@fugood/llama.node 1.4.11 → 1.4.13

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

@@ -306,7 +306,7 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
                 }
             }
         } else {
-            jmsg["content"] = json(); // null
+            jmsg["content"] = "";
         }
         if (!msg.reasoning_content.empty()) {
             jmsg["reasoning_content"] = msg.reasoning_content;
@@ -367,8 +367,8 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
                 const auto & function = tool.at("function");
                 result.push_back({
                     /* .name = */ function.at("name"),
-                    /* .description = */ function.at("description"),
-                    /* .parameters = */ function.at("parameters").dump(),
+                    /* .description = */ function.value("description", ""),
+                    /* .parameters = */ function.value("parameters", json::object()).dump(),
                 });
             }
         }
@@ -656,6 +656,7 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_QWEN3_CODER_XML: return "Qwen3 Coder";
         case COMMON_CHAT_FORMAT_APRIEL_1_5: return "Apriel 1.5";
         case COMMON_CHAT_FORMAT_XIAOMI_MIMO: return "Xiaomi MiMo";
+        case COMMON_CHAT_FORMAT_SOLAR_OPEN: return "Solar Open";
         case COMMON_CHAT_FORMAT_PEG_SIMPLE: return "peg-simple";
         case COMMON_CHAT_FORMAT_PEG_NATIVE: return "peg-native";
         case COMMON_CHAT_FORMAT_PEG_CONSTRUCTED: return "peg-constructed";
@@ -2051,7 +2052,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
             // Trigger on tool calls that appear in the commentary channel
             data.grammar_triggers.push_back({
                 COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-                "<\\|channel\\|>(commentary|analysis) to"
+                "<\\|channel\\|>(?:commentary|analysis) to"
             });
 
             // Trigger tool calls that appear in the role section, either at the
@@ -2384,17 +2385,17 @@ static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat
                 (inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call));
             // Trigger on some common known "good bad" outputs (only from the start and with a json that's about a specific argument name to avoid false positives)
             data.grammar_triggers.push_back({
-                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
                 // If thinking_forced_open, then we capture the </think> tag in the grammar,
                 // (important for required tool choice) and in the trigger's first capture (decides what is sent to the grammar)
-                std::string(data.thinking_forced_open ? "[\\s\\S]*?(</think>\\s*)" : "(?:<think>[\\s\\S]*?</think>\\s*)?") + (
+                std::string(data.thinking_forced_open ? "(</think>\\s*)" : "") + (
                     "\\s*("
                     "(?:<tool_call>"
                     "|<function"
                     "|(?:```(?:json|xml)?\n\\s*)?(?:<function_call>|<tools>|<xml><json>|<response>)?"
                     "\\s*\\{\\s*\"name\"\\s*:\\s*\"(?:" + string_join(escaped_names, "|") + ")\""
                     ")"
-                    ")[\\s\\S]*"
+                    ")"
                 ),
             });
             data.preserved_tokens = {
@@ -2504,6 +2505,27 @@ static common_chat_params common_chat_params_init_granite(const common_chat_temp
     return data;
 }
 
+static common_chat_params common_chat_params_init_solar_open(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+
+    // TODO: Reasoning effort
+    json additional_context = {};
+
+    data.prompt = apply(tmpl, inputs, std::nullopt, std::nullopt, additional_context);
+    data.format = COMMON_CHAT_FORMAT_SOLAR_OPEN;
+
+    data.preserved_tokens = {
+        "<|think|>",
+        "<|content|>",
+        "<|begin|>",
+        "<|end|>",
+    };
+
+    // TODO: Tool calling
+
+    return data;
+}
+
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     data.prompt = apply(tmpl, inputs);
@@ -2767,6 +2789,13 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_magistral(tmpl, params);
     }
 
+    // Solar Open
+    if (src.find("<|tool_response:begin|>") != std::string::npos &&
+        src.find("<|tool_response:name|>") != std::string::npos &&
+        src.find("<|tool_response:result|>") != std::string::npos) {
+        return common_chat_params_init_solar_open(tmpl, params);
+    }
+
     // Plain handler (no tools)
     if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
         return common_chat_params_init_without_tools(tmpl, params);

