session %zu\n", i);
              devices[i].context = nullptr;

package/src/LlamaContext.cpp

@@ -250,6 +250,8 @@ LlamaContext::LlamaContext(const Napi::CallbackInfo &info)
   }
 
   common_params params;
+  params.fit_params = false;
+
   params.model.path = get_option<std::string>(options, "model", "");
   if (params.model.path.empty()) {
     Napi::TypeError::New(env, "Model is required").ThrowAsJavaScriptException();

package/src/llama.cpp/common/arg.cpp

@@ -420,6 +420,8 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         }
     };
 
+    std::set<std::string> seen_args;
+
     for (int i = 1; i < argc; i++) {
         const std::string arg_prefix = "--";
 
@@ -430,6 +432,9 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         if (arg_to_options.find(arg) == arg_to_options.end()) {
             throw std::invalid_argument(string_format("error: invalid argument: %s", arg.c_str()));
         }
+        if (!seen_args.insert(arg).second) {
+            LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
+        }
         auto & tmp = arg_to_options[arg];
         auto opt = *tmp.first;
         bool is_positive = tmp.second;
@@ -750,6 +755,8 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
         }
     };
 
+    std::set<std::string> seen_args;
+
     for (int i = 1; i < argc; i++) {
         const std::string arg_prefix = "--";
 
@@ -760,6 +767,9 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
         if (arg_to_options.find(arg) == arg_to_options.end()) {
             throw std::invalid_argument(string_format("error: invalid argument: %s", arg.c_str()));
         }
+        if (!seen_args.insert(arg).second) {
+            LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
+        }
         auto opt = *arg_to_options[arg];
         std::string val;
         if (opt.value_hint != nullptr) {
@@ -863,7 +873,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         sampler_type_chars += common_sampler_type_to_chr(sampler);
         sampler_type_names += common_sampler_type_to_str(sampler) + ";";
     }
-    sampler_type_names.pop_back();
+    if (!sampler_type_names.empty()) {
+        sampler_type_names.pop_back(); // remove last semicolon
+    }
 
