Merge pull request #1 from ggerganov/master

Merge mainline

.editorconfig

@@ -26,3 +26,6 @@ indent_size = 2
 
 [examples/llama.swiftui/llama.swiftui.xcodeproj/*]
 indent_style = tab
+
+[examples/cvector-generator/*.txt]
+insert_final_newline = unset

.github/pull_request_template.md

@@ -1,5 +1,7 @@
-- Self Reported Review Complexity:
-    - [ ] Review Complexity : Low
-    - [ ] Review Complexity : Medium
-    - [ ] Review Complexity : High
-- [ ] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
+
+
+- [x] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
+- Self-reported review complexity:
+  - [ ] Low
+  - [ ] Medium
+  - [ ] High

CMakeLists.txt

@@ -119,6 +119,7 @@ option(LLAMA_HIP_UMA                         "llama: use HIP unified memory arch
 option(LLAMA_VULKAN                          "llama: use Vulkan"                                OFF)
 option(LLAMA_VULKAN_CHECK_RESULTS            "llama: run Vulkan op checks"                      OFF)
 option(LLAMA_VULKAN_DEBUG                    "llama: enable Vulkan debug output"                OFF)
+option(LLAMA_VULKAN_MEMORY_DEBUG             "llama: enable Vulkan memory debug output"         OFF)
 option(LLAMA_VULKAN_VALIDATE                 "llama: enable Vulkan validation"                  OFF)
 option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"                          OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
@@ -534,6 +535,10 @@ if (LLAMA_VULKAN)
             add_compile_definitions(GGML_VULKAN_DEBUG)
         endif()
 
+        if (LLAMA_VULKAN_MEMORY_DEBUG)
+            add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
+        endif()
+
         if (LLAMA_VULKAN_VALIDATE)
             add_compile_definitions(GGML_VULKAN_VALIDATE)
         endif()
@@ -684,7 +689,8 @@ if (LLAMA_SYCL)
     endif()
 
     set(GGML_HEADERS_SYCL ggml-sycl.h)
-    set(GGML_SOURCES_SYCL ggml-sycl.cpp)
+    file(GLOB GGML_SOURCES_SYCL "ggml-sycl/*.cpp")
+    list(APPEND GGML_SOURCES_SYCL "ggml-sycl.cpp")
 
     if (WIN32)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl sycl7 OpenCL mkl_sycl_blas_dll.lib mkl_intel_ilp64_dll.lib mkl_sequential_dll.lib mkl_core_dll.lib)

Makefile

@@ -38,6 +38,7 @@ BUILD_TARGETS = \
 	llama-tokenize \
 	llama-train-text-from-scratch \
 	llama-vdot \
+	llama-cvector-generator \
 	tests/test-c.o
 
 # Binaries only useful for tests
@@ -506,7 +507,7 @@ ifdef LLAMA_CUDA
 		CUDA_PATH ?= /usr/local/cuda
 	endif
 	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
 	OBJS         += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
 	OBJS         += $(OBJS_CUDA_TEMP_INST)
@@ -607,6 +608,10 @@ ifdef LLAMA_VULKAN_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_DEBUG
 endif
 
+ifdef LLAMA_VULKAN_MEMORY_DEBUG
+	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
+endif
+
 ifdef LLAMA_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
@@ -922,6 +927,10 @@ llama-eval-callback: examples/eval-callback/eval-callback.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

README-sycl.md

@@ -1,6 +1,7 @@
 # llama.cpp for SYCL
 
 - [Background](#background)
+- [Recommended Release](#recommended-release)
 - [News](#news)
 - [OS](#os)
 - [Hardware](#hardware)
@@ -31,8 +32,23 @@ When targeting **Intel CPU**, it is recommended to use llama.cpp for [Intel oneM
 
 It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, etc..*. In beginning work, the oneAPI's [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) open-source migration tool (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) was used for this purpose.
 
+## Recommended Release
+
+The SYCL backend would be broken by some PRs due to no online CI.
+
+The following release is verified with good quality:
+
+|Commit ID|Tag|Release|Verified  Platform|
+|-|-|-|-|
+|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1|
+
+
 ## News
 
+- 2024.5
+  - Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc770.
+  - Arch Linux is verified successfully.
+
 - 2024.4
   - Support data types: GGML_TYPE_IQ4_NL, GGML_TYPE_IQ4_XS, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ3_S, GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M.
 