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

@@ -135,6 +135,7 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_QWEN3_CODER_XML,
     COMMON_CHAT_FORMAT_APRIEL_1_5,
     COMMON_CHAT_FORMAT_XIAOMI_MIMO,
+    COMMON_CHAT_FORMAT_SOLAR_OPEN,
 
     // These are intended to be parsed by the PEG parser
     COMMON_CHAT_FORMAT_PEG_SIMPLE,

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

@@ -251,7 +251,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
         case GGML_SCHED_PRIO_REALTIME: p = -20; break;
     }
 
-    if (!setpriority(PRIO_PROCESS, 0, p)) {
+    if (setpriority(PRIO_PROCESS, 0, p) != 0) {
         LOG_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
         return false;
     }
@@ -1078,12 +1078,15 @@ struct common_init_result::impl {
     impl() = default;
     ~impl() = default;
 
+    // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
+
     llama_model_ptr   model;
     llama_context_ptr context;
 
     std::vector<llama_adapter_lora_ptr> lora;
 
     std::vector<common_sampler_ptr> samplers;
+    std::vector<llama_sampler_seq_config> samplers_seq_config;
 };
 
 common_init_result::common_init_result(common_params & params) :
@@ -1107,6 +1110,25 @@ common_init_result::common_init_result(common_params & params) :
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
+    // load and optionally apply lora adapters (must be loaded before context creation)
+    for (auto & la : params.lora_adapters) {
+        llama_adapter_lora_ptr lora;
+        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
+        if (lora == nullptr) {
+            LOG_ERR("%s: failed to load lora adapter '%s'\n", __func__, la.path.c_str());
+            pimpl->model.reset(model);
+            return;
+        }
+
+        char buf[1024];
+        la.ptr = lora.get();
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
+        la.task_name = buf;
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
+        la.prompt_prefix = buf;
+        pimpl->lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
+    }
+
     // updates params.sampling
     // TODO: fix naming
     common_init_sampler_from_model(model, params.sampling);
@@ -1141,10 +1163,19 @@ common_init_result::common_init_result(common_params & params) :
     //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
     //}
 
+    // init the backend samplers as part of the context creation
     pimpl->samplers.resize(cparams.n_seq_max);
+    pimpl->samplers_seq_config.resize(cparams.n_seq_max);
 
     for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
         pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
+        pimpl->samplers_seq_config[i] = { i, common_sampler_get(pimpl->samplers[i].get()) };
+    }
+
+    // TODO: temporarily gated behind a flag
+    if (params.sampling.backend_sampling) {
+        cparams.samplers   = pimpl->samplers_seq_config.data();
+        cparams.n_samplers = pimpl->samplers_seq_config.size();
     }
 
     llama_context * lctx = llama_init_from_model(model, cparams);
@@ -1168,6 +1199,12 @@ common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
     return pimpl->samplers[seq_id].get();
 }
 
+void common_init_result::reset_samplers() {
+    for (int i = 0; i < (int) pimpl->samplers.size(); ++i) {
+        llama_sampler_reset(common_sampler_get(pimpl->samplers[i].get()));
+    }
+}
+
 std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
     return pimpl->lora;
 }
@@ -1243,24 +1280,6 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         }
     }
 
-    // load and optionally apply lora adapters
-    for (auto & la : params.lora_adapters) {
-        llama_adapter_lora_ptr lora;
-        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
-        if (lora == nullptr) {
-            LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
-            return res;
-        }
-
-        char buf[1024];
-        la.ptr = lora.get();
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
-        la.task_name = buf;
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
-        la.prompt_prefix = buf;
-        res->lora().emplace_back(std::move(lora)); // copy to list of loaded adapters
-    }
-
     if (!params.lora_init_without_apply) {
         common_set_adapter_lora(lctx, params.lora_adapters);
     }
@@ -1301,6 +1320,9 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);
+
+        // reset samplers to reset RNG state after warmup to the seeded state
+        res->reset_samplers();
     }
 
     return res;
@@ -1339,11 +1361,8 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.devices = params.devices.data();
     }
 
-    if (params.n_gpu_layers != -1) {
-        mparams.n_gpu_layers = params.n_gpu_layers;
-    }
-
     mparams.vocab_only      = params.vocab_only;
+    mparams.n_gpu_layers    = params.n_gpu_layers;
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;

