Merge branch 'master' into auto-model-support

.github/labeler.yml

@@ -1,5 +1,16 @@
 # https://github.com/actions/labeler
-
+Kompute:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-kompute.h
+            - ggml-kompute.cpp
+            - README-kompute.md
+Apple Metal:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-metal.h
+            - ggml-metal.cpp
+            - README-metal.md
 SYCL:
     - changed-files:
         - any-glob-to-any-file:
@@ -9,6 +20,7 @@ SYCL:
 Nvidia GPU:
     - changed-files:
         - any-glob-to-any-file:
+            - ggml-cuda.h
             - ggml-cuda/**
 Vulkan:
     - changed-files:

.github/workflows/docker.yml

@@ -42,8 +42,9 @@ jobs:
           - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
+          # TODO: Disabled due to build issues https://github.com/ggerganov/llama.cpp/issues/7507
-          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
+          #- { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
+          #- { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4

CMakeLists.txt

@@ -72,6 +72,7 @@ else()
     set(INS_ENB ON)
 endif()
 
+option(LLAMA_SVE                             "llama: enable SVE"                                OFF)
 option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
 option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
 option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
@@ -1040,6 +1041,9 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
             list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()
+        if (LLAMA_SVE)
+            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+        endif()
     endif()
 elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
         (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND

README.md

@@ -203,6 +203,10 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
+**Tools:**
+
+- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+
 ---
 
 Here is a typical run using LLaMA v2 13B on M2 Ultra:

common/common.cpp

@@ -904,6 +904,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.interactive_specials = true;
         return true;
     }
+    if (arg == "--special") {
+        params.special = true;
+        return true;
+    }
     if (arg == "--embedding") {
         params.embedding = true;
         return true;
@@ -1362,6 +1366,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     printf("  -h, --help            show this help message and exit\n");
     printf("  --version             show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
+    printf("  --special             special tokens output enabled\n");
     printf("  --interactive-specials allow special tokens in user text, in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -cnv, --conversation  run in conversation mode (does not print special tokens and suffix/prefix)\n");
@@ -1855,11 +1860,15 @@ bool fs_create_directory_with_parents(const std::string & path) {
 
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
+    auto ensure_trailing_slash = [](std::string p) {
+        // Make sure to add trailing slash
+        if (p.back() != DIRECTORY_SEPARATOR) {
+            p += DIRECTORY_SEPARATOR;
+        }
+        return p;
+    };
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
-        if (cache_directory.back() != DIRECTORY_SEPARATOR) {
-            cache_directory += DIRECTORY_SEPARATOR;
-        }
     } else {
 #ifdef __linux__
         if (std::getenv("XDG_CACHE_HOME")) {
@@ -1870,12 +1879,12 @@ std::string fs_get_cache_directory() {
 #elif defined(__APPLE__)
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
-        cache_directory = std::getenv("APPDATA");
+        cache_directory = std::getenv("LOCALAPPDATA");
 #endif // __linux__
+        cache_directory = ensure_trailing_slash(cache_directory);
         cache_directory += "llama.cpp";
-        cache_directory += DIRECTORY_SEPARATOR;
     }
-    return cache_directory;
+    return ensure_trailing_slash(cache_directory);
 }
 
 
@@ -2840,6 +2849,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");
     fprintf(stream, "cpu_has_gpublas: %s\n",     ggml_cpu_has_gpublas()     ? "true" : "false");
     fprintf(stream, "cpu_has_neon: %s\n",        ggml_cpu_has_neon()        ? "true" : "false");
+    fprintf(stream, "cpu_has_sve: %s\n",         ggml_cpu_has_sve()         ? "true" : "false");
     fprintf(stream, "cpu_has_f16c: %s\n",        ggml_cpu_has_f16c()        ? "true" : "false");
     fprintf(stream, "cpu_has_fp16_va: %s\n",     ggml_cpu_has_fp16_va()     ? "true" : "false");
     fprintf(stream, "cpu_has_wasm_simd: %s\n",   ggml_cpu_has_wasm_simd()   ? "true" : "false");

common/common.h

@@ -146,6 +146,7 @@ struct gpt_params {
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
     bool interactive_specials = false; // whether to allow special tokens from user, during interactive mode
+    bool special           = false; // enable special token output
     bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
     bool chatml            = false; // chatml mode (used for models trained on chatml syntax)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache

common/train.cpp

@@ -1052,7 +1052,7 @@ struct train_params_common get_default_train_params_common() {
 
     params.custom_n_ctx = false;
 
-    params.use_flash              = true;
+    params.use_flash              = false;
     params.use_checkpointing      = true;
 
     params.sample_start           = "";

convert-hf-to-gguf-update.py

@@ -81,6 +81,7 @@ models = [
     {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
     {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
+    {"name": "smaug-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
 ]
 
 

convert-hf-to-gguf.py

@@ -312,11 +312,10 @@ class Model:
                         data = data.astype(np.float32)
                     data_qtype = gguf.GGMLQuantizationType.F32
 
-                block_size, type_size = gguf.GGML_QUANT_SIZES[data_qtype]
+                shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
+
                 # reverse shape to make it similar to the internal ggml dimension order
-                shape_str = f"""{{{', '.join(str(n) for n in reversed(
+                shape_str = f"{{{', '.join(str(n) for n in reversed(shape))}}}"
-                    (*data.shape[:-1], data.shape[-1] * data.dtype.itemsize // type_size * block_size))
-                )}}}"""
 
                 # n_dims is implicit in the shape
                 logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
@@ -2342,7 +2341,8 @@ class CommandR2Model(Model):
 
         # max_position_embeddings = 8192 in config.json but model was actually
         # trained on 128k context length
-        self.hparams["max_position_embeddings"] = self.hparams["model_max_length"]
+        # aya-23 models don't have model_max_length specified
+        self.hparams["max_position_embeddings"] = self.find_hparam(["model_max_length", "max_position_embeddings"])
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -2415,6 +2415,157 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_eos_token(True)
 
 
+@Model.register("ArcticForCausalLM")
+class ArcticModel(Model):
+    model_arch = gguf.MODEL_ARCH.ARCTIC
+
+    def set_vocab(self):
+        # The reason for using a custom implementation here is that the
+        # snowflake-arctic-instruct model redefined tokens 31998 and 31999 from
+        # tokenizer.model and used them as BOS and EOS instead of adding new tokens.
+        from sentencepiece import SentencePieceProcessor
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        if not tokenizer_path.is_file():
+            logger.error(f'Error: Missing {tokenizer_path}')
+            sys.exit(1)
+
+        # Read the whole vocabulary from the tokenizer.model file
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        # Use the added_tokens_decoder field from tokeniser_config.json as the source
+        # of information about added/redefined tokens and modify them accordingly.
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+
+                if "added_tokens_decoder" in tokenizer_config_json:
+                    added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"]
+                    for token_id, token_json in added_tokens_decoder.items():
+                        token_id = int(token_id)
+                        if (token_id >= vocab_size):
+                            logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                            continue
+
+                        token_content = token_json["content"]
+                        token_type = SentencePieceTokenTypes.USER_DEFINED
+                        token_score = -10000.0
+
+                        # Map unk_token to UNKNOWN, other special tokens to CONTROL
+                        # Set the score to 0.0 as in the original tokenizer.model
+                        if ("special" in token_json) and token_json["special"]:
+                            if token_content == tokenizer_config_json["unk_token"]:
+                                token_type = SentencePieceTokenTypes.UNKNOWN
+                            else:
+                                token_type = SentencePieceTokenTypes.CONTROL
+                            token_score = 0.0
+
+                        logger.info(f"Setting added token {token_id} to '{token_content}' (type: {token_type}, score: {token_score:.2f})")
+                        tokens[token_id] = token_content.encode("utf-8")
+                        toktypes[token_id] = token_type
+                        scores[token_id] = token_score
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+
+        if name.endswith("q_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith("k_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+
+        # process the experts separately
+        if name.find("block_sparse_moe.experts") != -1:
+            n_experts = self.hparams["num_local_experts"]
+
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for wid in ["w1", "w2", "w3"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def write_tensors(self):
+        super().write_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 ###### CONVERSION LOGIC ######
 
 

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -774,7 +774,7 @@ static struct train_params get_default_train_params() {
 
     params.samples_start_after_nl = false;
     params.use_adam               = true;
-    params.use_flash              = true;
+    params.use_flash              = false;
     params.use_scratch            = true;
 
     // only adam

examples/llama.android/app/build.gradle.kts

@@ -7,8 +7,6 @@ android {
     namespace = "com.example.llama"
     compileSdk = 34
 
-    ndkVersion = "26.1.10909125"
-
     defaultConfig {
         applicationId = "com.example.llama"
         minSdk = 33
@@ -20,17 +18,6 @@ android {
         vectorDrawables {
             useSupportLibrary = true
         }
-        ndk {
-            // Add NDK properties if wanted, e.g.
-            // abiFilters += listOf("arm64-v8a")
-        }
-        externalNativeBuild {
-            cmake {
-                arguments += "-DCMAKE_BUILD_TYPE=Release"
-                cppFlags += listOf()
-                arguments += listOf()
-            }
-        }
     }
 
     buildTypes {
@@ -55,17 +42,6 @@ android {
     composeOptions {
         kotlinCompilerExtensionVersion = "1.5.1"
     }
-    packaging {
-        resources {
-            excludes += "/META-INF/{AL2.0,LGPL2.1}"
-        }
-    }
-    externalNativeBuild {
-        cmake {
-            path = file("src/main/cpp/CMakeLists.txt")
-            version = "3.22.1"
-        }
-    }
 }
 
 dependencies {
@@ -78,6 +54,7 @@ dependencies {
     implementation("androidx.compose.ui:ui-graphics")
     implementation("androidx.compose.ui:ui-tooling-preview")
     implementation("androidx.compose.material3:material3")
+    implementation(project(":llama"))
     testImplementation("junit:junit:4.13.2")
     androidTestImplementation("androidx.test.ext:junit:1.1.5")
     androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")

examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt

@@ -1,5 +1,6 @@
 package com.example.llama
 
+import android.llama.cpp.LLamaAndroid
 import android.util.Log
 import androidx.compose.runtime.getValue
 import androidx.compose.runtime.mutableStateOf
@@ -9,7 +10,7 @@ import androidx.lifecycle.viewModelScope
 import kotlinx.coroutines.flow.catch
 import kotlinx.coroutines.launch
 
