feat: automatically adapt to current free VRAM state

.gitignore

@@ -11,6 +11,7 @@ node_modules
 /.eslintcache
 /.vitepress/.cache
 /test/.models
+/test/temp
 /coverage
 
 /llama/compile_commands.json
@@ -20,6 +21,8 @@ node_modules
 /llama/lastBuild.json
 /llama/gitRelease.bundle
 /llama/.temp
+/llama/.idea
+/llama/cmake-build-debug
 /llama/localBuilds
 /llama/Release
 /llama/Debug

.vitepress/config.ts

@@ -108,6 +108,22 @@ export default defineConfig({
             pageData.frontmatter.editLink = false;
             pageData.frontmatter.lastUpdated = false;
         }
+
+        let canonicalUrl = hostname + pageData.relativePath;
+        if (canonicalUrl.endsWith("/index.html"))
+            canonicalUrl = canonicalUrl.slice(0, -"index.html".length);
+        if (canonicalUrl.endsWith("/index.md"))
+            canonicalUrl = canonicalUrl.slice(0, -"index.md".length);
+        else if (canonicalUrl.endsWith(".html"))
+            canonicalUrl = canonicalUrl.slice(0, -".html".length);
+        else if (canonicalUrl.endsWith(".md"))
+            canonicalUrl = canonicalUrl.slice(0, -".md".length);
+
+        pageData.frontmatter.head ??= [];
+        pageData.frontmatter.head.push([
+            "link",
+            {rel: "canonical", href: canonicalUrl}
+        ])
     },
     themeConfig: {
         editLink: {
@@ -183,7 +199,16 @@ export default defineConfig({
                     {text: "Download", link: "/download"},
                     {text: "Complete", link: "/complete"},
                     {text: "Infill", link: "/infill"},
-                    {text: "Inspect", link: "/inspect"},
+                    {
+                        text: "Inspect",
+                        link: "/inspect",
+                        collapsed: true,
+                        items: [
+                            {text: "GPU", link: "/inspect/gpu"},
+                            {text: "GGUF", link: "/inspect/gguf"},
+                            {text: "Measure", link: "/inspect/measure"},
+                        ]
+                    },
                     {text: "Build", link: "/build"},
                     {text: "Clear", link: "/clear"}
                 ]
@@ -302,7 +327,7 @@ function orderClasses(sidebar: typeof typedocSidebar) {
         items: []
     };
     (classes.items as DefaultTheme.SidebarItem[]).push(LlamaTextGroup);
-    const LlamaTextGroupItemsOrder = ["SpecialToken", "BuiltinSpecialToken"];
+    const LlamaTextGroupItemsOrder = ["SpecialTokensText", "SpecialToken"];
 
     groupItems(
         classes.items,
@@ -327,7 +352,7 @@ function orderTypes(sidebar: typeof typedocSidebar) {
         (item) => (
             item.text === "BatchItem" ||
             item.text === "CustomBatchingDispatchSchedule" ||
-            item.text === "CustomBatchingPrioritizeStrategy" ||
+            item.text === "CustomBatchingPrioritizationStrategy" ||
             item.text === "PrioritizedBatchItem"
         ),
         {collapsed: false}

.vitepress/utils/getCommandHtmlDoc.ts

@@ -4,13 +4,54 @@ import {cliBinName, npxRunPrefix} from "../../src/config.js";
 import {buildHtmlTable} from "./buildHtmlTable.js";
 import {buildHtmlHeading} from "./buildHtmlHeading.js";
 
-export async function getCommandHtmlDoc(command: CommandModule<any, any>, cliName: string = cliBinName) {
-    const title = cliName + " " + (command.command ?? "");
+export async function getCommandHtmlDoc(command: CommandModule<any, any>, {
+    cliName = cliBinName,
+    parentCommand,
+    subCommandsParentPageLink
+}: {
+    cliName?: string,
+    parentCommand?: CommandModule<any, any>,
+    subCommandsParentPageLink?: string
+} = {}) {
+    const currentCommandCliCommand = resolveCommandCliCommand(command);
+    const resolvedParentCommandCliCommand = resolveCommandCliCommand(parentCommand);
+    const title = cliName + " " + (resolvedParentCommandCliCommand ?? "<command>").replace("<command>", currentCommandCliCommand ?? "");
     const description = command.describe ?? "";
-    const optionGroups = await getOptionsGroupFromCommand(command);
+    const {subCommands, optionGroups} = await parseCommandDefinition(command);
 
     let res = "";
 
