npm - @fugood/llama.node - Versions diffs - 0.3.13 → 0.3.15 - Mend

@fugood/llama.node 0.3.13 → 0.3.15

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (184) hide show

package/bin/darwin/arm64/llama-node.node +0 -0
package/bin/darwin/x64/llama-node.node +0 -0
package/bin/linux/arm64/llama-node.node +0 -0
package/bin/linux/x64/llama-node.node +0 -0
package/bin/linux-cuda/arm64/llama-node.node +0 -0
package/bin/linux-cuda/x64/llama-node.node +0 -0
package/bin/linux-vulkan/arm64/llama-node.node +0 -0
package/bin/linux-vulkan/x64/llama-node.node +0 -0
package/bin/win32/arm64/llama-node.node +0 -0
package/bin/win32/arm64/node.lib +0 -0
package/bin/win32/x64/llama-node.node +0 -0
package/bin/win32/x64/node.lib +0 -0
package/bin/win32-vulkan/arm64/llama-node.node +0 -0
package/bin/win32-vulkan/arm64/node.lib +0 -0
package/bin/win32-vulkan/x64/llama-node.node +0 -0
package/bin/win32-vulkan/x64/node.lib +0 -0
package/lib/binding.ts +1 -1
package/package.json +1 -1
package/src/LlamaContext.cpp +98 -76
package/src/LlamaContext.h +1 -1
package/src/common.hpp +1 -2
package/src/llama.cpp/.github/workflows/build.yml +89 -10
package/src/llama.cpp/.github/workflows/server.yml +2 -0
package/src/llama.cpp/CMakeLists.txt +9 -1
package/src/llama.cpp/cmake/common.cmake +2 -0
package/src/llama.cpp/common/CMakeLists.txt +3 -3
package/src/llama.cpp/common/arg.cpp +132 -13
package/src/llama.cpp/common/chat.cpp +960 -266
package/src/llama.cpp/common/chat.h +135 -0
package/src/llama.cpp/common/common.cpp +33 -174
package/src/llama.cpp/common/common.h +27 -67
package/src/llama.cpp/common/json-schema-to-grammar.cpp +4 -5
package/src/llama.cpp/common/json-schema-to-grammar.h +0 -1
package/src/llama.cpp/common/{minja.hpp → minja/minja.hpp} +37 -5
package/src/llama.cpp/common/ngram-cache.cpp +1 -0
package/src/llama.cpp/common/sampling.cpp +45 -7
package/src/llama.cpp/common/speculative.cpp +10 -9
package/src/llama.cpp/common/speculative.h +1 -1
package/src/llama.cpp/docs/build.md +45 -7
package/src/llama.cpp/examples/batched-bench/batched-bench.cpp +2 -2
package/src/llama.cpp/examples/cvector-generator/cvector-generator.cpp +4 -2
package/src/llama.cpp/examples/embedding/embedding.cpp +2 -1
package/src/llama.cpp/examples/export-lora/export-lora.cpp +4 -2
package/src/llama.cpp/examples/gritlm/gritlm.cpp +2 -2
package/src/llama.cpp/examples/imatrix/imatrix.cpp +3 -4
package/src/llama.cpp/examples/infill/infill.cpp +2 -2
package/src/llama.cpp/examples/llama-bench/llama-bench.cpp +2 -2
package/src/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp +5 -5
package/src/llama.cpp/examples/llava/CMakeLists.txt +7 -0
package/src/llama.cpp/examples/llava/clip.cpp +373 -107
package/src/llama.cpp/examples/llava/clip.h +19 -3
package/src/llama.cpp/examples/llava/gemma3-cli.cpp +341 -0
package/src/llama.cpp/examples/llava/llava.cpp +4 -2
package/src/llama.cpp/examples/llava/minicpmv-cli.cpp +30 -11
package/src/llama.cpp/examples/lookahead/lookahead.cpp +7 -6
package/src/llama.cpp/examples/lookup/lookup.cpp +1 -1
package/src/llama.cpp/examples/main/main.cpp +79 -34
package/src/llama.cpp/examples/parallel/parallel.cpp +6 -5
package/src/llama.cpp/examples/passkey/passkey.cpp +15 -14
package/src/llama.cpp/examples/perplexity/perplexity.cpp +6 -6
package/src/llama.cpp/examples/quantize/quantize.cpp +1 -0
package/src/llama.cpp/examples/quantize-stats/quantize-stats.cpp +2 -2
package/src/llama.cpp/examples/retrieval/retrieval.cpp +1 -1
package/src/llama.cpp/examples/run/linenoise.cpp/linenoise.cpp +882 -237