-class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
+class MainViewModel(private val llamaAndroid: LLamaAndroid = LLamaAndroid.instance()): ViewModel() {
     companion object {
         @JvmStatic
         private val NanosPerSecond = 1_000_000_000.0
@@ -28,7 +29,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
 
         viewModelScope.launch {
             try {
-                llm.unload()
+                llamaAndroid.unload()
             } catch (exc: IllegalStateException) {
                 messages += exc.message!!
             }
@@ -44,7 +45,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
         messages += ""
 
         viewModelScope.launch {
-            llm.send(text)
+            llamaAndroid.send(text)
                 .catch {
                     Log.e(tag, "send() failed", it)
                     messages += it.message!!
@@ -57,7 +58,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
         viewModelScope.launch {
             try {
                 val start = System.nanoTime()
-                val warmupResult = llm.bench(pp, tg, pl, nr)
+                val warmupResult = llamaAndroid.bench(pp, tg, pl, nr)
                 val end = System.nanoTime()
 
                 messages += warmupResult
@@ -70,7 +71,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
                     return@launch
                 }
 
-                messages += llm.bench(512, 128, 1, 3)
+                messages += llamaAndroid.bench(512, 128, 1, 3)
             } catch (exc: IllegalStateException) {
                 Log.e(tag, "bench() failed", exc)
                 messages += exc.message!!
@@ -81,7 +82,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
     fun load(pathToModel: String) {
         viewModelScope.launch {
             try {
-                llm.load(pathToModel)
+                llamaAndroid.load(pathToModel)
                 messages += "Loaded $pathToModel"
             } catch (exc: IllegalStateException) {
                 Log.e(tag, "load() failed", exc)

examples/llama.android/build.gradle.kts

@@ -2,4 +2,5 @@
 plugins {
     id("com.android.application") version "8.2.0" apply false
     id("org.jetbrains.kotlin.android") version "1.9.0" apply false
+    id("com.android.library") version "8.2.0" apply false
 }

examples/llama.android/llama/.gitignore

@@ -0,0 +1 @@
+/build

examples/llama.android/llama/CMakeLists.txt

@@ -42,7 +42,7 @@ add_subdirectory(../../../../../../ build-llama)
 # used in the AndroidManifest.xml file.
 add_library(${CMAKE_PROJECT_NAME} SHARED
     # List C/C++ source files with relative paths to this CMakeLists.txt.
-    llama-android.cpp)
+        llama-android.cpp)
 
 # Specifies libraries CMake should link to your target library. You
 # can link libraries from various origins, such as libraries defined in this

examples/llama.android/llama/build.gradle.kts

@@ -0,0 +1,68 @@
+plugins {
+    id("com.android.library")
+    id("org.jetbrains.kotlin.android")
+}
+
+android {
+    namespace = "android.llama.cpp"
+    compileSdk = 34
+
+    defaultConfig {
+        minSdk = 33
+
+        testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner"
+        consumerProguardFiles("consumer-rules.pro")
+        ndk {
+            // Add NDK properties if wanted, e.g.
+            // abiFilters += listOf("arm64-v8a")
+        }
+        externalNativeBuild {
+            cmake {
+                arguments += "-DCMAKE_BUILD_TYPE=Release"
+                cppFlags += listOf()
+                arguments += listOf()
+
+                cppFlags("")
+            }
+        }
+    }
+
+    buildTypes {
+        release {
+            isMinifyEnabled = false
+            proguardFiles(
+                getDefaultProguardFile("proguard-android-optimize.txt"),
+                "proguard-rules.pro"
+            )
+        }
+    }
+    externalNativeBuild {
+        cmake {
+            path("src/main/cpp/CMakeLists.txt")
+            version = "3.22.1"
+        }
+    }
+    compileOptions {
+        sourceCompatibility = JavaVersion.VERSION_1_8
+        targetCompatibility = JavaVersion.VERSION_1_8
+    }
+    kotlinOptions {
+        jvmTarget = "1.8"
+    }
+
+    packaging {
+        resources {
+            excludes += "/META-INF/{AL2.0,LGPL2.1}"
+        }
+    }
+}
+
+dependencies {
+
+    implementation("androidx.core:core-ktx:1.12.0")
+    implementation("androidx.appcompat:appcompat:1.6.1")
+    implementation("com.google.android.material:material:1.11.0")
+    testImplementation("junit:junit:4.13.2")
+    androidTestImplementation("androidx.test.ext:junit:1.1.5")
+    androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")
+}

examples/llama.android/llama/proguard-rules.pro

@@ -0,0 +1,21 @@
+# Add project specific ProGuard rules here.
+# You can control the set of applied configuration files using the
+# proguardFiles setting in build.gradle.
+#
+# For more details, see
+#   http://developer.android.com/guide/developing/tools/proguard.html
+
+# If your project uses WebView with JS, uncomment the following
+# and specify the fully qualified class name to the JavaScript interface
+# class:
+#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
+#   public *;
+#}
+
+# Uncomment this to preserve the line number information for
+# debugging stack traces.
+#-keepattributes SourceFile,LineNumberTable
+
+# If you keep the line number information, uncomment this to
+# hide the original source file name.
+#-renamesourcefileattribute SourceFile

examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt

@@ -0,0 +1,24 @@
+package android.llama.cpp
+
+import androidx.test.platform.app.InstrumentationRegistry
+import androidx.test.ext.junit.runners.AndroidJUnit4
+
+import org.junit.Test
+import org.junit.runner.RunWith
+
+import org.junit.Assert.*
+
+/**
+ * Instrumented test, which will execute on an Android device.
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+@RunWith(AndroidJUnit4::class)
+class ExampleInstrumentedTest {
+    @Test
+    fun useAppContext() {
+        // Context of the app under test.
+        val appContext = InstrumentationRegistry.getInstrumentation().targetContext
+        assertEquals("android.llama.cpp.test", appContext.packageName)
+    }
+}

examples/llama.android/llama/src/main/AndroidManifest.xml

@@ -0,0 +1,4 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android">
+
+</manifest>

examples/llama.android/llama/src/main/cpp/CMakeLists.txt

@@ -0,0 +1,49 @@
+# For more information about using CMake with Android Studio, read the
+# documentation: https://d.android.com/studio/projects/add-native-code.html.
+# For more examples on how to use CMake, see https://github.com/android/ndk-samples.
+
+# Sets the minimum CMake version required for this project.
+cmake_minimum_required(VERSION 3.22.1)
+
+# Declares the project name. The project name can be accessed via ${ PROJECT_NAME},
+# Since this is the top level CMakeLists.txt, the project name is also accessible
+# with ${CMAKE_PROJECT_NAME} (both CMake variables are in-sync within the top level
+# build script scope).
+project("llama-android")
+
+include(FetchContent)
+FetchContent_Declare(
+        llama
+        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
+        GIT_TAG        master
+)
+
+# Also provides "common"
+FetchContent_MakeAvailable(llama)
+
+# Creates and names a library, sets it as either STATIC
+# or SHARED, and provides the relative paths to its source code.
+# You can define multiple libraries, and CMake builds them for you.
+# Gradle automatically packages shared libraries with your APK.
+#
+# In this top level CMakeLists.txt, ${CMAKE_PROJECT_NAME} is used to define
+# the target library name; in the sub-module's CMakeLists.txt, ${PROJECT_NAME}
+# is preferred for the same purpose.
+#
+# In order to load a library into your app from Java/Kotlin, you must call
+# System.loadLibrary() and pass the name of the library defined here;
+# for GameActivity/NativeActivity derived applications, the same library name must be
+# used in the AndroidManifest.xml file.
+add_library(${CMAKE_PROJECT_NAME} SHARED
+        # List C/C++ source files with relative paths to this CMakeLists.txt.
+        llama-android.cpp)
+
+# Specifies libraries CMake should link to your target library. You
+# can link libraries from various origins, such as libraries defined in this
+# build script, prebuilt third-party libraries, or Android system libraries.
+target_link_libraries(${CMAKE_PROJECT_NAME}
+        # List libraries link to the target library
+        llama
+        common
+        android
+        log)

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -81,7 +81,7 @@ static void log_callback(ggml_log_level level, const char * fmt, void * data) {
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
+Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring filename) {
     llama_model_params model_params = llama_model_default_params();
 
     auto path_to_model = env->GetStringUTFChars(filename, 0);
@@ -101,13 +101,13 @@ Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1model(JNIEnv *, jobject, jlong model) {
+Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) {
     llama_free_model(reinterpret_cast<llama_model *>(model));
 }
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
+Java_android_llama_cpp_LLamaAndroid_new_1context(JNIEnv *env, jobject, jlong jmodel) {
     auto model = reinterpret_cast<llama_model *>(jmodel);
 
     if (!model) {
@@ -139,25 +139,25 @@ Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1context(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_free_1context(JNIEnv *, jobject, jlong context) {
     llama_free(reinterpret_cast<llama_context *>(context));
 }
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1free(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1free(JNIEnv *, jobject) {
     llama_backend_free();
 }
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_log_1to_1android(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_log_1to_1android(JNIEnv *, jobject) {
     llama_log_set(log_callback, NULL);
 }
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_bench_1model(
+Java_android_llama_cpp_LLamaAndroid_bench_1model(
         JNIEnv *env,
         jobject,
         jlong context_pointer,
@@ -271,13 +271,13 @@ Java_com_example_llama_Llm_bench_1model(
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
+Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
     llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
 }
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
+Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
 
     // Source: Copy of llama.cpp:llama_batch_init but heap-allocated.
 