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

@@ -80,6 +80,7 @@ int32_t cpu_get_num_math();
 //
 
 enum llama_example {
+    LLAMA_EXAMPLE_DEBUG,
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
     LLAMA_EXAMPLE_COMPLETION,
@@ -216,6 +217,8 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
+    bool backend_sampling = false;
+
     bool has_logit_bias() const {
         return !logit_bias.empty();
     }
@@ -330,7 +333,7 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM, -1 is auto, <= -2 is all
     int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
     float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
     bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory
@@ -371,6 +374,11 @@ struct common_params {
     std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
 
+    // llama-debug specific options
+    std::string logits_output_dir = "data"; // directory for saving logits output files                     // NOLINT
+    bool        save_logits       = false;  // whether to save logits to files                              // NOLINT
+    std::vector<std::string> tensor_filter; // filter tensor names for debug output (regex)                 // NOLINT
+
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
@@ -690,7 +698,9 @@ struct common_init_result {
 
     llama_model * model();
     llama_context * context();
+
     common_sampler * sampler(llama_seq_id seq_id);
+    void reset_samplers();
 
     std::vector<llama_adapter_lora_ptr> & lora();
 

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

@@ -106,12 +106,16 @@ static void llama_sampler_llg_free(llama_sampler * smpl) {
 }
 
 static llama_sampler_i llama_sampler_llg_i = {
-    /* .name   = */ llama_sampler_llg_name,
-    /* .accept = */ llama_sampler_llg_accept_impl,
-    /* .apply  = */ llama_sampler_llg_apply,
-    /* .reset  = */ llama_sampler_llg_reset,
-    /* .clone  = */ llama_sampler_llg_clone,
-    /* .free   = */ llama_sampler_llg_free,
+    /* .name              = */ llama_sampler_llg_name,
+    /* .accept            = */ llama_sampler_llg_accept_impl,
+    /* .apply             = */ llama_sampler_llg_apply,
+    /* .reset             = */ llama_sampler_llg_reset,
+    /* .clone             = */ llama_sampler_llg_clone,
+    /* .free              = */ llama_sampler_llg_free,
+    /* .backend_init      = */ NULL,
+    /* .backend_accept    = */ NULL,
+    /* .backend_apply     = */ NULL,
+    /* .backend_set_input = */ NULL,
 };
 
 static size_t llama_sampler_llg_tokenize_fn(const void * user_data, const uint8_t * bytes, size_t bytes_len,

package/src/llama.cpp/common/regex-partial.cpp

@@ -27,7 +27,7 @@ common_regex_match common_regex::search(const std::string & input, size_t pos, b
         return res;
     }
     std::match_results<std::string::const_reverse_iterator> srmatch;
-    if (std::regex_match(input.rbegin(), input.rend() - pos, srmatch, rx_reversed_partial)) {
+    if (std::regex_search(input.rbegin(), input.rend() - pos, srmatch, rx_reversed_partial, std::regex_constants::match_continuous)) {
         auto group = srmatch[1].str();
         if (group.length() != 0) {
             auto it = srmatch[1].second.base();
@@ -55,18 +55,18 @@ common_regex_match common_regex::search(const std::string & input, size_t pos, b
   to see if a string ends with a partial regex match, but but it's not in std::regex yet.
   Instead, we'll the regex into a partial match regex operating as a full match on the reverse iterators of the input.
 
-  - /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:(?:d)?c)?b)?a).*
-  - /a|b/ -> (a|b).*
+  - /abcd/ -> ^(dcba|cba|ba|a) -> ^((?:(?:(?:(?:d)?c)?b)?a)
+  - /a|b/ -> ^(a|b)
   - /a*?/ -> error, could match ""
-  - /a*b/ -> ((?:b)?a*+).* (final repetitions become eager)
-  - /.*?ab/ -> ((?:b)?a).* (merge .*)
-  - /a.*?b/ -> ((?:b)?.*?a).* (keep reluctant matches)
-  - /a(bc)d/ -> ((?:(?:d)?(?:(?:c)?b))?a).*
-  - /a(bc|de)/ -> ((?:(?:(?:e)?d)?|(?:(?:c)?b)?)?a).*
-  - /ab{2,4}c/ -> abbb?b?c -> ((?:(?:(?:(?:(?:c)?b)?b)?b?)?b?)?a).*
+  - /a*b/ -> ^((?:b)?a*+) (final repetitions become eager)
+  - /.*?ab/ -> ^((?:b)?a) (omit .*)
+  - /a.*?b/ -> ^((?:b)?.*?a) (keep reluctant matches)
+  - /a(bc)d/ -> ^((?:(?:d)?(?:(?:c)?b))?a)
+  - /a(bc|de)/ -> ^((?:(?:(?:e)?d)?|(?:(?:c)?b)?)?a)
+  - /ab{2,4}c/ -> ^cbbb?b?a -> ^((?:(?:(?:(?:(?:c)?b)?b)?b?)?b?)?a)
 