package/src/llama.cpp/examples/run/linenoise.cpp/linenoise.h +35 -26
package/src/llama.cpp/examples/run/run.cpp +196 -108
package/src/llama.cpp/examples/save-load-state/save-load-state.cpp +2 -2
package/src/llama.cpp/examples/server/server.cpp +113 -101
package/src/llama.cpp/examples/server/utils.hpp +94 -105
package/src/llama.cpp/examples/simple-chat/simple-chat.cpp +2 -2
package/src/llama.cpp/examples/speculative/speculative.cpp +14 -14
package/src/llama.cpp/examples/speculative-simple/speculative-simple.cpp +1 -1
package/src/llama.cpp/examples/sycl/run-llama2.sh +2 -2
package/src/llama.cpp/examples/tts/tts.cpp +263 -151
package/src/llama.cpp/ggml/CMakeLists.txt +14 -1
package/src/llama.cpp/ggml/cmake/common.cmake +26 -0
package/src/llama.cpp/ggml/include/ggml-alloc.h +1 -1
package/src/llama.cpp/ggml/include/ggml-backend.h +3 -3
package/src/llama.cpp/ggml/include/ggml-cpu.h +3 -0
package/src/llama.cpp/ggml/include/ggml.h +29 -1
package/src/llama.cpp/ggml/src/CMakeLists.txt +15 -34
package/src/llama.cpp/ggml/src/ggml-alloc.c +24 -15
package/src/llama.cpp/ggml/src/ggml-backend-impl.h +1 -1
package/src/llama.cpp/ggml/src/ggml-backend-reg.cpp +58 -54
package/src/llama.cpp/ggml/src/ggml-backend.cpp +10 -8
package/src/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp +6 -2
package/src/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp +3 -7
package/src/llama.cpp/ggml/src/ggml-cann/kernels/dup.cpp +3 -5
package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +139 -16
package/src/llama.cpp/ggml/src/ggml-cpu/amx/amx.cpp +2 -1
package/src/llama.cpp/ggml/src/ggml-cpu/cpu-feats-x86.cpp +4 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +2 -1
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +151 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c +1546 -387
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +1645 -113
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +22 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kernels.cpp +259 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kernels.h +61 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +288 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
package/src/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +15 -2
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h +2 -1
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/musa.h +3 -1
package/src/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +14 -0
package/src/llama.cpp/ggml/src/ggml-impl.h +1 -1
package/src/llama.cpp/ggml/src/ggml-metal/CMakeLists.txt +4 -5
package/src/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +242 -0
package/src/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt +6 -6
package/src/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt +1 -0
package/src/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +315 -138
package/src/llama.cpp/ggml/src/ggml-quants.c +114 -114
package/src/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +2 -1
package/src/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt +5 -0
package/src/llama.cpp/ggml/src/ggml-sycl/backend.hpp +2 -1
package/src/llama.cpp/ggml/src/ggml-sycl/common.cpp +17 -0
package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp +117 -36
package/src/llama.cpp/ggml/src/ggml-sycl/convert.cpp +33 -4
package/src/llama.cpp/ggml/src/ggml-sycl/convert.hpp +2 -2
package/src/llama.cpp/ggml/src/ggml-sycl/cpy.cpp +701 -0
package/src/llama.cpp/ggml/src/ggml-sycl/cpy.hpp +11 -0
package/src/llama.cpp/ggml/src/ggml-sycl/dequantize.hpp +55 -0