@@ -313,19 +313,19 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1init(JNIEnv *, jobject) {
     llama_backend_init();
 }
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_system_1info(JNIEnv *env, jobject) {
+Java_android_llama_cpp_LLamaAndroid_system_1info(JNIEnv *env, jobject) {
     return env->NewStringUTF(llama_print_system_info());
 }
 
 extern "C"
 JNIEXPORT jint JNICALL
-Java_com_example_llama_Llm_completion_1init(
+Java_android_llama_cpp_LLamaAndroid_completion_1init(
         JNIEnv *env,
         jobject,
         jlong context_pointer,
@@ -376,7 +376,7 @@ Java_com_example_llama_Llm_completion_1init(
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_completion_1loop(
+Java_android_llama_cpp_LLamaAndroid_completion_1loop(
         JNIEnv * env,
         jobject,
         jlong context_pointer,
@@ -438,6 +438,6 @@ Java_com_example_llama_Llm_completion_1loop(
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
     llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
 }

examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt

@@ -1,4 +1,4 @@
-package com.example.llama
+package android.llama.cpp
 
 import android.util.Log
 import kotlinx.coroutines.CoroutineDispatcher
@@ -10,7 +10,7 @@ import kotlinx.coroutines.withContext
 import java.util.concurrent.Executors
 import kotlin.concurrent.thread
 
-class Llm {
+class LLamaAndroid {
     private val tag: String? = this::class.simpleName
 
     private val threadLocalState: ThreadLocal<State> = ThreadLocal.withInitial { State.Idle }
@@ -165,8 +165,8 @@ class Llm {
         }
 
         // Enforce only one instance of Llm.
-        private val _instance: Llm = Llm()
+        private val _instance: LLamaAndroid = LLamaAndroid()
 
-        fun instance(): Llm = _instance
+        fun instance(): LLamaAndroid = _instance
     }
 }

examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt

@@ -0,0 +1,17 @@
+package android.llama.cpp
+
+import org.junit.Test
+
+import org.junit.Assert.*
+
+/**
+ * Example local unit test, which will execute on the development machine (host).
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+class ExampleUnitTest {
+    @Test
+    fun addition_isCorrect() {
+        assertEquals(4, 2 + 2)
+    }
+}

examples/llama.android/settings.gradle.kts

@@ -15,3 +15,4 @@ dependencyResolutionManagement {
 
 rootProject.name = "LlamaAndroid"
 include(":app")
+include(":llama")

examples/main/main.cpp

@@ -740,18 +740,26 @@ int main(int argc, char ** argv) {
         // display text
         if (input_echo && display) {
             for (auto id : embd) {
-                const std::string token_str = llama_token_to_piece(ctx, id, !params.conversation);
+                const std::string token_str = llama_token_to_piece(ctx, id, params.special);
-                printf("%s", token_str.c_str());
 
+                // Console/Stream Output
+                fprintf(stdout, "%s", token_str.c_str());
+
+                // Record Displayed Tokens To Log
+                // Note: Generated tokens are created one by one hence this check
                 if (embd.size() > 1) {
+                    // Incoming Requested Tokens
                     input_tokens.push_back(id);
                 } else {
+                    // Outgoing Generated Tokens
                     output_tokens.push_back(id);
                     output_ss << token_str;
                 }
+
+                fflush(stdout);
             }
-            fflush(stdout);
         }
+
         // reset color to default if there is no pending user input
         if (input_echo && (int) embd_inp.size() == n_consumed) {
             console::set_display(console::reset);

examples/server/public_simplechat/index.html

@@ -1,7 +1,7 @@
 <!DOCTYPE html>
 <html lang="en">
     <head>
-        <title>SimpleChat (LlamaCPP, ...) </title>
+        <title>SimpleChat LlamaCppEtal </title>
         <meta charset="UTF-8" />
         <meta name="viewport" content="width=device-width, initial-scale=1" />
         <meta name="message" content="Save Nature Save Earth" />
@@ -30,20 +30,17 @@
             <hr>
             <div class="sameline">
                 <label for="system-in">System</label>
-                <input type="text" name="system" id="system-in" class="flex-grow"/>
+                <input type="text" name="system" id="system-in" placeholder="e.g. you are a helpful ai assistant, who provides concise answers" class="flex-grow"/>
             </div>
 
             <hr>
             <div id="chat-div">
-                <p> Enter the system prompt above, before entering/submitting any user query.</p>
+                <p> You need to have javascript enabled.</p>
-                <p> Enter your text to the ai assistant below.</p>
-                <p> Use shift+enter for inserting enter.</p>
-                <p> Refresh the page to start over fresh.</p>
             </div>
 
             <hr>
             <div class="sameline">
-                <textarea id="user-in" class="flex-grow" rows="3"></textarea>
+                <textarea id="user-in" class="flex-grow" rows="3" placeholder="enter your query to the ai model here" ></textarea>
                 <button id="user-btn">submit</button>
             </div>
 

examples/server/public_simplechat/readme.md

@@ -14,11 +14,15 @@ own system prompts.
 The UI follows a responsive web design so that the layout can adapt to available display space in a usable
 enough manner, in general.
 
-NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
+Allows developer/end-user to control some of the behaviour by updating gMe members from browser's devel-tool
-culling of old messages from the chat.
+console.
 
-NOTE: It doesnt set any parameters other than temperature for now. However if someone wants they can update
+NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
-the js file as needed.
+culling of old messages from the chat by default. However by enabling the sliding window chat logic, a crude
+form of old messages culling can be achieved.
+
+NOTE: It doesnt set any parameters other than temperature and max_tokens for now. However if someone wants
+they can update the js file or equivalent member in gMe as needed.
 
 
 ## usage
@@ -43,11 +47,33 @@ next run this web front end in examples/server/public_simplechat
 ### using the front end
 
 Open this simple web front end from your local browser
+
 * http://127.0.0.1:PORT/index.html
 
 Once inside
+
 * Select between chat and completion mode. By default it is set to chat mode.
+
+* In completion mode
+  * logic by default doesnt insert any role specific "ROLE: " prefix wrt each role's message.
+    If the model requires any prefix wrt user role messages, then the end user has to
+    explicitly add the needed prefix, when they enter their chat message.
+    Similarly if the model requires any prefix to trigger assistant/ai-model response,
+    then the end user needs to enter the same.
+    This keeps the logic simple, while still giving flexibility to the end user to
+    manage any templating/tagging requirement wrt their messages to the model.
+  * the logic doesnt insert newline at the begining and end wrt the prompt message generated.
+    However if the chat being sent to /completions end point has more than one role's message,
+    then insert newline when moving from one role's message to the next role's message, so
+    that it can be clearly identified/distinguished.
+  * given that /completions endpoint normally doesnt add additional chat-templating of its
+    own, the above ensures that end user can create a custom single/multi message combo with
+    any tags/special-tokens related chat templating to test out model handshake. Or enduser
+    can use it just for normal completion related/based query.
+
 * If you want to provide a system prompt, then ideally enter it first, before entering any user query.
+  Normally Completion mode doesnt need system prompt, while Chat mode can generate better/interesting
+  responses with a suitable system prompt.
   * if chat.add_system_begin is used
     * you cant change the system prompt, after it is has been submitted once along with user query.
     * you cant set a system prompt, after you have submitted any user query
@@ -55,27 +81,121 @@ Once inside
     * one can change the system prompt any time during chat, by changing the contents of system prompt.
     * inturn the updated/changed system prompt will be inserted into the chat session.
     * this allows for the subsequent user chatting to be driven by the new system prompt set above.
+
 * Enter your query and either press enter or click on the submit button.
   If you want to insert enter (\n) as part of your chat/query to ai model, use shift+enter.
+
 * Wait for the logic to communicate with the server and get the response.
   * the user is not allowed to enter any fresh query during this time.
   * the user input box will be disabled and a working message will be shown in it.
+
 * just refresh the page, to reset wrt the chat history and or system prompt and start afresh.
+
 * Using NewChat one can start independent chat sessions.
   * two independent chat sessions are setup by default.
 
 
 ## Devel note
 
+### Reason behind this
+
+The idea is to be easy enough to use for basic purposes, while also being simple and easily discernable
+by developers who may not be from web frontend background (so inturn may not be familiar with template /
+end-use-specific-language-extensions driven flows) so that they can use it to explore/experiment things.
+
+And given that the idea is also to help explore/experiment for developers, some flexibility is provided
+to change behaviour easily using the devel-tools/console, for now. And skeletal logic has been implemented
+to explore some of the end points and ideas/implications around them.
+
+
+### General
+
+Me/gMe consolidates the settings which control the behaviour into one object.
+One can see the current settings, as well as change/update them using browsers devel-tool/console.
+
+  bCompletionFreshChatAlways - whether Completion mode collates complete/sliding-window history when
+  communicating with the server or only sends the latest user query/message.
+
+  bCompletionInsertStandardRolePrefix - whether Completion mode inserts role related prefix wrt the
+  messages that get inserted into prompt field wrt /Completion endpoint.
+
+  chatRequestOptions - maintains the list of options/fields to send along with chat request,
+  irrespective of whether /chat/completions or /completions endpoint.
+
+    If you want to add additional options/fields to send to the server/ai-model, and or
+    modify the existing options value or remove them, for now you can update this global var
+    using browser's development-tools/console.
+
+  iRecentUserMsgCnt - a simple minded SlidingWindow to limit context window load at Ai Model end.
+  This is disabled by default. However if enabled, then in addition to latest system message, only
+  the last/latest iRecentUserMsgCnt user messages after the latest system prompt and its responses
+  from the ai model will be sent to the ai-model, when querying for a new response. IE if enabled,
+  only user messages after the latest system message/prompt will be considered.
+
+    This specified sliding window user message count also includes the latest user query.
+    <0 : Send entire chat history to server
+     0 : Send only the system message if any to the server
+    >0 : Send the latest chat history from the latest system prompt, limited to specified cnt.
+
+
+By using gMe's iRecentUserMsgCnt and chatRequestOptions.max_tokens one can try to control the
+implications of loading of the ai-model's context window by chat history, wrt chat response to
+some extent in a simple crude way.
+
+
 Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js
 may not be visible. Also remember that just refreshing/reloading page in browser or for that
 matter clearing site data, dont directly override site caching in all cases. Worst case you may
 have to change port. Or in dev tools of browser, you may be able to disable caching fully.
 