-  The regex will match a reversed string fully, and the end of the first (And only) capturing group will indicate the reversed start of the original partial pattern
-  (i.e. just where the final .* starts in the inverted pattern; all other groups are turned into non-capturing groups, and reluctant quantifiers are ignored)
+  The regex will match a reversed string fully, and the end of the first (And only) capturing group will indicate the reversed start of the original partial pattern.
+  All other groups are turned into non-capturing groups, and reluctant quantifiers are ignored.
 */
 std::string regex_to_reversed_partial_regex(const std::string & pattern) {
     auto it = pattern.begin();
@@ -177,7 +177,7 @@ std::string regex_to_reversed_partial_regex(const std::string & pattern) {
             }
         }
 
-        // /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:d)?c)?b)?a).*
+        // /abcd/ -> ^(dcba|cba|ba|a) -> ^((?:(?:(?:d)?c)?b)?a)
         // if n(=4) parts, opening n-1(=3) non-capturing groups after the 1 capturing group
         // We'll do the outermost capturing group and final .* in the enclosing function.
         std::vector<std::string> res_alts;
@@ -200,5 +200,5 @@ std::string regex_to_reversed_partial_regex(const std::string & pattern) {
         throw std::runtime_error("Unmatched '(' in pattern");
     }
 
-    return "(" + res + ")[\\s\\S]*";
+    return "^(" + res + ")";
 }

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

@@ -120,17 +120,34 @@ struct common_sampler {
     }
 
     void set_logits(struct llama_context * ctx, int idx) {
-        const auto * logits = llama_get_logits_ith(ctx, idx);
+        const float *       sampled_probs  = llama_get_sampled_probs_ith     (ctx, idx);
+        const float *       sampled_logits = llama_get_sampled_logits_ith    (ctx, idx);
+        const llama_token * sampled_ids    = llama_get_sampled_candidates_ith(ctx, idx);
 
         const llama_model * model = llama_get_model(ctx);
         const llama_vocab * vocab = llama_model_get_vocab(model);
 
         const int n_vocab = llama_vocab_n_tokens(vocab);
 
-        cur.resize(n_vocab);
-
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+        if (sampled_probs) {
+            const uint32_t sampled_probs_count = llama_get_sampled_probs_count_ith(ctx, idx);
+            cur.resize(sampled_probs_count);
+            for (uint32_t i = 0; i < sampled_probs_count; ++i) {
+                cur[i] = llama_token_data{sampled_ids[i], sampled_logits[i], sampled_probs[i]};
+            }
+        } else if (sampled_logits) {
+            const uint32_t sampled_logits_count = llama_get_sampled_logits_count_ith(ctx, idx);
+            cur.resize(sampled_logits_count);
+            for (uint32_t i = 0; i < sampled_logits_count; i++) {
+                cur[i] = llama_token_data{sampled_ids[i], sampled_logits[i], 0.0f};
+            }
+        } else {
+            const auto * logits = llama_get_logits_ith(ctx, idx);
+            GGML_ASSERT(logits != nullptr);
+            cur.resize(n_vocab);
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+            }
         }
 
