minor : std::unordered_set over std::set

* lora: count lora nodes in graph_max_nodes

* 3 nodes per weight

* 4 nodes

* keep track n_lora_nodes from llama_model

* fix assert

* rm redundant header

* common: load adapters before context creation

* use 6 nodes

common/common.cpp

@@ -1109,6 +1109,25 @@ 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);
@@ -1245,24 +1264,6 @@ 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);
     }

include/llama-cpp.h

@@ -21,7 +21,9 @@ struct llama_sampler_deleter {
 };
 
 struct llama_adapter_lora_deleter {
-    void operator()(llama_adapter_lora * adapter) { llama_adapter_lora_free(adapter); }
+    void operator()(llama_adapter_lora *) {
+        // llama_adapter_lora_free is deprecated
+    }
 };
 
 typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;

include/llama.h

@@ -607,6 +607,8 @@ extern "C" {
     //
 
     // Load a LoRA adapter from file
+    // The adapter is valid as long as the associated model is not freed
+    // All adapters must be loaded before context creation
     LLAMA_API struct llama_adapter_lora * llama_adapter_lora_init(
             struct llama_model * model,
             const char * path_lora);
@@ -631,7 +633,8 @@ extern "C" {
 
     // Manually free a LoRA adapter
     // NOTE: loaded adapters will be free when the associated model is deleted
-    LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
+    LLAMA_API DEPRECATED(void llama_adapter_lora_free(struct llama_adapter_lora * adapter),
+            "adapters are now freed together with the associated model");
 
     // Get the invocation tokens if the current lora is an alora
     LLAMA_API uint64_t            llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter);

src/llama-adapter.cpp

@@ -411,6 +411,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
         }
     }
 
+    // register adapter with model
+    model.loras.insert(&adapter);
+
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
 
@@ -468,8 +471,8 @@ int32_t llama_adapter_meta_val_str_by_index(const llama_adapter_lora * adapter,
     return snprintf(buf, buf_size, "%s", it->second.c_str());
 }
 
-void llama_adapter_lora_free(llama_adapter_lora * adapter) {
-    delete adapter;
+void llama_adapter_lora_free(llama_adapter_lora *) {
+    // deprecated: adapters are freed by llama_model's destructor
 }
 
 uint64_t llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter) {

src/llama-adapter.h

@@ -77,6 +77,10 @@ struct llama_adapter_lora {
     ~llama_adapter_lora() = default;
 
     llama_adapter_lora_weight * get_weight(ggml_tensor * w);
+
+    uint32_t get_n_nodes() const {
+        return ab_map.size() * 6u; // a, b, scale, add, 2 x mul_mat
+    }
 };
 
 using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;

src/llama-context.cpp

@@ -1442,7 +1442,11 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
     if (model.arch == LLM_ARCH_QWEN3NEXT) {
         return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
     }
-    return std::max<uint32_t>(1024u, 8u*model.n_tensors());
+    uint32_t res = std::max<uint32_t>(1024u, 8u*model.n_tensors());
+    for (const auto & lora : model.loras) {
+        res += lora->get_n_nodes();
+    }
+    return res;
 }
 
 llm_graph_result * llama_context::get_gf_res_reserve() const {

src/llama-model.cpp

@@ -467,7 +467,11 @@ llama_model::llama_model(const llama_model_params & params) : params(params), pi
     pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern;
 }
 
-llama_model::~llama_model() = default;
+llama_model::~llama_model() {
+    for (auto * lora : loras) {
+        delete lora;
+    }
+}
 
 void llama_model::load_stats(llama_model_loader & ml) {
     pimpl->n_elements = ml.n_elements;

src/llama-model.h

@@ -11,6 +11,7 @@
 #include <memory>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 
 struct llama_cparams;
@@ -475,6 +476,9 @@ struct llama_model {
     // for quantize-stats only
     std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
 
+    // for keeping track of associated LoRA adapters
+    std::unordered_set<llama_adapter_lora *> loras;
+
     int64_t t_load_us  = 0;
     int64_t t_start_us = 0;
 

src/llama-sampling.cpp

@@ -421,39 +421,6 @@ void llama_sampler_free(struct llama_sampler * smpl) {
     delete smpl;
 }
 
-llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_context * ctx, int32_t idx) {
-    const auto * logits = llama_get_logits_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);
-
-    // TODO: do not allocate each time
-    std::vector<llama_token_data> cur;
-    cur.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-    }
-
-    llama_token_data_array cur_p = {
-        /* .data       = */ cur.data(),
-        /* .size       = */ cur.size(),
-        /* .selected   = */ -1,
-        /* .sorted     = */ false,
-    };
-
-    llama_sampler_apply(smpl, &cur_p);
-
-    GGML_ASSERT(cur_p.selected >= 0 && cur_p.selected < (int32_t) cur_p.size);
-
-    auto token = cur_p.data[cur_p.selected].id;
-
-    llama_sampler_accept(smpl, token);
-
-    return token;
-}
-
 // sampler chain
 
 static const char * llama_sampler_chain_name(const struct llama_sampler * /*smpl*/) {
@@ -527,12 +494,56 @@ struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_param
         /* .ctx   = */ new llama_sampler_chain {
             /* .params      = */ params,
             /* .samplers    = */ {},
+            /* .cur         = */ {},
             /* .t_sample_us = */ 0,
             /* .n_sample    = */ 0,
         }
     );
 }
 
+llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_context * ctx, int32_t idx) {
+    const auto * logits = llama_get_logits_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);
+
+    // use pre-allocated buffer from chain if available, otherwise allocate locally
+    std::vector<llama_token_data> * cur_ptr;
+    std::vector<llama_token_data> cur_local;
+
+    if (smpl->iface == &llama_sampler_chain_i) {
+        auto * chain = (llama_sampler_chain *) smpl->ctx;
+        cur_ptr = &chain->cur;
+    } else {
+        cur_ptr = &cur_local;
+    }
+
+    auto & cur = *cur_ptr;
+    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};
+    }
+
+    llama_token_data_array cur_p = {
+        /* .data       = */ cur.data(),
+        /* .size       = */ cur.size(),
+        /* .selected   = */ -1,
+        /* .sorted     = */ false,
+    };
+
+    llama_sampler_apply(smpl, &cur_p);
+
+    GGML_ASSERT(cur_p.selected >= 0 && cur_p.selected < (int32_t) cur_p.size);
+
+    auto token = cur_p.data[cur_p.selected].id;
+
+    llama_sampler_accept(smpl, token);
+
+    return token;
+}
+
 void llama_sampler_chain_add(struct llama_sampler * chain, struct llama_sampler * smpl) {
     auto * p = (llama_sampler_chain *) chain->ctx;
     p->samplers.push_back(smpl);

src/llama-sampling.h

@@ -16,6 +16,9 @@ struct llama_sampler_chain {
 
     std::vector<struct llama_sampler *> samplers;
 
+    // pre-allocated buffer for llama_sampler_sample to avoid repeated allocations
+    std::vector<llama_token_data> cur;
+
     // timing
 
     mutable int64_t t_sample_us;