+
 Concept of multiple chat sessions with different servers, as well as saving and restoring of
 those across browser usage sessions, can be woven around the SimpleChat/MultiChatUI class and
 its instances relatively easily, however given the current goal of keeping this simple, it has
 not been added, for now.
 
+
 By switching between chat.add_system_begin/anytime, one can control whether one can change
 the system prompt, anytime during the conversation or only at the beginning.
+
+
+read_json_early, is to experiment with reading json response data early on, if available,
+so that user can be shown generated data, as and when it is being generated, rather than
+at the end when full data is available.
+
+  the server flow doesnt seem to be sending back data early, atleast for request (inc options)
+  that is currently sent.
+
+  if able to read json data early on in future, as and when ai model is generating data, then
+  this helper needs to indirectly update the chat div with the recieved data, without waiting
+  for the overall data to be available.
+
+
+### Default setup
+
+By default things are setup to try and make the user experience a bit better, if possible.
+However a developer when testing the server of ai-model may want to change these value.
+
+Using iRecentUserMsgCnt reduce chat history context sent to the server/ai-model to be
+just the system-prompt, prev-user-request-and-ai-response and cur-user-request, instead of
+full chat history. This way if there is any response with garbage/repeatation, it doesnt
+mess with things beyond the next question/request/query, in some ways.
+
+Set max_tokens to 1024, so that a relatively large previous reponse doesnt eat up the space
+available wrt next query-response. However dont forget that the server when started should
+also be started with a model context size of 1k or more, to be on safe side.
+
+  The /completions endpoint of examples/server doesnt take max_tokens, instead it takes the
+  internal n_predict, for now add the same here on the client side, maybe later add max_tokens
+  to /completions endpoint handling code on server side.
+
+Frequency and presence penalty fields are set to 1.2 in the set of fields sent to server
+along with the user query. So that the model is partly set to try avoid repeating text in
+its response.
+
+A end-user can change these behaviour by editing gMe from browser's devel-tool/console.
+
+
+## At the end
+
+Also a thank you to all open source and open model developers, who strive for the common good.

examples/server/public_simplechat/simplechat.css

@@ -48,6 +48,13 @@ button {
     flex-direction: column;
 }
 
+.ul1 {
+    padding-inline-start: 2vw;
+}
+.ul2 {
+    padding-inline-start: 2vw;
+}
+
 * {
     margin: 0.6vmin;
 }

examples/server/public_simplechat/simplechat.js

@@ -14,23 +14,86 @@ class ApiEP {
 }
 
 let gUsageMsg = `
-    <p> Enter the system prompt above, before entering/submitting any user query.</p>
+    <p class="role-system">Usage</p>
-    <p> Enter your text to the ai assistant below.</p>
+    <ul class="ul1">
-    <p> Use shift+enter for inserting enter.</p>
+    <li> Set system prompt above, to try control ai response charactersitic, if model supports same.</li>
-    <p> Refresh the page to start over fresh.</p>
+        <ul class="ul2">
+        <li> Completion mode normally wont have a system prompt.</li>
+        </ul>
+    <li> Enter your query to ai assistant below.</li>
+        <ul class="ul2">
+        <li> Completion mode doesnt insert user/role: prefix implicitly.</li>
+        <li> Use shift+enter for inserting enter/newline.</li>
+        </ul>
+    <li> Default ContextWindow = [System, Last Query+Resp, Cur Query].</li>
+        <ul class="ul2">
+        <li> experiment iRecentUserMsgCnt, max_tokens, model ctxt window to expand</li>
+        </ul>
+    </ul>
 `;
 
+/** @typedef {{role: string, content: string}[]} ChatMessages */
+
 class SimpleChat {
 
     constructor() {
         /**
          * Maintain in a form suitable for common LLM web service chat/completions' messages entry
-         * @type {{role: string, content: string}[]}
+         * @type {ChatMessages}
          */
         this.xchat = [];
         this.iLastSys = -1;
     }
 
+    clear() {
+        this.xchat = [];
+        this.iLastSys = -1;
+    }
+
+    /**
+     * Recent chat messages.
+     * If iRecentUserMsgCnt < 0
+     *   Then return the full chat history
+     * Else
+     *   Return chat messages from latest going back till the last/latest system prompt.
+     *   While keeping track that the number of user queries/messages doesnt exceed iRecentUserMsgCnt.
+     * @param {number} iRecentUserMsgCnt
+     */
+    recent_chat(iRecentUserMsgCnt) {
+        if (iRecentUserMsgCnt < 0) {
+            return this.xchat;
+        }
+        if (iRecentUserMsgCnt == 0) {
+            console.warn("WARN:SimpleChat:SC:RecentChat:iRecentUsermsgCnt of 0 means no user message/query sent");
+        }
+        /** @type{ChatMessages} */
+        let rchat = [];
+        let sysMsg = this.get_system_latest();
+        if (sysMsg.length != 0) {
+            rchat.push({role: Roles.System, content: sysMsg});
+        }
+        let iUserCnt = 0;
+        let iStart = this.xchat.length;
+        for(let i=this.xchat.length-1; i > this.iLastSys; i--) {
+            if (iUserCnt >= iRecentUserMsgCnt) {
+                break;
+            }
+            let msg = this.xchat[i];
+            if (msg.role == Roles.User) {
+                iStart = i;
+                iUserCnt += 1;
+            }
+        }
+        for(let i = iStart; i < this.xchat.length; i++) {
+            let msg = this.xchat[i];
+            if (msg.role == Roles.System) {
+                continue;
+            }
+            rchat.push({role: msg.role, content: msg.content});
+        }
+        return rchat;
+    }
+
     /**
      * Add an entry into xchat
      * @param {string} role
@@ -57,7 +120,7 @@ class SimpleChat {
             div.replaceChildren();
         }
         let last = undefined;
-        for(const x of this.xchat) {
+        for(const x of this.recent_chat(gMe.iRecentUserMsgCnt)) {
             let entry = document.createElement("p");
             entry.className = `role-${x.role}`;
             entry.innerText = `${x.role}: ${x.content}`;
@@ -69,17 +132,21 @@ class SimpleChat {
         } else {
             if (bClear) {
                 div.innerHTML = gUsageMsg;
+                gMe.show_info(div);
             }
         }
     }
 
     /**
-     * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint
+     * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
+     * The needed fields/options are picked from a global object.
      * Convert the json into string.
      * @param {Object} obj
      */
     request_jsonstr(obj) {
-        obj["temperature"] = 0.7;
+        for(let k in gMe.chatRequestOptions) {
+            obj[k] = gMe.chatRequestOptions[k];
+        }
         return JSON.stringify(obj);
     }
 
@@ -88,18 +155,27 @@ class SimpleChat {
      */
     request_messages_jsonstr() {
         let req = {
-            messages: this.xchat,
+            messages: this.recent_chat(gMe.iRecentUserMsgCnt),
         }
         return this.request_jsonstr(req);
     }
 
     /**
      * Return a string form of json object suitable for /completions
+     * @param {boolean} bInsertStandardRolePrefix Insert "<THE_ROLE>: " as prefix wrt each role's message
      */
-    request_prompt_jsonstr() {
+    request_prompt_jsonstr(bInsertStandardRolePrefix) {
         let prompt = "";
-        for(const chat of this.xchat) {
+        let iCnt = 0;
-            prompt += `${chat.role}: ${chat.content}\n`;
+        for(const chat of this.recent_chat(gMe.iRecentUserMsgCnt)) {
+            iCnt += 1;
+            if (iCnt > 1) {
+                prompt += "\n";
+            }
+            if (bInsertStandardRolePrefix) {
+                prompt += `${chat.role}: `;
+            }
+            prompt += `${chat.content}`;
         }
         let req = {
             prompt: prompt,
@@ -171,7 +247,6 @@ let gChatURL = {
     'chat': `${gBaseURL}/chat/completions`,
     'completion': `${gBaseURL}/completions`,
 }
-const gbCompletionFreshChatAlways = true;
 
 
 /**
@@ -291,6 +366,8 @@ class MultiChatUI {
             // allow user to insert enter into their message using shift+enter.
             // while just pressing enter key will lead to submitting.
             if ((ev.key === "Enter") && (!ev.shiftKey)) {
+                let value = this.elInUser.value;
+                this.elInUser.value = value.substring(0,value.length-1);
                 this.elBtnUser.click();
                 ev.preventDefault();
             }
@@ -321,6 +398,29 @@ class MultiChatUI {
         }
     }
 
+    /**
+     * Try read json response early, if available.
+     * @param {Response} resp
+     */
+    async read_json_early(resp) {
+        if (!resp.body) {
+            throw Error("ERRR:SimpleChat:MCUI:ReadJsonEarly:No body...");
+        }
+        let tdUtf8 = new TextDecoder("utf-8");
+        let rr = resp.body.getReader();
+        let gotBody = "";
+        while(true) {
+            let { value: cur,  done: done} = await rr.read();
+            let curBody = tdUtf8.decode(cur);
+            console.debug("DBUG:SC:PART:", curBody);
+            gotBody += curBody;
+            if (done) {
+                break;
+            }
+        }
+        return JSON.parse(gotBody);
+    }
+
     /**
      * Handle user query submit request, wrt specified chat session.
      * @param {string} chatId
@@ -330,6 +430,14 @@ class MultiChatUI {
 
         let chat = this.simpleChats[chatId];
 