         cur_p = { cur.data(), cur.size(), -1, false };
@@ -159,7 +176,7 @@ std::string common_params_sampling::print() const {
     return std::string(result);
 }
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
+struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params) {
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
@@ -179,24 +196,30 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 #endif // LLAMA_USE_LLGUIDANCE
     } else {
         std::vector<std::string> trigger_patterns;
-        std::vector<std::string> patterns_anywhere;
         std::vector<llama_token> trigger_tokens;
         for (const auto & trigger : params.grammar_triggers) {
             switch (trigger.type) {
                 case COMMON_GRAMMAR_TRIGGER_TYPE_WORD:
                 {
                     const auto & word = trigger.value;
-                    patterns_anywhere.push_back(regex_escape(word));
+                    trigger_patterns.push_back(regex_escape(word));
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
                 {
-                    patterns_anywhere.push_back(trigger.value);
+                    trigger_patterns.push_back(trigger.value);
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL:
                 {
-                    trigger_patterns.push_back(trigger.value);
+                    const auto & pattern = trigger.value;
+                    std::string anchored = "^$";
+                    if (!pattern.empty()) {
+                        anchored = (pattern.front() != '^' ? "^" : "")
+                            + pattern
+                            + (pattern.back() != '$' ? "$" : "");
+                    }
+                    trigger_patterns.push_back(anchored);
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
@@ -210,10 +233,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
             }
         }
 
-        if (!patterns_anywhere.empty()) {
-            trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*");
-        }
-
         std::vector<const char *> trigger_patterns_c;
         trigger_patterns_c.reserve(trigger_patterns.size());
         for (const auto & regex : trigger_patterns) {
@@ -296,6 +315,12 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
         llama_sampler_chain_add(chain, smpl);
     }
 
+    if (grmr && params.backend_sampling) {
+        LOG_WRN("%s: backend sampling is not compatible with grammar, disabling\n", __func__);
+
+        params.backend_sampling = false;
+    }
+
     auto * result = new common_sampler {
         /* .params  = */ params,
         /* .grmr    = */ grmr,
@@ -405,6 +430,25 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
     auto & chain = gsmpl->chain;
     auto & cur_p = gsmpl->cur_p; // initialized by set_logits
 
+    // Check if a backend sampler has already sampled a token in which case we
+    // return that token id directly.
+    {
+        id = llama_get_sampled_token_ith(ctx, idx);
+
+        if (id != LLAMA_TOKEN_NULL) {
+            LOG_DBG("%s: Backend sampler selected token: '%d'. Will not run any CPU samplers\n", __func__, id);
+
+            GGML_ASSERT(!gsmpl->grmr && "using grammar in combination with backend sampling is not supported");
+
+            // TODO: simplify
+            gsmpl->cur.resize(1);
+            gsmpl->cur[0] = { id, 0.0f, 1.0f };
+            cur_p = { gsmpl->cur.data(), gsmpl->cur.size(), 0, true };
+
+            return id;
+        }
+    }
+
     gsmpl->set_logits(ctx, idx);
 
     if (grammar_first) {

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

@@ -36,7 +36,8 @@ struct common_sampler;
 
 // llama_sampler API overloads
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params);
+// note: can mutate params in some cases
+struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params);
 
 void common_sampler_free(struct common_sampler * gsmpl);
 
@@ -48,6 +49,7 @@ struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
 // arguments can be nullptr to skip printing
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
 
+// get the underlying llama_sampler_chain
 struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl);
 
 // extended sampling implementation:

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

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 9)
-set(GGML_VERSION_PATCH 4)
+set(GGML_VERSION_PATCH 5)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
@@ -430,10 +430,22 @@ if (MSVC)
     configure_msvc_target(ggml-cpu-x64)
     configure_msvc_target(ggml-cpu-sse42)
     configure_msvc_target(ggml-cpu-sandybridge)
+    # __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+    # skipping            ggml-cpu-ivybridge
+    # skipping            ggml-cpu-piledriver
     configure_msvc_target(ggml-cpu-haswell)
     configure_msvc_target(ggml-cpu-skylakex)
+    configure_msvc_target(ggml-cpu-cannonlake)
+    configure_msvc_target(ggml-cpu-cascadelake)
     configure_msvc_target(ggml-cpu-icelake)
+    # MSVC 2022 doesn't support BF16 intrinsics without `/arch:AVX10.1` ?!
+    # https://learn.microsoft.com/en-us/cpp/intrinsics/x64-amd64-intrinsics-list?view=msvc-170
+    # https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
+    # skipping            ggml-cpu-cooperlake
+    # skipping            ggml-cpu-zen4
     configure_msvc_target(ggml-cpu-alderlake)
+    # MSVC doesn't support AMX
+    # skipping            ggml-cpu-sapphirerapids
 
     if (GGML_BUILD_EXAMPLES)
         configure_msvc_target(common-ggml)

package/src/llama.cpp/ggml/include/ggml-backend.h

@@ -358,7 +358,7 @@ extern "C" {
     typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node);
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor const * const * test_nodes, size_t num_test_nodes);
 