package/src/llama.cpp/ggml/src/ggml-sycl/dmmv.cpp +147 -16
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.cpp +40 -40
package/src/llama.cpp/ggml/src/ggml-sycl/getrows.cpp +307 -0
package/src/llama.cpp/ggml/src/ggml-sycl/getrows.hpp +23 -0
package/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp +262 -746
package/src/llama.cpp/ggml/src/ggml-sycl/mmq.cpp +0 -1
package/src/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp +75 -78
package/src/llama.cpp/ggml/src/ggml-sycl/norm.cpp +114 -6
package/src/llama.cpp/ggml/src/ggml-sycl/norm.hpp +6 -0
package/src/llama.cpp/ggml/src/ggml-sycl/softmax.cpp +4 -1
package/src/llama.cpp/ggml/src/ggml-sycl/sycl_hw.cpp +13 -0
package/src/llama.cpp/ggml/src/ggml-sycl/sycl_hw.hpp +23 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv.cpp +305 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv.hpp +10 -0
package/src/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp +498 -188
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +0 -4
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +16 -3
package/src/llama.cpp/ggml/src/ggml.c +93 -5
package/src/llama.cpp/include/llama.h +105 -27
package/src/llama.cpp/models/ggml-vocab-gpt-4o.gguf.inp +112 -0
package/src/llama.cpp/models/ggml-vocab-gpt-4o.gguf.out +46 -0
package/src/llama.cpp/requirements/requirements-all.txt +1 -0
package/src/llama.cpp/requirements/requirements-tool_bench.txt +12 -0
package/src/llama.cpp/requirements.txt +1 -0
package/src/llama.cpp/src/CMakeLists.txt +5 -2
package/src/llama.cpp/src/llama-adapter.cpp +19 -20
package/src/llama.cpp/src/llama-adapter.h +11 -9
package/src/llama.cpp/src/llama-arch.cpp +123 -16
package/src/llama.cpp/src/llama-arch.h +19 -0
package/src/llama.cpp/src/llama-batch.h +2 -2
package/src/llama.cpp/src/llama-chat.cpp +1 -0
package/src/llama.cpp/src/llama-context.cpp +2253 -1222
package/src/llama.cpp/src/llama-context.h +214 -77
package/src/llama.cpp/src/llama-cparams.h +1 -0
package/src/llama.cpp/src/llama-grammar.cpp +182 -182
package/src/llama.cpp/src/llama-grammar.h +12 -3
package/src/llama.cpp/src/llama-graph.cpp +1662 -0
package/src/llama.cpp/src/llama-graph.h +574 -0
package/src/llama.cpp/src/llama-hparams.cpp +8 -0
package/src/llama.cpp/src/llama-hparams.h +9 -0
package/src/llama.cpp/src/llama-io.cpp +15 -0
package/src/llama.cpp/src/llama-io.h +35 -0
package/src/llama.cpp/src/llama-kv-cache.cpp +1006 -291
package/src/llama.cpp/src/llama-kv-cache.h +178 -109
package/src/llama.cpp/src/llama-memory.cpp +1 -0
package/src/llama.cpp/src/llama-memory.h +21 -0
package/src/llama.cpp/src/llama-mmap.cpp +11 -1
package/src/llama.cpp/src/llama-model.cpp +8230 -122
package/src/llama.cpp/src/llama-model.h +34 -1
package/src/llama.cpp/src/llama-quant.cpp +10 -1
package/src/llama.cpp/src/llama-sampling.cpp +43 -10
package/src/llama.cpp/src/llama-vocab.cpp +12 -0
package/src/llama.cpp/src/llama.cpp +51 -9837
package/src/llama.cpp/tests/test-backend-ops.cpp +247 -112
package/src/llama.cpp/tests/test-chat-template.cpp +32 -22
package/src/llama.cpp/tests/test-chat.cpp +593 -395
package/src/llama.cpp/tests/test-json-schema-to-grammar.cpp +63 -63
package/src/llama.cpp/tests/test-quantize-fns.cpp +1 -9
package/src/llama.cpp/Sources/llama/llama.h +0 -4
package/src/llama.cpp/common/chat.hpp +0 -55
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp +0 -143
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.hpp +0 -9
/package/src/llama.cpp/common/{chat-template.hpp → minja/chat-template.hpp} +0 -0