README.md

@@ -195,6 +195,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
 - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
+- [RAGNA Desktop](https://ragna.app/) (proprietary)
 - [RecurseChat](https://recurse.chat/) (proprietary)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
@@ -208,6 +209,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
 - [AIKit](https://github.com/sozercan/aikit) (MIT)
+- [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
@@ -386,6 +388,30 @@ brew install llama.cpp
 ```
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
 
+### Nix
+
+On Mac and Linux, the Nix package manager can be used via
+```
+nix profile install nixpkgs#llama-cpp
+```
+For flake enabled installs.
+
+Or
+```
+nix-env --file '<nixpkgs>' --install --attr llama-cpp
+```
+For non-flake enabled installs.
+
+This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
+
+#### Flox
+
+On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
+```
+flox install llama-cpp
+```
+Flox follows the nixpkgs build of llama.cpp.
+
 ### Metal Build
 
 On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.

common/common.cpp

@@ -1576,6 +1576,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
             return true;
         }
         params.out_file = argv[i];
+        params.cvector_outfile = argv[i];
         return true;
     }
     if (arg == "-ofreq" || arg == "--output-frequency") {
@@ -1610,6 +1611,55 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.i_chunk = std::stoi(argv[i]);
         return true;
     }
+    // cvector params
+    if (arg == "--completions-file") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.cvector_completions_file = argv[i];
+        return true;
+    }
+    if (arg == "--positive-file") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.cvector_positive_file = argv[i];
+        return true;
+    }
+    if (arg == "--negative-file") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.cvector_negative_file = argv[i];
+        return true;
+    }
+    if (arg == "--completions") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.n_completions = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--pca-batch") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.n_pca_batch = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--pca-iter") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.n_pca_iterations = std::stoi(argv[i]);
+        return true;
+    }
 #ifndef LOG_DISABLE_LOGS
     // Parse args for logging parameters
     if (log_param_single_parse(argv[i])) {
@@ -1931,6 +1981,16 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "logging",     "       --log-append",           "Don't truncate the old log file." });
 #endif // LOG_DISABLE_LOGS
 
+    options.push_back({ "cvector" });
+    options.push_back({ "cvector",     "-o,    --output FNAME",         "output file (default: '%s')", params.cvector_outfile.c_str() });
+    options.push_back({ "cvector",     "       --positive-file FNAME",  "positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str() });
+    options.push_back({ "cvector",     "       --negative-file FNAME",  "negative prompts file, one prompt per line (default: '%s')", params.cvector_negative_file.c_str() });
+    options.push_back({ "cvector",     "       --completions-file FNAME",
+                                                                        "completions file (default: '%s')", params.cvector_completions_file.c_str() });
+    options.push_back({ "cvector",     "       --completions N",        "number of lines of completions file to use (default: %d)", params.n_completions });
+    options.push_back({ "cvector",     "       --batch-pca N",          "batch size used for PCA. Larger batch runs faster, but uses more memory (default: %d)", params.n_pca_batch });
+    options.push_back({ "cvector",     "       --iter-pca N",           "number of iterations used for PCA (default: %d)", params.n_pca_iterations });
+
     printf("usage: %s [options]\n", argv[0]);
 
     for (const auto & o : options) {

common/common.h

@@ -73,7 +73,6 @@ struct gpt_params {
     int32_t n_gpu_layers_draft    =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
     int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
     float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
-    int32_t n_beams               =     0; // if non-zero then use beam search of given width.
     int32_t grp_attn_n            =     1; // group-attention factor
     int32_t grp_attn_w            =   512; // group-attention width
     int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
@@ -232,6 +231,15 @@ struct gpt_params {
 
     bool process_output = false; // collect data for the output tensor
     bool compute_ppl    = true;  // whether to compute perplexity
+
+    // cvector-generator params
+    int n_completions = 64;
+    int n_pca_batch = 20;
+    int n_pca_iterations = 1000;
+    std::string cvector_outfile          = "control_vector.gguf";
+    std::string cvector_completions_file = "examples/cvector-generator/completions.txt";
+    std::string cvector_positive_file    = "examples/cvector-generator/positive.txt";
+    std::string cvector_negative_file    = "examples/cvector-generator/negative.txt";
 };
 
 void gpt_params_handle_model_default(gpt_params & params);

convert-hf-to-gguf.py

@@ -1632,6 +1632,12 @@ class Qwen2MoeModel(Model):
         super().set_gguf_parameters()
         if (n_experts := self.hparams.get("num_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+            logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
+        if (shared_expert_intermediate_size := self.hparams.get('shared_expert_intermediate_size')) is not None:
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size)
+            logger.info(f"gguf: expert shared feed forward length = {shared_expert_intermediate_size}")
 
     _experts: list[dict[str, Tensor]] | None = None
 

examples/CMakeLists.txt

@@ -12,6 +12,7 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
+    add_subdirectory(cvector-generator)
     add_subdirectory(baby-llama)
     add_subdirectory(batched-bench)
     add_subdirectory(batched)

examples/cvector-generator/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-cvector-generator)
+add_executable(${TARGET} cvector-generator.cpp pca.hpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/cvector-generator/README.md

@@ -0,0 +1,34 @@
+# cvector-generator
+
+This example demonstrates how to generate a control vector using gguf models.
+
+Related PRs:
+- [Add support for control vectors](https://github.com/ggerganov/llama.cpp/pull/5970)
+- (Issue) [Generate control vector using llama.cpp](https://github.com/ggerganov/llama.cpp/issues/6880)
+- [Add cvector-generator example](https://github.com/ggerganov/llama.cpp/pull/7514)
+
+## Examples
+
+```sh
+# CPU only
+./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf
+
+# With GPU
+./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99
+
+# With advanced options
+./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99 --completions 128 --pca-iter 2000 --batch-pca 100
+
+# To see help message
+./cvector-generator -h
+# Then, have a look at "cvector" section
+```
+
+## Tips and tricks
+
+If you have multiple lines per prompt, you can escape the newline character (change it to `\n`). For example:
+
+```
+<|im_start|>system\nAct like a person who is extremely happy.<|im_end|>
+<|im_start|>system\nYou are in a very good mood today<|im_end|>
+```

examples/cvector-generator/completions.txt

@@ -0,0 +1,582 @@
+
+That game
+I can see
+Hmm, this
+I can relate to
+Who is
+I understand the
+Ugh,
+What the hell was
+Hey, did anyone
+Although
+Thank you for choosing
+What are you
+Oh w
+How dare you open
+It was my pleasure
+I'm hon
+I appreciate that you
+Are you k
+Whoever left this
+It's always
+Ew,
+Hey, I l
+Hello? Is someone
+I understand that
+That poem
+Aww, poor
+Hey, it
+Alright, who
+I didn't
+Well, life
+The document
+Oh no, this
+I'm concerned
+Hello, this is
