server : add OAI compat for /v1/completions (#10974)

* server : add OAI compat for /v1/completions

* add test

* add docs

* better docs

common/common.cpp

@@ -18,6 +18,7 @@
 #include <cstdarg>
 #include <cstring>
 #include <ctime>
+#include <filesystem>
 #include <fstream>
 #include <iostream>
 #include <iterator>
@@ -62,7 +63,9 @@
 #ifdef __linux__
 #include <linux/limits.h>
 #elif defined(_WIN32)
-#define PATH_MAX MAX_PATH
+#   if !defined(PATH_MAX)
+#   define PATH_MAX MAX_PATH
+#   endif
 #else
 #include <sys/syslimits.h>
 #endif
@@ -1148,8 +1151,7 @@ static bool common_download_file(const std::string & url, const std::string & pa
 #endif
 
     // Check if the file already exists locally
-    struct stat model_file_info;
-    auto file_exists = (stat(path.c_str(), &model_file_info) == 0);
+    auto file_exists = std::filesystem::exists(path);
 
     // If the file exists, check its JSON metadata companion file.
     std::string metadata_path = path + ".json";

convert_hf_to_gguf.py

@@ -1764,25 +1764,19 @@ class DeciModel(Model):
             self.gguf_writer.add_token_list(tokens)
             self.gguf_writer.add_token_types(toktypes)
 
-            special_vocab = gguf.SpecialVocab(
-                self.dir_model, load_merges=True,
-                special_token_types = ['bos', 'eos', 'eom', 'eot']
-            )
-            special_vocab._set_special_token("bos", 128000)
-            special_vocab._set_special_token("eos", 128001)
-            special_vocab._set_special_token("eom", 128008)
-            special_vocab._set_special_token("eot", 128009)
+            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
             special_vocab.add_to_gguf(self.gguf_writer)
         else:
             # DeciLM-7B
             self._set_vocab_llama_hf()
-#            self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
         if "block_configs" in self.hparams: # Llama-3_1-Nemotron-51B
             assert self.block_count == len(self._num_kv_heads)
             assert self.block_count == len(self._num_heads)
             assert self.block_count == len(self._ffn_dims)
+            if (rope_theta := self.hparams.get("rope_theta")) is not None:
+                self.gguf_writer.add_rope_freq_base(rope_theta)
             self.gguf_writer.add_head_count_kv(self._num_kv_heads)
             self.gguf_writer.add_head_count(self._num_heads)
             self.gguf_writer.add_feed_forward_length(self._ffn_dims)

examples/cvector-generator/mean.hpp

@@ -15,7 +15,7 @@ static void run(
     for (size_t il = 0; il < v_input.size(); ++il) {
         // prepare output vector
         struct ggml_tensor * ctrl_out = v_output[il];
-        ggml_format_name(ctrl_out, "direction.%ld", il+1);
+        ggml_format_name(ctrl_out, "direction.%zu", il+1);
 
         // calculate mean vector
         struct ggml_tensor * t_layer = v_input[il];

examples/cvector-generator/pca.hpp

@@ -302,7 +302,7 @@ static void run_pca(
 
         // prepare output vector
         struct ggml_tensor * ctrl_out = v_output[il];
-        ggml_format_name(ctrl_out, "direction.%ld", il+1);
+        ggml_format_name(ctrl_out, "direction.%zu", il+1);
 
         // run power_iteration
         params.i_layer = il;

examples/export-lora/export-lora.cpp

@@ -265,8 +265,8 @@ struct lora_merge_ctx {
             fout.write((const char *)data.data(), data.size());
         }
 
-        printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
-        printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
+        printf("%s : merged %zu tensors with lora adapters\n", __func__, n_merged);
+        printf("%s : wrote %zu tensors to output file\n", __func__, trans.size());
     }
 
     void copy_tensor(struct ggml_tensor * base) {
@@ -352,7 +352,7 @@ struct lora_merge_ctx {
                 const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale;
                 delta = ggml_scale(ctx0, delta, scale);
                 cur = ggml_add(ctx0, delta, cur);
-                printf("%s :   + merging from adapter[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
+                printf("%s :   + merging from adapter[%zu] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
                 printf("%s :     input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]);
             }
             cur = ggml_cast(ctx0, cur, out->type);

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

@@ -305,7 +305,9 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
-    llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
+    //llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
+    const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
+    delete batch;
 }
 
 extern "C"

examples/run/run.cpp

@@ -1,5 +1,6 @@
 #if defined(_WIN32)
 #    include <windows.h>
+#    include <io.h>
 #else
 #    include <sys/file.h>
 #    include <sys/ioctl.h>
@@ -253,7 +254,7 @@ class File {
                 return 1;
             }
 
-            OVERLAPPED overlapped = { 0 };
+            OVERLAPPED overlapped = {};
             if (!LockFileEx(hFile, LOCKFILE_EXCLUSIVE_LOCK | LOCKFILE_FAIL_IMMEDIATELY, 0, MAXDWORD, MAXDWORD,
                             &overlapped)) {
                 fd = -1;
@@ -277,7 +278,7 @@ class File {
         if (fd >= 0) {
 #    ifdef _WIN32
             if (hFile != INVALID_HANDLE_VALUE) {
-                OVERLAPPED overlapped = { 0 };
+                OVERLAPPED overlapped = {};
                 UnlockFileEx(hFile, 0, MAXDWORD, MAXDWORD, &overlapped);
             }
 #    else
@@ -293,7 +294,7 @@ class File {
   private:
     int fd = -1;
 #    ifdef _WIN32
-    HANDLE hFile;
+    HANDLE hFile = nullptr;
 #    endif
 };
 
@@ -464,7 +465,7 @@ class HttpClient {
         return (now_downloaded_plus_file_size * 100) / total_to_download;
     }
 
-    static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", percentage); }
+    static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", static_cast<long int>(percentage)); }
 
     static double calculate_speed(curl_off_t now_downloaded, const std::chrono::steady_clock::time_point & start_time) {
         const auto                          now             = std::chrono::steady_clock::now();

examples/server/CMakeLists.txt

@@ -34,6 +34,7 @@ endforeach()
 add_executable(${TARGET} ${TARGET_SRCS})
 install(TARGETS ${TARGET} RUNTIME)
 
+target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
 target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
 
 if (LLAMA_SERVER_SSL)

examples/server/README.md

@@ -345,7 +345,7 @@ node index.js
 
 > [!IMPORTANT]
 >
-> This endpoint is **not** OAI-compatible
+> This endpoint is **not** OAI-compatible. For OAI-compatible client, use `/v1/completions` instead.
 
 *Options:*
 
@@ -450,7 +450,7 @@ These words will not be included in the completion, so make sure to add them to
 
 `post_sampling_probs`: Returns the probabilities of top `n_probs` tokens after applying sampling chain.
 
-`response_fields`: A list of response fields, for example: `"response_fields": ["content", "generation_settings/n_predict"]`. If the specified field is missing, it will simply be omitted from the response without triggering an error.
+`response_fields`: A list of response fields, for example: `"response_fields": ["content", "generation_settings/n_predict"]`. If the specified field is missing, it will simply be omitted from the response without triggering an error. Note that fields with a slash will be unnested; for example, `generation_settings/n_predict` will move the field `n_predict` from the `generation_settings` object to the root of the response and give it a new name.
 
 **Response format**
 
@@ -523,6 +523,7 @@ These words will not be included in the completion, so make sure to add them to
 - `tokens_evaluated`: Number of tokens evaluated in total from the prompt
 - `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
+
 ### POST `/tokenize`: Tokenize a given text
 
 *Options:*
@@ -574,6 +575,10 @@ With input 'á' (utf8 hex: C3 A1) on tinyllama/stories260k
 
 ### POST `/embedding`: Generate embedding of a given text
 
+> [!IMPORTANT]
+>
+> This endpoint is **not** OAI-compatible. For OAI-compatible client, use `/v1/embeddings` instead.
+
 The same as [the embedding example](../embedding) does.
 
 *Options:*
@@ -744,96 +749,6 @@ To use this endpoint with POST method, you need to start server with `--props`
 
 - None yet
 
-### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
-
-Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
-
-*Options:*
-
-See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
-
-The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name",  "type": "string" }, "date": { "title": "Date",  "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants",  "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
-
-*Examples:*
-
-You can use either Python `openai` library with appropriate checkpoints:
-
-```python
-import openai
-
-client = openai.OpenAI(
-    base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
-    api_key = "sk-no-key-required"
-)
-
-completion = client.chat.completions.create(
-model="gpt-3.5-turbo",
-messages=[
-    {"role": "system", "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."},
-    {"role": "user", "content": "Write a limerick about python exceptions"}
-]
-)
-
-print(completion.choices[0].message)
-```
-
-... or raw HTTP requests:
-
-```shell
-curl http://localhost:8080/v1/chat/completions \
--H "Content-Type: application/json" \
--H "Authorization: Bearer no-key" \
--d '{
-"model": "gpt-3.5-turbo",
-"messages": [
-{
-    "role": "system",
-    "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."
-},
-{
-    "role": "user",
-    "content": "Write a limerick about python exceptions"
-}
-]
-}'
-```
-
-### POST `/v1/embeddings`: OpenAI-compatible embeddings API
-
-This endpoint requires that the model uses a pooling different than type `none`. The embeddings are normalized using the Eucledian norm.
-
-*Options:*
-
-See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-reference/embeddings).
-
-*Examples:*
-
-- input as string
-
-  ```shell
-  curl http://localhost:8080/v1/embeddings \
-  -H "Content-Type: application/json" \
-  -H "Authorization: Bearer no-key" \
-  -d '{
-          "input": "hello",
-          "model":"GPT-4",
-          "encoding_format": "float"
-  }'
-  ```
-
-- `input` as string array
-
-  ```shell
-  curl http://localhost:8080/v1/embeddings \
-  -H "Content-Type: application/json" \
-  -H "Authorization: Bearer no-key" \
-  -d '{
-          "input": ["hello", "world"],
-          "model":"GPT-4",
-          "encoding_format": "float"
-  }'
-  ```
-
 ### POST `/embeddings`: non-OpenAI-compatible embeddings API
 