package/src/llama.cpp/requirements.txt CHANGED Viewed

@@ -10,3 +10,4 @@
 -r ./requirements/requirements-convert_hf_to_gguf_update.txt
 -r ./requirements/requirements-convert_llama_ggml_to_gguf.txt
 -r ./requirements/requirements-convert_lora_to_gguf.txt
+-r ./requirements/requirements-tool_bench.txt

package/src/llama.cpp/src/CMakeLists.txt CHANGED Viewed

@@ -15,18 +15,21 @@ add_library(llama
             llama-chat.cpp
             llama-context.cpp
             llama-grammar.cpp
+            llama-graph.cpp
             llama-hparams.cpp
             llama-impl.cpp
+            llama-io.cpp
             llama-kv-cache.cpp
+            llama-memory.cpp
             llama-mmap.cpp
             llama-model-loader.cpp
             llama-model.cpp
             llama-quant.cpp
             llama-sampling.cpp
             llama-vocab.cpp
-            unicode.h
-            unicode.cpp
             unicode-data.cpp
+            unicode.cpp
+            unicode.h
             )
 target_include_directories(llama PUBLIC . ../include ../common)

package/src/llama.cpp/src/llama-adapter.cpp CHANGED Viewed

@@ -4,14 +4,13 @@
 #include "llama-mmap.h"
 #include "llama-model.h"
-#include <algorithm>
 #include <map>
 #include <cassert>
 #include <stdexcept>
 // vec
-struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
+ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
     if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) {
         return nullptr;
     }
@@ -19,7 +18,7 @@ struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
     return tensors[il];
 }
-struct ggml_tensor * llama_adapter_cvec::apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const {
+ggml_tensor * llama_adapter_cvec::apply_to(ggml_context * ctx, ggml_tensor * cur, int  il) const {
     ggml_tensor * layer_dir = tensor_for(il);
     if (layer_dir != nullptr) {
         cur = ggml_add(ctx, cur, layer_dir);
@@ -40,7 +39,7 @@ bool llama_adapter_cvec::init(const llama_model & model) {
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
-            struct ggml_init_params params = {
+            ggml_init_params params = {
                 /*.mem_size   =*/ hparams.n_layer*ggml_tensor_overhead(),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
@@ -91,7 +90,7 @@ bool llama_adapter_cvec::init(const llama_model & model) {
     return true;
 }
-int32_t llama_adapter_cvec::apply(
+bool llama_adapter_cvec::apply(
         const llama_model & model,
         const float * data,
         size_t len,
@@ -104,17 +103,17 @@ int32_t llama_adapter_cvec::apply(
         // disable the current control vector (but leave allocated for later)
         layer_start = -1;
         layer_end   = -1;
-        return 0;
+        return true;
     }
     if (n_embd != (int) hparams.n_embd) {
         LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__);
-        return 1;
+        return false;
     }
     if (tensors.empty()) {
         if (!init(model)) {
-            return 1;
+            return false;
         }
     }
@@ -130,12 +129,12 @@ int32_t llama_adapter_cvec::apply(
         }
     }
-    return 0;
+    return true;
 }
 // lora
-llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor * w) {
+llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) {
     const std::string name(w->name);
     const auto pos = ab_map.find(name);
@@ -146,11 +145,11 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor *
     return nullptr;
 }
-static void llama_adapter_lora_init_impl(struct llama_model & model, const char * path_lora, struct llama_adapter_lora & adapter) {
+static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) {
     LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
     ggml_context * ctx_init;
-    struct gguf_init_params meta_gguf_params = {
+    gguf_init_params meta_gguf_params = {
         /* .no_alloc = */ true,
         /* .ctx      = */ &ctx_init,
     };
@@ -201,7 +200,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
             // add a new context
-            struct ggml_init_params params = {
+            ggml_init_params params = {
                 /*.mem_size   =*/ n_tensors*ggml_tensor_overhead(),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
@@ -264,7 +263,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
             throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)");
         }
-        struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
+        ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
         // validate tensor shape
         if (is_token_embd) {
             // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd()
@@ -281,8 +280,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
         }
         // save tensor to adapter
-        struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
-        struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
+        ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
+        ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
         ggml_set_name(tensor_a, w.a->name);
         ggml_set_name(tensor_b, w.b->name);
         adapter.ab_map[name] = llama_adapter_lora_weight(tensor_a, tensor_b);
@@ -308,7 +307,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
     {
         llama_file gguf_file(path_lora, "rb");
         std::vector<uint8_t> read_buf;
-        auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) {
+        auto set_tensor = [&](ggml_tensor * orig, ggml_tensor * dev) {
             size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name));
             size_t size = ggml_nbytes(orig);
             read_buf.resize(size);
@@ -327,8 +326,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
-struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model, const char * path_lora) {
-    struct llama_adapter_lora * adapter = new llama_adapter_lora();
+llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) {
+    llama_adapter_lora * adapter = new llama_adapter_lora();
     try {
         llama_adapter_lora_init_impl(*model, path_lora, *adapter);
@@ -342,6 +341,6 @@ struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model,
     return nullptr;
 }
-void llama_adapter_lora_free(struct llama_adapter_lora * adapter) {
+void llama_adapter_lora_free(llama_adapter_lora * adapter) {
     delete adapter;
 }