+        // In completion mode, if configured, clear any previous chat history.
+        // So if user wants to simulate a multi-chat based completion query,
+        // they will have to enter the full thing, as a suitable multiline
+        // user input/query.
+        if ((apiEP == ApiEP.Completion) && (gMe.bCompletionFreshChatAlways)) {
+            chat.clear();
+        }
+
         chat.add_system_anytime(this.elInSystem.value, chatId);
 
         let content = this.elInUser.value;
@@ -344,7 +452,7 @@ class MultiChatUI {
         if (apiEP == ApiEP.Chat) {
             theBody = chat.request_messages_jsonstr();
         } else {
-            theBody = chat.request_prompt_jsonstr();
+            theBody = chat.request_prompt_jsonstr(gMe.bCompletionInsertStandardRolePrefix);
         }
 
         this.elInUser.value = "working...";
@@ -359,6 +467,7 @@ class MultiChatUI {
         });
 
         let respBody = await resp.json();
+        //let respBody = await this.read_json_early(resp);
         console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`);
         let assistantMsg;
         if (apiEP == ApiEP.Chat) {
@@ -376,13 +485,6 @@ class MultiChatUI {
         } else {
             console.debug(`DBUG:SimpleChat:MCUI:HandleUserSubmit:ChatId has changed:[${chatId}] [${this.curChatId}]`);
         }
-        // Purposefully clear at end rather than begin of this function
-        // so that one can switch from chat to completion mode and sequece
-        // in a completion mode with multiple user-assistant chat data
-        // from before to be sent/occur once.
-        if ((apiEP == ApiEP.Completion) && (gbCompletionFreshChatAlways)) {
-            chat.xchat.length = 0;
-        }
         this.ui_reset_userinput();
     }
 
@@ -462,17 +564,66 @@ class MultiChatUI {
 }
 
 
-let gMuitChat;
+class Me {
-const gChatIds = [ "Default", "Other" ];
+
+    constructor() {
+        this.defaultChatIds = [ "Default", "Other" ];
+        this.multiChat = new MultiChatUI();
+        this.bCompletionFreshChatAlways = true;
+        this.bCompletionInsertStandardRolePrefix = false;
+        this.iRecentUserMsgCnt = 2;
+        // Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
+        this.chatRequestOptions = {
+            "temperature": 0.7,
+            "max_tokens": 1024,
+            "frequency_penalty": 1.2,
+            "presence_penalty": 1.2,
+            "n_predict": 1024
+        };
+    }
+
+    /**
+     * @param {HTMLDivElement} elDiv
+     */
+    show_info(elDiv) {
+
+        var p = document.createElement("p");
+        p.innerText = "Settings (devel-tools-console gMe)";
+        p.className = "role-system";
+        elDiv.appendChild(p);
+
+        var p = document.createElement("p");
+        p.innerText = `bCompletionFreshChatAlways:${this.bCompletionFreshChatAlways}`;
+        elDiv.appendChild(p);
+
+        p = document.createElement("p");
+        p.innerText = `bCompletionInsertStandardRolePrefix:${this.bCompletionInsertStandardRolePrefix}`;
+        elDiv.appendChild(p);
+
+        p = document.createElement("p");
+        p.innerText = `iRecentUserMsgCnt:${this.iRecentUserMsgCnt}`;
+        elDiv.appendChild(p);
+
+        p = document.createElement("p");
+        p.innerText = `chatRequestOptions:${JSON.stringify(this.chatRequestOptions)}`;
+        elDiv.appendChild(p);
+
+    }
+
+}
+
+
+/** @type {Me} */
+let gMe;
 
 function startme() {
     console.log("INFO:SimpleChat:StartMe:Starting...");
-    gMuitChat = new MultiChatUI();
+    gMe = new Me();
-    for (let cid of gChatIds) {
+    for (let cid of gMe.defaultChatIds) {
-        gMuitChat.new_chat_session(cid);
+        gMe.multiChat.new_chat_session(cid);
     }
-    gMuitChat.setup_ui(gChatIds[0]);
+    gMe.multiChat.setup_ui(gMe.defaultChatIds[0], true);
-    gMuitChat.show_sessions();
+    gMe.multiChat.show_sessions();
 }
 
 document.addEventListener("DOMContentLoaded", startme);

examples/sycl/win-build-sycl.bat

@@ -13,10 +13,10 @@ if %errorlevel% neq 0 goto ERROR
 
 ::  for FP16
 ::  faster for long-prompt inference
-::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
 ::  for FP32
-cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
 if %errorlevel% neq 0 goto ERROR
 ::  build example/main only
 ::  make main

examples/tokenize/tokenize.cpp

@@ -3,40 +3,390 @@
 
 #include <cmath>
 #include <cstdio>
+#include <fstream>
 #include <string>
 #include <vector>
 
-int main(int argc, char ** argv) {
+#if defined(_WIN32)
-    if (argc < 3 || argv[1][0] == '-') {
+#define WIN32_LEAN_AND_MEAN
-        printf("usage: %s MODEL_PATH PROMPT [--ids]\n" , argv[0]);
+#include <windows.h>
+#include <shellapi.h>   // For CommandLineToArgvW
+#endif
+
+static void print_usage_information(const char * argv0, FILE * stream) {
+    fprintf(stream, "usage: %s [options]\n\n", argv0);
+    fprintf(stream, "The tokenize program tokenizes a prompt using a given model,\n");
+    fprintf(stream, "and prints the resulting tokens to standard output.\n\n");
+    fprintf(stream, "It needs a model file, a prompt, and optionally other flags\n");
+    fprintf(stream, "to control the behavior of the tokenizer.\n\n");
+    fprintf(stream, "    The possible options are:\n");
+    fprintf(stream, "\n");
+    fprintf(stream, "    -h, --help                           print this help and exit\n");
+    fprintf(stream, "    -m MODEL_PATH, --model MODEL_PATH    path to model.\n");
+    fprintf(stream, "    --ids                                if given, only print numerical token IDs, and not token strings.\n");
+    fprintf(stream, "                                         The output format looks like [1, 2, 3], i.e. parseable by Python.\n");
+    fprintf(stream, "    -f PROMPT_FNAME, --file PROMPT_FNAME read prompt from a file.\n");
+    fprintf(stream, "    -p PROMPT, --prompt PROMPT           read prompt from the argument.\n");
+    fprintf(stream, "    --stdin                              read prompt from standard input.\n");
+    fprintf(stream, "    --no-bos                             do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n");
+    fprintf(stream, "    --log-disable                        disable logs. Makes stderr quiet when loading the model.\n");
+}
+
+static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) text;
+    (void) user_data;
+}
+
+static std::string read_prompt_from_file(const char * filepath, bool & success) {
+    success = false;
+
+    std::ifstream in(filepath, std::ios::binary);
+    if (!in) {
+        fprintf(stderr, "%s: could not open file '%s' for reading: %s\n", __func__, filepath, strerror(errno));
+        return std::string();
+    }
+    // do not assume the file is seekable (e.g. /dev/stdin)
+    std::stringstream buffer;
+    buffer << in.rdbuf();
+    if (in.fail()) {
+        fprintf(stderr, "%s: could not read the entire file '%s': %s\n", __func__, filepath, strerror(errno));
+        return std::string();
+    }
+
+    success = true;
+    return buffer.str();
+}
+
+//
+// Function: ingest_args(...) -> vector<string>
+//
+//  Takes argc and argv arguments, and converts them to a vector of UTF-8 encoded
+//  strings, as an STL vector<string>.
+//
+//  In particular, it handles character encoding shenanigans on Windows.
+//
+// Note: raw_argc and raw_argv are not actually read at all on Windows.
+//       On Windows we call GetCommandLineW to get the arguments in wchar_t
+//       format, ignoring the regular argc/argv arguments to main().
+//
+// TODO: potential opportunity to roll common stuff into common/console.cpp
+//       in relation to Windows wchar_t shenanigans.
+static std::vector<std::string> ingest_args(int raw_argc, char ** raw_argv) {
+    std::vector<std::string> argv;
+
+    // Handle Windows, if given non-ASCII arguments.
+    // We convert wchar_t arguments into UTF-8 char* on this platform.
+    // Lets you invoke 'tokenize' on Windows cmd.exe with non-ASCII characters
+    // without throwing tantrums.
+#if defined(_WIN32)
+    int argc;
+    const LPWSTR cmdline_wargv = GetCommandLineW();
+    LPWSTR * wargv = CommandLineToArgvW(cmdline_wargv, &argc);
+
+    // silence unused arg warnings
+    (void) raw_argc;
+    (void) raw_argv;
+
+    for (int i = 0; i < argc; ++i) {
+        int length_needed = WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), 0, 0, NULL, NULL);
+        char * output_buf = (char *) calloc(length_needed+1, sizeof(char));
+        GGML_ASSERT(output_buf);
+
+        WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), output_buf, length_needed, NULL, NULL);
+        output_buf[length_needed] = '\0';
+
+        argv.push_back(output_buf);
+        free(output_buf);
+    }
+
+    LocalFree((HLOCAL) wargv);
+#else
+    int argc = raw_argc;
+    for (int i = 0; i < argc; ++i) {
+        argv.push_back(raw_argv[i]);
+    }
+#endif
+
+    GGML_ASSERT((unsigned int) argc == argv.size());
+
+    return argv;
+}
+
+//
+// Function: write_utf8_cstr_to_stdout(const char *) -> <writes to stdout>
+//
+// writes a string to standard output; taking into account that on Windows
+// to display correctly you have to use special handling. Works even if the
+// user has not set a unicode code page on a Windows cmd.exe.
+//
+// In case of invalid UTF-8, invalid_utf8 is set to true on Windows, and something
+// a human-readable is written instead.
+//
+// On non-Windows systems, simply printfs() the string.
+static void write_utf8_cstr_to_stdout(const char * str, bool & invalid_utf8) {
+        invalid_utf8 = false;
+
+#if defined(_WIN32)
+        // Are we in a console?
+        HANDLE hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
+        DWORD dwMode = 0;
+
+        // According to Microsoft docs:
+        // "WriteConsole fails if it is used with a standard handle that is redirected to a file."
+        // Also according to the docs, you can use GetConsoleMode to check for that.
+        if (hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(hConsole, &dwMode)) {
+            printf("%s", str);
+            return;
+        }
+
+        // MultiByteToWideChar reports an error if str is empty, don't report
+        // them as invalid_utf8.
+        if (*str == 0) {
+            return;
+        }
+        int length_needed = MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, str, strlen(str), NULL, 0);
+        if (length_needed == 0) {
+            DWORD err = GetLastError();
+            if (err == ERROR_NO_UNICODE_TRANSLATION) {
+                invalid_utf8 = true;
+                int len = strlen(str);
+                printf("<");
+                for (int i = 0; i < len; ++i) {
+                    if (i > 0) {
+                        printf(" ");
+                    }
+                    printf("%02x", (uint8_t) str[i]);
+                }
+                printf(">");
+                return;
+            }
+            GGML_ASSERT(false && "MultiByteToWideChar() failed in an unexpected way.");
+        }
+
+        LPWSTR wstr = (LPWSTR) calloc(length_needed+1, sizeof(*wstr));
+        GGML_ASSERT(wstr);
+
+        MultiByteToWideChar(CP_UTF8, 0, str, strlen(str), wstr, length_needed);
+        WriteConsoleW(hConsole, wstr, length_needed, NULL, NULL);
+
+        free(wstr);
+#else
+        // TODO: reporting invalid_utf8 would be useful on non-Windows too.
+        // printf will silently just write bad unicode.
+        printf("%s", str);
+#endif
+}
+
+int main(int raw_argc, char ** raw_argv) {
+    const std::vector<std::string> argv = ingest_args(raw_argc, raw_argv);
+    const int argc = argv.size();
+
+    if (argc <= 1) {
+        print_usage_information(argv[0].c_str(), stderr);
         return 1;
     }
 