 This endpoint supports all poolings, including `--pooling none`. When the pooling is `none`, the responses will contain the *unnormalized* embeddings for *all* input tokens. For all other pooling types, only the pooled embeddings are returned, normalized using Euclidian norm.
@@ -1064,6 +979,161 @@ To know the `id` of the adapter, use GET `/lora-adapters`
 ]
 ```
 
+## OpenAI-compatible API Endpoints
+
+### GET `/v1/models`: OpenAI-compatible Model Info API
+
+Returns information about the loaded model. See [OpenAI Models API documentation](https://platform.openai.com/docs/api-reference/models).
+
+The returned list always has one single element.
+
+By default, model `id` field is the path to model file, specified via `-m`. You can set a custom value for model `id` field via `--alias` argument. For example, `--alias gpt-4o-mini`.
+
+Example:
+
+```json
+{
+    "object": "list",
+    "data": [
+        {
+            "id": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
+            "object": "model",
+            "created": 1735142223,
+            "owned_by": "llamacpp",
+            "meta": {
+                "vocab_type": 2,
+                "n_vocab": 128256,
+                "n_ctx_train": 131072,
+                "n_embd": 4096,
+                "n_params": 8030261312,
+                "size": 4912898304
+            }
+        }
+    ]
+}
+```
+
+### POST `/v1/completions`: OpenAI-compatible Completions API
+
+Given an input `prompt`, it returns the predicted completion. Streaming mode is also supported. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps.
+
+*Options:*
+
+See [OpenAI Completions API documentation](https://platform.openai.com/docs/api-reference/completions).
+
+llama.cpp `/completion`-specific features such as `mirostat` are supported.
+
+*Examples:*
+
+Example usage with `openai` python library:
+
+```python
+import openai
+
+client = openai.OpenAI(
+    base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
+    api_key = "sk-no-key-required"
+)
+
+completion = client.completions.create(
+  model="davinci-002",
+  prompt="I believe the meaning of life is",
+  max_tokens=8
+)
+
+print(completion.choices[0].text)
+```
+
+### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
+
+Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
+
+*Options:*
+
+See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
+
+The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name",  "type": "string" }, "date": { "title": "Date",  "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants",  "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
+
+*Examples:*
+
+You can use either Python `openai` library with appropriate checkpoints:
+
+```python
+import openai
+
+client = openai.OpenAI(
+    base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
+    api_key = "sk-no-key-required"
+)
+
+completion = client.chat.completions.create(
+  model="gpt-3.5-turbo",
+  messages=[
+    {"role": "system", "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."},
+    {"role": "user", "content": "Write a limerick about python exceptions"}
+  ]
+)
+
+print(completion.choices[0].message)
+```
+
+... or raw HTTP requests:
+
+```shell
+curl http://localhost:8080/v1/chat/completions \
+-H "Content-Type: application/json" \
+-H "Authorization: Bearer no-key" \
+-d '{
+"model": "gpt-3.5-turbo",
+"messages": [
+{
+    "role": "system",
+    "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."
+},
+{
+    "role": "user",
+    "content": "Write a limerick about python exceptions"
+}
+]
+}'
+```
+
+### POST `/v1/embeddings`: OpenAI-compatible embeddings API
+
+This endpoint requires that the model uses a pooling different than type `none`. The embeddings are normalized using the Eucledian norm.
+
+*Options:*
+
+See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-reference/embeddings).
+
+*Examples:*
+
+- input as string
+
+  ```shell
+  curl http://localhost:8080/v1/embeddings \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer no-key" \
+  -d '{
+          "input": "hello",
+          "model":"GPT-4",
+          "encoding_format": "float"
+  }'
+  ```
+
+- `input` as string array
+
+  ```shell
+  curl http://localhost:8080/v1/embeddings \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer no-key" \
+  -d '{
+          "input": ["hello", "world"],
+          "model":"GPT-4",
+          "encoding_format": "float"
+  }'
+  ```
+
 ## More examples
 
 ### Interactive mode

examples/server/server.cpp

@@ -67,6 +67,13 @@ enum server_task_type {
     SERVER_TASK_TYPE_SET_LORA,
 };
 
+enum oaicompat_type {
+    OAICOMPAT_TYPE_NONE,
+    OAICOMPAT_TYPE_CHAT,
+    OAICOMPAT_TYPE_COMPLETION,
+    OAICOMPAT_TYPE_EMBEDDING,
+};
+
 // https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11
 enum error_type {
     ERROR_TYPE_INVALID_REQUEST,
@@ -101,11 +108,10 @@ struct slot_params {
     struct common_params_speculative speculative;
 
     // OAI-compat fields
-    bool        verbose        = false;
-    bool        oaicompat      = false;
-    bool        oaicompat_chat = true;
-    std::string oaicompat_model;
-    std::string oaicompat_cmpl_id;
+    bool           verbose        = false;
+    oaicompat_type oaicompat      = OAICOMPAT_TYPE_NONE;
+    std::string    oaicompat_model;
+    std::string    oaicompat_cmpl_id;
 
     json to_json() const {
         std::vector<std::string> samplers;
@@ -529,11 +535,10 @@ struct server_task_result_cmpl_final : server_task_result {
     slot_params generation_params;
 
     // OAI-compat fields
-    bool        verbose        = false;
-    bool        oaicompat      = false;
-    bool        oaicompat_chat = true; // TODO: support oaicompat for non-chat
-    std::string oaicompat_model;
-    std::string oaicompat_cmpl_id;
+    bool           verbose        = false;
+    oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
+    std::string    oaicompat_model;
+    std::string    oaicompat_cmpl_id;
 
     virtual int get_index() override {
         return index;
@@ -544,9 +549,16 @@ struct server_task_result_cmpl_final : server_task_result {
     }
 
     virtual json to_json() override {
-        return oaicompat
-            ? (stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat())
-            : to_json_non_oaicompat();
+        switch (oaicompat) {
+            case OAICOMPAT_TYPE_NONE:
+                return to_json_non_oaicompat();
+            case OAICOMPAT_TYPE_COMPLETION:
+                return to_json_oaicompat();
+            case OAICOMPAT_TYPE_CHAT:
+                return stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat();
+            default:
+                GGML_ASSERT(false && "Invalid oaicompat_type");
+        }
     }
 
     json to_json_non_oaicompat() {
@@ -574,6 +586,50 @@ struct server_task_result_cmpl_final : server_task_result {
         return response_fields.empty() ? res : json_get_nested_values(response_fields, res);
     }
 
+    json to_json_oaicompat() {
+        std::time_t t = std::time(0);
+        json logprobs = json(nullptr); // OAI default to null
+        if (!stream && probs_output.size() > 0) {
+            logprobs = json{
+                {"content", completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs)},
+            };
+        }
+        json finish_reason = "length";
+        if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
+            finish_reason = "stop";
+        }
+        json res = json {
+            {"choices",            json::array({
+                json{
+                    {"text",          stream ? "" : content}, // in stream mode, content is already in last partial chunk
+                    {"index",         index},
+                    {"logprobs",      logprobs},
+                    {"finish_reason", finish_reason},
+                }
+            })},
+            {"created",            t},
+            {"model",              oaicompat_model},
+            {"system_fingerprint", build_info},
+            {"object",             "text_completion"},
+            {"usage", json {
+                {"completion_tokens", n_decoded},
+                {"prompt_tokens",     n_prompt_tokens},
+                {"total_tokens",      n_decoded + n_prompt_tokens}
+            }},
+            {"id", oaicompat_cmpl_id}
+        };
+
+        // extra fields for debugging purposes
+        if (verbose) {
+            res["__verbose"] = to_json_non_oaicompat();
+        }
+        if (timings.prompt_n >= 0) {
+            res.push_back({"timings", timings.to_json()});
+        }
+
+        return res;
+    }
+
     json to_json_oaicompat_chat() {
         std::string finish_reason = "length";
         if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
@@ -671,11 +727,10 @@ struct server_task_result_cmpl_partial : server_task_result {
     result_timings timings;
 
     // OAI-compat fields
-    bool        verbose        = false;
-    bool        oaicompat      = false;
-    bool        oaicompat_chat = true; // TODO: support oaicompat for non-chat
-    std::string oaicompat_model;
-    std::string oaicompat_cmpl_id;
+    bool           verbose   = false;
+    oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
+    std::string    oaicompat_model;
+    std::string    oaicompat_cmpl_id;
 
     virtual int get_index() override {
         return index;
@@ -686,7 +741,16 @@ struct server_task_result_cmpl_partial : server_task_result {
     }
 
     virtual json to_json() override {
-        return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat();
+        switch (oaicompat) {
+            case OAICOMPAT_TYPE_NONE:
+                return to_json_non_oaicompat();
+            case OAICOMPAT_TYPE_COMPLETION:
+                return to_json_oaicompat();
+            case OAICOMPAT_TYPE_CHAT:
+                return to_json_oaicompat_chat();
+            default:
+                GGML_ASSERT(false && "Invalid oaicompat_type");
+        }
     }
 
     json to_json_non_oaicompat() {
@@ -711,6 +775,41 @@ struct server_task_result_cmpl_partial : server_task_result {
     }
 
     json to_json_oaicompat() {
+        std::time_t t = std::time(0);
+        json logprobs = json(nullptr); // OAI default to null
+        if (prob_output.probs.size() > 0) {
+            logprobs = json{
+                {"content", completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs)},
+            };
+        }
+        json res = json {
+            {"choices",            json::array({
+                json{
+                    {"text",          content},
+                    {"index",         index},
+                    {"logprobs",      logprobs},
+                    {"finish_reason", nullptr},
+                }
+            })},
+            {"created",            t},
+            {"model",              oaicompat_model},
+            {"system_fingerprint", build_info},
+            {"object",             "text_completion"},
+            {"id",                 oaicompat_cmpl_id}
+        };
+
+        // extra fields for debugging purposes
+        if (verbose) {
+            res["__verbose"] = to_json_non_oaicompat();
+        }
+        if (timings.prompt_n >= 0) {
+            res.push_back({"timings", timings.to_json()});
+        }
+
+        return res;
+    }
+
+    json to_json_oaicompat_chat() {
         bool first = n_decoded == 0;
         std::time_t t = std::time(0);
         json choices;
@@ -789,14 +888,16 @@ struct server_task_result_embd : server_task_result {
     int32_t n_tokens;
 
     // OAI-compat fields
-    bool oaicompat = false;
+    oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
 
     virtual int get_index() override {
         return index;
     }
 
     virtual json to_json() override {
-        return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat();
+        return oaicompat == OAICOMPAT_TYPE_EMBEDDING
+            ? to_json_oaicompat()
+            : to_json_non_oaicompat();
     }
 
     json to_json_non_oaicompat() {
@@ -1856,6 +1957,8 @@ struct server_context {
                 result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
                 slot.n_sent_text += result.text_to_send.size();
                 // add the token to slot queue and cache
+            } else {
+                result.text_to_send = "";
             }
 
             slot.add_token(result);
@@ -2042,7 +2145,6 @@ struct server_context {
 
         res->verbose           = slot.params.verbose;
         res->oaicompat         = slot.params.oaicompat;
-        res->oaicompat_chat    = slot.params.oaicompat_chat;
         res->oaicompat_model   = slot.params.oaicompat_model;
         res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id;
 
@@ -2083,7 +2185,6 @@ struct server_context {
         res->verbose           = slot.params.verbose;
         res->stream            = slot.params.stream;
         res->oaicompat         = slot.params.oaicompat;
-        res->oaicompat_chat    = slot.params.oaicompat_chat;
         res->oaicompat_model   = slot.params.oaicompat_model;
         res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id;
 
@@ -3504,12 +3605,11 @@ int main(int argc, char ** argv) {
 
     // handle completion-like requests (completion, chat, infill)
     // we can optionally provide a custom format for partial results and final results