package/src/llama.cpp/src/llama-adapter.h CHANGED Viewed

@@ -15,11 +15,11 @@
 //
 struct llama_adapter_cvec {
-    struct ggml_tensor * tensor_for(int il) const;
+    ggml_tensor * tensor_for(int il) const;
-    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const;
+    ggml_tensor * apply_to(ggml_context * ctx, ggml_tensor * cur, int  il) const;
-    int32_t apply(
+    bool apply(
             const llama_model & model,
             const float * data,
             size_t len,
@@ -36,7 +36,7 @@ private:
     std::vector<ggml_context_ptr> ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
-    std::vector<struct ggml_tensor *> tensors; // per layer
+    std::vector<ggml_tensor *> tensors; // per layer
 };
 //
@@ -44,8 +44,8 @@ private:
 //
 struct llama_adapter_lora_weight {
-    struct ggml_tensor * a = nullptr;
-    struct ggml_tensor * b = nullptr;
+    ggml_tensor * a = nullptr;
+    ggml_tensor * b = nullptr;
     // get actual scale based on rank and alpha
     float get_scale(float alpha, float adapter_scale) const {
@@ -55,12 +55,12 @@ struct llama_adapter_lora_weight {
     }
     llama_adapter_lora_weight() = default;
-    llama_adapter_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {}
+    llama_adapter_lora_weight(ggml_tensor * a, ggml_tensor * b) : a(a), b(b) {}
 };
 struct llama_adapter_lora {
     // map tensor name to lora_a_b
-    std::unordered_map<std::string, struct llama_adapter_lora_weight> ab_map;
+    std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
     std::vector<ggml_context_ptr> ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
@@ -70,5 +70,7 @@ struct llama_adapter_lora {
     llama_adapter_lora() = default;
     ~llama_adapter_lora() = default;
-    llama_adapter_lora_weight * get_weight(struct ggml_tensor * w);
+    llama_adapter_lora_weight * get_weight(ggml_tensor * w);
 };
+using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;