 
     /**
@@ -1184,7 +1196,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.system_prompt = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_DIFFUSION}));
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_DIFFUSION, LLAMA_EXAMPLE_MTMD}));
     add_opt(common_arg(
         {"--perf"},
         {"--no-perf"},
@@ -1226,13 +1238,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_DIFFUSION}));
     add_opt(common_arg(
         {"--in-file"}, "FNAME",
-        "an input file (repeat to specify multiple files)",
+        "an input file (use comma-separated values to specify multiple files)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            for (const auto & item : string_split<std::string>(value, ',')) {
+                std::ifstream file(item);
+                if (!file) {
+                    throw std::runtime_error(string_format("error: failed to open file '%s'\n", item.c_str()));
+                }
+                params.in_files.push_back(item);
             }
-            params.in_files.push_back(value);
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
@@ -1969,9 +1983,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_OFFLOAD"));
     add_opt(common_arg(
         {"--image", "--audio"}, "FILE",
-        "path to an image or audio file. use with multimodal models, can be repeated if you have multiple files\n",
+        "path to an image or audio file. use with multimodal models, use comma-separated values for multiple files\n",
         [](common_params & params, const std::string & value) {
-            params.image.emplace_back(value);
+            for (const auto & item : string_split<std::string>(value, ',')) {
+                params.image.emplace_back(item);
+            }
         }
     ).set_examples({LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
@@ -2218,12 +2234,39 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ));
     add_opt(common_arg(
-        {"--override-kv"}, "KEY=TYPE:VALUE",
-        "advanced option to override model metadata by key. may be specified multiple times.\n"
-        "types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false",
+        {"--override-kv"}, "KEY=TYPE:VALUE,...",
+        "advanced option to override model metadata by key. to specify multiple overrides, either use comma-separated or repeat this argument.\n"
+        "types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false",
         [](common_params & params, const std::string & value) {
-            if (!string_parse_kv_override(value.c_str(), params.kv_overrides)) {
-                throw std::runtime_error(string_format("error: Invalid type for KV override: %s\n", value.c_str()));
+            std::vector<std::string> kv_overrides;
+
+            std::string current;
+            bool escaping = false;
+
+            for (const char c : value) {
+                if (escaping) {
+                    current.push_back(c);
+                    escaping = false;
+                } else if (c == '\\') {
+                    escaping = true;
+                } else if (c == ',') {
+                    kv_overrides.push_back(current);
+                    current.clear();
+                } else {
+                    current.push_back(c);
+                }
+            }
+
+            if (escaping) {
+                current.push_back('\\');
+            }
+
+            kv_overrides.push_back(current);
+
+            for (const auto & kv_override : kv_overrides) {
+                if (!string_parse_kv_override(kv_override.c_str(), params.kv_overrides)) {
+                    throw std::runtime_error(string_format("error: Invalid type for KV override: %s\n", kv_override.c_str()));
+                }
             }
         }
     ));
@@ -2237,33 +2280,50 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ));
     add_opt(common_arg(
         {"--lora"}, "FNAME",
-        "path to LoRA adapter (can be repeated to use multiple adapters)",
+        "path to LoRA adapter (use comma-separated values to load multiple adapters)",
         [](common_params & params, const std::string & value) {
-            params.lora_adapters.push_back({ std::string(value), 1.0, "", "", nullptr });
+            for (const auto & item : string_split<std::string>(value, ',')) {
+                params.lora_adapters.push_back({ item, 1.0, "", "", nullptr });
+            }
         }
         // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
     add_opt(common_arg(
-        {"--lora-scaled"}, "FNAME", "SCALE",
-        "path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters)",
-        [](common_params & params, const std::string & fname, const std::string & scale) {
-            params.lora_adapters.push_back({ fname, std::stof(scale), "", "", nullptr });
+        {"--lora-scaled"}, "FNAME:SCALE,...",
+        "path to LoRA adapter with user defined scaling (format: FNAME:SCALE,...)\n"
+        "note: use comma-separated values",
+        [](common_params & params, const std::string & value) {
+            for (const auto & item : string_split<std::string>(value, ',')) {
+                auto parts = string_split<std::string>(item, ':');
+                if (parts.size() != 2) {
+                    throw std::invalid_argument("lora-scaled format: FNAME:SCALE");
+                }
+                params.lora_adapters.push_back({ parts[0], std::stof(parts[1]), "", "", nullptr });
+            }
         }
         // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
     add_opt(common_arg(
         {"--control-vector"}, "FNAME",
-        "add a control vector\nnote: this argument can be repeated to add multiple control vectors",
+        "add a control vector\nnote: use comma-separated values to add multiple control vectors",
         [](common_params & params, const std::string & value) {
-            params.control_vectors.push_back({ 1.0f, value, });
+            for (const auto & item : string_split<std::string>(value, ',')) {
+                params.control_vectors.push_back({ 1.0f, item, });
+            }
         }
     ));
     add_opt(common_arg(
-        {"--control-vector-scaled"}, "FNAME", "SCALE",
+        {"--control-vector-scaled"}, "FNAME:SCALE,...",
         "add a control vector with user defined scaling SCALE\n"
-        "note: this argument can be repeated to add multiple scaled control vectors",
-        [](common_params & params, const std::string & fname, const std::string & scale) {
-            params.control_vectors.push_back({ std::stof(scale), fname });
+        "note: use comma-separated values (format: FNAME:SCALE,...)",
+        [](common_params & params, const std::string & value) {
+            for (const auto & item : string_split<std::string>(value, ',')) {
+                auto parts = string_split<std::string>(item, ':');
+                if (parts.size() != 2) {
+                    throw std::invalid_argument("control-vector-scaled format: FNAME:SCALE");
+                }
+                params.control_vectors.push_back({ std::stof(parts[1]), parts[0] });
+            }
         }
     ));
     add_opt(common_arg(
@@ -2353,13 +2413,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("HF_TOKEN"));
     add_opt(common_arg(
         {"--context-file"}, "FNAME",
-        "file to load context from (repeat to specify multiple files)",
+        "file to load context from (use comma-separated values to specify multiple files)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value, std::ios::binary);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            for (const auto & item : string_split<std::string>(value, ',')) {
+                std::ifstream file(item, std::ios::binary);
+                if (!file) {
+                    throw std::runtime_error(string_format("error: failed to open file '%s'\n", item.c_str()));
+                }
+                params.context_files.push_back(item);
             }
-            params.context_files.push_back(value);
         }
     ).set_examples({LLAMA_EXAMPLE_RETRIEVAL}));
     add_opt(common_arg(
@@ -2550,6 +2612,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.api_prefix = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_API_PREFIX"));
+    add_opt(common_arg(
+        {"--webui-config"}, "JSON",
+        "JSON that provides default WebUI settings (overrides WebUI defaults)",
+        [](common_params & params, const std::string & value) {
+            params.webui_config_json = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_CONFIG"));
+    add_opt(common_arg(
+        {"--webui-config-file"}, "PATH",
+        "JSON file that provides default WebUI settings (overrides WebUI defaults)",
+        [](common_params & params, const std::string & value) {
+            params.webui_config_json = read_file(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_CONFIG_FILE"));
     add_opt(common_arg(
         {"--webui"},
         {"--no-webui"},