-    const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](
+    const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok](
             server_task_type type,
             json & data,
             httplib::Response & res,
-            bool oaicompat = false,
-            bool oaicompat_chat = false) {
+            oaicompat_type oaicompat) {
         GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL);
 
         if (ctx_server.params_base.embedding) {
@@ -3534,9 +3634,8 @@ int main(int argc, char ** argv) {
                 task.id_selected_slot = json_value(data, "id_slot", -1);
 
                 // OAI-compat
-                task.params.oaicompat           = oaicompat;
-                task.params.oaicompat_chat      = oaicompat_chat;
-                task.params.oaicompat_cmpl_id   = completion_id;
+                task.params.oaicompat         = oaicompat;
+                task.params.oaicompat_cmpl_id = completion_id;
                 // oaicompat_model is already populated by params_from_json_cmpl
 
                 tasks.push_back(task);
@@ -3587,7 +3686,7 @@ int main(int argc, char ** argv) {
                 }, [&](const json & error_data) {
                     server_sent_event(sink, "error", error_data);
                 });
-                if (oaicompat) {
+                if (oaicompat != OAICOMPAT_TYPE_NONE) {
                     static const std::string ev_done = "data: [DONE]\n\n";
                     sink.write(ev_done.data(), ev_done.size());
                 }
@@ -3603,17 +3702,25 @@ int main(int argc, char ** argv) {
         }
     };
 
-    const auto handle_completions = [&handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_completions = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
         json data = json::parse(req.body);
-        return handle_completions_generic(
+        return handle_completions_impl(
             SERVER_TASK_TYPE_COMPLETION,
             data,
             res,
-            /* oaicompat */ false,
-            /* oaicompat_chat */ false);
+            OAICOMPAT_TYPE_NONE);
     };
 
-    const auto handle_infill = [&ctx_server, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_completions_oai = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
+        json data = oaicompat_completion_params_parse(json::parse(req.body));
+        return handle_completions_impl(
+            SERVER_TASK_TYPE_COMPLETION,
+            data,
+            res,
+            OAICOMPAT_TYPE_COMPLETION);
+    };
+
+    const auto handle_infill = [&ctx_server, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
         // check model compatibility
         std::string err;
         if (llama_token_fim_pre(ctx_server.model) == LLAMA_TOKEN_NULL) {
@@ -3682,22 +3789,25 @@ int main(int argc, char ** argv) {
             tokenized_prompts[0]
         );
 
-        return handle_completions_generic(SERVER_TASK_TYPE_INFILL, data, res);
+        return handle_completions_impl(
+            SERVER_TASK_TYPE_INFILL,
+            data,
+            res,
+            OAICOMPAT_TYPE_NONE); // infill is not OAI compatible
     };
 
-    const auto handle_chat_completions = [&ctx_server, &params, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_chat_completions = [&ctx_server, &params, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
         if (ctx_server.params_base.embedding) {
             res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
-        json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), params.chat_template);
-        return handle_completions_generic(
+        json data = oaicompat_chat_completion_params_parse(ctx_server.model, json::parse(req.body), params.chat_template);
+        return handle_completions_impl(
             SERVER_TASK_TYPE_COMPLETION,
             data,
             res,
-            /* oaicompat */ true,
-            /* oaicompat_chat */ true);
+            OAICOMPAT_TYPE_CHAT);
     };
 
     const auto handle_models = [&params, &ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
@@ -3770,10 +3880,10 @@ int main(int argc, char ** argv) {
         res_ok(res, data);
     };
 
-    const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, bool oaicompat) {
+    const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, oaicompat_type oaicompat) {
         const json body = json::parse(req.body);
 
-        if (oaicompat && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) {
+        if (oaicompat != OAICOMPAT_TYPE_NONE && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) {
             res_error(res, format_error_response("Pooling type 'none' is not OAI compatible. Please use a different pooling type", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
@@ -3783,13 +3893,24 @@ int main(int argc, char ** argv) {
         if (body.count("input") != 0) {
             prompt = body.at("input");
         } else if (body.contains("content")) {
-            oaicompat = false;
+            oaicompat = OAICOMPAT_TYPE_NONE; // "content" field is not OAI compatible
             prompt = body.at("content");
         } else {
             res_error(res, format_error_response("\"input\" or \"content\" must be provided", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
 
+        bool use_base64 = false;
+        if (body.count("encoding_format") != 0) {
+            const std::string& format = body.at("encoding_format");
+            if (format == "base64") {
+                use_base64 = true;
+            } else if (format != "float") {
+                res_error(res, format_error_response("The format to return the embeddings in. Can be either float or base64", ERROR_TYPE_INVALID_REQUEST));
+                return;
+            }
+        }
+
         std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true);
         for (const auto & tokens : tokenized_prompts) {
             // this check is necessary for models that do not add BOS token to the input
@@ -3841,16 +3962,18 @@ int main(int argc, char ** argv) {
         }
 
         // write JSON response
-        json root = oaicompat ? format_embeddings_response_oaicompat(body, responses) : json(responses);
+        json root = oaicompat == OAICOMPAT_TYPE_EMBEDDING
+            ? format_embeddings_response_oaicompat(body, responses, use_base64)
+            : json(responses);
         res_ok(res, root);
     };
 
     const auto handle_embeddings = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) {
-        handle_embeddings_impl(req, res, false);
+        handle_embeddings_impl(req, res, OAICOMPAT_TYPE_NONE);
     };
 
     const auto handle_embeddings_oai = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) {
-        handle_embeddings_impl(req, res, true);
+        handle_embeddings_impl(req, res, OAICOMPAT_TYPE_EMBEDDING);
     };
 
     const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
@@ -4020,7 +4143,7 @@ int main(int argc, char ** argv) {
     svr->Get ("/v1/models",           handle_models); // public endpoint (no API key check)
     svr->Post("/completion",          handle_completions); // legacy
     svr->Post("/completions",         handle_completions);
-    svr->Post("/v1/completions",      handle_completions);
+    svr->Post("/v1/completions",      handle_completions_oai);
     svr->Post("/chat/completions",    handle_chat_completions);
     svr->Post("/v1/chat/completions", handle_chat_completions);
     svr->Post("/infill",              handle_infill);

examples/server/tests/unit/test_chat_completion.py

@@ -83,7 +83,7 @@ def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_conte
 def test_chat_completion_with_openai_library():
     global server
     server.start()
-    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
     res = client.chat.completions.create(
         model="gpt-3.5-turbo-instruct",
         messages=[
@@ -170,7 +170,7 @@ def test_chat_completion_with_timings_per_token():
 def test_logprobs():
     global server
     server.start()
-    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
     res = client.chat.completions.create(
         model="gpt-3.5-turbo-instruct",
         temperature=0.0,
@@ -197,7 +197,7 @@ def test_logprobs():
 def test_logprobs_stream():
     global server
     server.start()
-    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
     res = client.chat.completions.create(
         model="gpt-3.5-turbo-instruct",
         temperature=0.0,

examples/server/tests/unit/test_completion.py

@@ -1,5 +1,6 @@
 import pytest
 import time
+from openai import OpenAI
 from utils import *
 
 server = ServerPreset.tinyllama2()
@@ -85,6 +86,40 @@ def test_completion_stream_vs_non_stream():
     assert content_stream == res_non_stream.body["content"]
 
 
+def test_completion_stream_with_openai_library():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
+    res = client.completions.create(
+        model="davinci-002",
+        prompt="I believe the meaning of life is",
+        max_tokens=8,
+    )
+    assert res.system_fingerprint is not None and res.system_fingerprint.startswith("b")
+    assert res.choices[0].finish_reason == "length"
+    assert res.choices[0].text is not None
+    assert match_regex("(going|bed)+", res.choices[0].text)
+
+
+def test_completion_with_openai_library():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
+    res = client.completions.create(
+        model="davinci-002",
+        prompt="I believe the meaning of life is",
+        max_tokens=8,
+        stream=True,
+    )
+    output_text = ''
+    for data in res:
+        choice = data.choices[0]
+        if choice.finish_reason is None:
+            assert choice.text is not None
+            output_text += choice.text
+    assert match_regex("(going|bed)+", output_text)
+
+
 @pytest.mark.parametrize("n_slots", [1, 2])
 def test_consistent_result_same_seed(n_slots: int):
     global server

examples/server/tests/unit/test_embedding.py

@@ -1,3 +1,5 @@
+import base64
+import struct
 import pytest
 from openai import OpenAI
 from utils import *
@@ -194,3 +196,42 @@ def test_embedding_usage_multiple():
     assert res.status_code == 200
     assert res.body['usage']['prompt_tokens'] == res.body['usage']['total_tokens']
     assert res.body['usage']['prompt_tokens'] == 2 * 9
+
+
+def test_embedding_openai_library_base64():
+    server.start()
+    test_input = "Test base64 embedding output"
+
+    # get embedding in default format
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": test_input
+    })
+    assert res.status_code == 200
+    vec0 = res.body["data"][0]["embedding"]
+
+    # get embedding in base64 format
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": test_input,
+        "encoding_format": "base64"
+    })
+
+    assert res.status_code == 200
+    assert "data" in res.body
+    assert len(res.body["data"]) == 1
+
+    embedding_data = res.body["data"][0]
+    assert "embedding" in embedding_data
+    assert isinstance(embedding_data["embedding"], str)
+
+    # Verify embedding is valid base64
+    decoded = base64.b64decode(embedding_data["embedding"])
+    # Verify decoded data can be converted back to float array
+    float_count = len(decoded) // 4  # 4 bytes per float
+    floats = struct.unpack(f'{float_count}f', decoded)
+    assert len(floats) > 0
+    assert all(isinstance(x, float) for x in floats)
+    assert len(floats) == len(vec0)
+
+    # make sure the decoded data is the same as the original
+    for x, y in zip(floats, vec0):
+        assert abs(x - y) < EPSILON

examples/server/utils.hpp

@@ -3,6 +3,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
+#include "common/base64.hpp"
 
 #ifndef NDEBUG
 // crash the server in debug mode, otherwise send an http 500 error
@@ -548,10 +549,49 @@ static bool server_sent_event(httplib::DataSink & sink, const char * event, cons
 // OAI utils
 //
 
-static json oaicompat_completion_params_parse(
-    const struct llama_model * model,
-    const json & body, /* openai api json semantics */
-    const std::string & chat_template) {
+static json oaicompat_completion_params_parse(const json & body) {
+    json llama_params;
+
+    if (!body.contains("prompt")) {
+        throw std::runtime_error("\"prompt\" is required");
+    }
+
+    // Handle "stop" field
+    if (body.contains("stop") && body.at("stop").is_string()) {
+        llama_params["stop"] = json::array({body.at("stop").get<std::string>()});
+    } else {
+        llama_params["stop"] = json_value(body, "stop", json::array());
+    }
+
+    // Handle "n" field
+    int n_choices = json_value(body, "n", 1);
+    if (n_choices != 1) {
+        throw std::runtime_error("Only one completion choice is allowed");
+    }
+
+    // Params supported by OAI but unsupported by llama.cpp
+    static const std::vector<std::string> unsupported_params { "best_of", "echo", "suffix" };
+    for (const auto & param : unsupported_params) {
+        if (body.contains(param)) {
+            throw std::runtime_error("Unsupported param: " + param);
+        }
+    }
+
+    // Copy remaining properties to llama_params
+    for (const auto & item : body.items()) {
+        // Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens"
+        if (!llama_params.contains(item.key()) || item.key() == "n_predict") {