+This art
+Hmm, this drink
+Hi there!
+It seems
+Is
+Good
+I can't
+Ex
+Who are
+I can see that
+Wow,
+Today is a
+Hey friend
+Sometimes friends
+Oh, this old
+The weather outside
+This place is sur
+I appreciate your input
+Thank you for the
+Look at
+I'm disappoint
+To my
+How dare you
+That's an
+This piece of art
+Eww
+This park is
+This is incredible
+Oh no, someone
+Exc
+Well, it'
+I warned
+Hey, I understand
+Hey, I saw
+How dare you go
+What the he
+Hey
+It's
+Hello? Hello?
+It
+Oh no!
+This is the perfect
+Good morning,
+Oh no, there
+It's so
+Yeah
+Uh,
+Hello everyone
+Who turned off
+The weather
+Who'
+Hey, this
+Wait,
+Eww, gross
+Excuse
+It seems like you
+Thank you so
+What happened?
+Oh my g
+I am deeply sad
+I war
+Okay, let'
+Hey, that
+That was a beautiful
+Oh no! That
+What happened
+Hey there
+The artist'
+What?!
+Hey, it'
+I am disappoint
+It seems like
+Oh no! The
+This park is a
+If you
+Yes! I did
+It sounds
+What
+Who is it
+Hmm, that
+That's strange
+Yeah, that was
+That's interesting
+This park
+What the hell
+Who is that
+I feel like my
+Oh well
+What the hell is
+Hello? Hello
+To my dearest
+Bless you!\"
+Thank you for
+Oh, looks like
+Can you please
+This place is
+Eww, what
+Bless you
+Is everything
+Hey, I just
+Whoever left these
+Well, that'
+I feel
+Hey, do you
+It's sad
+Oh no, it
+Hey, that'
+Oh my god,
+Thank you,
+Hello little one,
+I apolog
+Hey team, I
+How dare you read
+Who is this and
+Whoever left
+Hi there! W
+A
+If you have
+I was
+U
+Bless
+Well, this
+Oh, I'
+It's a
+Eww,
+Is everything okay?
+Oh, I
+Hello, can you
+Al
+That was a great
+What are
+I understand that not
+Oh no, not
+Who is it?\"
+Hey, can we
+Whoever is taking
+I would love to
+Hey, I noticed
+Hey, could
+I understand that there
+Hello?
+D
+Oh man, I
+Thank you so much
+Oh no, my
+Dear [Name
+Uh
+I remember
+Hey, who
+Well, it
+Are you
+I understand that it
+Hey, is
+I would
+Who is this
+Excuse me
+Alright
+I am thrilled
+Sometimes friends have
+Who the
+It's interesting
+I would love
+E
+Hello? Is anyone
+Well, this is
+This place
+Well,
+I warned you
+Hey, watch where
+Oh my
+That'
+Sometimes friends have different
+I understand that everyone
+What?
+What do these notes
+I can relate
+I'm not
+I understand
+To my dear
+Guys
+Well
+Hey, I appreciate
+Wow, what
+Dear
+That melody
+Who the hell
+Today is
+Hello little
+Wow, look
+That's great
+Love is never wrong
+I'm having
+Whoa, did
+Ugh
+Can you please provide
+I miss you,
+I feel uncom
+I know
+Ugh, this
+Hey, watch
+Oh great, a
+I didn
+Okay
+That game of char
+Oh
+I appreciate
+Who's there
+I am so
+Oh great, someone
+Hey, could you
+I remember wondering
+Wait, what?
+What do
+Hello? Can
+Hey there,
+That game of
+This is incred
+Oh my gosh
+Oh great, f
+I appreciate your
+It sounds like
+What the heck
+Okay, I understand
+Ew
+I understand that this
+Uh, hi
+Hi everyone!
+What the hell?
+Thank you for your
+Oh no, the
+Wow, I
+Who turned
+Dear [
+Whoever
+This is a
+Whoa, he
+What in the world
+Although the physical
+Hello, who is
+That's amaz
+Hey, I know
+Okay, that
+Hi everyone
+Hey, is everything
+I understand your fr
+Oh no, poor
+Oh, look
+Good morning
+Ew, gross
+Oh no, did
+Look at the family
+Hey team
+Yes!
+Hey, can I
+Okay, that'
+It's great
+Love is
+Hey, what
+Good morning, world
+Who is it?
+That poem really reson
+I
+That's
+I understand the task
+Gu
+Hello? Who'
+This postcard is
+Whoa,
+Oh, that
+I understand that I
+Whoever is
+Hello? Who is
+I'm really
+Wow, this
+Can
+This artwork really
+This is a shame
+I miss you too
+Who are you?
+Today is a difficult
+Hey, just
+Are you okay
+I am
+Hi,
+Wow, that
+Hey there! Can
+Okay, stay
+Oh great, just
+Yeah,
+Hello? Can you
+Oh, looks
+Thank you for sharing
+I'm glad
+Hey, is that
+Hmm
+It was my
+It sounds like you
+Wow, your
+I was promised certain
+That was such a
+Thank
+Excuse you
+That was
+Hey team,
+I feel un
+It was
+What'
+Hey friend, I
+How
+Saying goodbye
+That
+It's heart
+How dare
+Oh,
+Hello, may
+What's this
+Thank you for recogn
+Aww, that
+Oh, I remember
+Hmm, that'
+I miss
+I know this
+Wait
+Is everything okay
+Who is that person
+Wow, you
+Oh great
+I'm sad
+Wow, the
+I am very disappoint
+Who turned off the
+I understand that things
+I'm very
+Hi
+That's very
+Okay, I
+Oh no,
+Wow, there
+What's wrong
+I apologize for
+Hey, I
+Can I help you
+Oh, I didn
+Alright,
+Oh wow,
+Oh my goodness
+I know this event
+What in the
+Saying
+Yeah, that
+Guys, I
+Hey, this v
+This post
+Are
+Hey, can
+Hello? Is
+I can only imagine
+Oh, that sounds
+Hey, is anyone
+I am disappointed
+Hello,
+Hey everyone, I
+That was such
+It's okay
+The artist
+Whoa
+I understand that mistakes
+Can I help
+Who
+Hi everyone! I
+Hey, can you
+Wow, how
+Today
+Oh no, I
+Oh well, I
+Well, that
+This is the
+Yes! I finally
+Hey there little
+Hello everyone!
+Love is never
+Look at the
+This postcard
+Oh great,
+Can I
+Hmm, this is
+I understand your
+Oh, look at
+B
+I'm so
+Whoa, this
+W
+Oh, this
+Sometimes
+This piece of
+What the
+That was a
+Hey, do
+Oh no
+Whoa, what
+I feel like I
+The documentary
+Hello
+Hello little one
+I understand that my
+Eww, that
+Wow, an
+Yes! Finally,
+Although the physical location
+Whoever is watching
+That movie
+I remember wondering about
+Hey there, little
+Who's
+Hello, who
+Hello everyone! Thank
+Hello, can
+That's too
+Hey, just wanted
+Hey there, I
+Saying good
+Hey there!
+Who is there?
+Oh my good
+I am very
+Oh no, what
+Wow, thank
+I was promised
+Hi, is
+Hey, I'
+Guys, the
+Oh no, that
+Who is there
+Hello, this
+That movie really touched
+If you have something
+The documentary was
+I'm starting
+Are you kidd
+That movie really
+Hey everyone,
+Thank you for considering
+I didn'
+Yes! I
+Can you
+Oh my god
+Hey, whoever
+That melody really
+Thank you, little
+Hello, may I
+Look
+Wow, we
+It looks
+What do these
+Oh wow
+I apologize
+What are you all
+It's such
+It's clear
+Hey, I was
+Hey friend,
+I can only
+The weather outside is
+Eww, this
+I miss you
+Wow
+Aww,
+Hi, is there
+This artwork
+Okay,
+Oh well,
+This
+I'
+Say
+Hey there little gu
+Hmm,
+Whoa, who
+I am thr
+Oh man
+Okay, stay calm
+I'm happy
+Oh, this cur
+Oh man,
+I'm sorry
+Hello? Who
+What?! That
+This piece
+Hey everyone
+That's so
+Are you okay?
+What happened? Where
+Hi there
+The
+Who the hell entered
+I can
+Guys,
+What's
+What in
+It's important
+I'm
+I'm coming
+It'
+Yes! Finally
+Wait, what
+Wow, reading
+I'm surprised
+Hey, did
+Hey,
+Okay, let
+I understand that you
+Who the hell threw
+Eww, who
+Thank you for thinking
+Who is this?\"
+I am deeply
+Thank you for including
+Oh no, an
+It looks like you
+Aww
+I'm confused
+Wow, it
+That poem really
+Yes
+Hey there, is
+Hey, what'
+Thank you for remember
+To
+This is
+Thank you for making
+I can'
+That mel
+Wow, they
+I feel like
+Although the
+Who are you
+Love
+If
+What the hell are
+I am so sad
+Oh, I found
+Thank you
+It looks like
+Well, life is
+I appreciate that
+The artist's
+Whoa, that
+It's never

examples/cvector-generator/cvector-generator.cpp

@@ -0,0 +1,499 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+#include "pca.hpp"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#include <cstdio>
+#include <string>
+#include <tuple>
+#include <vector>
+#include <algorithm>
+#include <iostream>
+#include <fstream>
+#include <climits>
+
+
+//////////////////////////////////////////////////
+// utils
+
+template <class Iter>
+static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
+    std::string ret;
+    for (; begin != end; ++begin) {
+        ret += llama_token_to_piece(ctx, *begin);
+    }
+
+    return ret;
+}
+
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
+
+    printf("\nexample usage:\n");
+    printf("\n    CPU only:   %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf\n", argv[0]);
+    printf("\n    with GPU:   %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99\n", argv[0]);
+    printf("\n    advanced:   %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99 --completions 128 --pca-iter 2000 --batch-pca 100\n", argv[0]);