+    if (subCommands.length > 0) {
+        res += buildHtmlHeading("h2", htmlEscape("Commands"), "commands");
+
+        res += buildHtmlTable(
+            [
+                "Command",
+                "Description"
+            ].map(htmlEscape),
+            subCommands
+                .map((subCommand) => {
+                    if (subCommand.command == null || subCommand.describe === false)
+                        return null;
+
+                    const resolvedCommandCliCommand = resolveCommandCliCommand(subCommand) ?? "";
+                    const commandPageLink = resolveCommandPageLink(subCommand);
+
+                    let cliCommand = resolvedCommandCliCommand;
+                    cliCommand = (currentCommandCliCommand ?? "<command>").replace("<command>", cliCommand);
+
+                    if (parentCommand != null)
+                        cliCommand = (resolvedParentCommandCliCommand ?? "<command>").replace("<command>", cliCommand);
+
+                    return [
+                        `<a href="${subCommandsParentPageLink != null ? (subCommandsParentPageLink + "/") : ""}${commandPageLink}"><code>` + htmlEscape(cliName + " " + cliCommand) + "</code></a>",
+                        htmlEscape(String(subCommand.describe ?? ""))
+                    ];
+                })
+                .filter((row): row is string[] => row != null)
+        );
+    }
+
     if (optionGroups.length !== 0) {
         res += buildHtmlHeading("h2", htmlEscape("Options"), "options");
 
@@ -37,7 +78,10 @@ export async function getCommandHtmlDoc(command: CommandModule<any, any>, cliNam
 }
 
 
-async function getOptionsGroupFromCommand(command: CommandModule<any, any>): Promise<OptionsGroup[]> {
+async function parseCommandDefinition(command: CommandModule<any, any>): Promise<{
+    subCommands: CommandModule<any, any>[],
+    optionGroups: OptionsGroup[]
+}> {
     const yargsStub = getYargsStub();
     function getYargsStub() {
         function option(name: string, option: Options) {
@@ -57,10 +101,16 @@ async function getOptionsGroupFromCommand(command: CommandModule<any, any>): Pro
             return yargsStub;
         }
 
-        return {option};
+        function command(subCommand: CommandModule<any, any>) {
+            subCommands.push(subCommand);
+            return yargsStub;
+        }
+
+        return {option, command};
     }
 
     const options: Record<string, {name: string, option: Options}[]> = {};
+    const subCommands: CommandModule<any, any>[] = [];
     const groups: string[] = [];
 
     if (command.builder instanceof Function)
@@ -97,10 +147,13 @@ async function getOptionsGroupFromCommand(command: CommandModule<any, any>): Pro
         return 0;
     });
 
-    return groups.map((group) => ({
-        name: normalizeGroupName(group),
-        options: options[group]!
-    }));
+    return {
+        subCommands,
+        optionGroups: groups.map((group) => ({
+            name: normalizeGroupName(group),
+            options: options[group]!
+        }))
+    };
 }
 
 function normalizeGroupName(groupName: string): string {
@@ -156,8 +209,12 @@ function renderOptionsGroupOptionsTable(options: {name: string, option: Options}
 
         let optionDescription: string[] = option.description != null ? [htmlEscape(option.description)] : [];
 