+            llama_params[item.key()] = item.value();
+        }
+    }
+
+    return llama_params;
+}
+
+static json oaicompat_chat_completion_params_parse(
+        const struct llama_model * model,
+        const json & body, /* openai api json semantics */
+        const std::string & chat_template) {
     json llama_params;
 
     // Apply chat template to the list of messages
@@ -613,16 +653,31 @@ static json oaicompat_completion_params_parse(
     return llama_params;
 }
 
-static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
+static json format_embeddings_response_oaicompat(const json & request, const json & embeddings, bool use_base64 = false) {
     json data = json::array();
     int32_t n_tokens = 0;
     int i = 0;
     for (const auto & elem : embeddings) {
-        data.push_back(json{
-            {"embedding", json_value(elem, "embedding", json::array())},
-            {"index",     i++},
-            {"object",    "embedding"}
-        });
+        json embedding_obj;
+
+        if (use_base64) {
+            const auto& vec = json_value(elem, "embedding", json::array()).get<std::vector<float>>();
+            const char* data_ptr = reinterpret_cast<const char*>(vec.data());
+            size_t data_size = vec.size() * sizeof(float);
+            embedding_obj = {
+                {"embedding", base64::encode(data_ptr, data_size)},
+                {"index", i++},
+                {"object", "embedding"},
+                {"encoding_format", "base64"}
+            };
+        } else {
+            embedding_obj = {
+                {"embedding", json_value(elem, "embedding", json::array())},
+                {"index", i++},
+                {"object", "embedding"}
+            };
+        }
+        data.push_back(embedding_obj);
 
         n_tokens += json_value(elem, "tokens_evaluated", 0);
     }

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -2744,13 +2744,13 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     cl_image_format     img_fmt_1d;
     cl_image_desc       img_desc_1d;
     cl_buffer_region    region;
-    cl_mem              A_image1d;
-    cl_mem              B_image1d;
-    cl_mem              B_sub_buffer;
-    cl_mem              C_d;
+    cl_mem              A_image1d = nullptr;
+    cl_mem              B_image1d = nullptr;
+    cl_mem              B_sub_buffer = nullptr;
+    cl_mem              C_d = nullptr;
     // for B transpose
-    cl_mem B_d;
-    cl_mem B_d_input_image;
+    cl_mem B_d = nullptr;
+    cl_mem B_d_input_image = nullptr;
     // <--------------------------------------------> //
 
     // define matrix dimensions

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

@@ -145,6 +145,8 @@ class vk_perf_logger;
 #endif
 static void ggml_vk_destroy_buffer(vk_buffer& buf);
 
+static constexpr uint32_t mul_mat_vec_max_cols = 8;
+
 struct vk_device_struct {
     std::mutex mutex;
 
@@ -202,8 +204,8 @@ struct vk_device_struct {
     vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_mul_mat_vec_p021_f16_f32;
@@ -411,7 +413,7 @@ struct vk_op_unary_push_constants {
     uint32_t ne;
     uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
-    uint32_t d_offset;
+    uint32_t misalign_offsets;
     float param1; float param2;
     uint32_t ne0_012mp; uint32_t ne0_012L;
     uint32_t ne0_01mp;  uint32_t ne0_01L;
@@ -459,7 +461,7 @@ struct vk_op_binary_push_constants {
     uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
     uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23;
-    uint32_t d_offset;
+    uint32_t misalign_offsets;
     float param1; float param2; int32_t param3;
 };
 
@@ -546,7 +548,7 @@ struct vk_staging_memcpy {
 };
 
 struct vk_op_upscale_push_constants {
-    uint32_t ne; uint32_t d_offset;
+    uint32_t ne; uint32_t a_offset; uint32_t d_offset;
     uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13;
     float sf0; float sf1; float sf2; float sf3;
@@ -1404,10 +1406,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
         // spec constants and tile sizes for non-quant matmul/matmul_id
         l_warptile = { 256, 128, 256, 64 };
         m_warptile = { 256, 128, 128, 64 };
-        s_warptile = { 128,  32,  16, 64 };
+        s_warptile = { 128,  64,  64, 64 };
         l_wg_denoms = {128, 256, 1 };
         m_wg_denoms = {128, 128, 1 };
-        s_wg_denoms = { 32,  16, 1 };
+        s_wg_denoms = { 64,  64, 1 };
 
         // spec constants and tile sizes for quant matmul (non-Qi_K)
         l_warptile_mmq = { 256, 128, 256, 64 };
@@ -1855,53 +1857,60 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     // mul mat vec
 
-    // AMD GCN and Intel graphics cards perform best when the number of rows per shader is doubled
-    uint32_t rm = 1;
-    if ((device->vendor_id == VK_VENDOR_ID_AMD && device->subgroup_min_size == 64 && device->subgroup_max_size == 64) || device->vendor_id == VK_VENDOR_ID_INTEL)
-        rm = 2;
+    // the number of rows computed per shader depends on GPU model and quant
+    uint32_t rm_stdq = 1;
+    uint32_t rm_kq = 2;
+    if (device->vendor_id == VK_VENDOR_ID_AMD) {
+        if (device->subgroup_min_size == 64 && device->subgroup_max_size == 64) { // GCN
+            rm_stdq = 2;
+            rm_kq = 4;
+        }
+    } else if (device->vendor_id == VK_VENDOR_ID_INTEL)
+        rm_stdq = 2;
 
-    // computing additional rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+    for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
 
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
+    }
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true);
 
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -2012,11 +2021,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);
     if (device->float_controls_rte_fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
@@ -2887,9 +2896,10 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
 }
 
-static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols) {
     VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
     GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
+    GGML_ASSERT(num_cols >= 1 && num_cols <= mul_mat_vec_max_cols);
 
     switch (a_type) {
         case GGML_TYPE_F32:
@@ -2910,7 +2920,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
             return nullptr;
     }
 
-    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type];
+    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type][num_cols-1];
 }
 
 static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) {
@@ -3920,8 +3930,6 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const uint64_t ne12 = src1->ne[2];
     const uint64_t ne13 = src1->ne[3];
 
-    GGML_ASSERT(ne11 == 1);
-
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
     const uint64_t ne22 = dst->ne[2];
@@ -3930,6 +3938,11 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const uint64_t r2 = ne12 / ne02;
     const uint64_t r3 = ne13 / ne03;
 
+    // batch_n indicates that we need to compute a few vector results, and this assumes
+    // ne12 and ne13 are 1. It overloads the batch_strides to hold the row strides.
+    GGML_ASSERT(ne11 == 1 || ne12 * ne13 == 1);
+    bool batch_n = ne11 > 1;
+
     ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
@@ -3980,7 +3993,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
-    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type);
+    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11);
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
     GGML_ASSERT(dmmv != nullptr);
@@ -4052,8 +4065,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
     }
 
-    uint32_t stride_batch_x = ne00*ne01;
-    uint32_t stride_batch_y = ne10*ne11;
+    // For batch_n, the A matrix is the same for each batch, and B/D use the row stride as the batch stride
+    uint32_t stride_batch_x = batch_n ? 0 : ne00*ne01;
+    uint32_t stride_batch_y = batch_n ? ne10 : (ne10*ne11);
+    uint32_t stride_batch_d = batch_n ? ne20 : (ne20*ne21);
 
     if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) {
         stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
@@ -4076,7 +4091,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     // compute
     const vk_mat_vec_push_constants pc = {
         (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        stride_batch_x, stride_batch_y, (uint32_t)(ne20*ne21),
+        stride_batch_x, stride_batch_y, stride_batch_d,
         (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
     };
     ggml_vk_sync_buffers(subctx);
@@ -4256,7 +4271,10 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
     } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1 &&
                !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
         ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
-    } else if (dst->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
+    // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four)
+    // when ne12 and ne13 are one.
+    } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 1)) &&
+               (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
         ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun);
     } else {
         ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun);
@@ -5071,6 +5089,57 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     }
 }
 
+static uint32_t get_misalign_bytes(ggml_backend_vk_context * ctx, const ggml_tensor * t)
+{
+    return ((vk_tensor_offset(t) + t->view_offs) & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1));;
+}
+
+template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    GGML_UNUSED(p);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+    GGML_UNUSED(dst);
+    static_assert(!std::is_const<T>::value, "unexpected type");
+    GGML_ASSERT(!src0 || get_misalign_bytes(ctx, src0) == 0);
+    GGML_ASSERT(!src1 || get_misalign_bytes(ctx, src1) == 0);
+    GGML_ASSERT(!src2 || get_misalign_bytes(ctx, src2) == 0);
+    GGML_ASSERT(!dst  || get_misalign_bytes(ctx, dst) == 0);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    p.misalign_offsets = (a_offset << 16) | d_offset;
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    GGML_ASSERT(dst->op != GGML_OP_GET_ROWS || (a_offset == 0 && b_offset == 0 && d_offset == 0));
+
+    p.misalign_offsets = (a_offset << 16) | (b_offset << 8) | d_offset;
+
+    GGML_UNUSED(src2);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_upscale_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    p.a_offset = a_offset;
+    p.d_offset = d_offset;
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+}
+
 template<typename PC>
 static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
@@ -5174,8 +5243,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     }
 
     GGML_ASSERT(d_D != nullptr);
-    uint64_t d_buf_offset = ((vk_tensor_offset(dst) + dst->view_offs) / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
-    GGML_ASSERT(d_buf_offset == vk_tensor_offset(dst) || op == GGML_OP_CPY);  // NOLINT
+    uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
     if(!src0_uma) {
         d_X = src0_buf_ctx->dev_buffer;
         x_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
@@ -5191,6 +5259,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         z_buf_offset = vk_tensor_offset(src2) + src2->view_offs;
         GGML_ASSERT(d_Z != nullptr);
     }
+    // Compute misalignment offset for descriptors and store it in in push constants, then align the descriptor offsets.
+    init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, dst);
+    x_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    y_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    z_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
 
     if (op_supports_incontiguous) {
         x_sz = ggml_nbytes(src0);
@@ -5378,7 +5452,6 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
-    const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
 
     int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
     int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
@@ -5390,7 +5463,7 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t) dst->nb[3] /  dst_type_size,
-        d_offset,
+        0,
         0.0f, 0.0f, offset,
     }, dryrun);
 }
@@ -5594,7 +5667,7 @@ static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, c
     const float sf3 = (float)dst->ne[3] / src0->ne[3];
 
     ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
-        (uint32_t)ggml_nelements(dst), 0,
+        (uint32_t)ggml_nelements(dst), 0, 0,
         (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], (uint32_t)dst->ne[2],(uint32_t)dst->ne[3],
         sf0, sf1, sf2, sf3,
@@ -5704,13 +5777,12 @@ static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, co
 static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
-    const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
 
     ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        d_offset,
+        0,
         0.0f, 0.0f,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);

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

@@ -21,9 +21,9 @@ void main() {
     get_indices(idx, i00, i01, i02, i03);
 
     if (ox < p.ne10 && oy < p.ne11 && oz < p.ne12) {
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
     } else {
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]));
     }
 }
 

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

@@ -22,7 +22,7 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
 
         idx += num_threads;
     }

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

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
 }

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

@@ -30,12 +30,12 @@ void main() {
     const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
+    data_d[get_doffset() + dst_idx] = D_TYPE(is_src0 ? data_a[get_aoffset() + src0_idx] : data_b[get_boffset() + src1_idx]);
 #else
     if (is_src0) {
-        data_d[p.d_offset + dst_idx] = data_a[src0_idx];
+        data_d[get_doffset() + dst_idx] = data_a[get_aoffset() + src0_idx];
     } else {
-        data_d[p.d_offset + dst_idx] = data_b[src1_idx];
+        data_d[get_doffset() + dst_idx] = data_b[get_boffset() + src1_idx];
     }
 #endif
 }

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

@@ -19,9 +19,9 @@ void main() {
     if (idx + (num_iter-1)*num_threads < p.ne) {
         [[unroll]] for (uint i = 0; i < num_iter; ++i) {
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-            data_d[p.d_offset + idx] = D_TYPE(data_a[idx]);
+            data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
-            data_d[p.d_offset + idx] = data_a[idx];
+            data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
 #endif
             idx += num_threads;
         }
@@ -32,9 +32,9 @@ void main() {
             }
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-            data_d[p.d_offset + idx] = D_TYPE(data_a[idx]);
+            data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
-            data_d[p.d_offset + idx] = data_a[idx];
+            data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
 #endif
             idx += num_threads;
         }

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

@@ -13,8 +13,8 @@ void main() {
     }
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
 #else
-    data_d[p.d_offset + dst_idx(idx)] = data_a[src0_idx(idx)];
+    data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)];
 #endif
 }

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

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(cos(val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(cos(val));
 }

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

@@ -20,7 +20,7 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
 
         idx += num_threads;
     }

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

@@ -7,7 +7,7 @@ layout (push_constant) uniform parameter
     uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
-    uint d_offset;
+    uint misalign_offsets;
     float param1; float param2; int param3;
 } p;
 
@@ -22,6 +22,10 @@ uint get_idx() {
     return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 }
 
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
+uint get_doffset() { return p.misalign_offsets & 0xFF; }
+
 // mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
 uint fastmod(uint a, uint b) {
     if ((b & (b-1)) == 0) {

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

@@ -6,7 +6,7 @@ layout (push_constant) uniform parameter
     uint ne;
     uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
-    uint d_offset;
+    uint misalign_offsets;
     float param1; float param2;
 
     uint ne0_012mp; uint ne0_012L;
@@ -24,6 +24,9 @@ uint get_idx() {
     return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 }
 
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
 // see init_fastdiv_values in ggml-vulkan.cpp
 uint fastdiv(uint n, uint mp, uint L) {
     uint msbs, lsbs;

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

@@ -15,10 +15,10 @@ void main() {
         return;
     }
 
-    const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+    const uint i01 = data_b[get_boffset() + i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
 
-    const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
-    const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+    const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+    const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
     data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);

ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp

@@ -2,6 +2,7 @@
 
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_spirv_intrinsics: enable
+#extension GL_EXT_control_flow_attributes : require
 
 #if RTE16
 spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
@@ -23,40 +24,64 @@ layout (push_constant) uniform parameter
 
 #include "types.comp"
 
-#define BLOCK_SIZE 256
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
 
-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+const uint NUM_ITER = 512 / BLOCK_SIZE;
+
+layout(local_size_x_id = 0, local_size_y = 1, 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[];};
 
 void main() {
-    const uint i = gl_GlobalInvocationID.x;
-    if (i >= p.pelements) {
-        return;
-    }
-
-    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
-    const uint kx = i / ksize;
-    const uint kd = kx * ksize;
-    const uint ky = (i - kd) / p.OW;
-    const uint ix = i % p.OW;
+    const uint gidx = gl_GlobalInvocationID.x;
 
     const uint oh = gl_GlobalInvocationID.y;
     const uint batch = gl_GlobalInvocationID.z / p.IC;
     const uint ic = gl_GlobalInvocationID.z % p.IC;
 
-    const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
-    const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
-
-    const uint offset_dst =
-        ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
-        (ic * (p.KW * p.KH) + ky * p.KW + kx);
-
-    if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) {
-        data_d[offset_dst] = D_TYPE(0.0f);
-    } else {
-        const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
-        data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]);
+    A_TYPE values[NUM_ITER];
+    uint offset_dst[NUM_ITER];
+    [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
+        values[idx] = A_TYPE(0);
     }
+
+    [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
+
+        const uint i = gidx * NUM_ITER + idx;
+
+        const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+        const uint kx = i / ksize;
+        const uint kd = kx * ksize;
+        const uint ky = (i - kd) / p.OW;
+        const uint ix = i % p.OW;
+
+        const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
+        const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
+
+        offset_dst[idx] =
+            ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
+            (ic * (p.KW * p.KH) + ky * p.KW + kx);
+
+        if (i >= p.pelements) {
+            continue;
+        }
+
+        if (iih < p.IH && iiw < p.IW) {
+            const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
+            values[idx] = data_a[offset_src + iih * p.IW + iiw];
+        }
+    }
+
+    [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
+
+        const uint i = gidx * NUM_ITER + idx;
+
+        if (i >= p.pelements) {
+            continue;
+        }
+
+        data_d[offset_dst[idx]] = D_TYPE(values[idx]);
+    }
+
 }

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

@@ -20,7 +20,7 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
 
         idx += num_threads;
     }

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

@@ -9,9 +9,6 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-layout (constant_id = 1) const uint NUM_ROWS = 1;
-
 #if !defined(DATA_A_F32) && !defined(DATA_A_F16)
 #define K_PER_ITER 8
 #else
@@ -21,70 +18,70 @@ layout (constant_id = 1) const uint NUM_ROWS = 1;
 
 uint a_offset, b_offset, d_offset, y_offset;
 
-shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
-
-void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
+void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
 {
-    const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
-    const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
-    const uint iybs = col - col%QUANT_K; // y block start index
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
+        const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
+        const uint iybs = col - col%QUANT_K; // y block start index
 
 #if K_PER_ITER == 8
 #if QUANT_R == 2
-    const B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4];
-    const B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4];
-    const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
-    const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w);
+        const B_TYPE_VEC4 bv02 = data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4];
+        const B_TYPE_VEC4 bv13 = data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs + y_offset) / 4];
+        const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
+        const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w);
 #else
-    const vec4 bv0 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4]);
-    const vec4 bv1 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4 + 1]);
+        const vec4 bv0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4]);
+        const vec4 bv1 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4 + 1]);
 #endif
 #else
-    // Check if the second of the pair of elements is OOB, and don't fetch B or
-    // accumulate it. We still fetch a pair of elements for A, which is fine for
-    // quantized formats since they'll be within the same block. We should
-    // probably skip fetching the second element for F16/F32, but as of now we
-    // still do.
-    const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
+        // Check if the second of the pair of elements is OOB, and don't fetch B or
+        // accumulate it. We still fetch a pair of elements for A, which is fine for
+        // quantized formats since they'll be within the same block. We should
+        // probably skip fetching the second element for F16/F32, but as of now we
+        // still do.
+        const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
 
-    FLOAT_TYPE b0 = 0, b1 = 0;
-    b0 = FLOAT_TYPE(data_b[b_offset + iybs + iqs]);
-    if (!OOB) {
-        b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
-    }
+        FLOAT_TYPE b0 = 0, b1 = 0;
+        b0 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]);
+        if (!OOB) {
+            b1 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset]);
+        }
 #endif
-    uint ibi = first_row*p.ncols;
-    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib = (ibi + col)/QUANT_K; // block index
-        ibi += p.ncols;
+        uint ibi = first_row*p.ncols;
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib = (ibi + col)/QUANT_K; // block index
+            ibi += p.ncols;
 
 #if K_PER_ITER == 8
-        vec4 v = dequantize4(ib, iqs, a_offset);
-        vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
+            vec4 v = dequantize4(ib, iqs, a_offset);
+            vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
 
-        const vec2 dm = get_dm(ib, a_offset);
-        if (dm.y != 0) { // quant has min component
-            v = v * dm.x + dm.y;
-            v2 = v2 * dm.x + dm.y;
-        }
+            const vec2 dm = get_dm(ib, a_offset);
+            if (dm.y != 0) { // quant has min component
+                v = v * dm.x + dm.y;
+                v2 = v2 * dm.x + dm.y;
+            }
 
-        // matrix multiplication
-        FLOAT_TYPE rowtmp = dot(bv0, v);
-        rowtmp += dot(bv1, v2);
+            // matrix multiplication
+            FLOAT_TYPE rowtmp = dot(bv0, v);
+            rowtmp += dot(bv1, v2);
 
-        if (dm.y == 0)
-            rowtmp *= dm.x;
+            if (dm.y == 0)
+                rowtmp *= dm.x;
 
-        temp[n] += rowtmp;
+            temp[j][n] += rowtmp;
 #else
-        const vec2 v = dequantize(ib, iqs, a_offset);
+            const vec2 v = dequantize(ib, iqs, a_offset);
 
-        // matrix multiplication
-        temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]);
-        if (!OOB) {
-            temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
-        }
+            // matrix multiplication
+            temp[j][n] = fma(FLOAT_TYPE(v.x), b0, temp[j][n]);
+            if (!OOB) {
+                temp[j][n] = fma(FLOAT_TYPE(v.y), b1, temp[j][n]);
+            }
 #endif
+        }
     }
 }
 
@@ -96,10 +93,12 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 
     y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 
-    FLOAT_TYPE temp[NUM_ROWS];
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
 
-    for (uint i = 0; i < NUM_ROWS; ++i) {
-        temp[i] = FLOAT_TYPE(0);
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
     }
 
     uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
@@ -131,24 +130,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
         i++;
     }
 
-    // sum up partial sums and write back result
-    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        tmpsh[n][tid] = temp[n];
-    }
-    barrier();
-    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-                tmpsh[n][tid] += tmpsh[n][tid + s];
-            }
-        }
-        barrier();
-    }
-    if (tid == 0) {
-        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
-        }
-    }
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
 }
 
 void main() {

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

@@ -83,3 +83,36 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
             batch_idx * p.batch_stride_d;
 #endif
 }
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+layout (constant_id = 1) const uint NUM_ROWS = 1;
+layout (constant_id = 2) const uint NUM_COLS = 1;
+
+shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
+
+void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+    // sum up partial sums and write back result
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            tmpsh[j][n][tid] = temp[j][n];
+        }
+    }
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                    tmpsh[j][n][tid] += tmpsh[j][n][tid + s];
+                }
+            }
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]);
+            }
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -5,22 +5,11 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
-
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
     const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -38,76 +27,89 @@ void main() {
     const uint s_offset = 8*v_im;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y_idx = i * QUANT_K + y_offset;
 
-        f16vec2 d = data_a[ib0 + i].d;
-        const FLOAT_TYPE dall = d.x;
-        const FLOAT_TYPE dmin = d.y;
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            f16vec2 d = data_a[ib0 + i].d;
+            const FLOAT_TYPE dall = d.x;
+            const FLOAT_TYPE dmin = d.y;
 
-        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
-        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
-        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
-        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
-        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
-        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
-        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
-        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
+            uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
+            uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
 
-        uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
-        uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
+            uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
+            uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
+            uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
 
-        uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
-        uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
-        uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
-        uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
+            uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
+            uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
+            uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
+            uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
 
-        uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
-        uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
-        uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
-        uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
+            uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
+            uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
+            uvec2 qs0 =  uvec2(unpack8(qs0_u16));
+            uvec2 qs16 = uvec2(unpack8(qs16_u16));
 
-        uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
-        uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
-        uvec2 qs0 =  uvec2(unpack8(qs0_u16));
-        uvec2 qs16 = uvec2(unpack8(qs16_u16));
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0];
+                B_TYPE_VEC2 b16 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8];
+                B_TYPE_VEC2 b32 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16];
+                B_TYPE_VEC2 b48 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24];
+                B_TYPE_VEC2 b64 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32];
+                B_TYPE_VEC2 b80 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40];
+                B_TYPE_VEC2 b96 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48];
+                B_TYPE_VEC2 b112 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56];
 
-        FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
-        FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
-        [[unroll]] for (int l = 0; l < 2; ++l) {
-            sum1 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 0) & 3),
-                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
-                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 2) & 3),
-                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
-                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 4) & 3),
-                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
-                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 6) & 3),
-                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
-            sum2 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_hi4[0]),
-                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_hi4[1]),
-                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_hi4[2]),
-                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_hi4[3]),
-                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_hi4[0]),
-                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_hi4[1]),
-                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_hi4[2]),
-                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
+                FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
+                FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
+                [[unroll]] for (int l = 0; l < 2; ++l) {
+                    sum1 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 0) & 3),
+                           fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
+                           fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 2) & 3),
+                           fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
+                           fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 4) & 3),
+                           fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
+                           fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 6) & 3),
+                           fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
+                    sum2 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_hi4[0]),
+                           fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_hi4[1]),
+                           fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_hi4[2]),
+                           fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_hi4[3]),
+                           fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_hi4[0]),
+                           fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_hi4[1]),
+                           fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_hi4[2]),
+                           fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
+                }
+                temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
+            }
         }
-        temp = fma(dall, sum1, fma(-dmin, sum2, temp));
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
 
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
         }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        compute_outputs(first_row, p.stride_d - first_row);
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -5,22 +5,11 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
-
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
     const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -35,66 +24,80 @@ void main() {
 
     const uint8_t m = uint8_t(1 << (4 * v_im));
 
-    const uint l0 = 2*v_in;                                // 0...15
+    const uint l0 = 2*v_in;                                 // 0...15
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
 
     const uint s_shift = 4 * v_im;
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y_idx = i * QUANT_K + y_offset;
 
-        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
-        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
-        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
-        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
-        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
-        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
-        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
-        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
-        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
+            uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
+            uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
+            uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
+            uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
+            uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
+            uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
+            u8vec2 s0 = unpack8(s0_16);
+            u8vec2 s2 = unpack8(s2_16);
+            u8vec2 s4 = unpack8(s4_16);
+            u8vec2 s6 = unpack8(s6_16);
+            u8vec2 s8 = unpack8(s8_16);
+            u8vec2 s10 = unpack8(s10_16);
 
-        uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
-        uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
-        uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
-        uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
-        uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
-        uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
-        u8vec2 s0 = unpack8(s0_16);
-        u8vec2 s2 = unpack8(s2_16);
-        u8vec2 s4 = unpack8(s4_16);
-        u8vec2 s6 = unpack8(s6_16);
-        u8vec2 s8 = unpack8(s8_16);
-        u8vec2 s10 = unpack8(s10_16);
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
 
-        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-        [[unroll]] for (int l = 0; l < 2; ++l) {
-            sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b96[l])  * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b16[l])  * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b48[l])  * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
+                B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0];
+                B_TYPE_VEC2 b16 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8];
+                B_TYPE_VEC2 b32 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16];
+                B_TYPE_VEC2 b48 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24];
+                B_TYPE_VEC2 b64 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32];
+                B_TYPE_VEC2 b80 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40];
+                B_TYPE_VEC2 b96 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48];
+                B_TYPE_VEC2 b112 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56];
+
+                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+                [[unroll]] for (int l = 0; l < 2; ++l) {
+                    sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
+                          fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
+                          fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
+                          fma(FLOAT_TYPE(b96[l])  * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
+                          fma(FLOAT_TYPE(b16[l])  * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
+                          fma(FLOAT_TYPE(b48[l])  * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
+                          fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
+                          fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
+                }
+                temp[j][n] = fma(d, sum, temp[j][n]);
+            }
         }
-        temp = fma(d, sum, temp);
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
 
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
         }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        compute_outputs(first_row, p.stride_d - first_row);
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -6,22 +6,11 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
-
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
     const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -31,8 +20,8 @@ void main() {
 
     const uint step = 4;
 
-    const uint il = itid/step;                               // 0...3