+    printf("\n");
+}
+
+//////////////////////////////////////////////////
+
+
+// cb_eval is reused for each pair of positive - negative prompt
+struct callback_data {
+    ggml_context * ctx_ggml = nullptr;   // holds v_pos, v_neg, v_diff_filtered
+
+    int n_layers = 0;
+    int n_tokens = 0;
+    bool is_eval_pos = true;
+
+    // each element of the vector correspond to one layer
+    std::vector<struct ggml_tensor *> v_pos; // vector of matrices of size [n_embd, n_tokens]
+    std::vector<struct ggml_tensor *> v_neg; // vector of matrices of size [n_embd, n_tokens]
+    std::vector<struct ggml_tensor *> v_diff_filtered;   // vector of matrices of size [n_embd, n_nonzero_rows]. NOTE: n_nonzero_rows maybe different for each layer
+
+    // save a tensor into either v_pos or v_neg (decided by is_eval_pos)
+    void save_tensor_for_layer(struct ggml_tensor * t) {
+        GGML_ASSERT(t->type == GGML_TYPE_F32);
+
+        if (ctx_ggml == nullptr) {
+            // alloc a new ctx_ggml if needed
+            struct ggml_init_params params_ggml = {
+                /*.mem_size   =*/ ggml_tensor_overhead() * n_layers * 3u,
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+            ctx_ggml = ggml_init(params_ggml);
+        }
+
+        // copy tensor data
+        auto n_bytes = ggml_nbytes(t);
+        struct ggml_tensor * t_layer = ggml_new_tensor_2d(ctx_ggml, t->type, t->ne[0], t->ne[1]);
+        t_layer->data = malloc(n_bytes); // TODO @ngxson : get rid of this malloc somehow
+        ggml_backend_tensor_get(t, t_layer->data, 0, n_bytes);
+        ggml_set_name(t_layer, ggml_get_name(t));
+        //print_debug_tensor(t_layer);
+
+        if (is_eval_pos) {
+            v_pos.push_back(t_layer);
+        } else {
+            v_neg.push_back(t_layer);
+        }
+    }
+
+    // calculate diff (v_pos - v_neg) and place the result back to v_pos
+    // all zero rows in the diff tensor will also be removed
+    // NOTE: final layer is ignored. we only have (n_layers - 1) to process
+    std::vector<struct ggml_tensor *> calc_diff() {
+        for (float il = 0; il < v_pos.size(); il++) {
+            float * a = (float *) v_pos[il]->data;
+            float * b = (float *) v_neg[il]->data;
+            size_t n_elem = ggml_nelements(v_pos[il]);
+            for (size_t j = 0; j < n_elem; j++) {
+                a[j] -= b[j];
+            }
+            //print_debug_tensor(v_pos[i]);
+            auto diff_filtered = filter_nonzero_rows(v_pos[il]);
+            v_diff_filtered.push_back(diff_filtered);
+        }
+        return v_diff_filtered; // for convinient, we return the result std::vector
+    }
+
+    // delete zero rows from a given 2D tensor
+    struct ggml_tensor * filter_nonzero_rows(struct ggml_tensor * a) {
+        //printf("filter_nonzero_rows\n");
+        auto is_row_all_zeros = [](struct ggml_tensor * t, int row, float eps) -> bool {
+            // check if given row containing all zero elements
+            int n_cols = t->ne[0]; // hint: should be equal to n_embd
+            for (int col = 0; col < n_cols; ++col) {
+                if (ggml_get_f32_nd(t, col, row, 0, 0) > eps) {
+                    return false;
+                }
+            }
+            return true;
+        };
+        std::vector<int> rows_to_copy; // the idx of non-zero cols (to be copied to row of diff_filtered)
+        for (int i_row = 0; i_row < a->ne[1]; i_row++) {
+            if (!is_row_all_zeros(a, i_row, 1e-6)) {
+                rows_to_copy.push_back(i_row);
+            }
+        }
+
+        // get "n_nonzero_rows" for the output "diff_filtered"
+        int n_nonzero_rows = rows_to_copy.size();
+        //printf("n_nonzero_rows: %d\n", n_nonzero_rows);
+        int n_embd = a->ne[0];
+        GGML_ASSERT(n_nonzero_rows > 0);
+
+        // diff_filtered: [n_embd, n_nonzero_rows]
+        struct ggml_tensor * diff_filtered = ggml_new_tensor_2d(
+            ctx_ggml, GGML_TYPE_F32, n_embd, n_nonzero_rows);
+        ggml_format_name(diff_filtered, "diff_filtered_%s", a->name);
+        diff_filtered->data = malloc(ggml_nbytes(diff_filtered));
+
+        // copy non-zero rows
+        for (int dest_row = 0; dest_row < n_nonzero_rows; dest_row++) {
+            int src_row = rows_to_copy[dest_row];
+            for (int i = 0; i < n_embd; i++) {
+                float src_elem = ggml_get_f32_nd(a, i, src_row, 0, 0);
+                ggml_set_f32_nd(diff_filtered, i, dest_row, 0, 0, src_elem);
+            }
+        }
+
+        //print_debug_tensor(diff_filtered);
+
+        return diff_filtered;
+    }
+
+    // we don't implement destructor, because we want to reuse callback_data. we just want to free the tensors
+    void reset() {
+        for (auto ptr : v_pos) free(ptr->data);
+        for (auto ptr : v_neg) free(ptr->data);
+        for (auto ptr : v_diff_filtered) free(ptr->data);
+        v_pos.clear();
+        v_neg.clear();
+        v_diff_filtered.clear();
+        if (ctx_ggml) {
+            ggml_free(ctx_ggml);
+        }
+        ctx_ggml = nullptr;
+    }
+};
+
+/**
+ * process_ctx is used to store the ggml context for pre-post processing the diff vectors
+ * in short, input => v_diff and output => v_final
+ */
+struct train_context {
+    ggml_context * ctx_ggml;
+    int n_embd;
+    int n_layers;
+
+    /* pair of prompts to be used for generating final vector */
+    std::vector<std::string> positive_entries;
+    std::vector<std::string> negative_entries;
+
+    // each element of the vector correspond to one layer
+    // NOTE: the last layer is discard. therefore, we will have (n_layers - 1) elements here
+    // NOTE (2): v_diff is transposed from v_diff_tmp
+    std::vector<struct ggml_tensor *> v_diff;  // vector of matrices of size [m, n_embd] where m ~ n_tokens * n_completions (v_diff contains no zero-rows)
+    std::vector<struct ggml_tensor *> v_final; // vector of vectors of size [n_embd] to be written to file
+
+    // to easily re-alloc when concat v_diff, we temporary store v_diff in a vector instead of a tensor
+    // v_diff_tmp will get converted unto v_diff later on
+    std::vector<std::vector<uint8_t>> v_diff_tmp;
+
+    train_context(int n_embd_, int n_layers_) {
+        n_embd = n_embd_;
+        n_layers = n_layers_;
+        struct ggml_init_params params_ggml = {
+            /*.mem_size   =*/ ggml_tensor_overhead() * (n_layers - 1) * 2u,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ctx_ggml = ggml_init(params_ggml);
+        for (int il = 0; il < n_layers - 1; il++) {
+            std::vector<uint8_t> empty;
+            v_diff_tmp.push_back(empty);
+            auto t = ggml_new_tensor_1d(ctx_ggml, GGML_TYPE_F32, n_embd);
+            t->data = malloc(ggml_nbytes(t)); // TODO: get rid of malloc if possible
+            v_final.push_back(t);
+        }
+    }
+
+    // add new rows into existing tensor in v_diff_tmp
+    void concat_diff_tmp(const std::vector<struct ggml_tensor *> & diff_filtered) {
+        GGML_ASSERT((int) diff_filtered.size() == n_layers - 1);
+        for (int il = 0; il < n_layers - 1; il++) {
+            auto t = diff_filtered[il];
+            auto & diff_tmp = v_diff_tmp[il];
+            size_t curr_size = diff_tmp.size();
+            diff_tmp.resize(curr_size + ggml_nbytes(t));
+            memcpy(diff_tmp.data() + curr_size, t->data, ggml_nbytes(t));
+        }
+    }
+
+    // build the v_diff tensors from v_diff_tmp (v_diff need to be transposed)
+    // TODO @ngxson : maybe add option NOT to transpose v_diff; will be useful for "mean" method
+    void build_v_diff() {
+        printf("build_v_diff\n");
+        for (int il = 0; il < n_layers - 1; il++) {
+            auto & diff_tmp = v_diff_tmp[il];
+            int n_elem = diff_tmp.size() / sizeof(float);
+            GGML_ASSERT(n_elem % n_embd == 0);
+            int n_rows = n_elem / n_embd;
+            struct ggml_tensor * diff = ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_rows, n_embd);