package/src/llama.cpp/src/llama-arch.cpp CHANGED Viewed

@@ -36,6 +36,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_MINICPM3,         "minicpm3"         },
     { LLM_ARCH_GEMMA,            "gemma"            },
     { LLM_ARCH_GEMMA2,           "gemma2"           },
+    { LLM_ARCH_GEMMA3,           "gemma3"           },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
     { LLM_ARCH_XVERSE,           "xverse"           },
@@ -58,6 +59,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
     { LLM_ARCH_RWKV6QWEN2,       "rwkv6qwen2"       },
+    { LLM_ARCH_RWKV7,            "rwkv7"            },
+    { LLM_ARCH_ARWKV7,           "arwkv7"           },
     { LLM_ARCH_GRANITE,          "granite"          },
     { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
     { LLM_ARCH_CHAMELEON,        "chameleon"        },
@@ -109,22 +112,26 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EMBEDDING_SCALE,                   "%s.embedding_scale"                   },
     { LLM_KV_TOKEN_SHIFT_COUNT,                 "%s.token_shift_count"                 },
-    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
-    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
-    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"         },
-    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"              },
-    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"             },
-    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"           },
-    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"     },
-    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon" },
-    { LLM_KV_ATTENTION_GROUPNORM_EPS,          "%s.attention.group_norm_epsilon"     },
-    { LLM_KV_ATTENTION_GROUPNORM_GROUPS,       "%s.attention.group_norm_groups"      },
-    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                 },
-    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"            },
-    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
-    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
-    { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"         },
-    { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                  },
+    { LLM_KV_ATTENTION_HEAD_COUNT,                   "%s.attention.head_count"                   },
+    { LLM_KV_ATTENTION_HEAD_COUNT_KV,                "%s.attention.head_count_kv"                },
+    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,               "%s.attention.max_alibi_bias"               },
+    { LLM_KV_ATTENTION_CLAMP_KQV,                    "%s.attention.clamp_kqv"                    },
+    { LLM_KV_ATTENTION_KEY_LENGTH,                   "%s.attention.key_length"                   },
+    { LLM_KV_ATTENTION_VALUE_LENGTH,                 "%s.attention.value_length"                 },
+    { LLM_KV_ATTENTION_LAYERNORM_EPS,                "%s.attention.layer_norm_epsilon"           },
+    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,            "%s.attention.layer_norm_rms_epsilon"       },
+    { LLM_KV_ATTENTION_GROUPNORM_EPS,                "%s.attention.group_norm_epsilon"           },
+    { LLM_KV_ATTENTION_GROUPNORM_GROUPS,             "%s.attention.group_norm_groups"            },
+    { LLM_KV_ATTENTION_CAUSAL,                       "%s.attention.causal"                       },
+    { LLM_KV_ATTENTION_Q_LORA_RANK,                  "%s.attention.q_lora_rank"                  },
+    { LLM_KV_ATTENTION_KV_LORA_RANK,                 "%s.attention.kv_lora_rank"                 },
+    { LLM_KV_ATTENTION_DECAY_LORA_RANK,              "%s.attention.decay_lora_rank"              },
+    { LLM_KV_ATTENTION_ICLR_LORA_RANK,               "%s.attention.iclr_lora_rank"               },
+    { LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, "%s.attention.value_residual_mix_lora_rank" },
+    { LLM_KV_ATTENTION_GATE_LORA_RANK,               "%s.attention.gate_lora_rank"               },
+    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
+    { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
+    { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -766,6 +773,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
         },
     },
+    {
+        LLM_ARCH_GEMMA3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_STARCODER2,
         {
@@ -1217,6 +1244,74 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,                    "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_RWKV7,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,                "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,           "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,               "output_norm" },
+            { LLM_TENSOR_OUTPUT,                    "output" },
+            { LLM_TENSOR_ATTN_NORM,                 "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_NORM_2,               "blk.%d.attn_norm_2" },
+            { LLM_TENSOR_TIME_MIX_W0,               "blk.%d.time_mix_w0" },
+            { LLM_TENSOR_TIME_MIX_W1,               "blk.%d.time_mix_w1" },
+            { LLM_TENSOR_TIME_MIX_W2,               "blk.%d.time_mix_w2" },
+            { LLM_TENSOR_TIME_MIX_A0,               "blk.%d.time_mix_a0" },
+            { LLM_TENSOR_TIME_MIX_A1,               "blk.%d.time_mix_a1" },
+            { LLM_TENSOR_TIME_MIX_A2,               "blk.%d.time_mix_a2" },
+            { LLM_TENSOR_TIME_MIX_V0,               "blk.%d.time_mix_v0" },