-        if (option.default != null) {
-            optionDescription.push(`(${htmlEscape("default: ")}<code>${htmlEscape(option.default)}</code>)`);
+        const hasDefaultDescription = option.defaultDescription != null && option.defaultDescription.trim().length > 0;
+        if (option.default != null || hasDefaultDescription) {
+            if (hasDefaultDescription && option.defaultDescription != null)
+                optionDescription.push(`<span style="opacity: 0.72">(${htmlEscape("default: ")}${htmlEscape(option.defaultDescription.trim())})</span>`);
+            else
+                optionDescription.push(`<span style="opacity: 0.72">(${htmlEscape("default: ")}<code>${htmlEscape(option.default)}</code>)</span>`);
         }
 
         if (option.type != null) {
@@ -184,6 +241,19 @@ function renderOptionsGroupOptionsTable(options: {name: string, option: Options}
     return buildHtmlTable(tableHeaders, tableRows);
 }
 
+function resolveCommandCliCommand(command?: CommandModule<any, any>) {
+    if (command == null)
+        return undefined;
+
+    return command.command instanceof Array
+        ? command.command[0]
+        : command.command;
+}
+
+function resolveCommandPageLink(command: CommandModule<any, any>) {
+    return resolveCommandCliCommand(command)?.split(" ")?.[0];
+}
+
 type OptionsGroup = {
     name: string,
     options: Array<{

README.md

@@ -20,7 +20,7 @@
 ## Features
 * Run a text generation model locally on your machine
 * Metal, CUDA and Vulkan support
-* Pre-built binaries are provided, with a fallback to building from source without `node-gyp` or Python
+* Pre-built binaries are provided, with a fallback to building from source _**without**_ `node-gyp` or Python
 * Chat with a model using a chat wrapper
 * Use the CLI to chat with a model without writing any code
 * Up-to-date with the latest version of `llama.cpp`. Download and compile the latest release with a single CLI command.

docs/guide/cli/cli.data.ts

@@ -4,14 +4,17 @@ import {BuildCommand} from "../../../src/cli/commands/BuildCommand.js";
 import {ChatCommand} from "../../../src/cli/commands/ChatCommand.js";
 import {CompleteCommand} from "../../../src/cli/commands/CompleteCommand.js";
 import {InfillCommand} from "../../../src/cli/commands/InfillCommand.js";
-import {InspectCommand} from "../../../src/cli/commands/InspectCommand.js";
+import {InspectCommand} from "../../../src/cli/commands/inspect/InspectCommand.js";
+import {InspectGpuCommand} from "../../../src/cli/commands/inspect/commands/InspectGpuCommand.js";
+import {InspectGgufCommand} from "../../../src/cli/commands/inspect/commands/InspectGgufCommand.js";
 import {DownloadCommand} from "../../../src/cli/commands/DownloadCommand.js";
 import {ClearCommand} from "../../../src/cli/commands/ClearCommand.js";
 import {htmlEscape} from "../../../.vitepress/utils/htmlEscape.js";
 import {cliBinName, npxRunPrefix} from "../../../src/config.js";
 import {buildHtmlHeading} from "../../../.vitepress/utils/buildHtmlHeading.js";
 import {buildHtmlTable} from "../../../.vitepress/utils/buildHtmlTable.js";
 import {setIsInDocumentationMode} from "../../../src/state.js";
+import {InspectMeasureCommand} from "../../../src/cli/commands/inspect/commands/InspectMeasureCommand.js";
 
 export default {
     async load() {
@@ -31,7 +34,20 @@ export default {
             chat: await getCommandHtmlDoc(ChatCommand),
             complete: await getCommandHtmlDoc(CompleteCommand),
             infill: await getCommandHtmlDoc(InfillCommand),
-            inspect: await getCommandHtmlDoc(InspectCommand),
+            inspect: {
+                index: await getCommandHtmlDoc(InspectCommand, {
+                    subCommandsParentPageLink: "inspect"
+                }),
+                gpu: await getCommandHtmlDoc(InspectGpuCommand, {
+                    parentCommand: InspectCommand
+                }),
+                gguf: await getCommandHtmlDoc(InspectGgufCommand, {
+                    parentCommand: InspectCommand
+                }),
+                measure: await getCommandHtmlDoc(InspectMeasureCommand, {
+                    parentCommand: InspectCommand
+                })
+            },
             download: await getCommandHtmlDoc(DownloadCommand),
             build: await getCommandHtmlDoc(BuildCommand),
             clear: await getCommandHtmlDoc(ClearCommand)

docs/guide/cli/inspect.md

@@ -5,7 +5,7 @@ outline: deep
 
 <script setup lang="ts">
 import {data as docs} from "./cli.data.js";
-const commandDoc = docs.inspect;
+const commandDoc = docs.inspect.index;
 </script>
 