-    const uint ir = itid - step*il;                          // 0...7 or 0...3
+    const uint il = itid/step;                      // 0...3
+    const uint ir = itid - step*il;                 // 0...7 or 0...3
     const uint n =  4;
 
     const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
@@ -42,90 +31,103 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 64*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y1_idx = i * QUANT_K + y_offset;
         const uint y2_idx = y1_idx + 128;
 
-        f16vec2 d = data_a[ib0 + i].d;
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            f16vec2 d = data_a[ib0 + i].d;
+            const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+            const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
 
-        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
-        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
-        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
-        uvec4 scale0 = uvec4(unpack8(scale0_u32));
-        uvec4 scale4 = uvec4(unpack8(scale4_u32));
-        uvec4 scale8 = uvec4(unpack8(scale8_u32));
+            uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+            uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+            uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+            uvec4 scale0 = uvec4(unpack8(scale0_u32));
+            uvec4 scale4 = uvec4(unpack8(scale4_u32));
+            uvec4 scale8 = uvec4(unpack8(scale8_u32));
 
-        const uint32_t sc0 = (  scale0.x       & 0x3f);
-        const uint32_t sc1 = (  scale0.y       & 0x3f);
-        const uint32_t sc2 = (  scale4.x       & 0x3f);
-        const uint32_t sc3 = (  scale4.y       & 0x3f);
-        const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
-        const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
-        const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
-        const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
+            const uint32_t sc0 = (  scale0.x       & 0x3f);
+            const uint32_t sc1 = (  scale0.y       & 0x3f);
+            const uint32_t sc2 = (  scale4.x       & 0x3f);
+            const uint32_t sc3 = (  scale4.y       & 0x3f);
+            const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
+            const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
+            const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
+            const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
 
-        uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
-        uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
+            uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
+            uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
 
-        uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
-        uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
-        uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
-        uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
+            uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
+            uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
 
-        uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
-        uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
-        uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
-        uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
+            uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
+            uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
+            uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
+            uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
 
-        const uint32_t q4_0  = qs0_lo4.x;
-        const uint32_t q4_1  = qs0_lo4.y;
-        const uint32_t q4_2  = qs0_lo4.z;
-        const uint32_t q4_3  = qs0_lo4.w;
-        const uint32_t q4_4  = qs0_hi4.x;
-        const uint32_t q4_5  = qs0_hi4.y;
-        const uint32_t q4_6  = qs0_hi4.z;
-        const uint32_t q4_7  = qs0_hi4.w;
-        const uint32_t q4_8  = qs64_lo4.x;
-        const uint32_t q4_9  = qs64_lo4.y;
-        const uint32_t q4_10 = qs64_lo4.z;
-        const uint32_t q4_11 = qs64_lo4.w;
-        const uint32_t q4_12 = qs64_hi4.x;
-        const uint32_t q4_13 = qs64_hi4.y;
-        const uint32_t q4_14 = qs64_hi4.z;
-        const uint32_t q4_15 = qs64_hi4.w;
+            const uint32_t q4_0  = qs0_lo4.x;
+            const uint32_t q4_1  = qs0_lo4.y;
+            const uint32_t q4_2  = qs0_lo4.z;
+            const uint32_t q4_3  = qs0_lo4.w;
+            const uint32_t q4_4  = qs0_hi4.x;
+            const uint32_t q4_5  = qs0_hi4.y;
+            const uint32_t q4_6  = qs0_hi4.z;
+            const uint32_t q4_7  = qs0_hi4.w;
+            const uint32_t q4_8  = qs64_lo4.x;
+            const uint32_t q4_9  = qs64_lo4.y;
+            const uint32_t q4_10 = qs64_lo4.z;
+            const uint32_t q4_11 = qs64_lo4.w;
+            const uint32_t q4_12 = qs64_hi4.x;
+            const uint32_t q4_13 = qs64_hi4.y;
+            const uint32_t q4_14 = qs64_hi4.z;
+            const uint32_t q4_15 = qs64_hi4.w;
 
-        B_TYPE_VEC4 by10 =  data_b_v4[(b_offset + y1_idx) / 4];
-        B_TYPE_VEC4 by132 = data_b_v4[(b_offset + y1_idx) / 4 + 8];
-        B_TYPE_VEC4 by20 =  data_b_v4[(b_offset + y2_idx) / 4];
-        B_TYPE_VEC4 by232 = data_b_v4[(b_offset + y2_idx) / 4 + 8];
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                B_TYPE_VEC4 by10 =  data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4];
+                B_TYPE_VEC4 by132 = data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8];
+                B_TYPE_VEC4 by20 =  data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4];
+                B_TYPE_VEC4 by232 = data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8];
 
-        const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
-        const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
-        const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x),      q4_8,  fma(FLOAT_TYPE(by20.y),  q4_9,  fma(FLOAT_TYPE(by20.z),  q4_10, FLOAT_TYPE(by20.w) *  q4_11)));
-        const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x),     q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
-        const FLOAT_TYPE smin =
-            fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
-            fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
-            fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
-            fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
-        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
-    }
-
-    tmp[gl_LocalInvocationID.x] = temp;
-
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+                const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
+                const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
+                const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x),      q4_8,  fma(FLOAT_TYPE(by20.y),  q4_9,  fma(FLOAT_TYPE(by20.z),  q4_10, FLOAT_TYPE(by20.w) *  q4_11)));
+                const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x),     q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
+                const FLOAT_TYPE smin =
+                    fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
+                    fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
+                    fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
+                    fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
+                temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
+            }
         }
-        barrier();
     }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -6,22 +6,11 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
-
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
     const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -39,122 +28,135 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 64*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y1_idx = i * QUANT_K + y_offset;
         const uint y2_idx = y1_idx + 128;
 
-        f16vec2 d = data_a[ib0 + i].d;
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            f16vec2 d = data_a[ib0 + i].d;
+            const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+            const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
 
-        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
-        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
-        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
-        uvec4 scale0 = uvec4(unpack8(scale0_u32));
-        uvec4 scale4 = uvec4(unpack8(scale4_u32));
-        uvec4 scale8 = uvec4(unpack8(scale8_u32));
+            uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+            uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+            uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+            uvec4 scale0 = uvec4(unpack8(scale0_u32));
+            uvec4 scale4 = uvec4(unpack8(scale4_u32));
+            uvec4 scale8 = uvec4(unpack8(scale8_u32));
 
-        const uint32_t sc0 = (  scale0.x       & 0x3f);
-        const uint32_t sc1 = (  scale0.y       & 0x3f);
-        const uint32_t sc2 = (  scale4.x       & 0x3f);
-        const uint32_t sc3 = (  scale4.y       & 0x3f);
-        const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
-        const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
-        const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
-        const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
+            const uint32_t sc0 = (  scale0.x       & 0x3f);
+            const uint32_t sc1 = (  scale0.y       & 0x3f);
+            const uint32_t sc2 = (  scale4.x       & 0x3f);
+            const uint32_t sc3 = (  scale4.y       & 0x3f);
+            const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
+            const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
+            const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
+            const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
 
-        uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
-        uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
+            uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
+            uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
 
-        uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
-        uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
-        uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
-        uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
+            uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
+            uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
 
-        uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
+            uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
 
-        uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
-        uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
-        uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
-        uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
+            uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
+            uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
+            uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
+            uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
 
-        qs0_16_u32_lo4 += qs0_16_lo4_offset16;
-        qs0_16_u32_hi4 += qs0_16_hi4_offset16;
-        qs64_80_u32_lo4 += qs64_80_lo4_offset16;
-        qs64_80_u32_hi4 += qs64_80_hi4_offset16;
+            qs0_16_u32_lo4 += qs0_16_lo4_offset16;
+            qs0_16_u32_hi4 += qs0_16_hi4_offset16;
+            qs64_80_u32_lo4 += qs64_80_lo4_offset16;
+            qs64_80_u32_hi4 += qs64_80_hi4_offset16;
 
-        uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
-        uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
-        uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
-        uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
+            uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
+            uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
+            uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
+            uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
 
-        const uint32_t q4_0  = qs0_16_lo4.x;
-        const uint32_t q4_1  = qs0_16_lo4.y;
-        const uint32_t q4_2  = qs0_16_lo4.z;
-        const uint32_t q4_3  = qs0_16_lo4.w;
-        const uint32_t q4_4  = qs0_16_hi4.x;
-        const uint32_t q4_5  = qs0_16_hi4.y;
-        const uint32_t q4_6  = qs0_16_hi4.z;
-        const uint32_t q4_7  = qs0_16_hi4.w;
-        const uint32_t q4_8  = qs64_80_lo4.x;
-        const uint32_t q4_9  = qs64_80_lo4.y;
-        const uint32_t q4_10 = qs64_80_lo4.z;
-        const uint32_t q4_11 = qs64_80_lo4.w;
-        const uint32_t q4_12 = qs64_80_hi4.x;
-        const uint32_t q4_13 = qs64_80_hi4.y;
-        const uint32_t q4_14 = qs64_80_hi4.z;
-        const uint32_t q4_15 = qs64_80_hi4.w;
+            const uint32_t q4_0  = qs0_16_lo4.x;
+            const uint32_t q4_1  = qs0_16_lo4.y;