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

@@ -1092,7 +1092,7 @@ common_init_result::common_init_result(common_params & params) :
     auto cparams = common_context_params_to_llama(params);
 
     if (params.fit_params) {
-        LOG_INF("%s: fitting params to device memory, to report bugs during this step use -fit off (or --verbose if you can't)\n", __func__);
+        LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
         llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
             params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
             params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);

package/src/llama.cpp/common/common.h

@@ -485,8 +485,11 @@ struct common_params {
 
     std::map<std::string, std::string> default_template_kwargs;
 
+    // webui configs
+    bool webui = true;
+    std::string webui_config_json;
+
     // "advanced" endpoints are disabled by default for better security
-    bool webui            = true;
     bool endpoint_slots   = true;
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;

package/src/llama.cpp/common/sampling.cpp

@@ -104,10 +104,9 @@ struct ring_buffer {
 struct common_sampler {
     common_params_sampling params;
 
+    struct llama_sampler * grmr;
     struct llama_sampler * chain;
 
-    bool grammar;
-
     ring_buffer<llama_token> prev;
 
     std::vector<llama_token_data> cur;
@@ -167,15 +166,14 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
     lparams.no_perf = params.no_perf;
 
+    llama_sampler * grmr = nullptr;
     llama_sampler * chain = llama_sampler_chain_init(lparams);
 
-    bool grammar = false;
     std::vector<llama_sampler *> samplers;
 
     if (params.grammar.compare(0, 11, "%llguidance") == 0) {
 #ifdef LLAMA_USE_LLGUIDANCE
-        samplers.push_back(llama_sampler_init_llg(vocab, "lark", params.grammar.c_str()));
-        grammar = true;
+        grmr = llama_sampler_init_llg(vocab, "lark", params.grammar.c_str());
 #else
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
@@ -224,15 +222,12 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
         if (!params.grammar.empty()) {
              if (params.grammar_lazy) {
-                 samplers.push_back(
-                         llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
-                             trigger_patterns_c.data(), trigger_patterns_c.size(),
-                             trigger_tokens.data(),     trigger_tokens.size()));
+                 grmr = llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
+                         trigger_patterns_c.data(), trigger_patterns_c.size(),
+                         trigger_tokens.data(), trigger_tokens.size());
              } else {
-                 samplers.push_back(llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"));
+                 grmr = llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
              }
-
-             grammar = true;
         }
     }
 
@@ -303,8 +298,8 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
     auto * result = new common_sampler {
         /* .params  = */ params,
+        /* .grmr    = */ grmr,
         /* .chain   = */ chain,
-        /* .grammar = */ grammar,
         /* .prev    = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
         /* .cur     = */ {},
         /* .cur_p   = */ {},
@@ -315,6 +310,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
 void common_sampler_free(struct common_sampler * gsmpl) {
     if (gsmpl) {
+        llama_sampler_free(gsmpl->grmr);
         llama_sampler_free(gsmpl->chain);
 
         delete gsmpl;
@@ -324,25 +320,12 @@ void common_sampler_free(struct common_sampler * gsmpl) {
 void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
     const auto tm = gsmpl->tm();
 
-    if (gsmpl->grammar) {
-        const int n_smpl = llama_sampler_chain_n(gsmpl->chain);
-
-        for (int i = 0; i < n_smpl; i++) {
-            auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
-
-            // the grammar sampler is always the first one
-            if (i == 0) {
-                if (accept_grammar) {
-                    llama_sampler_accept(smpl, token);
-                }
-            } else {
-                llama_sampler_accept(smpl, token);
-            }
-        }
-    } else {
-        llama_sampler_accept(gsmpl->chain, token);
+    if (gsmpl->grmr && accept_grammar) {
+        llama_sampler_accept(gsmpl->grmr, token);
     }
 
+    llama_sampler_accept(gsmpl->chain, token);
+
     gsmpl->prev.push_back(token);
 }
 