-    const char * model_path = argv[1];
+    //////
-    const char * prompt     = argv[2];
+    // Read out all the command line arguments.
+    //////
 
-    const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids";
+    // variables where to put any arguments we see.
+    bool printing_ids = false;
+    bool no_bos = false;
+    bool disable_logging = false;
+    const char * model_path = NULL;
+    const char * prompt_path = NULL;
+    const char * prompt_arg = NULL;
+
+    // track which arguments were explicitly given
+    // used for sanity checking down the line
+    bool model_path_set = false;
+    bool prompt_path_set = false;
+    bool prompt_set = false;
+    bool stdin_set = false;
+
+    int iarg = 1;
+    for (; iarg < argc; ++iarg) {
+        std::string arg{argv[iarg]};
+        if (arg == "-h" || arg == "--help") {
+            print_usage_information(argv[0].c_str(), stdout);
+            return 0;
+        }
+        else if (arg == "--ids") {
+            printing_ids = true;
+        }
+        else if (arg == "-m" || arg == "--model") {
+            if (model_path_set) {
+                fprintf(stderr, "Error: -m or --model specified multiple times.\n");
+                return 1;
+            }
+            model_path = argv[++iarg].c_str();
+            model_path_set = true;
+        }
+        else if (arg == "--no-bos") {
+            no_bos = true;
+        }
+        else if (arg == "-p" || arg == "--prompt") {
+            if (prompt_set) {
+                fprintf(stderr, "Error: -p or --prompt specified multiple times.\n");
+                return 1;
+            }
+            prompt_arg = argv[++iarg].c_str();
+            prompt_set = true;
+        }
+        else if (arg == "-f" || arg == "--file") {
+            if (prompt_path_set) {
+                fprintf(stderr, "Error: -f or --file specified multiple times.\n");
+                return 1;
+            }
+            prompt_path = argv[++iarg].c_str();
+            prompt_path_set = true;
+        }
+        else if (arg == "--stdin") {
+            stdin_set = true;
+        }
+        else if (arg == "--log-disable") {
+            disable_logging = true;
+        }
+        else {
+            fprintf(stderr, "Error: unknown option '%s'\n", argv[iarg].c_str());
+            return 1;
+        }
+    }
+
+    //////
+    // Sanity check the command line arguments.
+    //////
+
+    // Check that we have the required stuff set.
+    if (model_path_set && model_path == NULL) {
+        fprintf(stderr, "Error: --model requires an argument.\n");
+        return 1;
+    }
+    if (!model_path_set) {
+        fprintf(stderr, "Error: must specify --model.\n");
+        return 1;
+    }
+    if (prompt_path_set && prompt_path == NULL) {
+        fprintf(stderr, "Error: --file requires an argument.\n");
+        return 1;
+    }
+    if (prompt_set && prompt_arg == NULL) {
+        fprintf(stderr, "Error: --prompt requires an argument.\n");
+        return 1;
+    }
+    const int prompts_set = !!(prompt_path_set) + !!(prompt_set) + !!(stdin_set);
+    if (prompts_set > 1) {
+        fprintf(stderr, "Error: --stdin, --file and --prompt are mutually exclusive.\n");
+        return 1;
+    }
+    // Must have some prompt.
+    if (prompts_set == 0) {
+        fprintf(stderr, "Error: must specify one of: --stdin, --file or --prompt.\n");
+        return 1;
+    }
+
+    GGML_ASSERT(model_path);
+    GGML_ASSERT(prompt_path || prompt_arg || stdin_set);
+
+    //////
+    // Figure out where will the prompt come from.
+    //////
+
+    std::string prompt;
+    if (prompt_path_set) {
+        bool success = false;
+        prompt = read_prompt_from_file(prompt_path, success);
+        if (!success) {
+            return 1;
+        }
+    } else if (prompt_set) {
+        prompt = prompt_arg;
+    } else {
+        GGML_ASSERT(stdin_set);
+        // we read stdin *after* loading model (early exit if model cannot
+        // be loaded, which can be a nicer user experience)
+    }
+
+    //////
+    // Start actually doing the tokenizing stuff.
+    //////
+
+#ifdef LOG_DISABLE_LOGS
+    disable_logging = true;
+#endif
+
+    if (disable_logging) {
+        llama_log_set(llama_log_callback_null, NULL);
+    }
 
     llama_backend_init();
 
     llama_model_params model_params = llama_model_default_params();
     model_params.vocab_only = true;
     llama_model * model = llama_load_model_from_file(model_path, model_params);
+    if (!model) {
+        fprintf(stderr, "Error: could not load model from file '%s'.\n", model_path);
+        return 1;
+    }
 
     llama_context_params ctx_params = llama_context_default_params();
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+    if (!ctx) {
+        fprintf(stderr, "Error: could not create context.\n");
+        return 1;
+    }
+
+    // read entire prompt from stdin?
+    if (stdin_set) {
+        GGML_ASSERT(!prompt_path_set && !prompt_set);
+
+        std::stringstream stdin_buffer;
+        stdin_buffer << std::cin.rdbuf();
+        if (std::cin.fail()) {
+            fprintf(stderr, "Error: could not read the entire standard input.\n");
+            return 1;
+        }
+
+        prompt = stdin_buffer.str();
+    }
+
+    const bool model_wants_add_bos = llama_should_add_bos_token(model);
+    const bool add_bos = model_wants_add_bos && !no_bos;
 
     std::vector<llama_token> tokens;
+    tokens = ::llama_tokenize(model, prompt, add_bos, true);
 
-    tokens = ::llama_tokenize(model, prompt, true, true);
+    if (printing_ids) {
+        printf("[");
+    }
 
     for (int i = 0; i < (int) tokens.size(); i++) {
         if (printing_ids) {
-            printf("%d\n", tokens[i]);
+            if (i > 0) {
+                printf(", ");
+            }
+            printf("%d", tokens[i]);
         } else {
-            printf("%6d -> '%s'\n", tokens[i], llama_token_to_piece(ctx, tokens[i]).c_str());
+            bool invalid_utf8 = false;
+            printf("%6d -> '", tokens[i]);
+            write_utf8_cstr_to_stdout(llama_token_to_piece(ctx, tokens[i]).c_str(), invalid_utf8);
+            if (invalid_utf8) {
+                printf("' (utf-8 decode failure)\n");
+            } else {
+                printf("'\n");
+            }
         }
     }
 
+    if (printing_ids) {
+        printf("]\n");
+    }
+
+    // silence valgrind
+    llama_free(ctx);
+    llama_free_model(model);
+
     return 0;
 }

