metal : improve clarity (minor) (#10171)

* ggml : add ggml_flash_attn_ext_get_prec

* metal : use F16 precision in FA kernels

ggml-ci

* metal : minor clean-up

* metal : compile-guard bf16 FA kernels

ggml-ci

* build : remove obsolete compile flag [no ci]

* metal : prevent int overflows [no ci]

* cuda : disable BF16 FA

ggml-ci

* metal : fix BF16 requirement for FA kernels

ggml-ci

* make : clean-up [no ci]

.editorconfig

@@ -24,6 +24,16 @@ insert_final_newline = unset
 [examples/server/public/*]
 indent_size = 2
 
+[examples/server/public/deps_*]
+trim_trailing_whitespace = unset
+indent_style = unset
+indent_size = unset
+
+[examples/server/deps_*]
+trim_trailing_whitespace = unset
+indent_style = unset
+indent_size = unset
+
 [examples/llama.swiftui/llama.swiftui.xcodeproj/*]
 indent_style = tab
 

Makefile

@@ -1455,22 +1455,13 @@ llama-server: \
 	examples/server/server.cpp \
 	examples/server/utils.hpp \
 	examples/server/httplib.h \
-	examples/server/colorthemes.css.hpp \
-	examples/server/style.css.hpp \
-	examples/server/theme-beeninorder.css.hpp \
-	examples/server/theme-ketivah.css.hpp \
-	examples/server/theme-mangotango.css.hpp \
-	examples/server/theme-playground.css.hpp \
-	examples/server/theme-polarnight.css.hpp \
-	examples/server/theme-snowstorm.css.hpp \
 	examples/server/index.html.hpp \
-	examples/server/index-new.html.hpp \
-	examples/server/index.js.hpp \
 	examples/server/completion.js.hpp \
-	examples/server/system-prompts.js.hpp \
-	examples/server/prompt-formats.js.hpp \
-	examples/server/json-schema-to-grammar.mjs.hpp \
 	examples/server/loading.html.hpp \
+	examples/server/deps_daisyui.min.css.hpp \
+	examples/server/deps_markdown-it.js.hpp \
+	examples/server/deps_tailwindcss.js.hpp \
+	examples/server/deps_vue.esm-browser.js.hpp \
 	common/json.hpp \
 	common/stb_image.h \
 	$(OBJ_ALL)

Package.swift

@@ -61,13 +61,15 @@ let package = Package(
             name: "llama",
             path: ".",
             exclude: [
+               "build",
                "cmake",
                "examples",
                "scripts",
                "models",
                "tests",
                "CMakeLists.txt",
-               "Makefile"
+               "Makefile",
+               "ggml/src/ggml-metal-embed.metal"
             ],
             sources: sources,
             resources: resources,

examples/llama-bench/llama-bench.cpp

@@ -256,6 +256,9 @@ static ggml_type ggml_type_from_name(const std::string & s) {
     if (s == "f16") {
         return GGML_TYPE_F16;
     }
+    if (s == "bf16") {
+        return GGML_TYPE_BF16;
+    }
     if (s == "q8_0") {
         return GGML_TYPE_Q8_0;
     }

examples/server/CMakeLists.txt

@@ -15,22 +15,13 @@ set(TARGET_SRCS
     httplib.h
 )
 set(PUBLIC_ASSETS
-    colorthemes.css
-    style.css
-    theme-beeninorder.css
-    theme-ketivah.css
-    theme-mangotango.css
-    theme-playground.css
-    theme-polarnight.css
-    theme-snowstorm.css
     index.html
-    index-new.html
-    index.js
     completion.js
-    system-prompts.js
-    prompt-formats.js
-    json-schema-to-grammar.mjs
     loading.html
+    deps_daisyui.min.css
+    deps_markdown-it.js
+    deps_tailwindcss.js
+    deps_vue.esm-browser.js
 )
 
 foreach(asset ${PUBLIC_ASSETS})