+            { LLM_TENSOR_TIME_MIX_V1,               "blk.%d.time_mix_v1" },
+            { LLM_TENSOR_TIME_MIX_V2,               "blk.%d.time_mix_v2" },
+            { LLM_TENSOR_TIME_MIX_G1,               "blk.%d.time_mix_g1" },
+            { LLM_TENSOR_TIME_MIX_G2,               "blk.%d.time_mix_g2" },
+            { LLM_TENSOR_TIME_MIX_K_K,              "blk.%d.time_mix_k_k" },
+            { LLM_TENSOR_TIME_MIX_K_A,              "blk.%d.time_mix_k_a" },
+            { LLM_TENSOR_TIME_MIX_R_K,              "blk.%d.time_mix_r_k" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
+            { LLM_TENSOR_TIME_MIX_KEY,              "blk.%d.time_mix_key" },
+            { LLM_TENSOR_TIME_MIX_VALUE,            "blk.%d.time_mix_value" },
+            { LLM_TENSOR_TIME_MIX_RECEPTANCE,       "blk.%d.time_mix_receptance" },
+            { LLM_TENSOR_TIME_MIX_LN,               "blk.%d.time_mix_ln" },
+            { LLM_TENSOR_TIME_MIX_OUTPUT,           "blk.%d.time_mix_output" },
+            { LLM_TENSOR_CHANNEL_MIX_LERP_K,        "blk.%d.channel_mix_lerp_k" },
+            { LLM_TENSOR_CHANNEL_MIX_KEY,           "blk.%d.channel_mix_key" },
+            { LLM_TENSOR_CHANNEL_MIX_VALUE,         "blk.%d.channel_mix_value" },
+        },
+    },
+    {
+        LLM_ARCH_ARWKV7,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,                "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,           "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,               "output_norm" },
+            { LLM_TENSOR_OUTPUT,                    "output" },
+            { LLM_TENSOR_ATTN_NORM,                 "blk.%d.attn_norm" },
+            { LLM_TENSOR_TIME_MIX_W0,               "blk.%d.time_mix_w0" },
+            { LLM_TENSOR_TIME_MIX_W1,               "blk.%d.time_mix_w1" },
+            { LLM_TENSOR_TIME_MIX_W2,               "blk.%d.time_mix_w2" },
+            { LLM_TENSOR_TIME_MIX_A0,               "blk.%d.time_mix_a0" },
+            { LLM_TENSOR_TIME_MIX_A1,               "blk.%d.time_mix_a1" },
+            { LLM_TENSOR_TIME_MIX_A2,               "blk.%d.time_mix_a2" },
+            { LLM_TENSOR_TIME_MIX_V0,               "blk.%d.time_mix_v0" },
+            { LLM_TENSOR_TIME_MIX_V1,               "blk.%d.time_mix_v1" },
+            { LLM_TENSOR_TIME_MIX_V2,               "blk.%d.time_mix_v2" },
+            { LLM_TENSOR_TIME_MIX_G1,               "blk.%d.time_mix_g1" },
+            { LLM_TENSOR_TIME_MIX_G2,               "blk.%d.time_mix_g2" },
+            { LLM_TENSOR_TIME_MIX_K_K,              "blk.%d.time_mix_k_k" },
+            { LLM_TENSOR_TIME_MIX_K_A,              "blk.%d.time_mix_k_a" },
+            { LLM_TENSOR_TIME_MIX_R_K,              "blk.%d.time_mix_r_k" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
+            { LLM_TENSOR_TIME_MIX_KEY,              "blk.%d.time_mix_key" },
+            { LLM_TENSOR_TIME_MIX_VALUE,            "blk.%d.time_mix_value" },
+            { LLM_TENSOR_TIME_MIX_RECEPTANCE,       "blk.%d.time_mix_receptance" },
+            { LLM_TENSOR_TIME_MIX_LN,               "blk.%d.time_mix_ln" },
+            { LLM_TENSOR_TIME_MIX_OUTPUT,           "blk.%d.time_mix_output" },
+            { LLM_TENSOR_FFN_NORM,                  "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,                  "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,                  "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,                    "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_GRANITE,
         {
@@ -1376,6 +1471,12 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_OUT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_A1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_A2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_V1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_V2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_G1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_G2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_DECAY_W1,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_DECAY_W2,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_KEY,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1394,6 +1495,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TIME_MIX_LN,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CHANNEL_MIX_LERP_K,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CHANNEL_MIX_LERP_R,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_K_K,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_K_A,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_R_K,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LERP_W,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_K,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_V,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
@@ -1401,6 +1505,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TIME_MIX_LERP_G,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_FUSED,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_DECAY,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_W0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_A0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_V0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_FIRST,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}},
     {LLM_TENSOR_ATTN_NORM,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_ATTN_NORM_2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},