+            ggml_set_name(diff, (std::string("diff_") + std::to_string(il)).c_str());
+            // copy data & transpose
+            diff->data = malloc(ggml_nbytes(diff)); // TODO: get rid of this malloc if possible
+            float * arr = (float *) diff_tmp.data();
+            for (int ir = 0; ir < n_rows; ++ir) {
+                for (int ic = 0; ic < n_embd; ++ic) {
+                    float f = arr[ir*n_embd + ic];
+                    ggml_set_f32_nd(diff, ir, ic, 0, 0, f);
+                }
+            }
+            v_diff.push_back(diff);
+            print_debug_tensor(diff);
+            // free memory of diff_tmp
+            diff_tmp.resize(0);
+        }
+    }
+
+    ~train_context() {
+        for (auto ptr : v_final) free(ptr->data);
+        for (auto ptr : v_diff) free(ptr->data);
+        // no need to free v_diff_tmp, since we didn't use malloc
+        ggml_free(ctx_ggml);
+    }
+};
+
+struct tokenized_prompt {
+    std::vector<llama_token> tokens_pos;
+    std::vector<llama_token> tokens_neg;
+    size_t max_seq_len;
+
+    tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
+        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+        tokens_pos = ::llama_tokenize(ctx, pos, add_bos);
+        tokens_neg = ::llama_tokenize(ctx, neg, add_bos);
+        max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
+        padding_seq(ctx, tokens_pos, max_seq_len);
+        padding_seq(ctx, tokens_neg, max_seq_len);
+    }
+
+    void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
+        // TODO: customize padding token
+        std::vector<llama_token> pad_tokens = ::llama_tokenize(ctx, " ", false);
+        llama_token pad_tok = pad_tokens.back();
+        while (tokens.size() < len) {
+            tokens.push_back(pad_tok);
+        }
+    }
+};
+
+//////////////////////////////////////////////////
+
+template <typename T>
+static std::string to_string(const T & val) {
+    std::stringstream ss;
+    ss << val;
+    return ss.str();
+}
+
+static std::vector<std::string> ctrlvec_load_prompt_file(std::string path, bool skip_empty_lines) {
+    std::vector<std::string> output;
+    std::ifstream file(path);
+    if (!file.is_open()) {
+        fprintf(stderr, "error: unable to open file: %s\n", path.c_str());
+        exit(1);
+    }
+    std::string line;
+    while (std::getline(file, line)) {
+        bool is_skip = skip_empty_lines && line.empty();
+        if (!is_skip) {
+            string_process_escapes(line);
+            output.push_back(line);
+        }
+    }
+    file.close();
+    return output;
+}
+
+//////////////////////////////////////////////////
+
+static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
+    auto * cb_data = (callback_data *) user_data;
+    static const char * l_out_name = "l_out";
+    const bool is_l_out = strncmp(t->name, l_out_name, strlen(l_out_name)) == 0;
+
+    if (ask) {
+        return is_l_out;
+    }
+
+    if (!is_l_out || t->ne[1] != cb_data->n_tokens) {
+        return true;
+    }
+
+    // save the tensor to current context
+    cb_data->save_tensor_for_layer(t);
+    return true;
+}
+
+static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
+    llama_kv_cache_clear(ctx);
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return false;
+    }
+    return true;
+}
+
+static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const std::string fname, const std::string model_hint) {
+    struct gguf_context * ctx = gguf_init_empty();
+
+    const std::string arch = "controlvector";
+    gguf_set_val_str(ctx, "general.architecture", arch.c_str());
+    gguf_set_val_str(ctx, (arch + ".model_hint").c_str(), model_hint.c_str());
+    gguf_set_val_i32(ctx, (arch + ".layer_count").c_str(), v_ctrl.size());
+
+    for (size_t i = 0; i < v_ctrl.size(); ++i) {
+        gguf_add_tensor(ctx, v_ctrl[i]);
+        print_debug_tensor(v_ctrl[i]);
+        printf("Added tensor: %s\n", v_ctrl[i]->name);
+    }
+
+    printf("%s: writing file...\n", __func__);
+    gguf_write_to_file(ctx, fname.c_str(), false);
+    printf("%s: wrote file '%s'\n", __func__, fname.c_str());
+    gguf_free(ctx);
+}
+
+/**
+ * Load prompt files and completion file.
+ * Then format each pair of prompt + completion to make an entry.
+ */
+static int prepare_entries(gpt_params & params, train_context & ctx_train) {
+    // load prompts
+    std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
+    std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
+    if (positive_prompts.size() != negative_prompts.size()) {
+        fprintf(stderr, "number of positive and negative prompts must be equal\n");
+        return 1;
+    }
+    if (positive_prompts.empty()) {
+        fprintf(stderr, "must provide at least one prompt pair\n");
+        return 1;
+    }
+
+    // create templated prompts
+    std::vector<std::string> completions = ctrlvec_load_prompt_file(params.cvector_completions_file, false);
+    auto format_template = [](std::string persona, std::string suffix) {
+        // entry in positive/negative.txt must already be formatted i.e. "[INST] Act as if you're extremely happy. [/INST]"
+        return persona + " " + suffix;
+    };
+    for (size_t i = 0; i < positive_prompts.size(); ++i) {
+        for (int j = 0; j < std::min((int) completions.size(), params.n_completions); ++j) {
+            // TODO replicate the truncations done by the python implementation
+            ctx_train.positive_entries.push_back(format_template(positive_prompts[i], completions[j]));
+            ctx_train.negative_entries.push_back(format_template(negative_prompts[i], completions[j]));
+        }
+    }
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
+        return 1;
+    }
+
+    if (params.n_pca_iterations % params.n_pca_batch != 0) {
+        fprintf(stderr, "PCA iterations must by multiply of PCA batch size\n");
+        return 1;
+    }
+
+
+    callback_data cb_data;
+
+    // pass the callback to the backend scheduler
+    // it will be executed for each node during the graph computation
+    params.cb_eval = cb_eval;
+    params.cb_eval_user_data = &cb_data;
+    params.warmup = false;
+
+    print_build_info();
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model to get hparams
+    llama_model * model;
+    llama_context * ctx;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+
+    // int n_ctx = llama_n_ctx(ctx);
+    int n_layers = llama_n_layer(model);
+    int n_embd = llama_n_embd(model);
+    // get model hint param (a.k.a model arch name)
+    char model_hint[128];
+    llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
+
+    // init train_context
+    train_context ctx_train(n_embd, n_layers);
+
+    // load and prepare entries for training
+    prepare_entries(params, ctx_train);
+
+    // we have to pretokenize everything because otherwise we don't know how much overhead to allocate ctx_diffs_wrapped
+    std::vector<tokenized_prompt> tokenized_prompts;
+    size_t n_total_tokens = 0;
+    for (size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
+        tokenized_prompt t(ctx, ctx_train.positive_entries[i], ctx_train.negative_entries[i]);
+        n_total_tokens += 2 * t.max_seq_len;
+        tokenized_prompts.push_back(std::move(t));
+    }
+
+    std::cout << "n_total_tokens: " << n_total_tokens << std::endl;
+
+    for(size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
+        bool success = false;
+        tokenized_prompt t = tokenized_prompts[i];
+        cb_data.n_layers = n_layers;
+        cb_data.n_tokens = t.max_seq_len;
+
+        printf("Evaluating prompt[%d/%d]: \"%s\" - \"%s\" (%d tokens)\n",
+            (int) i+1, (int) ctx_train.positive_entries.size(),
+            tokens_to_str(ctx, t.tokens_pos.cbegin(), t.tokens_pos.cend()).c_str(),
+            tokens_to_str(ctx, t.tokens_neg.cbegin(), t.tokens_neg.cend()).c_str(),
+            (int) t.max_seq_len);
+
+        cb_data.is_eval_pos = true;
+        success = get_hidden_layers(ctx, t.tokens_pos);