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1714641030,
+        "lastModified": 1715865404,
-        "narHash": "sha256-yzcRNDoyVP7+SCNX0wmuDju1NUCt8Dz9+lyUXEI0dbI=",
+        "narHash": "sha256-/GJvTdTpuDjNn84j82cU6bXztE0MSkdnTWClUCRub78=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "e5d10a24b66c3ea8f150e47dfdb0416ab7c3390e",
+        "rev": "8dc45382d5206bd292f9c2768b8058a8fd8311d9",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1714635257,
+        "lastModified": 1716509168,
-        "narHash": "sha256-4cPymbty65RvF1DWQfc+Bc8B233A1BWxJnNULJKQ1EY=",
+        "narHash": "sha256-4zSIhSRRIoEBwjbPm3YiGtbd8HDWzFxJjw5DYSDy1n8=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "63c3a29ca82437c87573e4c6919b09a24ea61b0f",
+        "rev": "bfb7a882678e518398ce9a31a881538679f6f092",
         "type": "github"
       },
       "original": {

ggml-impl.h

@@ -144,6 +144,10 @@ extern "C" {
 #endif
 #endif
 
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+#endif
+
 // 16-bit float
 // on Arm, we use __fp16
 // on x86, we use uint16_t

ggml-quants.c

@@ -3813,7 +3813,44 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         return;
     }
 #endif
-#if defined(__ARM_NEON)
+#if defined(__ARM_FEATURE_SVE)
+    const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
+    const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);
+
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+
+    assert(nb % 2 == 0); // TODO: handle odd nb
+
+    for (int i = 0; i < nb; i += 2) {
+        const block_q4_0 * restrict x0 = &x[i + 0];
+        const block_q4_0 * restrict x1 = &x[i + 1];
+        const block_q8_0 * restrict y0 = &y[i + 0];
+        const block_q8_0 * restrict y1 = &y[i + 1];
+
+        // load x
+        const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
+        const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+
+        // 4-bit -> 8-bit
+        const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04));
+        const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04));
+
+        // sub 8
+        const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
+        const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
+
+        // load y
+        const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+        const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+        // dot product
+        sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+#elif defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
@@ -5384,7 +5421,32 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         return;
     }
 #endif
-#if defined(__ARM_NEON)
+#if defined(__ARM_FEATURE_SVE)
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+
+    assert(nb % 2 == 0); // TODO: handle odd nb
+
+    for (int i = 0; i < nb; i += 2) {
+        const block_q8_0 * restrict x0 = &x[i + 0];
+        const block_q8_0 * restrict x1 = &x[i + 1];
+        const block_q8_0 * restrict y0 = &y[i + 0];
+        const block_q8_0 * restrict y1 = &y[i + 1];
+
+        // load x
+        const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
+        const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
+
+        // load y
+        const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+        const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+        sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+#elif defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 

ggml.c

@@ -22742,6 +22742,16 @@ int ggml_cpu_has_neon(void) {
 #endif
 }
 
+int ggml_cpu_has_sve(void) {
+#if defined(__ARM_FEATURE_SVE)
+    // TODO: Currently, SVE 256 bit is only supported.
+    GGML_ASSERT(svcntb() == QK8_0);
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_arm_fma(void) {
 #if defined(__ARM_FEATURE_FMA)
     return 1;

ggml.h

@@ -2404,6 +2404,7 @@ extern "C" {
     GGML_API int ggml_cpu_has_avx512_bf16(void);
     GGML_API int ggml_cpu_has_fma        (void);
     GGML_API int ggml_cpu_has_neon       (void);
+    GGML_API int ggml_cpu_has_sve        (void);
     GGML_API int ggml_cpu_has_arm_fma    (void);
     GGML_API int ggml_cpu_has_metal      (void);
     GGML_API int ggml_cpu_has_f16c       (void);

gguf-py/gguf/constants.py

@@ -139,6 +139,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R  = auto()
     DBRX       = auto()
     OLMO       = auto()
+    ARCTIC     = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -167,6 +168,7 @@ class MODEL_TENSOR(IntEnum):
     FFN_DOWN           = auto()
     FFN_UP             = auto()
     FFN_ACT            = auto()
+    FFN_NORM_EXP       = auto()
     FFN_GATE_EXP       = auto()
     FFN_DOWN_EXP       = auto()
     FFN_UP_EXP         = auto()
@@ -218,6 +220,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
+    MODEL_ARCH.ARCTIC:         "arctic",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -251,6 +254,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_DOWN_SHEXP:     "blk.{bid}.ffn_down_shexp",
     MODEL_TENSOR.FFN_UP_SHEXP:       "blk.{bid}.ffn_up_shexp",
     MODEL_TENSOR.FFN_ACT:            "blk.{bid}.ffn",
+    MODEL_TENSOR.FFN_NORM_EXP:       "blk.{bid}.ffn_norm_exps",
     MODEL_TENSOR.FFN_GATE_EXP:       "blk.{bid}.ffn_gate_exps",
     MODEL_TENSOR.FFN_DOWN_EXP:       "blk.{bid}.ffn_down_exps",
     MODEL_TENSOR.FFN_UP_EXP:         "blk.{bid}.ffn_up_exps",
@@ -732,6 +736,27 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.ARCTIC: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_NORM_EXP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
     # TODO
 }
 

gguf-py/gguf/gguf_reader.py

@@ -12,6 +12,8 @@ from typing import Any, Literal, NamedTuple, TypeVar, Union
 import numpy as np
 import numpy.typing as npt
 
+from .quants import quant_shape_to_byte_shape
+
 if __name__ == "__main__":
     import sys
     from pathlib import Path
@@ -251,6 +253,7 @@ class GGUFReader:
             tensor_names.add(tensor_name)
             ggml_type = GGMLQuantizationType(raw_dtype[0])
             n_elems = int(np.prod(dims))
+            np_dims = tuple(reversed(dims.tolist()))
             block_size, type_size = GGML_QUANT_SIZES[ggml_type]
             n_bytes = n_elems * type_size // block_size
             data_offs = int(start_offs + offset_tensor[0])
@@ -279,6 +282,7 @@ class GGUFReader:
             else:
                 item_count = n_bytes
                 item_type = np.uint8
+                np_dims = quant_shape_to_byte_shape(np_dims, ggml_type)
             tensors.append(ReaderTensor(
                 name = tensor_name,
                 tensor_type = ggml_type,
@@ -286,7 +290,7 @@ class GGUFReader:
                 n_elements = n_elems,
                 n_bytes = n_bytes,
                 data_offset = data_offs,
-                data = self._get(data_offs, item_type, item_count),
+                data = self._get(data_offs, item_type, item_count).reshape(np_dims),
                 field = field,
             ))
         self.tensors = tensors

gguf-py/gguf/gguf_writer.py

@@ -13,7 +13,6 @@ from string import ascii_letters, digits
 import numpy as np
 
 from .constants import (
-    GGML_QUANT_SIZES,
     GGUF_DEFAULT_ALIGNMENT,
     GGUF_MAGIC,
     GGUF_VERSION,
@@ -26,6 +25,8 @@ from .constants import (
     TokenType,
 )
 
+from .quants import quant_shape_from_byte_shape
+
 logger = logging.getLogger(__name__)
 
 
@@ -229,10 +230,7 @@ class GGUFWriter:
         else:
             dtype = raw_dtype
             if tensor_dtype == np.uint8:
-                block_size, type_size = GGML_QUANT_SIZES[raw_dtype]
+                tensor_shape = quant_shape_from_byte_shape(tensor_shape, raw_dtype)
-                if tensor_shape[-1] % type_size != 0:
-                    raise ValueError(f"Quantized tensor row size ({tensor_shape[-1]}) is not a multiple of {dtype.name} type size ({type_size})")
-                tensor_shape = tuple(tensor_shape[:-1]) + (tensor_shape[-1] // type_size * block_size,)
         n_dims = len(tensor_shape)
         self.ti_data += self._pack("I", n_dims)
         for i in range(n_dims):

gguf-py/gguf/quants.py

@@ -1,5 +1,5 @@
 from __future__ import annotations
-from typing import Callable
+from typing import Callable, Sequence
 
 from numpy.typing import DTypeLike
 
@@ -9,6 +9,20 @@ from .lazy import LazyNumpyTensor
 import numpy as np
 
 
+def quant_shape_to_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType):
+    block_size, type_size = GGML_QUANT_SIZES[quant_type]
+    if shape[-1] % block_size != 0:
+        raise ValueError(f"Quantized tensor row size ({shape[-1]}) is not a multiple of {quant_type.name} block size ({block_size})")
+    return (*shape[:-1], shape[-1] // block_size * type_size)
+
+
+def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType):
+    block_size, type_size = GGML_QUANT_SIZES[quant_type]
+    if shape[-1] % type_size != 0:
+        raise ValueError(f"Quantized tensor bytes per row ({shape[-1]}) is not a multiple of {quant_type.name} type size ({type_size})")
+    return (*shape[:-1], shape[-1] // type_size * block_size)
+
+
 # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
 def __compute_fp32_to_bf16(n: np.ndarray) -> np.ndarray:
     n = n.astype(np.float32, copy=False).view(np.int32)

gguf-py/gguf/tensor_mapping.py

@@ -244,6 +244,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
             "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
+            "model.layers.{bid}.residual_mlp.w3",                     # arctic
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -272,6 +273,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
             "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2
             "transformer.h.{bid}.mlp.linear_1",           # refact
+            "model.layers.{bid}.residual_mlp.w1",         # arctic
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
@@ -306,6 +308,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc2",                           # nomic-bert
             "model.layers.{bid}.mlp.c_proj",                          # starcoder2
             "encoder.layer.{bid}.mlp.wo",                             # jina-bert-v2
+            "model.layers.{bid}.residual_mlp.w2",                     # arctic
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -382,6 +385,18 @@ class TensorNameMap:
         ),
     }
 
+    # architecture-specific block mappings
+    arch_block_mappings_cfg: dict[MODEL_ARCH, dict[MODEL_TENSOR, tuple[str, ...]]] = {
+        MODEL_ARCH.ARCTIC: {
+            MODEL_TENSOR.FFN_NORM: (
+                "model.layers.{bid}.residual_layernorm",
+            ),
+            MODEL_TENSOR.FFN_NORM_EXP: (
+                "model.layers.{bid}.post_attention_layernorm",
+            ),
+        },
+    }
+
     mapping: dict[str, tuple[MODEL_TENSOR, str]]
 
     def __init__(self, arch: MODEL_ARCH, n_blocks: int):
@@ -393,12 +408,14 @@ class TensorNameMap:
             self.mapping[tensor_name] = (tensor, tensor_name)
             for key in keys:
                 self.mapping[key] = (tensor, tensor_name)
+        if arch in self.arch_block_mappings_cfg:
+            self.block_mappings_cfg.update(self.arch_block_mappings_cfg[arch])
         for bid in range(n_blocks):
             for tensor, keys in self.block_mappings_cfg.items():
                 if tensor not in MODEL_TENSORS[arch]:
                     continue
                 # TODO: make this configurable
-                n_experts = 60
+                n_experts = 128
                 for xid in range(n_experts):
                     tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
                     self.mapping[tensor_name] = (tensor, tensor_name)

gguf-py/scripts/gguf-new-metadata.py

@@ -118,9 +118,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
 
     for tensor in reader.tensors:
         total_bytes += tensor.n_bytes
-        # Dimensions are written in reverse order, so flip them first
+        writer.add_tensor_info(tensor.name, tensor.data.shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type)
-        shape = np.flipud(tensor.shape).tolist()
-        writer.add_tensor_info(tensor.name, shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type)
 
     bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True)
 

llama.cpp

@@ -103,7 +103,7 @@
 #endif
 
 #define LLAMA_MAX_NODES   8192
-#define LLAMA_MAX_EXPERTS 60
+#define LLAMA_MAX_EXPERTS 128
 