examples/server/README.md

@@ -928,6 +928,16 @@ Apart from error types supported by OAI, we also have custom types that are spec
 }
 ```
 
+### Legacy completion web UI
+
+A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggerganov/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy`
+
+For example:
+
+```sh
+./llama-server -m my_model.gguf -c 8192 --path ./examples/server/public_legacy
+```
+
 ### Extending or building alternative Web Front End
 
 You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method.

examples/server/chat.mjs

@@ -1,7 +1,7 @@
 import * as readline from 'node:readline'
 import { stdin, stdout } from 'node:process'
 import { readFileSync } from 'node:fs'
-import { SchemaConverter }  from './public/json-schema-to-grammar.mjs'
+import { SchemaConverter }  from './public_legacy/json-schema-to-grammar.mjs'
 
 const args = process.argv.slice(2);
 const grammarJsonSchemaFile = args.find(

examples/server/deps.sh

@@ -6,5 +6,20 @@ DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 PUBLIC=$DIR/public
 
 echo "download js bundle files"
-curl https://npm.reversehttp.com/@preact/signals-core,@preact/signals,htm/preact,preact,preact/hooks > $PUBLIC/index.js
-echo >> $PUBLIC/index.js # add newline
+
+# Note for contributors: Always pin to a specific version "maj.min.patch" to avoid breaking the CI
+
+curl -L https://cdn.tailwindcss.com/3.4.14 > $PUBLIC/deps_tailwindcss.js
+echo >> $PUBLIC/deps_tailwindcss.js # add newline
+
+curl -L https://cdnjs.cloudflare.com/ajax/libs/daisyui/4.12.14/styled.min.css > $PUBLIC/deps_daisyui.min.css
+curl -L https://cdnjs.cloudflare.com/ajax/libs/daisyui/4.12.14/themes.min.css >> $PUBLIC/deps_daisyui.min.css
+echo >> $PUBLIC/deps_daisyui.min.css # add newline
+
+curl -L https://unpkg.com/vue@3.5.12/dist/vue.esm-browser.js > $PUBLIC/deps_vue.esm-browser.js
+echo >> $PUBLIC/deps_vue.esm-browser.js # add newline
+
+curl -L https://cdnjs.cloudflare.com/ajax/libs/markdown-it/13.0.2/markdown-it.js > $PUBLIC/deps_markdown-it.js
+echo >> $PUBLIC/deps_markdown-it.js # add newline
+
+ls -lah $PUBLIC

examples/server/public/completion.js

@@ -1,12 +1,16 @@
 const paramDefaults = {
   stream: true,
-  n_predict: 500,
   temperature: 0.2,
-  stop: ["</s>"]
 };
 
 let generation_settings = null;
 
+export class CompletionError extends Error {
+  constructor(message, name, data) {
+    super(message);
+    this.name = name;
+  }
+};
 
 // Completes the prompt as a generator. Recommended for most use cases.
 //
@@ -29,7 +33,7 @@ export async function* llama(prompt, params = {}, config = {}) {
 
   const completionParams = { ...paramDefaults, ...params, prompt };
 
-  const response = await fetch(`${api_url}/completion`, {
+  const response = await fetch(`${api_url}${config.endpoint || '/completion'}`, {
     method: 'POST',
     body: JSON.stringify(completionParams),
     headers: {
@@ -41,6 +45,18 @@ export async function* llama(prompt, params = {}, config = {}) {
     signal: controller.signal,
   });
 
+  const status = response.status;
+  if (status !== 200) {
+    try {
+      const body = await response.json();
+      if (body && body.error && body.error.message) {
+        throw new CompletionError(body.error.message, 'ServerError');
+      }
+    } catch (err) {
+      throw new CompletionError(err.message, 'ServerError');
+    }
+  }
+
   const reader = response.body.getReader();
   const decoder = new TextDecoder();
 
@@ -78,7 +94,12 @@ export async function* llama(prompt, params = {}, config = {}) {
       for (const line of lines) {
         const match = regex.exec(line);
         if (match) {
-          result[match[1]] = match[2]
+          result[match[1]] = match[2];
+          if (result.data === '[DONE]') {
+            cont = false;
+            break;
+          }
+
           // since we know this is llama.cpp, let's just decode the json in data
           if (result.data) {
             result.data = JSON.parse(result.data);

examples/server/public_legacy/completion.js

@@ -0,0 +1,209 @@
+const paramDefaults = {
+  stream: true,
+  n_predict: 500,
+  temperature: 0.2,
+  stop: ["</s>"]
+};
+
+let generation_settings = null;
+
+
+// Completes the prompt as a generator. Recommended for most use cases.
+//
+// Example:
+//
+//    import { llama } from '/completion.js'
+//
+//    const request = llama("Tell me a joke", {n_predict: 800})
+//    for await (const chunk of request) {
+//      document.write(chunk.data.content)
+//    }
+//
+export async function* llama(prompt, params = {}, config = {}) {
+  let controller = config.controller;
+  const api_url = config.api_url?.replace(/\/+$/, '') || "";
+
+  if (!controller) {
+    controller = new AbortController();
+  }
+
+  const completionParams = { ...paramDefaults, ...params, prompt };
+
+  const response = await fetch(`${api_url}${config.endpoint || '/completion'}`, {
+    method: 'POST',
+    body: JSON.stringify(completionParams),
+    headers: {
+      'Connection': 'keep-alive',
+      'Content-Type': 'application/json',
+      'Accept': 'text/event-stream',
+      ...(params.api_key ? {'Authorization': `Bearer ${params.api_key}`} : {})
+    },
+    signal: controller.signal,
+  });
+
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+
+  let content = "";
+  let leftover = ""; // Buffer for partially read lines
+
+  try {
+    let cont = true;
+
+    while (cont) {
+      const result = await reader.read();
+      if (result.done) {
+        break;
+      }
+
+      // Add any leftover data to the current chunk of data
+      const text = leftover + decoder.decode(result.value);
+
+      // Check if the last character is a line break
+      const endsWithLineBreak = text.endsWith('\n');
+
+      // Split the text into lines
+      let lines = text.split('\n');
+
+      // If the text doesn't end with a line break, then the last line is incomplete
+      // Store it in leftover to be added to the next chunk of data
+      if (!endsWithLineBreak) {
+        leftover = lines.pop();
+      } else {
+        leftover = ""; // Reset leftover if we have a line break at the end
+      }
+
+      // Parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm;
+      for (const line of lines) {
+        const match = regex.exec(line);
+        if (match) {
+          result[match[1]] = match[2];
+          if (result.data === '[DONE]') {
+            cont = false;
+            break;
+          }
+
+          // since we know this is llama.cpp, let's just decode the json in data
+          if (result.data) {
+            result.data = JSON.parse(result.data);
+            content += result.data.content;
+
+            // yield
+            yield result;
+
+            // if we got a stop token from server, we will break here
+            if (result.data.stop) {
+              if (result.data.generation_settings) {
+                generation_settings = result.data.generation_settings;
+              }
+              cont = false;
+              break;
+            }
+          }
+          if (result.error) {
+            try {
+              result.error = JSON.parse(result.error);
+              if (result.error.message.includes('slot unavailable')) {
+                // Throw an error to be caught by upstream callers
+                throw new Error('slot unavailable');
+              } else {
+                console.error(`llama.cpp error [${result.error.code} - ${result.error.type}]: ${result.error.message}`);
+              }
+            } catch(e) {
+              console.error(`llama.cpp error ${result.error}`)
+            }
+          }
+        }
+      }
+    }
+  } catch (e) {
+    if (e.name !== 'AbortError') {
+      console.error("llama error: ", e);
+    }
+    throw e;
+  }
+  finally {
+    controller.abort();
+  }
+
+  return content;
+}
+
+// Call llama, return an event target that you can subscribe to
+//
+// Example:
+//
+//    import { llamaEventTarget } from '/completion.js'
+//
+//    const conn = llamaEventTarget(prompt)
+//    conn.addEventListener("message", (chunk) => {
+//      document.write(chunk.detail.content)
+//    })
+//
+export const llamaEventTarget = (prompt, params = {}, config = {}) => {
+  const eventTarget = new EventTarget();
+  (async () => {
+    let content = "";
+    for await (const chunk of llama(prompt, params, config)) {
+      if (chunk.data) {
+        content += chunk.data.content;
+        eventTarget.dispatchEvent(new CustomEvent("message", { detail: chunk.data }));
+      }
+      if (chunk.data.generation_settings) {
+        eventTarget.dispatchEvent(new CustomEvent("generation_settings", { detail: chunk.data.generation_settings }));
+      }
+      if (chunk.data.timings) {
+        eventTarget.dispatchEvent(new CustomEvent("timings", { detail: chunk.data.timings }));
+      }
+    }
+    eventTarget.dispatchEvent(new CustomEvent("done", { detail: { content } }));
+  })();
+  return eventTarget;
+}
+
+// Call llama, return a promise that resolves to the completed text. This does not support streaming
+//
+// Example:
+//
+//     llamaPromise(prompt).then((content) => {
+//       document.write(content)
+//     })
+//
+//     or
+//
+//     const content = await llamaPromise(prompt)
+//     document.write(content)
+//
+export const llamaPromise = (prompt, params = {}, config = {}) => {
+  return new Promise(async (resolve, reject) => {
+    let content = "";
+    try {
+      for await (const chunk of llama(prompt, params, config)) {
+        content += chunk.data.content;
+      }
+      resolve(content);
+    } catch (error) {
+      reject(error);
+    }
+  });
+};
+
+/**
+ * (deprecated)
+ */
+export const llamaComplete = async (params, controller, callback) => {
+  for await (const chunk of llama(params.prompt, params, { controller })) {
+    callback(chunk);
+  }
+}
+
+// Get the model info from the server. This is useful for getting the context window and so on.
+export const llamaModelInfo = async (config = {}) => {
+  if (!generation_settings) {
+    const api_url = config.api_url?.replace(/\/+$/, '') || "";
+    const props = await fetch(`${api_url}/props`).then(r => r.json());
+    generation_settings = props.default_generation_settings;
+  }
+  return generation_settings;
+}

examples/server/public_legacy/loading.html

@@ -0,0 +1,12 @@
+<!DOCTYPE html>
+<html>
+    <head>
+        <meta http-equiv="refresh" content="5">
+    </head>
+    <body>
+        <div id="loading">
+            The model is loading. Please wait.<br/>
+            The user interface will appear soon.
+        </div>
+    </body>
+</html>

examples/server/server.cpp

@@ -14,22 +14,13 @@
 #define MIMETYPE_JSON "application/json; charset=utf-8"
 
 // auto generated files (update with ./deps.sh)
-#include "colorthemes.css.hpp"
-#include "style.css.hpp"
-#include "theme-beeninorder.css.hpp"
-#include "theme-ketivah.css.hpp"
-#include "theme-mangotango.css.hpp"
-#include "theme-playground.css.hpp"
-#include "theme-polarnight.css.hpp"
-#include "theme-snowstorm.css.hpp"
 #include "index.html.hpp"
-#include "index-new.html.hpp"
-#include "index.js.hpp"
 #include "completion.js.hpp"
-#include "system-prompts.js.hpp"
-#include "prompt-formats.js.hpp"
-#include "json-schema-to-grammar.mjs.hpp"
 #include "loading.html.hpp"
+#include "deps_daisyui.min.css.hpp"
+#include "deps_markdown-it.js.hpp"
+#include "deps_tailwindcss.js.hpp"
+#include "deps_vue.esm-browser.js.hpp"
 
 #include <atomic>
 #include <condition_variable>
@@ -2285,16 +2276,6 @@ int main(int argc, char ** argv) {
     std::atomic<server_state> state{SERVER_STATE_LOADING_MODEL};
 
     svr->set_default_headers({{"Server", "llama.cpp"}});
-
-    // CORS preflight
-    svr->Options(R"(.*)", [](const httplib::Request &, httplib::Response & res) {
-        // Access-Control-Allow-Origin is already set by middleware
-        res.set_header("Access-Control-Allow-Credentials", "true");
-        res.set_header("Access-Control-Allow-Methods",     "POST");
-        res.set_header("Access-Control-Allow-Headers",     "*");
-        return res.set_content("", "text/html"); // blank response, no data
-    });
-
     svr->set_logger(log_server_request);
 
     auto res_error = [](httplib::Response & res, const json & error_data) {
@@ -2407,6 +2388,14 @@ int main(int argc, char ** argv) {
     // register server middlewares
     svr->set_pre_routing_handler([&middleware_validate_api_key, &middleware_server_state](const httplib::Request & req, httplib::Response & res) {
         res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        // If this is OPTIONS request, skip validation because browsers don't include Authorization header
+        if (req.method == "OPTIONS") {
+            res.set_header("Access-Control-Allow-Credentials", "true");
+            res.set_header("Access-Control-Allow-Methods",     "GET, POST");
+            res.set_header("Access-Control-Allow-Headers",     "*");
+            res.set_content("", "text/html"); // blank response, no data
+            return httplib::Server::HandlerResponse::Handled; // skip further processing
+        }
         if (!middleware_server_state(req, res)) {
             return httplib::Server::HandlerResponse::Handled;
         }
@@ -3116,33 +3105,19 @@ int main(int argc, char ** argv) {
     // register static assets routes
     if (!params.public_path.empty()) {
         // Set the base directory for serving static files
-        svr->set_base_dir(params.public_path);
-    }
-
-    if (!params.api_keys.empty()) {
-        // for now, if API key is set, web UI is unusable
-        svr->Get("/", [&](const httplib::Request &, httplib::Response & res) {
-            return res.set_content("Web UI is disabled because API key is set.", "text/html; charset=utf-8");
-        });
+        bool is_found = svr->set_mount_point("/", params.public_path);
+        if (!is_found) {
+            LOG_ERR("%s: static assets path not found: %s\n", __func__, params.public_path.c_str());
+            return 1;
+        }
     } else {
         // using embedded static files
-        svr->Get("/",                           handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
-        svr->Get("/index.js",                   handle_static_file(index_js, index_js_len, "text/javascript; charset=utf-8"));
-        svr->Get("/completion.js",              handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));
-        svr->Get("/json-schema-to-grammar.mjs", handle_static_file(json_schema_to_grammar_mjs, json_schema_to_grammar_mjs_len, "text/javascript; charset=utf-8"));
-
-        // add new-ui files
-        svr->Get("/colorthemes.css",       handle_static_file(colorthemes_css, colorthemes_css_len, "text/css; charset=utf-8"));
-        svr->Get("/style.css",             handle_static_file(style_css, style_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-beeninorder.css", handle_static_file(theme_beeninorder_css, theme_beeninorder_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-ketivah.css",     handle_static_file(theme_ketivah_css, theme_ketivah_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-mangotango.css",  handle_static_file(theme_mangotango_css, theme_mangotango_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-playground.css",  handle_static_file(theme_playground_css, theme_playground_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-polarnight.css",  handle_static_file(theme_polarnight_css, theme_polarnight_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-snowstorm.css",   handle_static_file(theme_snowstorm_css, theme_snowstorm_css_len, "text/css; charset=utf-8"));
-        svr->Get("/index-new.html",        handle_static_file(index_new_html, index_new_html_len, "text/html; charset=utf-8"));
-        svr->Get("/system-prompts.js",     handle_static_file(system_prompts_js, system_prompts_js_len, "text/javascript; charset=utf-8"));
-        svr->Get("/prompt-formats.js",     handle_static_file(prompt_formats_js, prompt_formats_js_len, "text/javascript; charset=utf-8"));
+        svr->Get("/",                        handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
+        svr->Get("/completion.js",           handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));
+        svr->Get("/deps_daisyui.min.css",    handle_static_file(deps_daisyui_min_css, deps_daisyui_min_css_len, "text/css; charset=utf-8"));
+        svr->Get("/deps_markdown-it.js",     handle_static_file(deps_markdown_it_js, deps_markdown_it_js_len, "text/javascript; charset=utf-8"));
+        svr->Get("/deps_tailwindcss.js",     handle_static_file(deps_tailwindcss_js, deps_tailwindcss_js_len, "text/javascript; charset=utf-8"));
+        svr->Get("/deps_vue.esm-browser.js", handle_static_file(deps_vue_esm_browser_js, deps_vue_esm_browser_js_len, "text/javascript; charset=utf-8"));
     }
 
     // register API routes

examples/server/tests/features/security.feature

@@ -64,5 +64,5 @@ Feature: Security
       | localhost       | Access-Control-Allow-Origin      | localhost         |
       | web.mydomain.fr | Access-Control-Allow-Origin      | web.mydomain.fr   |
       | origin          | Access-Control-Allow-Credentials | true              |
-      | web.mydomain.fr | Access-Control-Allow-Methods     | POST              |
+      | web.mydomain.fr | Access-Control-Allow-Methods     | GET, POST         |
       | web.mydomain.fr | Access-Control-Allow-Headers     | *                 |

ggml/include/ggml.h

@@ -1746,6 +1746,9 @@ extern "C" {
             struct ggml_tensor * a,
             enum ggml_prec       prec);
 
+    GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
+            const struct ggml_tensor * a);
+
     // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,

ggml/src/ggml-cuda.cu

@@ -3159,6 +3159,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 #ifndef FLASH_ATTN_AVAILABLE
             return false;
 #endif
+            if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
+                return false;
+            }
             if (op->src[0]->ne[0] ==  64 && op->src[1]->type == GGML_TYPE_F16) {
                 return true;
             }

ggml/src/ggml-cuda/fattn.cu

@@ -13,9 +13,9 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
     const ggml_tensor * KQV = dst;
     const ggml_tensor * Q   = dst->src[0];
 
-    const int32_t precision = KQV->op_params[3];
+    const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
-    if (precision != GGML_PREC_DEFAULT) {
+    if (prec != GGML_PREC_DEFAULT) {
         if (Q->ne[1] <= 32 || Q->ne[0] > 128) {
             constexpr int cols_per_block = 16;
             switch (Q->ne[0]) {
@@ -301,11 +301,11 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     ggml_cuda_set_device(ctx.device);
     const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    const int32_t precision = KQV->op_params[3];
+    const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
     // On AMD the tile kernels perform poorly, use the vec kernel instead:
     if (cc >= CC_OFFSET_AMD) {
-        if (precision == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+        if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
             ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
         } else {
             ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
@@ -332,7 +332,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     }
 
     if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) {
-        if (precision == GGML_PREC_DEFAULT) {
+        if (prec == GGML_PREC_DEFAULT) {
             ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
             return;
         } else if(Q->ne[0] <= 128) {

ggml/src/ggml-metal.m

@@ -269,6 +269,12 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,
@@ -300,12 +306,14 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,
@@ -585,6 +593,9 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
             struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \
             id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
             kernel->pipeline = [device newComputePipelineStateWithFunction:metal_function error:&error]; \
+            GGML_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
+                    (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
+                    (int) kernel->pipeline.threadExecutionWidth); \
             [metal_function release]; \
             if (error) { \
                 GGML_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
@@ -777,6 +788,12 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,       flash_attn_ext_f16_h112,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,       flash_attn_ext_f16_h128,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,       flash_attn_ext_f16_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,       flash_attn_ext_bf16_h64,        has_simdgroup_mm && has_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,       flash_attn_ext_bf16_h80,        has_simdgroup_mm && has_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,       flash_attn_ext_bf16_h96,        has_simdgroup_mm && has_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112,      flash_attn_ext_bf16_h112,       has_simdgroup_mm && has_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128,      flash_attn_ext_bf16_h128,       has_simdgroup_mm && has_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,      flash_attn_ext_bf16_h256,       has_simdgroup_mm && has_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,       flash_attn_ext_q4_0_h64,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,       flash_attn_ext_q4_0_h80,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,       flash_attn_ext_q4_0_h96,        has_simdgroup_mm);
@@ -808,12 +825,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128,      flash_attn_ext_q8_0_h128,       has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,      flash_attn_ext_q8_0_h256,       has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,   flash_attn_ext_vec_f16_h128,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128,  flash_attn_ext_vec_bf16_h128,   has_simdgroup_reduction && has_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128,  flash_attn_ext_vec_q4_0_h128,   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H128,  flash_attn_ext_vec_q4_1_h128,   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H128,  flash_attn_ext_vec_q5_0_h128,   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H128,  flash_attn_ext_vec_q5_1_h128,   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H128,  flash_attn_ext_vec_q8_0_h128,   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,   flash_attn_ext_vec_f16_h256,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H256,  flash_attn_ext_vec_bf16_h256,   has_simdgroup_reduction && has_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H256,  flash_attn_ext_vec_q4_0_h256,   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H256,  flash_attn_ext_vec_q4_1_h256,   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,  flash_attn_ext_vec_q5_0_h256,   has_simdgroup_reduction);
@@ -1111,7 +1130,7 @@ static void ggml_metal_encode_node(
     const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
     const uint64_t nb21 = src2 ? src2->nb[1] : 0;
     const uint64_t nb22 = src2 ? src2->nb[2] : 0;
-    const uint64_t nb23 = src2 ? src2->nb[3] : 0;
+    const uint64_t nb23 = src2 ? src2->nb[3] : 0; GGML_UNUSED(nb23);
 
     const int64_t  ne0  =  dst ?  dst->ne[0] : 0;
     const int64_t  ne1  =  dst ?  dst->ne[1] : 0;
@@ -3033,6 +3052,23 @@ static void ggml_metal_encode_node(
                                               }
                                 }
                             } break;
+                        case GGML_TYPE_BF16:
+                            {
+                                switch (ne00) {
+                                    case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64 ].pipeline; break;
+                                    case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80 ].pipeline; break;
+                                    case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96 ].pipeline; break;
+                                    case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112].pipeline; break;
+                                    case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128].pipeline; break;
+                                    case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256].pipeline; break;
+                                    default:
+                                              {
+                                                  GGML_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                                  GGML_LOG_ERROR("add template specialization for this size\n");
+                                                  GGML_ABORT("add template specialization for this size");
+                                              }
+                                }
+                            } break;
                         case GGML_TYPE_Q4_0:
                             {
                                 switch (ne00) {
@@ -3133,6 +3169,7 @@ static void ggml_metal_encode_node(
                             {
                                 switch (src1->type) {
                                     case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break;
+                                    case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128].pipeline; break;
                                     case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128].pipeline; break;
                                     case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H128].pipeline; break;
                                     case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H128].pipeline; break;
@@ -3150,6 +3187,7 @@ static void ggml_metal_encode_node(
                             {
                                 switch (src1->type) {
                                     case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
+                                    case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H256].pipeline; break;
                                     case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H256].pipeline; break;
                                     case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H256].pipeline; break;
                                     case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256].pipeline; break;
@@ -3194,18 +3232,15 @@ static void ggml_metal_encode_node(
                 [encoder setBytes:&nb11          length:sizeof(uint64_t)      atIndex:14];
                 [encoder setBytes:&nb12          length:sizeof(uint64_t)      atIndex:15];
                 [encoder setBytes:&nb13          length:sizeof(uint64_t)      atIndex:16];
-                [encoder setBytes:&nb21          length:sizeof(uint64_t)      atIndex:17];
-                [encoder setBytes:&nb22          length:sizeof(uint64_t)      atIndex:18];
-                [encoder setBytes:&nb23          length:sizeof(uint64_t)      atIndex:19];
-                [encoder setBytes:&nb31          length:sizeof(uint64_t)      atIndex:20];
-                [encoder setBytes:&ne1           length:sizeof( int64_t)      atIndex:21];
-                [encoder setBytes:&ne2           length:sizeof( int64_t)      atIndex:22];
-                [encoder setBytes:&scale         length:sizeof(   float)      atIndex:23];
-                [encoder setBytes:&max_bias      length:sizeof(   float)      atIndex:24];
-                [encoder setBytes:&m0            length:sizeof(m0)            atIndex:25];
-                [encoder setBytes:&m1            length:sizeof(m1)            atIndex:26];
-                [encoder setBytes:&n_head_log2   length:sizeof(n_head_log2)   atIndex:27];
-                [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:28];
+                [encoder setBytes:&nb31          length:sizeof(uint64_t)      atIndex:17];
+                [encoder setBytes:&ne1           length:sizeof( int64_t)      atIndex:18];
+                [encoder setBytes:&ne2           length:sizeof( int64_t)      atIndex:19];
+                [encoder setBytes:&scale         length:sizeof(   float)      atIndex:20];
+                [encoder setBytes:&max_bias      length:sizeof(   float)      atIndex:21];
+                [encoder setBytes:&m0            length:sizeof(m0)            atIndex:22];
+                [encoder setBytes:&m1            length:sizeof(m1)            atIndex:23];
+                [encoder setBytes:&n_head_log2   length:sizeof(n_head_log2)   atIndex:24];
+                [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:25];
 
                 if (!use_vec_kernel) {
                     // half8x8 kernel
@@ -3216,11 +3251,14 @@ static void ggml_metal_encode_node(
                     GGML_ASSERT(nqptg  % 8  == 0);
                     GGML_ASSERT(ncpsg  % 32 == 0);
 
+                    // 2*(2*ncpsg + nqptg)*(nsg)
+                    // ncpsg soft_max values + ncpsg mask values + a diagonal scaling matrix (in float)
+                    //
                     // 16*32*(nsg)
                     // the shared memory needed for the simdgroups to load the KV cache
                     // each thread loads (dequantizes) 16 head elements, there are 32 threads in th SG
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*(ncpsg + nqptg)*(nsg)) + 16*32*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + 16*32*(nsg))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
 
@@ -3254,12 +3292,12 @@ static void ggml_metal_encode_node(
 
                     // ne00 + 2*ncpsg*(nsg)
                     // for each query, we load it as f16 in shared memory (ne00)
-                    // and store the attention scores (nqptg x ncpsg) as f32
+                    // and store the soft_max values and the mask
                     //
-                    // 2*ne00*(nsg)
-                    // each simdgroup has a full f32 head vector in shared mem to accumulate results
+                    // ne00*(nsg)
+                    // each simdgroup has a full f16 head vector in shared mem to accumulate results
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*ncpsg*(nsg)) + 2*ne00*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*ncpsg*(nsg)) + ne00*(nsg))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
 

ggml/src/ggml.c

@@ -4228,6 +4228,15 @@ void ggml_flash_attn_ext_set_prec(
     ggml_set_op_params_i32(a, 3, prec_i32); // scale is on first pos, max_bias on second
 }
 
+enum ggml_prec ggml_flash_attn_ext_get_prec(
+        const struct ggml_tensor * a) {
+    GGML_ASSERT(a->op == GGML_OP_FLASH_ATTN_EXT);
+
+    const int32_t prec_i32 = ggml_get_op_params_i32(a, 3);
+
+    return (enum ggml_prec) prec_i32;
+}
+
 // ggml_flash_attn_back
 
 struct ggml_tensor * ggml_flash_attn_back(

grammars/README.md

@@ -124,7 +124,7 @@ You can use GBNF grammars:
 - In [llama-cli](../examples/main), passed as the `--json` / `-j` flag
 - To convert to a grammar ahead of time:
     - in CLI, with [examples/json_schema_to_grammar.py](../examples/json_schema_to_grammar.py)
-    - in JavaScript with [json-schema-to-grammar.mjs](../examples/server/public/json-schema-to-grammar.mjs) (this is used by the [server](../examples/server)'s Web UI)
+    - in JavaScript with [json-schema-to-grammar.mjs](../examples/server/public_legacy/json-schema-to-grammar.mjs) (this is used by the [server](../examples/server)'s Web UI)
 
 Take a look at [tests](../tests/test-json-schema-to-grammar.cpp) to see which features are likely supported (you'll also find usage examples in https://github.com/ggerganov/llama.cpp/pull/5978, https://github.com/ggerganov/llama.cpp/pull/6659 & https://github.com/ggerganov/llama.cpp/pull/6555).
 

scripts/sync-ggml-am.sh

@@ -114,46 +114,22 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
 
     # replace filenames:
     #
-    # CMakelists.txt          -> ggml/CMakeLists.txt
-    # src/CMakeLists.txt      -> ggml/src/CMakeLists.txt
-    # cmake/FindSIMD.cmake    -> ggml/cmake/FindSIMD.cmake
+    # CMakelists.txt       -> ggml/CMakeLists.txt
+    # src/CMakeLists.txt   -> ggml/src/CMakeLists.txt
+    # cmake/FindSIMD.cmake -> ggml/cmake/FindSIMD.cmake
     #
-    # src/ggml.c              -> ggml/src/ggml.c
-    # src/ggml-aarch64.c      -> ggml/src/ggml-aarch64.c
-    # src/ggml-aarch64.h      -> ggml/src/ggml-aarch64.h
-    # src/ggml-alloc.c        -> ggml/src/ggml-alloc.c
-    # src/ggml-amx/*          -> ggml/src/ggml-amx/
-    # src/ggml-amx.cpp        -> ggml/src/ggml-amx.cpp
-    # src/ggml-backend-impl.h -> ggml/src/ggml-backend-impl.h
-    # src/ggml-backend.cpp    -> ggml/src/ggml-backend.cpp
-    # src/ggml-cann/*         -> ggml/src/ggml-cann/
-    # src/ggml-cann.cpp       -> ggml/src/ggml-cann.cpp
-    # src/ggml-common.h       -> ggml/src/ggml-common.h
-    # src/ggml-cuda/*         -> ggml/src/ggml-cuda/
-    # src/ggml-cuda.cu        -> ggml/src/ggml-cuda.cu
-    # src/ggml-impl.h         -> ggml/src/ggml-impl.h
-    # src/ggml-kompute.cpp    -> ggml/src/ggml-kompute.cpp
-    # src/ggml-metal.m        -> ggml/src/ggml-metal.m
-    # src/ggml-quants.c       -> ggml/src/ggml-quants.c
-    # src/ggml-quants.h       -> ggml/src/ggml-quants.h
-    # src/ggml-rpc.cpp        -> ggml/src/ggml-rpc.cpp
-    # src/ggml-sycl/*         -> ggml/src/ggml-sycl/
-    # src/ggml-sycl.cpp       -> ggml/src/ggml-sycl.cpp
-    # src/ggml-vulkan.cpp     -> ggml/src/ggml-vulkan.cpp
-    # src/vulkan-shaders/*    -> ggml/src/vulkan-shaders/
+    # src/ggml*.c          -> ggml/src/ggml*.c
+    # src/ggml*.cpp        -> ggml/src/ggml*.cpp
+    # src/ggml*.h          -> ggml/src/ggml*.h
+    # src/ggml*.cu         -> ggml/src/ggml*.cu
+    # src/ggml*.m          -> ggml/src/ggml*.m
+    # src/ggml-amx/*       -> ggml/src/ggml-amx/
+    # src/ggml-cann/*      -> ggml/src/ggml-cann/
+    # src/ggml-cuda/*      -> ggml/src/ggml-cuda/
+    # src/ggml-sycl/*      -> ggml/src/ggml-sycl/
+    # src/vulkan-shaders/* -> ggml/src/vulkan-shaders/
     #
-    # include/ggml.h         -> ggml/include/ggml.h
-    # include/ggml-alloc.h   -> ggml/include/ggml-alloc.h
-    # include/ggml-amx.h     -> ggml/include/ggml-amx.h
-    # include/ggml-backend.h -> ggml/include/ggml-backend.h
-    # include/ggml-blas.h    -> ggml/include/ggml-blas.h
-    # include/ggml-cann.h    -> ggml/include/ggml-cann.h
-    # include/ggml-cuda.h    -> ggml/include/ggml-cuda.h
-    # include/ggml-kompute.h -> ggml/include/ggml-kompute.h
-    # include/ggml-metal.h   -> ggml/include/ggml-metal.h
-    # include/ggml-rpc.h     -> ggml/include/ggml-rpc.h
-    # include/ggml-sycl.h    -> ggml/include/ggml-sycl.h
-    # include/ggml-vulkan.h  -> ggml/include/ggml-vulkan.h
+    # include/ggml*.h -> ggml/include/ggml*.h
     #
     # tests/test-opt.cpp           -> tests/test-opt.cpp
     # tests/test-grad0.cpp         -> tests/test-grad0.cpp
@@ -168,41 +144,17 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)CMakeLists.txt/\1ggml\/CMakeLists.txt/g' \
         -e 's/([[:space:]]|[ab]\/)src\/CMakeLists.txt/\1ggml\/src\/CMakeLists.txt/g' \
         -e 's/([[:space:]]|[ab]\/)cmake\/FindSIMD.cmake/\1ggml\/cmake\/FindSIMD.cmake/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml\.c/\1ggml\/src\/ggml.c/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.c/\1ggml\/src\/ggml-aarch64.c/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.h/\1ggml\/src\/ggml-aarch64.h/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-alloc\.c/\1ggml\/src\/ggml-alloc.c/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.c/\1ggml\/src\/ggml\1.c/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.cpp/\1ggml\/src\/ggml\1.cpp/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.h/\1ggml\/src\/ggml\1.h/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.cu/\1ggml\/src\/ggml\1.cu/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.m/\1ggml\/src\/ggml\1.m/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-amx\//\1ggml\/src\/ggml-amx\//g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-amx\.cpp/\1ggml\/src\/ggml-amx.cpp/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-backend-impl\.h/\1ggml\/src\/ggml-backend-impl.h/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-backend\.cpp/\1ggml\/src\/ggml-backend.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\.cpp/\1ggml\/src\/ggml-cann.cpp/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-common\.h/\1ggml\/src\/ggml-common.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-cuda\.cu/\1ggml\/src\/ggml-cuda.cu/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-impl\.h/\1ggml\/src\/ggml-impl.h/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-kompute\.cpp/\1ggml\/src\/ggml-kompute.cpp/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-metal\.m/\1ggml\/src\/ggml-metal.m/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-quants\.c/\1ggml\/src\/ggml-quants.c/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-quants\.h/\1ggml\/src\/ggml-quants.h/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-rpc\.cpp/\1ggml\/src\/ggml-rpc.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\.cpp/\1ggml\/src\/ggml-sycl.cpp/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-vulkan\.cpp/\1ggml\/src\/ggml-vulkan.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/vulkan-shaders\//\1ggml\/src\/vulkan-shaders\//g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml\.h/\1ggml\/include\/ggml.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-alloc\.h/\1ggml\/include\/ggml-alloc.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-amx\.h/\1ggml\/include\/ggml-amx.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-backend\.h/\1ggml\/include\/ggml-backend.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-blas\.h/\1ggml\/include\/ggml-blas.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-cann\.h/\1ggml\/include\/ggml-cann.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-cuda\.h/\1ggml\/include\/ggml-cuda.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-kompute\.h/\1ggml\/include\/ggml-kompute.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-metal\.h/\1ggml\/include\/ggml-metal.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-rpc\.h/\1ggml\/include\/ggml-rpc.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-sycl\.h/\1ggml\/include\/ggml-sycl.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-vulkan\.h/\1ggml\/include\/ggml-vulkan.h/g' \
+        -e 's/([[:space:]]|[ab]\/)include\/ggml(.*)\.h/\1ggml\/include\/ggml\1.h/g' \
         -e 's/([[:space:]]|[ab]\/)examples\/common\.h/\1examples\/common.h/g' \
         -e 's/([[:space:]]|[ab]\/)examples\/common\.cpp/\1examples\/common.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.h/\1examples\/common-ggml.h/g' \

scripts/sync-ggml.last

@@ -1 +1 @@
-a099cb514d6687e436a5a423d1fb0448be0feb20
+89952d649e0c5cabbb9ff8c4906f5a843a789fb2

scripts/sync-ggml.sh

@@ -4,43 +4,18 @@ cp -rpv ../ggml/CMakeLists.txt       ./ggml/CMakeLists.txt
 cp -rpv ../ggml/src/CMakeLists.txt   ./ggml/src/CMakeLists.txt
 cp -rpv ../ggml/cmake/FindSIMD.cmake ./ggml/cmake/FindSIMD.cmake
 
-cp -rpv ../ggml/src/ggml.c              ./ggml/src/ggml.c
-cp -rpv ../ggml/src/ggml-aarch64.c      ./ggml/src/ggml-aarch64.c
-cp -rpv ../ggml/src/ggml-aarch64.h      ./ggml/src/ggml-aarch64.h
-cp -rpv ../ggml/src/ggml-alloc.c        ./ggml/src/ggml-alloc.c
-cp -rpv ../ggml/src/ggml-amx/*          ./ggml/src/ggml-amx/
-cp -rpv ../ggml/src/ggml-amx.cpp        ./ggml/src/ggml-amx.cpp
-cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml/src/ggml-backend-impl.h
-cp -rpv ../ggml/src/ggml-backend.cpp    ./ggml/src/ggml-backend.cpp
-cp -rpv ../ggml/src/ggml-cann/*         ./ggml/src/ggml-cann/
-cp -rpv ../ggml/src/ggml-cann.cpp       ./ggml/src/ggml-cann.cpp
-cp -rpv ../ggml/src/ggml-common.h       ./ggml/src/ggml-common.h
-cp -rpv ../ggml/src/ggml-cuda/*         ./ggml/src/ggml-cuda/
-cp -rpv ../ggml/src/ggml-cuda.cu        ./ggml/src/ggml-cuda.cu
-cp -rpv ../ggml/src/ggml-impl.h         ./ggml/src/ggml-impl.h
-cp -rpv ../ggml/src/ggml-kompute.cpp    ./ggml/src/ggml-kompute.cpp
-cp -rpv ../ggml/src/ggml-metal.m        ./ggml/src/ggml-metal.m
-cp -rpv ../ggml/src/ggml-metal.metal    ./ggml/src/ggml-metal.metal
-cp -rpv ../ggml/src/ggml-quants.c       ./ggml/src/ggml-quants.c
-cp -rpv ../ggml/src/ggml-quants.h       ./ggml/src/ggml-quants.h
-cp -rpv ../ggml/src/ggml-rpc.cpp        ./ggml/src/ggml-rpc.cpp
-cp -rpv ../ggml/src/ggml-sycl/*         ./ggml/src/ggml-sycl/
-cp -rpv ../ggml/src/ggml-sycl.cpp       ./ggml/src/ggml-sycl.cpp
-cp -rpv ../ggml/src/ggml-vulkan.cpp     ./ggml/src/ggml-vulkan.cpp
-cp -rpv ../ggml/src/vulkan-shaders/*    ./ggml/src/vulkan-shaders/
+cp -rpv ../ggml/src/ggml*.c          ./ggml/src/
+cp -rpv ../ggml/src/ggml*.cpp        ./ggml/src/
+cp -rpv ../ggml/src/ggml*.h          ./ggml/src/
+cp -rpv ../ggml/src/ggml*.cu         ./ggml/src/
+cp -rpv ../ggml/src/ggml*.m          ./ggml/src/
+cp -rpv ../ggml/src/ggml-amx/*       ./ggml/src/ggml-amx/
+cp -rpv ../ggml/src/ggml-cann/*      ./ggml/src/ggml-cann/
+cp -rpv ../ggml/src/ggml-cuda/*      ./ggml/src/ggml-cuda/
+cp -rpv ../ggml/src/ggml-sycl/*      ./ggml/src/ggml-sycl/
+cp -rpv ../ggml/src/vulkan-shaders/* ./ggml/src/vulkan-shaders/
 
-cp -rpv ../ggml/include/ggml.h         ./ggml/include/ggml.h
-cp -rpv ../ggml/include/ggml-alloc.h   ./ggml/include/ggml-alloc.h
-cp -rpv ../ggml/include/ggml-amx.h     ./ggml/include/ggml-amx.h
-cp -rpv ../ggml/include/ggml-backend.h ./ggml/include/ggml-backend.h
-cp -rpv ../ggml/include/ggml-blas.h    ./ggml/include/ggml-blas.h
-cp -rpv ../ggml/include/ggml-cann.h    ./ggml/include/ggml-cann.h
-cp -rpv ../ggml/include/ggml-cuda.h    ./ggml/include/ggml-cuda.h
-cp -rpv ../ggml/include/ggml-kompute.h ./ggml/include/ggml-kompute.h
-cp -rpv ../ggml/include/ggml-metal.h   ./ggml/include/ggml-metal.h
-cp -rpv ../ggml/include/ggml-rpc.h     ./ggml/include/ggml-rpc.h
-cp -rpv ../ggml/include/ggml-sycl.h    ./ggml/include/ggml-sycl.h
-cp -rpv ../ggml/include/ggml-vulkan.h  ./ggml/include/ggml-vulkan.h
+cp -rpv ../ggml/include/ggml*.h ./ggml/include/
 
 cp -rpv ../ggml/tests/test-opt.cpp           ./tests/test-opt.cpp
 cp -rpv ../ggml/tests/test-grad0.cpp         ./tests/test-grad0.cpp

tests/run-json-schema-to-grammar.mjs

@@ -1,5 +1,5 @@
 import { readFileSync } from "fs"
-import { SchemaConverter } from "../examples/server/public/json-schema-to-grammar.mjs"
+import { SchemaConverter } from "../examples/server/public_legacy/json-schema-to-grammar.mjs"
 
 const [, , file] = process.argv
 const url = `file://${file}`

tests/test-backend-ops.cpp

@@ -3745,7 +3745,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                     for (int nh : { 32, }) {
                         for (int kv : { 512, 1024, }) {
                             for (int nb : { 1, 3, 32, 35, }) {
-                                for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
+                                for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
                                     test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV));
                                 }
                             }