+        if (!success) break;
+
+        cb_data.is_eval_pos = false;
+        success = get_hidden_layers(ctx, t.tokens_neg);
+        if (!success) break;
+
+        // calculate diff and remove all zero rows
+        auto v_diff_filtered = cb_data.calc_diff();
+
+        // save & concat the filtered v_diff to ctx_train
+        ctx_train.concat_diff_tmp(v_diff_filtered);
+
+        // reset for next iteration
+        cb_data.reset();
+    }
+
+    // done with the model, we can now free it to make gain some memory
+    printf("Done evaluate prompts, unload model...\n");
+    llama_free(ctx);
+    llama_free_model(model);
+
+    // prepare ctx_train for PCA
+    ctx_train.build_v_diff();
+
+    // run PCA
+    PCA::pca_params pca_params;
+    pca_params.n_threads = params.n_threads;
+    pca_params.n_batch = params.n_pca_batch;
+    pca_params.n_iterations = params.n_pca_iterations;
+    PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
+
+    // write output vectors to gguf
+    export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/cvector-generator/negative.txt

@@ -0,0 +1 @@
+[INST] Act like a person who is extremely sad. [/INST]

examples/cvector-generator/pca.hpp

@@ -0,0 +1,322 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#include <cstdio>
+#include <ctime>
+#include <string>
+#include <tuple>
+#include <vector>
+#include <algorithm>
+#include <iostream>
+#include <fstream>
+
+#define DEBUG_POS 5
+
+static void print_debug_tensor(struct ggml_tensor * t, bool with_data = true) {
+    printf("%s: %s (%s): [%d, %d]\n", __func__, t->name, ggml_type_name(t->type), (int) t->ne[0], (int) t->ne[1]);
+    if (!with_data) return;
+    printf("%s: %s[0] = [", __func__, t->name);
+    for (size_t i = 0; i <= DEBUG_POS; i++) {
+        printf(" %f,", ggml_get_f32_nd(t, i, 0, 0, 0));
+    }
+    printf(" ... ]\n");
+}
+
+namespace PCA {
+
+// input params for PCA computations
+struct pca_params {
+    int n_threads = 1;
+    int n_batch = 20; // number of iterations do to in one batch. larger the batch, more memory is used
+    int n_iterations = 1000;
+    float tolerance = 1e-7;
+
+    // for debugging
+    int i_layer = 0;
+    int n_layers = 0;
+};
+
+// result from each iteration
+struct pca_result {
+    struct ggml_tensor * calculated_square = NULL;
+    std::vector<struct ggml_tensor *> eigenvectors;
+    std::vector<float> distances;
+};
+
+struct pca_model {
+    ggml_backend_t backend = NULL;
+    ggml_backend_buffer_t buffer;
+    struct ggml_context * ctx;      // context to compute graph on target device
+    struct ggml_context * ctx_host; // host context to store results
+
+    // tensors on target device
+    struct ggml_tensor * dev_input;
+    struct ggml_tensor * dev_square;
+    struct ggml_tensor * dev_eigenvector;
+
+    pca_model(struct ggml_tensor * t_input) {
+#ifdef GGML_USE_CUDA
+        fprintf(stderr, "%s: using CUDA backend\n", __func__);
+        backend = ggml_backend_cuda_init(0); // init device 0
+        if (!backend) {
+            fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+        }
+#endif
+
+// TODO: enable Metal support when support for GGML_OP_SQRT is added
+// #ifdef GGML_USE_METAL
+//         fprintf(stderr, "%s: using Metal backend\n", __func__);
+//         backend = ggml_backend_metal_init();
+//         if (!backend) {
+//             fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
+//         }
+// #endif
+
+        // if there aren't GPU Backends fallback to CPU backend
+        if (!backend) {
+            backend = ggml_backend_cpu_init();
+        }
+
+        const int num_tensors = 4;
+        struct ggml_init_params params {
+            /*.mem_size   =*/ ggml_tensor_overhead() * num_tensors,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ctx = ggml_init(params);
+
+        auto n_samples = t_input->ne[0];
+        auto n_embd    = t_input->ne[1];
+
+        dev_input       = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_samples, n_embd);
+        dev_square      = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,    n_embd);
+        dev_eigenvector = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        ggml_set_name(dev_input,       "dev_input");
+        ggml_set_name(dev_square,      "dev_square");
+        ggml_set_name(dev_eigenvector, "dev_eigenvector");
+        buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
+        ggml_backend_tensor_set(dev_input, t_input->data, 0, ggml_nbytes(t_input));
+
+        // initialize eigenvector to random normalized vector
+        {
+            std::vector<float> random_vec(ggml_nelements(dev_eigenvector), 0.0);
+            std::default_random_engine generator(static_cast<unsigned int>(std::time(0)));
+            std::uniform_real_distribution<float> distribution(0.0, 1.0);
+            float sum_sqr = 0.0; // for normalizing random_vec
+            for (size_t i = 0; i < random_vec.size(); ++i) {
+                float f = distribution(generator);
+                sum_sqr += f * f;
+                random_vec[i] = f;
+            }
+            // normalize it
+            float random_vec_norm = std::sqrt(sum_sqr);
+            for (size_t i = 0; i < random_vec.size(); ++i) {
+                random_vec[i] /= random_vec_norm;
+            }
+            ggml_backend_tensor_set(dev_eigenvector, random_vec.data(), 0, ggml_nbytes(dev_eigenvector));
+        }
+    }
+
+    ~pca_model() {
+        ggml_free(ctx);
+        ggml_backend_buffer_free(buffer);
+        ggml_backend_free(backend);
+    }
+};
+
+static struct ggml_cgraph * build_graph_piter(
+        const struct pca_params & params,
+        const pca_model & model,
+        bool calc_square = false) {
+    GGML_ASSERT(params.n_batch > 0);
+    // TODO: buf_size must be able to scale with params.n_batch
+    static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
+    static std::vector<uint8_t> buf(buf_size);
+
+    struct ggml_init_params params0 = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf.data(),
+        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_allocr_alloc_graph()
+    };
+    // create a temporally context to build the graph
+    struct ggml_context * ctx0 = ggml_init(params0);
+    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    // turn v_diff_original into square matrix if needed
+    struct ggml_tensor * tmp_square;
+    if (calc_square) {
+        tmp_square = ggml_mul_mat(ctx0, model.dev_input, model.dev_input);
+        ggml_set_name(tmp_square, "tmp_square");
+    }
+
+    struct ggml_tensor * b_tensor;
+    struct ggml_tensor * distance;
+    struct ggml_tensor * old_eigen    = model.dev_eigenvector;
+    struct ggml_tensor * input_square = calc_square ? tmp_square : model.dev_square;
+
+    for (int i = 0; i < params.n_batch; ++i) {
+        // b_tensor = square * eigenvector^T
+        b_tensor = ggml_mul_mat(ctx0, input_square, old_eigen);
+        ggml_set_name(b_tensor, "b_tensor");
+
+        // normalize
+        b_tensor = ggml_div_inplace(ctx0,
+            b_tensor,
+            ggml_sqrt_inplace(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, b_tensor)))
+        );
+        ggml_format_name(b_tensor, "b_tensor_norm_%d", i);
+
+        // calculate distance(new eigenvector - old eigenvector)
+        // we don't use ggml_sub because it may not be implemented on GPU backend
+        struct ggml_tensor * new_sub_old = ggml_add(ctx0, old_eigen, ggml_scale(ctx0, b_tensor, -1));
+        distance = ggml_sqrt_inplace(ctx0,
+            ggml_sum_rows(ctx0, ggml_sqr_inplace(ctx0, new_sub_old)));
+        ggml_format_name(distance, "distance_%d", i);
+
+        old_eigen = b_tensor;
+
+        // build operations nodes
+        ggml_build_forward_expand(gf, distance);
+    }
+
+    // delete the temporally context used to build the graph
+    ggml_free(ctx0);
+    return gf;
+}
+
+static ggml_status compute_piter(
+        const struct pca_params & params,
+        const pca_model & model,