     // Tensor initialization
     GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

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

@@ -357,15 +357,29 @@ if (GGML_CPU_ALL_VARIANTS)
     endif()
     if (GGML_SYSTEM_ARCH STREQUAL "x86")
         ggml_add_cpu_backend_variant(x64)
-        ggml_add_cpu_backend_variant(sse42        SSE42)
-        ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
-        ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
-        ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
-        ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
-        ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
+        ggml_add_cpu_backend_variant(sse42              SSE42)
+        ggml_add_cpu_backend_variant(sandybridge        SSE42 AVX)
+        if (NOT MSVC)
+            # __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+            ggml_add_cpu_backend_variant(ivybridge      SSE42 AVX F16C)
+            ggml_add_cpu_backend_variant(piledriver     SSE42 AVX F16C FMA)
+        endif()
+        ggml_add_cpu_backend_variant(haswell            SSE42 AVX F16C FMA AVX2 BMI2)
+        ggml_add_cpu_backend_variant(skylakex           SSE42 AVX F16C FMA AVX2 BMI2 AVX512)
+        ggml_add_cpu_backend_variant(cannonlake         SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI)
+        ggml_add_cpu_backend_variant(cascadelake        SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VNNI)
+        ggml_add_cpu_backend_variant(icelake            SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI AVX512_VNNI)
+        if (NOT MSVC)
+            # MSVC 2022 doesn't support BF16 intrinsics without `/arch:AVX10.1` ?!
+            # https://learn.microsoft.com/en-us/cpp/intrinsics/x64-amd64-intrinsics-list?view=msvc-170
+            # https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
+            ggml_add_cpu_backend_variant(cooperlake     SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VNNI AVX512_BF16)
+            ggml_add_cpu_backend_variant(zen4           SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16)
+        endif()
+        ggml_add_cpu_backend_variant(alderlake          SSE42 AVX F16C FMA AVX2 BMI2 AVX_VNNI)
         if (NOT MSVC)
             # MSVC doesn't support AMX
-            ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+            ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
         endif()
     elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
         if (CMAKE_SYSTEM_NAME MATCHES "Linux")
@@ -387,8 +401,8 @@ if (GGML_CPU_ALL_VARIANTS)
             ggml_add_cpu_backend_variant(android_armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
             ggml_add_cpu_backend_variant(android_armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
             ggml_add_cpu_backend_variant(android_armv9.0_1    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE2)
-            ggml_add_cpu_backend_variant(android_armv9.2_1    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SME)
-            ggml_add_cpu_backend_variant(android_armv9.2_2    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE SME)
+            ggml_add_cpu_backend_variant(android_armv9.2_1    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE SME)
+            ggml_add_cpu_backend_variant(android_armv9.2_2    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE SVE2 SME)
         elseif (APPLE)
             ggml_add_cpu_backend_variant(apple_m1             DOTPROD)
             ggml_add_cpu_backend_variant(apple_m2_m3          DOTPROD MATMUL_INT8)

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

@@ -561,9 +561,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.14.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.16.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "45e110675d93f99f82c23a1afcca76bc")
+        set(KLEIDIAI_ARCHIVE_MD5  "0a9e9008adb6031f9e8cf70dff4a3321")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -615,6 +615,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED)
         string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
         string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
+        string(FIND "${ARCH_FLAGS_TEMP}" "+sve" SVE_ENABLED)
 
         set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
 
@@ -659,6 +660,15 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
         endif()
 
+        if (NOT SVE_ENABLED MATCHES -1)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/kai_common_sve_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.c)
+        endif()
+
         set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")
         list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES})
     endif()

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -328,7 +328,7 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <intrin.h>
-#elif defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
+#elif defined(__SSE__) || defined(__SSE3__) || defined(__SSSE3__) || defined(__AVX__) || defined(__F16C__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX512BF16__)
 #include <immintrin.h>
 #endif