 //
 // logging
@@ -221,6 +221,7 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
+    LLM_ARCH_ARCTIC,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -257,6 +258,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
+    { LLM_ARCH_ARCTIC,          "arctic"       },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -459,6 +461,7 @@ enum llm_tensor {
     LLM_TENSOR_FFN_DOWN_EXP,  // split experts for backward compatibility
     LLM_TENSOR_FFN_GATE_EXP,
     LLM_TENSOR_FFN_UP_EXP,
+    LLM_TENSOR_FFN_NORM_EXPS,
     LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
     LLM_TENSOR_FFN_GATE_EXPS,
     LLM_TENSOR_FFN_UP_EXPS,
@@ -1036,6 +1039,28 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_ARCTIC,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { 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_NORM_EXPS,   "blk.%d.ffn_norm_exps" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1736,6 +1761,7 @@ enum e_model {
     MODEL_8x7B,
     MODEL_8x22B,
     MODEL_16x12B,
+    MODEL_10B_128x3_66B,
 };
 
 static const size_t kiB = 1024;
@@ -1911,6 +1937,7 @@ struct llama_layer {
     struct ggml_tensor * ffn_norm_b;
     struct ggml_tensor * layer_out_norm;
     struct ggml_tensor * layer_out_norm_b;
+    struct ggml_tensor * ffn_norm_exps;
 
     // ff
     struct ggml_tensor * ffn_gate; // w1
@@ -3785,47 +3812,48 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
 
 static const char * llama_model_type_name(e_model type) {
     switch (type) {
-        case MODEL_14M:    return "14M";
+        case MODEL_14M:           return "14M";
-        case MODEL_17M:    return "17M";
+        case MODEL_17M:           return "17M";
-        case MODEL_22M:    return "22M";
+        case MODEL_22M:           return "22M";
-        case MODEL_33M:    return "33M";
+        case MODEL_33M:           return "33M";
-        case MODEL_70M:    return "70M";
+        case MODEL_70M:           return "70M";
-        case MODEL_109M:   return "109M";
+        case MODEL_109M:          return "109M";
-        case MODEL_137M:   return "137M";
+        case MODEL_137M:          return "137M";
-        case MODEL_160M:   return "160M";
+        case MODEL_160M:          return "160M";
-        case MODEL_335M:   return "335M";
+        case MODEL_335M:          return "335M";
-        case MODEL_410M:   return "410M";
+        case MODEL_410M:          return "410M";
-        case MODEL_0_5B:   return "0.5B";
+        case MODEL_0_5B:          return "0.5B";
-        case MODEL_1B:     return "1B";
+        case MODEL_1B:            return "1B";
-        case MODEL_1_4B:   return "1.4B";
+        case MODEL_1_4B:          return "1.4B";
-        case MODEL_2B:     return "2B";
+        case MODEL_2B:            return "2B";
-        case MODEL_2_8B:   return "2.8B";
+        case MODEL_2_8B:          return "2.8B";
-        case MODEL_3B:     return "3B";
+        case MODEL_3B:            return "3B";
-        case MODEL_4B:     return "4B";
+        case MODEL_4B:            return "4B";
-        case MODEL_6_9B:   return "6.9B";
+        case MODEL_6_9B:          return "6.9B";
-        case MODEL_7B:     return "7B";
+        case MODEL_7B:            return "7B";
-        case MODEL_8B:     return "8B";
+        case MODEL_8B:            return "8B";
-        case MODEL_12B:    return "12B";
+        case MODEL_12B:           return "12B";
-        case MODEL_13B:    return "13B";
+        case MODEL_13B:           return "13B";
-        case MODEL_14B:    return "14B";
+        case MODEL_14B:           return "14B";
-        case MODEL_15B:    return "15B";
+        case MODEL_15B:           return "15B";
-        case MODEL_20B:    return "20B";
+        case MODEL_20B:           return "20B";
-        case MODEL_30B:    return "30B";
+        case MODEL_30B:           return "30B";
-        case MODEL_34B:    return "34B";
+        case MODEL_34B:           return "34B";
-        case MODEL_35B:    return "35B";
+        case MODEL_35B:           return "35B";
-        case MODEL_40B:    return "40B";
+        case MODEL_40B:           return "40B";
-        case MODEL_65B:    return "65B";
+        case MODEL_65B:           return "65B";
-        case MODEL_70B:    return "70B";
+        case MODEL_70B:           return "70B";
-        case MODEL_314B:   return "314B";
+        case MODEL_314B:          return "314B";
-        case MODEL_SMALL:  return "0.1B";
+        case MODEL_SMALL:         return "0.1B";
-        case MODEL_MEDIUM: return "0.4B";
+        case MODEL_MEDIUM:        return "0.4B";
-        case MODEL_LARGE:  return "0.8B";
+        case MODEL_LARGE:         return "0.8B";
-        case MODEL_XL:     return "1.5B";
+        case MODEL_XL:            return "1.5B";
-        case MODEL_A2_7B:  return "A2.7B";
+        case MODEL_A2_7B:         return "A2.7B";
-        case MODEL_8x7B:   return "8x7B";
+        case MODEL_8x7B:          return "8x7B";
-        case MODEL_8x22B:  return "8x22B";
+        case MODEL_8x22B:         return "8x22B";
-        case MODEL_16x12B: return "16x12B";
+        case MODEL_16x12B:        return "16x12B";
-        default:           return "?B";
+        case MODEL_10B_128x3_66B: return "10B+128x3.66B";
+        default:                  return "?B";
     }
 }
 
@@ -4347,6 +4375,19 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_ARCTIC:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                if (hparams.n_expert == 128) {
+                    switch (hparams.n_layer) {
+                        case 35: model.type = e_model::MODEL_10B_128x3_66B; break;
+                        default: model.type = e_model::MODEL_UNKNOWN;
+                    }
+                } else {
+                    model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -4556,6 +4597,9 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "dbrx") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DBRX;
+            } else if (
+                tokenizer_pre == "smaug-bpe") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMAUG;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -6133,6 +6177,46 @@ static bool llm_load_tensors(
                         layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff});
                     }
                 } break;
+            case LLM_ARCH_ARCTIC:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_embd});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_embd, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
+                        layer.ffn_norm_exps = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM_EXPS, "weight", i), {n_embd});
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, false);
+                        layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
+                        layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -10794,6 +10878,140 @@ struct llm_build_context {
 
         return gf;
     }
+
+    struct ggml_cgraph * build_arctic() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, cur,
+                    model.layers[il].ffn_up,   NULL,
+                    model.layers[il].ffn_gate, NULL,
+                    model.layers[il].ffn_down, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            struct ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
+            cb(ffn_out, "ffn_out", il);
+
+            // MoE
+            cur = llm_build_norm(ctx0, inpSA, hparams,
+                    model.layers[il].ffn_norm_exps, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm_exps", il);
+
+            cur = llm_build_moe_ffn(ctx0, cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    cb, il);
+            cb(cur, "ffn_moe_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_out);
+            cb(cur, "ffn_out", il);
+
+            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+            if (layer_dir != nullptr) {
+                cur = ggml_add(ctx0, cur, layer_dir);
+            }
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
 };
 
 static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@@ -11008,6 +11226,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_gptneox();
             } break;
+        case LLM_ARCH_ARCTIC:
+            {
+                result = llm.build_arctic();
+            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -12297,6 +12519,7 @@ struct llm_tokenizer_bpe {
                         });
                         break;
                     case LLAMA_VOCAB_PRE_TYPE_DBRX:
+                    case LLAMA_VOCAB_PRE_TYPE_SMAUG:
                         word_collection = unicode_regex_split(text, {
                             // same as llama3
                             "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -16019,6 +16242,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_XVERSE:
         case LLM_ARCH_COMMAND_R:
         case LLM_ARCH_OLMO:
+        case LLM_ARCH_ARCTIC:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
@@ -17645,6 +17869,10 @@ bool llama_token_is_eog(const struct llama_model * model, llama_token token) {
     );
 }
 
+bool llama_token_is_control(const struct llama_model * model, llama_token token) {
+    return llama_is_control_token(model->vocab, token);
+}
+
 llama_token llama_token_bos(const struct llama_model * model) {
     return model->vocab.special_bos_id;
 }
@@ -18121,6 +18349,7 @@ const char * llama_print_system_info(void) {
     s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | ";
     s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
     s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
+    s += "SVE = "         + std::to_string(ggml_cpu_has_sve())         + " | ";
     s += "ARM_FMA = "     + std::to_string(ggml_cpu_has_arm_fma())     + " | ";
     s += "F16C = "        + std::to_string(ggml_cpu_has_f16c())        + " | ";
     s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";

llama.h

@@ -85,6 +85,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,
         LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
         LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
+        LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
     };
 
     // note: these values should be synchronized with ggml_rope
@@ -823,6 +824,9 @@ extern "C" {
     // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
     LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
 
+    // Identify if Token Id is a control token or a render-able token
+    LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token);
+
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence