release : pack /lib and /include in the packages

For consistency

Signed-off-by: Eric Curtin <ecurtin@redhat.com>

.github/workflows/build.yml

@@ -88,7 +88,7 @@ jobs:
         run: |
           cp LICENSE ./build/bin/
           cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/* ./build/lib/* ./build/include/*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -151,7 +151,7 @@ jobs:
         run: |
           cp LICENSE ./build/bin/
           cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/* ./build/lib/* ./build/include/*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -220,7 +220,7 @@ jobs:
         run: |
           cp LICENSE ./build/bin/
           cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip ./build/bin/*
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip ./build/bin/* ./build/lib/* ./build/include/*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -1026,7 +1026,7 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
 
           echo "cp oneAPI running time dll files to ./build/bin done"
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/* ./build/lib/* ./build/include/*
 
       - name: Upload the release package
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}

README.md

@@ -16,7 +16,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- **Introducing GGUF-my-LoRA** https://github.com/ggerganov/llama.cpp/discussions/10123
+- **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode
+- Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
+- Introducing GGUF-my-LoRA https://github.com/ggerganov/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669
 - Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 

common/chat-template.hpp

@@ -25,6 +25,7 @@ class chat_template {
     // Meta-Llama-3.1-8B-Instruct's template expects arguments to be an object.
     // Most other templates (and OpenAI's API) expect the arguments object to be stringified.
     bool requires_object_arguments_ = false;
+    bool requires_typed_content_ = false;
     bool supports_system_role_ = true;
     bool supports_parallel_tool_calls_ = false;
     std::string source_;
@@ -32,14 +33,14 @@ class chat_template {
     std::string eos_token_;
     std::shared_ptr<minja::TemplateNode> template_root_;
 
-    std::string try_render(
+    std::string try_raw_render(
         const nlohmann::ordered_json & messages,
         const nlohmann::ordered_json & tools,
         bool add_generation_prompt,
         const nlohmann::ordered_json & extra_context = nlohmann::ordered_json()) const
     {
         try {
-            auto prompt = apply(messages, tools, add_generation_prompt, extra_context);
+            auto prompt = apply(messages, tools, add_generation_prompt, extra_context, /* adjust_inputs= */ false);
             // fprintf(stderr, "Prompt: %s\n", prompt.c_str());
             return prompt;
         } catch (const std::exception & e) {
@@ -60,7 +61,7 @@ class chat_template {
         supports_tools_ = source.find("tools") != std::string::npos;
 
         auto renders_string_arguments =
-            try_render({
+            try_raw_render({
                 {
                     {"role", "user"},
                     {"content", "Hey"}
@@ -81,7 +82,7 @@ class chat_template {
             }, {}, false).find("{\"code\": \"print") != std::string::npos;
         if (!renders_string_arguments) {
             auto renders_object_arguments =
-                try_render({
+                try_raw_render({
                     {
                         {"role", "user"},
                         {"content", "Hey"}
@@ -106,10 +107,13 @@ class chat_template {
         }
         supports_parallel_tool_calls_ = source.find("tool_call_id") != std::string::npos;
 
-        supports_system_role_ = try_render({
+        supports_system_role_ = try_raw_render({
             {{"role", "system"}, {"content", "<System Needle>"}},
             {{"role", "user"},   {"content", "Hey"}}
         }, {}, false).find("<System Needle>") != std::string::npos;
+
+        requires_typed_content_ = try_raw_render({{{"role", "user"},   {"content", "Hey"}}}, {}, false).find("Hey") == std::string::npos
+            && try_raw_render({{{"role", "user"},   {"content", {{{"type", "text"}, {"text", "Hey"}}}}}}, {}, false).find("Hey") != std::string::npos;
     }
 
     const std::string & source() const { return source_; }
@@ -122,19 +126,34 @@ class chat_template {
         const nlohmann::ordered_json & messages,
         const nlohmann::ordered_json & tools,
         bool add_generation_prompt,
-        const nlohmann::ordered_json & extra_context = nlohmann::ordered_json()) const
+        const nlohmann::ordered_json & extra_context = nlohmann::ordered_json(),
+        bool adjust_inputs = true) const
     {
         json actual_messages;
 
         // First, "fix" messages so they have a chance to be rendered correctly by the template
 
-        if (requires_object_arguments_ || !supports_system_role_ || !supports_tools_) {
+        if (adjust_inputs && (requires_object_arguments_ || !supports_system_role_ || !supports_tools_ || requires_typed_content_)) {
             actual_messages = json::array();
 
+            auto add_message = [&](const json & msg) {
+                if (requires_typed_content_ && msg.contains("content") && !msg.at("content").is_null() && msg.at("content").is_string()) {
+                    actual_messages.push_back({
+                        {"role", msg.at("role")},
+                        {"content", {{
+                            {"type", "text"},
+                            {"text", msg.at("content")},
+                        }}},
+                    });
+                } else {
+                    actual_messages.push_back(msg);
+                }
+            };
+
             std::string pending_system;
             auto flush_sys = [&]() {
                 if (!pending_system.empty()) {
-                    actual_messages.push_back({
+                    add_message({
                         {"role", "user"},
                         {"content", pending_system},
                     });
@@ -217,7 +236,7 @@ class chat_template {
                         }
                     }
                 }
-                actual_messages.push_back(message);
+                add_message(message);
             }
             flush_sys();
         } else {

common/minja.hpp

@@ -693,7 +693,7 @@ enum SpaceHandling { Keep, Strip, StripSpaces, StripNewline };
 
 class TemplateToken {
 public:
-    enum class Type { Text, Expression, If, Else, Elif, EndIf, For, EndFor, Set, EndSet, Comment, Macro, EndMacro, Filter, EndFilter };
+    enum class Type { Text, Expression, If, Else, Elif, EndIf, For, EndFor, Generation, EndGeneration, Set, EndSet, Comment, Macro, EndMacro, Filter, EndFilter };
 
     static std::string typeToString(Type t) {
         switch (t) {
@@ -712,6 +712,8 @@ public:
             case Type::EndMacro: return "endmacro";
             case Type::Filter: return "filter";
             case Type::EndFilter: return "endfilter";
+            case Type::Generation: return "generation";
+            case Type::EndGeneration: return "endgeneration";
         }
         return "Unknown";
     }
@@ -788,6 +790,14 @@ struct EndForTemplateToken : public TemplateToken {
     EndForTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFor, location, pre, post) {}
 };
 
+struct GenerationTemplateToken : public TemplateToken {
+    GenerationTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Generation, location, pre, post) {}
+};
+
+struct EndGenerationTemplateToken : public TemplateToken {
+    EndGenerationTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndGeneration, location, pre, post) {}
+};
+
 struct SetTemplateToken : public TemplateToken {
     std::string ns;
     std::vector<std::string> var_names;
@@ -2149,7 +2159,7 @@ private:
       static std::regex comment_tok(R"(\{#([-~]?)(.*?)([-~]?)#\})");
       static std::regex expr_open_regex(R"(\{\{([-~])?)");
       static std::regex block_open_regex(R"(^\{%([-~])?[\s\n\r]*)");
-      static std::regex block_keyword_tok(R"((if|else|elif|endif|for|endfor|set|endset|block|endblock|macro|endmacro|filter|endfilter)\b)");
+      static std::regex block_keyword_tok(R"((if|else|elif|endif|for|endfor|generation|endgeneration|set|endset|block|endblock|macro|endmacro|filter|endfilter)\b)");
       static std::regex non_text_open_regex(R"(\{\{|\{%|\{#)");
       static std::regex expr_close_regex(R"([\s\n\r]*([-~])?\}\})");
       static std::regex block_close_regex(R"([\s\n\r]*([-~])?%\})");
@@ -2229,6 +2239,12 @@ private:
             } else if (keyword == "endfor") {
               auto post_space = parseBlockClose();
               tokens.push_back(std::make_unique<EndForTemplateToken>(location, pre_space, post_space));
+            } else if (keyword == "generation") {
+              auto post_space = parseBlockClose();
+              tokens.push_back(std::make_unique<GenerationTemplateToken>(location, pre_space, post_space));
+            } else if (keyword == "endgeneration") {
+              auto post_space = parseBlockClose();
+              tokens.push_back(std::make_unique<EndGenerationTemplateToken>(location, pre_space, post_space));
             } else if (keyword == "set") {
               static std::regex namespaced_var_regex(R"((\w+)[\s\n\r]*\.[\s\n\r]*(\w+))");
 
@@ -2330,6 +2346,13 @@ private:
                   throw unterminated(**start);
               }
               children.emplace_back(std::make_shared<ForNode>(token->location, std::move(for_token->var_names), std::move(for_token->iterable), std::move(for_token->condition), std::move(body), for_token->recursive, std::move(else_body)));
+          } else if (dynamic_cast<GenerationTemplateToken*>(token.get())) {
+              auto body = parseTemplate(begin, it, end);
+              if (it == end || (*(it++))->type != TemplateToken::Type::EndGeneration) {
+                  throw unterminated(**start);
+              }
+              // Treat as a no-op, as our scope is templates for inference, not training (`{% generation %}` wraps generated tokens for masking).
+              children.emplace_back(std::move(body));
           } else if (auto text_token = dynamic_cast<TextTemplateToken*>(token.get())) {
               SpaceHandling pre_space = (it - 1) != begin ? (*(it - 2))->post_space : SpaceHandling::Keep;
               SpaceHandling post_space = it != end ? (*it)->pre_space : SpaceHandling::Keep;
@@ -2397,6 +2420,7 @@ private:
                   || dynamic_cast<EndFilterTemplateToken*>(token.get())
                   || dynamic_cast<EndIfTemplateToken*>(token.get())
                   || dynamic_cast<ElseTemplateToken*>(token.get())
+                  || dynamic_cast<EndGenerationTemplateToken*>(token.get())
                   || dynamic_cast<ElifTemplateToken*>(token.get())) {
               it--;  // unconsume the token
               break;  // exit the loop

examples/main/README.md

@@ -310,9 +310,9 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Batch Size
 
--   `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
+- `-ub N`, `--ubatch-size N`: Physical batch size. This is the maximum number of tokens that may be processed at a time. Increasing this value may improve performance during prompt processing, at the expense of higher memory usage. Default: `512`.
 
-- `-ub N`, `--ubatch-size N`: physical maximum batch size. This is for pipeline parallelization. Default: `512`.
+- `-b N`, `--batch-size N`: Logical batch size. Increasing this value above the value of the physical batch size may improve prompt processing performance when using multiple GPUs with pipeline parallelism. Default: `2048`.
 
 ### Prompt Caching
 

examples/run/README.md

@@ -3,11 +3,10 @@
 The purpose of this example is to demonstrate a minimal usage of llama.cpp for running models.
 
 ```bash
-llama-run granite-code
+llama-run granite3-moe
 ```
 
 ```bash
-llama-run -h
 Description:
   Runs a llm
 
@@ -17,7 +16,7 @@ Usage:
 Options:
   -c, --context-size <value>
       Context size (default: 2048)
-  -n, --ngl <value>
+  -n, -ngl, --ngl <value>
       Number of GPU layers (default: 0)
   --temp <value>
       Temperature (default: 0.8)

examples/run/run.cpp

@@ -147,7 +147,8 @@ class Opt {
                 if (handle_option_with_value(argc, argv, i, context_size) == 1) {
                     return 1;
                 }
-            } else if (options_parsing && (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "--ngl") == 0)) {
+            } else if (options_parsing &&
+                       (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) {
                 if (handle_option_with_value(argc, argv, i, ngl) == 1) {
                     return 1;
                 }
@@ -194,7 +195,7 @@ class Opt {
             "Options:\n"
             "  -c, --context-size <value>\n"
             "      Context size (default: %d)\n"
-            "  -n, --ngl <value>\n"
+            "  -n, -ngl, --ngl <value>\n"
             "      Number of GPU layers (default: %d)\n"
             "  --temp <value>\n"
             "      Temperature (default: %.1f)\n"
@@ -634,20 +635,20 @@ class LlamaData {
         return path.substr(pos + 1);
     }
 
-    int remove_proto(std::string & model_) {
-        const std::string::size_type pos = model_.find("://");
+    int rm_until_substring(std::string & model_, const std::string & substring) {
+        const std::string::size_type pos = model_.find(substring);
         if (pos == std::string::npos) {
             return 1;
         }
 
-        model_ = model_.substr(pos + 3);  // Skip past "://"
+        model_ = model_.substr(pos + substring.size());  // Skip past the substring
         return 0;
     }
 
     int resolve_model(std::string & model_) {
         int                            ret     = 0;
         if (string_starts_with(model_, "file://") || std::filesystem::exists(model_)) {
-            remove_proto(model_);
+            rm_until_substring(model_, "://");
 
             return ret;
         }
@@ -656,13 +657,16 @@ class LlamaData {
         const std::vector<std::string> headers = { "--header",
                                                    "Accept: application/vnd.docker.distribution.manifest.v2+json" };
         if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://")) {
-            remove_proto(model_);
+            rm_until_substring(model_, "://");
+            ret = huggingface_dl(model_, headers, bn);
+        } else if (string_starts_with(model_, "hf.co/")) {
+            rm_until_substring(model_, "hf.co/");
             ret = huggingface_dl(model_, headers, bn);
         } else if (string_starts_with(model_, "ollama://")) {
-            remove_proto(model_);
+            rm_until_substring(model_, "://");
             ret = ollama_dl(model_, headers, bn);
         } else if (string_starts_with(model_, "https://")) {
-            download(model_, headers, bn, true);
+            ret = download(model_, headers, bn, true);
         } else {
             ret = ollama_dl(model_, headers, bn);
         }

examples/server/server.cpp

@@ -267,6 +267,11 @@ struct server_task {
         params.speculative.n_min = std::max(params.speculative.n_min, 2);
         params.speculative.n_max = std::max(params.speculative.n_max, 0);
 
+        // Use OpenAI API logprobs only if n_probs wasn't provided
+        if (data.contains("logprobs") && params.sampling.n_probs == defaults.sampling.n_probs){
+            params.sampling.n_probs = json_value(data, "logprobs", defaults.sampling.n_probs);
+        }
+
         if (data.contains("lora")) {
             if (data.at("lora").is_array()) {
                 params.lora = parse_lora_request(params_base.lora_adapters, data.at("lora"));
@@ -1428,6 +1433,10 @@ struct server_queue {
         } else {
             queue_tasks.push_back(std::move(task));
         }
+        // if this is cancel task make sure to clean up pending tasks
+        if (task.type == SERVER_TASK_TYPE_CANCEL) {
+            cleanup_pending_task(task.id_target);
+        }
         condition_tasks.notify_one();
         return task.id;
     }
@@ -1445,6 +1454,10 @@ struct server_queue {
             } else {
                 queue_tasks.push_back(std::move(task));
             }
+            // if this is cancel task make sure to clean up pending tasks
+            if (task.type == SERVER_TASK_TYPE_CANCEL) {
+                cleanup_pending_task(task.id_target);
+            }
         }
         condition_tasks.notify_one();
         return 0;
@@ -1539,6 +1552,20 @@ struct server_queue {
             }
         }
     }
+
+private:
+    void cleanup_pending_task(int id_task) {
+        // no need lock because this is called exclusively by post()
+        auto rm_func = [id_task](const server_task & task) {
+            return task.id_target == id_task;
+        };
+        queue_tasks.erase(
+            std::remove_if(queue_tasks.begin(),          queue_tasks.end(),          rm_func),
+            queue_tasks.end());
+        queue_tasks_deferred.erase(
+            std::remove_if(queue_tasks_deferred.begin(), queue_tasks_deferred.end(), rm_func),
+            queue_tasks_deferred.end());
+    }
 };
 
 struct server_response {
@@ -1574,6 +1601,12 @@ struct server_response {
 
         std::unique_lock<std::mutex> lock(mutex_results);
         waiting_task_ids.erase(id_task);
+        // make sure to clean up all pending results
+        queue_results.erase(
+            std::remove_if(queue_results.begin(), queue_results.end(), [id_task](const server_task_result_ptr & res) {
+                return res->id == id_task;
+            }),
+            queue_results.end());
     }
 
     void remove_waiting_task_ids(const std::unordered_set<int> & id_tasks) {
@@ -1593,7 +1626,7 @@ struct server_response {
                 return !queue_results.empty();
             });
 
-            for (int i = 0; i < (int) queue_results.size(); i++) {
+            for (size_t i = 0; i < queue_results.size(); i++) {
                 if (id_tasks.find(queue_results[i]->id) != id_tasks.end()) {
                     server_task_result_ptr res = std::move(queue_results[i]);
                     queue_results.erase(queue_results.begin() + i);
@@ -1610,12 +1643,6 @@ struct server_response {
     server_task_result_ptr recv_with_timeout(const std::unordered_set<int> & id_tasks, int timeout) {
         while (true) {
             std::unique_lock<std::mutex> lock(mutex_results);
-            bool cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout), [&]{
-                return !queue_results.empty();
-            });
-            if (!cr_res) {
-                return nullptr;
-            }
 
             for (int i = 0; i < (int) queue_results.size(); i++) {
                 if (id_tasks.find(queue_results[i]->id) != id_tasks.end()) {
@@ -1624,6 +1651,11 @@ struct server_response {
                     return res;
                 }
             }
+
+            std::cv_status cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout));
+            if (cr_res == std::cv_status::timeout) {
+                return nullptr;
+            }
         }
 
         // should never reach here
@@ -1767,6 +1799,9 @@ struct server_context {
             // force F16 KV cache for the draft model for extra performance
             cparams_dft.type_k = GGML_TYPE_F16;
             cparams_dft.type_v = GGML_TYPE_F16;
+
+            // the context is not needed - we will create one for each slot
+            llama_init_dft.context.reset();
         }
 
         chat_templates = common_chat_templates_from_model(model, params_base.chat_template);
@@ -2368,8 +2403,8 @@ struct server_context {
 
             server_task task(SERVER_TASK_TYPE_CANCEL);
             task.id_target = id_task;
-            cancel_tasks.push_back(task);
             queue_results.remove_waiting_task_id(id_task);
+            cancel_tasks.push_back(task);
         }
         // push to beginning of the queue, so it has highest priority
         queue_tasks.post(cancel_tasks, true);

examples/server/webui/index.html

@@ -141,6 +141,7 @@
               :msg="pendingMsg"
               :key="pendingMsg.id"
               :is-generating="isGenerating"
+              :show-thought-in-progress="config.showThoughtInProgress"
               :edit-user-msg-and-regenerate="() => {}"
               :regenerate-msg="() => {}"></message-bubble>
           </div>
@@ -202,6 +203,20 @@
               </template>
             </div>
           </details>
+          <!-- Section: Reasoning models -->
+          <details class="collapse collapse-arrow bg-base-200 mb-2 overflow-visible">
+            <summary class="collapse-title font-bold">Reasoning models</summary>
+            <div class="collapse-content">
+              <div class="flex flex-row items-center mb-2">
+                <input type="checkbox" class="checkbox" v-model="config.showThoughtInProgress" />
+                <span class="ml-4">Expand though process by default for generating message</span>
+              </div>
+              <div class="flex flex-row items-center mb-2">
+                <input type="checkbox" class="checkbox" v-model="config.excludeThoughtOnReq" />
+                <span class="ml-4">Exclude thought process when sending request to API (Recommended for DeepSeek-R1)</span>
+              </div>
+            </div>
+          </details>
           <!-- Section: Advanced config -->
           <details class="collapse collapse-arrow bg-base-200 mb-2 overflow-visible">
             <summary class="collapse-title font-bold">Advanced config</summary>
@@ -261,7 +276,17 @@
           <span v-if="msg.content === null" class="loading loading-dots loading-md"></span>
           <!-- render message as markdown -->
           <div v-else dir="auto">
-            <vue-markdown :source="msg.content"></vue-markdown>
+            <details v-if="msg.role === 'assistant' && splitMsgContent.cot" class="collapse bg-base-200 collapse-arrow mb-4" :open="splitMsgContent.isThinking && showThoughtInProgress">
+              <summary class="collapse-title">
+                <span v-if="splitMsgContent.isThinking">
+                  <span v-if="isGenerating" class="loading loading-spinner loading-md mr-2" style="vertical-align: middle;"></span>
+                  <b>Thinking</b>
+                </span>
+                <b v-else>Thought Process</b>
+              </summary>
+              <vue-markdown :source="splitMsgContent.cot" dir="auto" class="collapse-content"></vue-markdown>
+            </details>
+            <vue-markdown :source="splitMsgContent.content"></vue-markdown>
           </div>
           <!-- render timings if enabled -->
           <div class="dropdown dropdown-hover dropdown-top mt-2" v-if="timings && config.showTokensPerSecond">

examples/server/webui/src/main.js

@@ -17,6 +17,11 @@ import { asyncIterator } from '@sec-ant/readable-stream/ponyfill/asyncIterator';
 
 const isDev = import.meta.env.MODE === 'development';
 
+// types
+/** @typedef {{ id: number, role: 'user' | 'assistant', content: string, timings: any }} Message */
+/** @typedef {{ role: 'user' | 'assistant', content: string }} APIMessage */
+/** @typedef {{ id: string, lastModified: number, messages: Array<Message> }} Conversation */
+
 // utility functions
 const isString = (x) => !!x.toLowerCase;
 const isBoolean = (x) => x === true || x === false;
@@ -50,6 +55,8 @@ const CONFIG_DEFAULT = {
   apiKey: '',
   systemMessage: 'You are a helpful assistant.',
   showTokensPerSecond: false,
+  showThoughtInProgress: false,
+  excludeThoughtOnReq: true,
   // make sure these default values are in sync with `common.h`
   samplers: 'edkypmxt',
   temperature: 0.8,
@@ -172,6 +179,7 @@ const MessageBubble = defineComponent({
     config: Object,
     msg: Object,
     isGenerating: Boolean,
+    showThoughtInProgress: Boolean,
     editUserMsgAndRegenerate: Function,
     regenerateMsg: Function,
   },
@@ -188,7 +196,31 @@ const MessageBubble = defineComponent({
         prompt_per_second: this.msg.timings.prompt_n / (this.msg.timings.prompt_ms / 1000),
         predicted_per_second: this.msg.timings.predicted_n / (this.msg.timings.predicted_ms / 1000),
       };
-    }
+    },
+    splitMsgContent() {
+      const content = this.msg.content;
+      if (this.msg.role !== 'assistant') {
+        return { content };
+      }
+      let actualContent = '';
+      let cot = '';
+      let isThinking = false;
+      let thinkSplit = content.split('<think>', 2);
+      actualContent += thinkSplit[0];
+      while (thinkSplit[1] !== undefined) {
+        // <think> tag found
+        thinkSplit = thinkSplit[1].split('</think>', 2);
+        cot += thinkSplit[0];
+        isThinking = true;
+        if (thinkSplit[1] !== undefined) {
+          // </think> closing tag found
+          isThinking = false;
+          thinkSplit = thinkSplit[1].split('<think>', 2);
+          actualContent += thinkSplit[0];
+        }
+      }
+      return { content: actualContent, cot, isThinking };
+    },
   },
   methods: {
     copyMsg() {
@@ -208,7 +240,10 @@ const MessageBubble = defineComponent({
 // format: { [convId]: { id: string, lastModified: number, messages: [...] } }
 // convId is a string prefixed with 'conv-'
 const StorageUtils = {
-  // manage conversations
+  /**
+   * manage conversations
+   * @returns {Array<Conversation>}
+   */
   getAllConversations() {
     const res = [];
     for (const key in localStorage) {
@@ -219,11 +254,19 @@ const StorageUtils = {
     res.sort((a, b) => b.lastModified - a.lastModified);
     return res;
   },
-  // can return null if convId does not exist
+  /**
+   * can return null if convId does not exist
+   * @param {string} convId
+   * @returns {Conversation | null}
+   */
   getOneConversation(convId) {
     return JSON.parse(localStorage.getItem(convId) || 'null');
   },
-  // if convId does not exist, create one
+  /**
+   * if convId does not exist, create one
+   * @param {string} convId
+   * @param {Message} msg
+   */
   appendMsg(convId, msg) {
     if (msg.content === null) return;
     const conv = StorageUtils.getOneConversation(convId) || {
@@ -235,12 +278,24 @@ const StorageUtils = {
     conv.lastModified = Date.now();
     localStorage.setItem(convId, JSON.stringify(conv));
   },
+  /**
+   * Get new conversation id
+   * @returns {string}
+   */
   getNewConvId() {
     return `conv-${Date.now()}`;
   },
+  /**
+   * remove conversation by id
+   * @param {string} convId
+   */
   remove(convId) {
     localStorage.removeItem(convId);
   },
+  /**
+   * remove all conversations
+   * @param {string} convId
+   */
   filterAndKeepMsgs(convId, predicate) {
     const conv = StorageUtils.getOneConversation(convId);
     if (!conv) return;
@@ -248,6 +303,11 @@ const StorageUtils = {
     conv.lastModified = Date.now();
     localStorage.setItem(convId, JSON.stringify(conv));
   },
+  /**
+   * remove last message from conversation
+   * @param {string} convId
+   * @returns {Message | undefined}
+   */
   popMsg(convId) {
     const conv = StorageUtils.getOneConversation(convId);
     if (!conv) return;
@@ -322,10 +382,12 @@ const mainApp = createApp({
   data() {
     return {
       conversations: StorageUtils.getAllConversations(),
-      messages: [], // { id: number, role: 'user' | 'assistant', content: string }
+      /** @type {Array<Message>} */
+      messages: [],
       viewingConvId: StorageUtils.getNewConvId(),
       inputMsg: '',
       isGenerating: false,
+      /** @type {Array<Message> | null} */
       pendingMsg: null, // the on-going message from assistant
       stopGeneration: () => {},
       selectedTheme: StorageUtils.getTheme(),
@@ -333,6 +395,7 @@ const mainApp = createApp({
       showConfigDialog: false,
       // const
       themes: THEMES,
+      /** @type {CONFIG_DEFAULT} */
       configDefault: {...CONFIG_DEFAULT},
       configInfo: {...CONFIG_INFO},
       isDev,
@@ -425,42 +488,50 @@ const mainApp = createApp({
       this.isGenerating = true;
 
       try {
+        /** @type {CONFIG_DEFAULT} */
+        const config = this.config;
         const abortController = new AbortController();
         this.stopGeneration = () => abortController.abort();
+        /** @type {Array<APIMessage>} */
+        let messages = [
+          { role: 'system', content: config.systemMessage },
+          ...normalizeMsgsForAPI(this.messages),
+        ];
+        if (config.excludeThoughtOnReq) {
+          messages = filterThoughtFromMsgs(messages);
+        }
+        if (isDev) console.log({messages});
         const params = {
-          messages: [
-            { role: 'system', content: this.config.systemMessage },
-            ...this.messages,
-          ],
+          messages,
           stream: true,
           cache_prompt: true,
-          samplers: this.config.samplers,
-          temperature: this.config.temperature,
-          dynatemp_range: this.config.dynatemp_range,
-          dynatemp_exponent: this.config.dynatemp_exponent,
-          top_k: this.config.top_k,
-          top_p: this.config.top_p,
-          min_p: this.config.min_p,
-          typical_p: this.config.typical_p,
-          xtc_probability: this.config.xtc_probability,
-          xtc_threshold: this.config.xtc_threshold,
-          repeat_last_n: this.config.repeat_last_n,
-          repeat_penalty: this.config.repeat_penalty,
-          presence_penalty: this.config.presence_penalty,
-          frequency_penalty: this.config.frequency_penalty,
-          dry_multiplier: this.config.dry_multiplier,
-          dry_base: this.config.dry_base,
-          dry_allowed_length: this.config.dry_allowed_length,
-          dry_penalty_last_n: this.config.dry_penalty_last_n,
-          max_tokens: this.config.max_tokens,
-          timings_per_token: !!this.config.showTokensPerSecond,
-          ...(this.config.custom.length ? JSON.parse(this.config.custom) : {}),
+          samplers: config.samplers,
+          temperature: config.temperature,
+          dynatemp_range: config.dynatemp_range,
+          dynatemp_exponent: config.dynatemp_exponent,
+          top_k: config.top_k,
+          top_p: config.top_p,
+          min_p: config.min_p,
+          typical_p: config.typical_p,
+          xtc_probability: config.xtc_probability,
+          xtc_threshold: config.xtc_threshold,
+          repeat_last_n: config.repeat_last_n,
+          repeat_penalty: config.repeat_penalty,
+          presence_penalty: config.presence_penalty,
+          frequency_penalty: config.frequency_penalty,
+          dry_multiplier: config.dry_multiplier,
+          dry_base: config.dry_base,
+          dry_allowed_length: config.dry_allowed_length,
+          dry_penalty_last_n: config.dry_penalty_last_n,
+          max_tokens: config.max_tokens,
+          timings_per_token: !!config.showTokensPerSecond,
+          ...(config.custom.length ? JSON.parse(config.custom) : {}),
         };
         const chunks = sendSSEPostRequest(`${BASE_URL}/v1/chat/completions`, {
           method: 'POST',
           headers: {
             'Content-Type': 'application/json',
-            ...(this.config.apiKey ? {'Authorization': `Bearer ${this.config.apiKey}`} : {})
+            ...(config.apiKey ? {'Authorization': `Bearer ${config.apiKey}`} : {})
           },
           body: JSON.stringify(params),
           signal: abortController.signal,
@@ -477,7 +548,7 @@ const mainApp = createApp({
             };
           }
           const timings = chunk.timings;
-          if (timings && this.config.showTokensPerSecond) {
+          if (timings && config.showTokensPerSecond) {
             // only extract what's really needed, to save some space
             this.pendingMsg.timings = {
               prompt_n: timings.prompt_n,
@@ -598,3 +669,33 @@ try {
     <button class="btn" onClick="localStorage.clear(); window.location.reload();">Clear localStorage</button>
   </div>`;
 }
+
+/**
+ * filter out redundant fields upon sending to API
+ * @param {Array<APIMessage>} messages
+ * @returns {Array<APIMessage>}
+ */
+function normalizeMsgsForAPI(messages) {
+  return messages.map((msg) => {
+    return {
+      role: msg.role,
+      content: msg.content,
+    };
+  });
+}
+
+/**
+ * recommended for DeepsSeek-R1, filter out content between <think> and </think> tags
+ * @param {Array<APIMessage>} messages
+ * @returns {Array<APIMessage>}
+ */
+function filterThoughtFromMsgs(messages) {
+  return messages.map((msg) => {
+    return {
+      role: msg.role,
+      content: msg.role === 'assistant'
+        ? msg.content.split('</think>').at(-1).trim()
+        : msg.content,
+    };
+  });
+}

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1673,31 +1673,31 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
 
         if (device->coopmat_acc_f16_support) {
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
 
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         } else {
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
 
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         }
 
         // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
@@ -1707,31 +1707,31 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
 
             if (device->coopmat_acc_f16_support) {
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             } else {
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             }
         }
 #undef CREATE_MM2
@@ -2012,7 +2012,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_tanh_f32, "tanh_f32", tanh_f32_len, tanh_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true);
 
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);

ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp

@@ -12,7 +12,7 @@ layout (push_constant) uniform parameter
 
 #include "types.comp"
 
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x = 1, local_size_y = 512, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -17,13 +17,13 @@
 #include <cstring>
 #include <cstdlib>
 #include <cassert>
+#include <algorithm>
 #include <sys/stat.h>
 #include <sys/types.h>
 
 #ifdef _WIN32
     #include <windows.h>
     #include <direct.h> // For _mkdir on Windows
-    #include <algorithm> // For std::replace on w64devkit
 #else
     #include <unistd.h>
     #include <sys/wait.h>
@@ -502,6 +502,7 @@ void write_output_files() {
     fprintf(hdr, "#include <cstdint>\n\n");
     fprintf(src, "#include \"%s\"\n\n", basename(target_hpp).c_str());
 
+    std::sort(shader_fnames.begin(), shader_fnames.end());
     for (const auto& pair : shader_fnames) {
         const std::string& name = pair.first;
         #ifdef _WIN32

tests/test-backend-ops.cpp

@@ -3909,17 +3909,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 4));
     test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 128, 4));
 
-    for (int i = 1; i < 9; ++i) {
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16,    GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0,   GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_1,   GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q5_0,   GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q5_1,   GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0,   GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_K,   GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q5_K,   GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q6_K,   GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
-        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_IQ4_NL, GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
+    for (ggml_type type_a : all_types) {
+        for (int i = 1; i < 10; ++i) {
+            test_cases.emplace_back(new test_mul_mat(type_a,    GGML_TYPE_F32, 16,  i, 256, { 1,  1}, {1, 1}));
+        }
     }
 
 #if 1