package/src/llama.cpp/src/llama-arch.h CHANGED Viewed

@@ -40,6 +40,7 @@ enum llm_arch {
     LLM_ARCH_MINICPM3,
     LLM_ARCH_GEMMA,
     LLM_ARCH_GEMMA2,
+    LLM_ARCH_GEMMA3,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
     LLM_ARCH_XVERSE,
@@ -62,6 +63,8 @@ enum llm_arch {
     LLM_ARCH_EXAONE,
     LLM_ARCH_RWKV6,
     LLM_ARCH_RWKV6QWEN2,
+    LLM_ARCH_RWKV7,
+    LLM_ARCH_ARWKV7,
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
     LLM_ARCH_CHAMELEON,
@@ -126,6 +129,10 @@ enum llm_kv {
     LLM_KV_ATTENTION_CAUSAL,
     LLM_KV_ATTENTION_Q_LORA_RANK,
     LLM_KV_ATTENTION_KV_LORA_RANK,
+    LLM_KV_ATTENTION_DECAY_LORA_RANK,
+    LLM_KV_ATTENTION_ICLR_LORA_RANK,
+    LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK,
+    LLM_KV_ATTENTION_GATE_LORA_RANK,
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
@@ -249,8 +256,20 @@ enum llm_tensor {
     LLM_TENSOR_SSM_A,
     LLM_TENSOR_SSM_D,
     LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
     LLM_TENSOR_TIME_MIX_W2,
+    LLM_TENSOR_TIME_MIX_A0,
+    LLM_TENSOR_TIME_MIX_A1,
+    LLM_TENSOR_TIME_MIX_A2,
+    LLM_TENSOR_TIME_MIX_V0,
+    LLM_TENSOR_TIME_MIX_V1,
+    LLM_TENSOR_TIME_MIX_V2,
+    LLM_TENSOR_TIME_MIX_G1,
+    LLM_TENSOR_TIME_MIX_G2,
+    LLM_TENSOR_TIME_MIX_K_K,
+    LLM_TENSOR_TIME_MIX_K_A,
+    LLM_TENSOR_TIME_MIX_R_K,
     LLM_TENSOR_TIME_MIX_LERP_X,
     LLM_TENSOR_TIME_MIX_LERP_W,
     LLM_TENSOR_TIME_MIX_LERP_K,

package/src/llama.cpp/src/llama-batch.h CHANGED Viewed

@@ -42,9 +42,9 @@ struct llama_sbatch {
     bool logits_all; // TODO: remove once lctx.logits_all is removed too
     // sorted indices into the batch
-    std::vector<size_t> ids;
+    std::vector<int64_t> ids;
     // batch indices of the output
-    std::vector<size_t> out_ids;
+    std::vector<int64_t> out_ids;
     std::vector<llama_sbatch_seq> seq;
     const llama_batch * batch = nullptr;

package/src/llama.cpp/src/llama-chat.cpp CHANGED Viewed

@@ -4,6 +4,7 @@
 #include <map>
 #include <sstream>
+#include <algorithm>
 #if __cplusplus >= 202000L
     #define LU8(x) (const char*)(u8##x)