 {{commandDoc.description}}

docs/guide/cli/inspect/gguf.md

@@ -0,0 +1,17 @@
+---
+outline: deep
+---
+# `inspect gguf` command
+
+<script setup lang="ts">
+import {data as docs} from "../cli.data.js";
+const commandDoc = docs.inspect.gguf;
+</script>
+
+{{commandDoc.description}}
+
+## Usage
+```shell-vue
+{{commandDoc.usage}}
+```
+<div v-html="commandDoc.options"></div>

docs/guide/cli/inspect/gpu.md

@@ -0,0 +1,17 @@
+---
+outline: deep
+---
+# `inspect gpu` command
+
+<script setup lang="ts">
+import {data as docs} from "../cli.data.js";
+const commandDoc = docs.inspect.gpu;
+</script>
+
+{{commandDoc.description}}
+
+## Usage
+```shell-vue
+{{commandDoc.usage}}
+```
+<div v-html="commandDoc.options"></div>

docs/guide/cli/inspect/measure.md

@@ -0,0 +1,17 @@
+---
+outline: deep
+---
+# `inspect measure` command
+
+<script setup lang="ts">
+import {data as docs} from "../cli.data.js";
+const commandDoc = docs.inspect.measure;
+</script>
+
+{{commandDoc.description}}
+
+## Usage
+```shell-vue
+{{commandDoc.usage}}
+```
+<div v-html="commandDoc.options"></div>

llama/CMakeLists.txt

@@ -29,19 +29,19 @@ include_directories("gpuInfo")
 include_directories("llama.cpp")
 include_directories("./llama.cpp/common")
 
-if (LLAMA_CUBLAS)
+if (LLAMA_CUDA)
     cmake_minimum_required(VERSION 3.17)
 
     find_package(CUDAToolkit)
     if (CUDAToolkit_FOUND)
-        message(STATUS "Using cuBLAS for GPU info")
+        message(STATUS "Using CUDA for GPU info")
 
         enable_language(CUDA)
 
         set(GPU_INFO_HEADERS ${GPU_INFO_HEADERS} gpuInfo/cuda-gpu-info.h)
         set(GPU_INFO_SOURCES ${GPU_INFO_SOURCES} gpuInfo/cuda-gpu-info.cu)
 
-        add_compile_definitions(GPU_INFO_USE_CUBLAS)
+        add_compile_definitions(GPU_INFO_USE_CUDA)
 
         if (LLAMA_STATIC)
             set(LLAMA_EXTRA_LIBS ${GPU_INFO_EXTRA_LIBS} CUDA::cudart_static)
@@ -60,7 +60,7 @@ if (LLAMA_CUBLAS)
             endif()
         endif()
     else()
-        message(FATAL_ERROR "cuBLAS was not found")
+        message(FATAL_ERROR "CUDA was not found")
     endif()
 endif()
 
@@ -100,7 +100,7 @@ if (LLAMA_HIPBLAS)
 
     if (${hipblas_FOUND} AND ${hip_FOUND})
         message(STATUS "Using HIP and hipBLAS for GPU info")
-        add_compile_definitions(GPU_INFO_USE_HIPBLAS GPU_INFO_USE_CUBLAS)
+        add_compile_definitions(GPU_INFO_USE_HIPBLAS GPU_INFO_USE_CUDA)
         add_library(gpu-info-rocm OBJECT gpuInfo/cuda-gpu-info.cu gpuInfo/cuda-gpu-info.h)
         set_source_files_properties(gpuInfo/cuda-gpu-info.cu PROPERTIES LANGUAGE CXX)
         target_link_libraries(gpu-info-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)