@@ -353,8 +336,8 @@ void common_sampler_reset(struct common_sampler * gsmpl) {
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
     return new common_sampler {
         /* .params  = */ gsmpl->params,
+        /* .grmr    = */ llama_sampler_clone(gsmpl->grmr),
         /* .chain   = */ llama_sampler_clone(gsmpl->chain),
-        /* .grammar = */ gsmpl->grammar,
         /* .prev    = */ gsmpl->prev,
         /* .cur     = */ gsmpl->cur,
         /* .cur_p   = */ gsmpl->cur_p,
@@ -410,7 +393,7 @@ struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl) {
     return gsmpl->chain;
 }
 
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx) {
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
     llama_synchronize(ctx);
 
     // start measuring sampling time after the llama_context synchronization in order to not measure any ongoing async operations
@@ -418,11 +401,42 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 
     llama_token id = LLAMA_TOKEN_NULL;
 
+    auto & grmr  = gsmpl->grmr;
     auto & chain = gsmpl->chain;
     auto & cur_p = gsmpl->cur_p; // initialized by set_logits
 
     gsmpl->set_logits(ctx, idx);
 
+    if (grammar_first) {
+        llama_sampler_apply(grmr, &cur_p);
+    }
+
+    llama_sampler_apply(chain, &cur_p);
+
+    id = cur_p.data[cur_p.selected].id;
+
+    if (grammar_first) {
+        return id;
+    }
+
+    // check if it the sampled token fits the grammar (grammar-based rejection sampling)
+    {
+        llama_token_data       single_token_data       = { id, 1.0f, 0.0f };
+        llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
+
+        llama_sampler_apply(grmr, &single_token_data_array);
+
+        const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
+        if (is_valid) {
+            return id;
+        }
+    }
+
+    // resampling:
+    // if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
+    gsmpl->set_logits(ctx, idx);
+
+    llama_sampler_apply(grmr,  &cur_p);
     llama_sampler_apply(chain, &cur_p);
 
     GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
@@ -432,7 +446,7 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
     return id;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft) {
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first) {
     GGML_ASSERT(idxs.size() == draft.size() + 1 && "idxs.size() must be draft.size() + 1");
 
     std::vector<llama_token> result;
@@ -440,7 +454,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
 
     size_t i = 0;
     for (; i < draft.size(); i++) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -452,7 +466,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     }
 
     if (i == draft.size()) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -462,13 +476,13 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     return result;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft) {
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first) {
     std::vector<int> idxs(draft.size() + 1);
     for (size_t i = 0; i < idxs.size(); ++i) {
         idxs[i] = i;
     }
 
-    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft);
+    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft, grammar_first);
 }
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {

package/src/llama.cpp/common/sampling.h

@@ -57,7 +57,10 @@ struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl);
 // - check if the token fits the grammar (if any)
 // - if not: resample by first applying the grammar constraints and then sampling again (slower path)
 //
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx);
+// if grammar_first is true, the grammar is applied before the samplers (slower)
+// useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
+//
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
 
 // generalized version of common_sampler_sample
 //
@@ -75,10 +78,10 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 //
 // returns at least 1 token, up to idxs.size()
 //
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first = false);
 
 // assume idxs == [ 0, 1, 2, ..., draft.size() ]
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first = false);
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 

package/src/llama.cpp/common/speculative.cpp

@@ -315,7 +315,7 @@ llama_tokens common_speculative_gen_draft(
     for (int i = 0; i < params.n_draft; ++i) {
         common_batch_clear(batch);
 
-        common_sampler_sample(smpl, ctx_dft, 0);
+        common_sampler_sample(smpl, ctx_dft, 0, true);
 
         const auto * cur_p = common_sampler_get_candidates(smpl, true);
 

package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt

@@ -458,6 +458,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_RV_ZFH)
                 string(APPEND MARCH_STR "_zfh")
             endif()
+
             if (GGML_XTHEADVECTOR)
                 string(APPEND MARCH_STR "_xtheadvector")
             elseif (GGML_RVV)
@@ -465,6 +466,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 if (GGML_RV_ZVFH)
                     string(APPEND MARCH_STR "_zvfh")
                 endif()
+                if (GGML_RV_ZVFBFWMA)
+                    string(APPEND MARCH_STR "_zvfbfwma")
+                endif()
             endif()
             if (GGML_RV_ZICBOP)
                 string(APPEND MARCH_STR "_zicbop")