+        struct ggml_cgraph * gf,
+        ggml_gallocr_t allocr,
+        struct pca_result & result) {
+    // allocate tensors
+    ggml_gallocr_alloc_graph(allocr, gf);
+
+    if (ggml_backend_is_cpu(model.backend)) {
+        ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
+    }
+
+// TODO: enable GPU support when support for GGML_OP_SQRT is added
+//#ifdef GGML_USE_METAL
+//    if (ggml_backend_is_metal(model.backend)) {
+//        ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
+//    }
+//#endif
+
+    ggml_status res = ggml_backend_graph_compute(model.backend, gf);
+    if (res == GGML_STATUS_SUCCESS) {
+        auto extract_i = [](std::string prefix, std::string str) -> int {
+            int i = -1;
+            if (str.rfind(prefix, 0) == 0) {
+                sscanf(str.c_str(), (prefix + "%d").c_str(), &i);
+            }
+            return i;
+        };
+        result.calculated_square = NULL;
+        result.eigenvectors.clear();
+        result.distances.clear();
+        result.eigenvectors.resize(params.n_batch);
+        result.distances.resize(params.n_batch);
+        // get output nodes
+        for (int i = 0; i < gf->n_nodes; ++i) {
+            auto node = gf->nodes[i];
+            int iter = -1;
+            // find b_tensor (without copying data from device)
+            if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {
+                result.eigenvectors[iter] = node;
+            }
+            // find distances, then copy data from device
+            if ((iter = extract_i("distance_", node->name)) > -1) {
+                float d;
+                ggml_backend_tensor_get(node, &d, 0, sizeof(float));
+                result.distances[iter] = d;
+                // std::cout << node->name << " = " << d << "\n";
+            }
+            // find tmp_square if it exists (without copying data from device)
+            if (std::string(node->name) == "tmp_square") {
+                result.calculated_square = node;
+            }
+        }
+    }
+    return res;
+}
+
+static void power_iteration(
+        const struct pca_params & params,
+        struct ggml_tensor * input, // shape of input: [n_samples, n_embd]
+        struct ggml_tensor * output) {
+    //printf("in power iteration\n");
+    struct pca_model model(input);
+
+    ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
+    struct pca_result result;
+    struct ggml_tensor * last_eigenvector = NULL;
+
+    int n_iters = params.n_iterations / params.n_batch; // more batch, fewer iterations
+    for (int iter = 0; iter < n_iters; ++iter) {
+        bool calc_square = (iter == 0); // only need to calculate square for first iteration
+        struct ggml_cgraph * gf = build_graph_piter(params, model, calc_square);
+        // ggml_graph_dump_dot(gf, nullptr, "/tmp/_cgraph.dot");
+        compute_piter(params, model, gf, allocr, result);
+
+        for (size_t k = 0; k < result.distances.size(); ++k) {
+            last_eigenvector = result.eigenvectors[k];
+            if (result.distances[k] < params.tolerance) {
+                break; // done
+            }
+        }
+
+        if (calc_square) {
+            // copy and store the square matrix if needed
+            GGML_ASSERT(result.calculated_square != NULL);
+            ggml_backend_tensor_copy(result.calculated_square, model.dev_square);
+        }
+
+        {
+            // copy last eigen vector and store as input for next iteration
+            GGML_ASSERT(last_eigenvector != NULL);
+            ggml_backend_tensor_copy(last_eigenvector, model.dev_eigenvector);
+        }
+
+        printf("%s: layer %d/%d, iteration: %d / total: %d (batch = %d) ...\n",
+            __func__, params.i_layer+1, params.n_layers, iter, n_iters, params.n_batch);
+    }
+
+    // get output tensor
+    GGML_ASSERT(last_eigenvector);
+    ggml_backend_tensor_get(last_eigenvector, output->data, 0, ggml_nbytes(last_eigenvector));
+    //print_debug_tensor(output);
+    ggml_gallocr_free(allocr);
+}
+
+static void run_pca(
+        struct pca_params & params,
+        const std::vector<struct ggml_tensor *> & v_input, // shape of v_input[0]: [n_samples, n_embd]
+        const std::vector<struct ggml_tensor *> & v_output) {
+    printf("%s: Running PCA...\n", __func__);
+    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);
+
+        // run power_iteration
+        params.i_layer = il;
+        params.n_layers = v_input.size();
+        power_iteration(params, v_input[il], ctrl_out);
+        printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size());
+    }
+}
+
+}

examples/cvector-generator/positive.txt

@@ -0,0 +1 @@
+[INST] Act like a person who is extremely happy. [/INST]

examples/infill/infill.cpp

@@ -223,7 +223,11 @@ int main(int argc, char ** argv) {
     inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
     embd_inp = inp_pfx;
     embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-    embd_inp.push_back(llama_token_middle(model));
+
+    const llama_token middle_token = llama_token_middle(model);
+    if (middle_token >= 0) {
+        embd_inp.push_back(middle_token);
+    }
 
     LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
     LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
@@ -528,7 +532,12 @@ int main(int argc, char ** argv) {
                 inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
                 embd_inp = inp_pfx;
                 embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-                embd_inp.push_back(llama_token_middle(model));
+
+                const llama_token middle_token = llama_token_middle(model);
+                if (middle_token >= 0) {
+                    embd_inp.push_back(middle_token);
+                }
+
                 embd.clear();
                 n_remain = params.n_predict;
                 n_past = 0;

examples/server/server.cpp

@@ -2040,7 +2040,12 @@ struct server_context {
                             prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
                             prefix_tokens.insert(prefix_tokens.end(),   llama_token_suffix(model));
                             prefix_tokens.insert(prefix_tokens.end(),   suffix_tokens.begin(), suffix_tokens.end());
-                            prefix_tokens.push_back(llama_token_middle(model));
+
+                            const llama_token middle_token = llama_token_middle(model);
+                            if (middle_token >= 0) {
+                                prefix_tokens.push_back(middle_token);
+                            }
+
                             prompt_tokens = prefix_tokens;
                         } else {
                             prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1717786204,
-        "narHash": "sha256-4q0s6m0GUcN7q+Y2DqD27iLvbcd1G50T2lv08kKxkSI=",
+        "lastModified": 1718318537,
+        "narHash": "sha256-4Zu0RYRcAY/VWuu6awwq4opuiD//ahpc2aFHg2CWqFY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "051f920625ab5aabe37c920346e3e69d7d34400e",
+        "rev": "e9ee548d90ff586a6471b4ae80ae9cfcbceb3420",
         "type": "github"
       },
       "original": {

ggml-backend.c

@@ -1172,7 +1172,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
             // check if a backend with higher prio wants to offload the op
             if (src_backend_id == sched->n_backends - 1) {
                 for (int b = 0; b < src_backend_id; b++) {
-                    if (ggml_backend_offload_op(sched->backends[b], tensor)) {
+                    if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");
                         return b;
                     }
@@ -1706,14 +1706,16 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     bool backend_ids_changed = false;
     for (int i = 0; i < sched->graph->n_nodes; i++) {
-        if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i]) {
+        if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i] &&
+            sched->bufts[sched->node_backend_ids[i]] != sched->bufts[sched->prev_node_backend_ids[i]]) {