+            const uint32_t q4_2  = qs0_16_lo4.z;
+            const uint32_t q4_3  = qs0_16_lo4.w;
+            const uint32_t q4_4  = qs0_16_hi4.x;
+            const uint32_t q4_5  = qs0_16_hi4.y;
+            const uint32_t q4_6  = qs0_16_hi4.z;
+            const uint32_t q4_7  = qs0_16_hi4.w;
+            const uint32_t q4_8  = qs64_80_lo4.x;
+            const uint32_t q4_9  = qs64_80_lo4.y;
+            const uint32_t q4_10 = qs64_80_lo4.z;
+            const uint32_t q4_11 = qs64_80_lo4.w;
+            const uint32_t q4_12 = qs64_80_hi4.x;
+            const uint32_t q4_13 = qs64_80_hi4.y;
+            const uint32_t q4_14 = qs64_80_hi4.z;
+            const uint32_t q4_15 = qs64_80_hi4.w;
 
-        B_TYPE_VEC2 by10 =  data_b_v2[(b_offset + y1_idx) / 2];
-        B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8];
-        B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16];
-        B_TYPE_VEC2 by148 = data_b_v2[(b_offset + y1_idx) / 2 + 24];
-        B_TYPE_VEC2 by20 =  data_b_v2[(b_offset + y2_idx) / 2];
-        B_TYPE_VEC2 by216 = data_b_v2[(b_offset + y2_idx) / 2 + 8];
-        B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
-        B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                B_TYPE_VEC2 by10 =  data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2];
+                B_TYPE_VEC2 by116 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 8];
+                B_TYPE_VEC2 by132 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16];
+                B_TYPE_VEC2 by148 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24];
+                B_TYPE_VEC2 by20 =  data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2];
+                B_TYPE_VEC2 by216 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 8];
+                B_TYPE_VEC2 by232 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16];
+                B_TYPE_VEC2 by248 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24];
 
-        const FLOAT_TYPE sx =
-          fma(FLOAT_TYPE(by10.x), q4_0,
-          fma(FLOAT_TYPE(by10.y), q4_1,
-          fma(FLOAT_TYPE(by116.x), q4_2,
-             FLOAT_TYPE(by116.y) * q4_3)));
-        const FLOAT_TYPE sy =
-          fma(FLOAT_TYPE(by132.x), q4_4,
-          fma(FLOAT_TYPE(by132.y), q4_5,
-          fma(FLOAT_TYPE(by148.x), q4_6,
-             FLOAT_TYPE(by148.y) * q4_7)));
-        const FLOAT_TYPE sz =
-          fma(FLOAT_TYPE(by20.x), q4_8,
-          fma(FLOAT_TYPE(by20.y), q4_9,
-          fma(FLOAT_TYPE(by216.x), q4_10,
-             FLOAT_TYPE(by216.y) * q4_11)));
-        const FLOAT_TYPE sw =
-          fma(FLOAT_TYPE(by232.x), q4_12,
-          fma(FLOAT_TYPE(by232.y), q4_13,
-          fma(FLOAT_TYPE(by248.x), q4_14,
-             FLOAT_TYPE(by248.y) * q4_15)));
-        const FLOAT_TYPE smin =
-          fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
-          fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
-          fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
-              (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
-        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
-    }
-
-    tmp[gl_LocalInvocationID.x] = temp;
-
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+                const FLOAT_TYPE sx =
+                  fma(FLOAT_TYPE(by10.x), q4_0,
+                  fma(FLOAT_TYPE(by10.y), q4_1,
+                  fma(FLOAT_TYPE(by116.x), q4_2,
+                     FLOAT_TYPE(by116.y) * q4_3)));
+                const FLOAT_TYPE sy =
+                  fma(FLOAT_TYPE(by132.x), q4_4,
+                  fma(FLOAT_TYPE(by132.y), q4_5,
+                  fma(FLOAT_TYPE(by148.x), q4_6,
+                     FLOAT_TYPE(by148.y) * q4_7)));
+                const FLOAT_TYPE sz =
+                  fma(FLOAT_TYPE(by20.x), q4_8,
+                  fma(FLOAT_TYPE(by20.y), q4_9,
+                  fma(FLOAT_TYPE(by216.x), q4_10,
+                     FLOAT_TYPE(by216.y) * q4_11)));
+                const FLOAT_TYPE sw =
+                  fma(FLOAT_TYPE(by232.x), q4_12,
+                  fma(FLOAT_TYPE(by232.y), q4_13,
+                  fma(FLOAT_TYPE(by248.x), q4_14,
+                     FLOAT_TYPE(by248.y) * q4_15)));
+                const FLOAT_TYPE smin =
+                  fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
+                  fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
+                  fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
+                      (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
+                temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
+            }
         }
-        barrier();
     }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -6,22 +6,11 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
-
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
     const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -42,69 +31,82 @@ void main() {
     const uint s_offset  =  8*v_im + is;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
-        const uint y_idx   = i * QUANT_K + y_offset;
+        const uint y_idx = i * QUANT_K + y_offset;
 
-        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
-        FLOAT_TYPE scales[4];
-        scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
-        scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
-        scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
-        scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
+            FLOAT_TYPE scales[4];
+            scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
+            scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
+            scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
+            scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
 
-        uint32_t ql0_u32 =  uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
-        uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
+            uint32_t ql0_u32 =  uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
+            uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
 
-        uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
-        uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
-        uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
-        uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
+            uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
+            uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
 
-        uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
-        uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
-        uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
-        uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
-        uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
+            uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
+            uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
+            uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
+            uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
+            uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
 
-        uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
-        uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
-        uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
-        uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
+            uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
+            uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
+            uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
+            uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
 
-        uvec4 q0 = uvec4(unpack8(q0_u32));
-        uvec4 q1 = uvec4(unpack8(q1_u32));
-        uvec4 q2 = uvec4(unpack8(q2_u32));
-        uvec4 q3 = uvec4(unpack8(q3_u32));
+            uvec4 q0 = uvec4(unpack8(q0_u32));
+            uvec4 q1 = uvec4(unpack8(q1_u32));
+            uvec4 q2 = uvec4(unpack8(q2_u32));
+            uvec4 q3 = uvec4(unpack8(q3_u32));
 
-        B_TYPE_VEC4 by0  = data_b_v4[(b_offset + y_idx) / 4];
-        B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8];
-        B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16];
-        B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24];
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                B_TYPE_VEC4 by0  = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4];
+                B_TYPE_VEC4 by32 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 8];
+                B_TYPE_VEC4 by64 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 16];
+                B_TYPE_VEC4 by96 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 24];
 
-        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-        [[unroll]] for (int l = 0; l < 4; ++l) {
-            sum = fma(FLOAT_TYPE(by0[l])  * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
-                  fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
-                  fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
-                  fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
+                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+                [[unroll]] for (int l = 0; l < 4; ++l) {
+                    sum = fma(FLOAT_TYPE(by0[l])  * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
+                          fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
+                          fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
+                          fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
+                }
+                temp[j][n] += sum * d;
+            }
         }
-        temp += sum * d;
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
-    // sum up partial sums and write back result
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
 
-    barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
         }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        compute_outputs(first_row, p.stride_d - first_row);
     }
 }

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

@@ -24,5 +24,5 @@ void main() {
 
     const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
 
-    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : 0.0f);
+    data_d[get_doffset() + dst_idx] = D_TYPE(is_src0 ? data_a[get_aoffset() + src0_idx] : 0.0f);
 }

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

@@ -22,5 +22,5 @@ void main() {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx_mod(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx_mod(idx)]);
 }

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

@@ -18,7 +18,7 @@ void main() {
             continue;
         }
 
-        data_d[p.d_offset + idx] = D_TYPE(FLOAT_TYPE(data_a[idx]) * FLOAT_TYPE(p.param1));
+        data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]) * FLOAT_TYPE(p.param1));
         idx += num_threads;
     }
 }

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

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(sin(val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sin(val));
 }

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

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val * val);
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val * val);
 }

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

@@ -2,7 +2,7 @@
 
 layout (push_constant) uniform parameter
 {
-    uint ne; uint d_offset;
+    uint ne; uint a_offset; uint d_offset;
     uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13;
     float sf0; float sf1; float sf2; float sf3;
@@ -32,5 +32,5 @@ void main() {
     const uint i02 = uint(i12 / p.sf2);
     const uint i03 = uint(i13 / p.sf3);
 
-    data_d[p.d_offset + idx] = D_TYPE(data_a[i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
+    data_d[p.d_offset + idx] = D_TYPE(data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
 }

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

@@ -78,7 +78,8 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
     }
 
     PROCESS_INFORMATION pi;
-    STARTUPINFOA si = { sizeof(STARTUPINFOA) };
+    STARTUPINFOA si = {};
+    si.cb = sizeof(STARTUPINFOA);
     si.dwFlags = STARTF_USESTDHANDLES;
     si.hStdOutput = stdout_write;
     si.hStdError = stderr_write;

tests/test-backend-ops.cpp

@@ -3937,7 +3937,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32000, 512, 1, 1}));
 
-    for (int bs : {1, 512}) {
+    for (int bs : {1, 2, 3, 4, 5, 8, 512}) {
         for (ggml_type type_a : all_types) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {
                 test_cases.emplace_back(new test_mul_mat(type_a, type_b, 4096, bs, 14336, {1,  1}, {1, 1}));
@@ -3945,6 +3945,18 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         }
     }
 
+    for (int K : {3, 5}) {
+        for (int IC : {256, 2560}) {
+            for (int IW_IH : {32, 64, 256}) {
+                if (IC == 2560 && IW_IH == 256) {
+                    // too big
+                    continue;
+                }
+                test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {IW_IH, IW_IH, IC, 1}, {K, K, IC, 1}, 1, 1, 1, 1, 1, 1, true));
+            }
+        }
+    }
+
     return test_cases;
 }