             backend_ids_changed = true;
             break;
         }
     }
     if (!backend_ids_changed) {
         for (int i = 0; i < sched->graph->n_leafs; i++) {
-            if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i]) {
+            if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i] &&
+                sched->bufts[sched->leaf_backend_ids[i]] != sched->bufts[sched->prev_leaf_backend_ids[i]]) {
                 backend_ids_changed = true;
                 break;
             }
@@ -1977,6 +1979,15 @@ int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
     return sched->n_copies;
 }
 
+int ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched) {
+    return sched->n_backends;
+}
+
+ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i) {
+    GGML_ASSERT(i >= 0 && i < sched->n_backends);
+    return sched->backends[i];
+}
+
 size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
     int backend_index = ggml_backend_sched_backend_id(sched, backend);
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);

ggml-backend.h

@@ -182,6 +182,9 @@ extern "C" {
     // Initialize backend buffers from a measure graph
     GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
 
+    GGML_API int                  ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i);
+
     // Get the number of splits of the last graph
     GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
     GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);

ggml-cuda.cu

@@ -2267,6 +2267,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SQR:
             ggml_cuda_op_sqr(ctx, dst);
             break;
+        case GGML_OP_SQRT:
+            ggml_cuda_op_sqrt(ctx, dst);
+            break;
         case GGML_OP_CLAMP:
             ggml_cuda_op_clamp(ctx, dst);
             break;
@@ -2830,6 +2833,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_RMS_NORM:
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
+        case GGML_OP_SQRT:
         case GGML_OP_CLAMP:
         case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:

ggml-cuda/mmvq.cu

@@ -117,7 +117,7 @@ static __global__ void mul_mat_vec_q(
             tmp[j][i] = warp_reduce_sum(tmp[j][i]);
         }
 
-        if (threadIdx.x < rows_per_cuda_block) {
+        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < nrows_dst)) {
             dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
         }
     }

ggml-cuda/unary.cu

@@ -92,6 +92,15 @@ static __global__ void sqr_f32(const float * x, float * dst, const int k) {
     dst[i] = x[i] * x[i];
 }
 
+static __global__ void sqrt_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = sqrtf(x[i]);
+}
+
 static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
     gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -142,6 +151,11 @@ static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t
     sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void sqrt_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SQRT_BLOCK_SIZE - 1) / CUDA_SQRT_BLOCK_SIZE;
+    sqrt_f32<<<num_blocks, CUDA_SQRT_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;
@@ -284,3 +298,17 @@ void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     sqr_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
 }
+
+void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sqrt_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}

ggml-cuda/unary.cuh

@@ -8,6 +8,7 @@
 #define CUDA_HARDSIGMOID_BLOCK_SIZE 256
 #define CUDA_HARDSWISH_BLOCK_SIZE 256
 #define CUDA_SQR_BLOCK_SIZE 256
+#define CUDA_SQRT_BLOCK_SIZE 256
 
 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
@@ -28,3 +29,5 @@ void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-impl.h

@@ -17,7 +17,7 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
-#if defined(_WIN32)
+#if defined(_MSC_VER)
 
 #define m512bh(p) p
 #define m512i(p) p

ggml-rpc.cpp

@@ -73,9 +73,13 @@ struct rpc_tensor {
     uint64_t view_offs;
     uint64_t data;
     char name[GGML_MAX_NAME];
+
+    char padding[4];
 };
 #pragma pack(pop)
 
+static_assert(sizeof(rpc_tensor) % 8 == 0, "rpc_tensor size must be multiple of 8");
+
 // RPC commands
 enum rpc_cmd {
     ALLOC_BUFFER = 0,
@@ -599,9 +603,8 @@ static void serialize_graph(const ggml_cgraph * cgraph, std::vector<uint8_t> & o
     int output_size = sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(rpc_tensor);
     output.resize(output_size, 0);
     memcpy(output.data(), &n_nodes, sizeof(n_nodes));
-    uint64_t * out_nodes = (uint64_t *)(output.data() + sizeof(n_nodes));
     for (uint32_t i = 0; i < n_nodes; i++) {
-        out_nodes[i] = reinterpret_cast<uint64_t>(cgraph->nodes[i]);
+        memcpy(output.data() + sizeof(n_nodes) + i * sizeof(uint64_t), &cgraph->nodes[i], sizeof(uint64_t));
     }
     uint32_t * out_ntensors = (uint32_t *)(output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t));
     *out_ntensors = n_tensors;
@@ -1036,7 +1039,9 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, std::vector<u
     }
     std::unordered_map<uint64_t, ggml_tensor*> tensor_map;
     for (uint32_t i = 0; i < n_nodes; i++) {
-        graph->nodes[i] = create_node(nodes[i], ctx, tensor_ptrs, tensor_map);
+        int64_t id;
+        memcpy(&id, &nodes[i], sizeof(id));
+        graph->nodes[i] = create_node(id, ctx, tensor_ptrs, tensor_map);
     }
     ggml_status status = ggml_backend_graph_compute(backend, graph);
     // output serialization format: | status (1 byte) |

ggml-sycl.h

@@ -8,14 +8,12 @@
 
 #include "ggml.h"
 #include "ggml-backend.h"
+#include "ggml-sycl/presets.hpp"
 
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
-#define GGML_SYCL_MAX_DEVICES       48
-#define GGML_SYCL_NAME "SYCL"
-
 // backend API
 GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
 
@@ -33,13 +31,6 @@ GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
 GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
 GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
 GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
-
-// TODO: these are temporary
-//       ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index);
-GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id);
-GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode();
 
 // SYCL doesn't support registering host memory, keep here for reference
 // GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);

ggml-sycl/backend.hpp

@@ -0,0 +1,23 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_BACKEND_HPP
+#define GGML_SYCL_BACKEND_HPP
+
+#include "common.hpp"
+#include "convert.hpp"
+#include "dequantize.hpp"
+#include "dmmv.hpp"
+#include "mmq.hpp"
+#include "mmvq.hpp"
+
+#endif // GGML_SYCL_BACKEND_HPP

ggml-sycl/common.cpp

@@ -0,0 +1,53 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "common.hpp"
+
+int get_current_device_id() {
+  return dpct::dev_mgr::instance().current_device_id();
+}
+
+void* ggml_sycl_host_malloc(size_t size) try {
+  if (getenv("GGML_SYCL_NO_PINNED") != nullptr) {
+    return nullptr;
+  }
+
+  void* ptr = nullptr;
+  // allow to use dpct::get_in_order_queue() for host malloc
+  dpct::err0 err = CHECK_TRY_ERROR(
+      ptr = (void*)sycl::malloc_host(size, dpct::get_in_order_queue()));
+
+  if (err != 0) {
+    // clear the error
+    fprintf(
+        stderr,
+        "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
+        size / 1024.0 / 1024.0,
+        "syclGetErrorString is not supported");
+    return nullptr;
+  }
+
+  return ptr;
+} catch (sycl::exception const& exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+void ggml_sycl_host_free(void* ptr) try {
+  // allow to use dpct::get_in_order_queue() for host malloc
+  SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, dpct::get_in_order_queue())));
+} catch (sycl::exception const& exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}

ggml-sycl/common.hpp

@@ -0,0 +1,298 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_COMMON_HPP
+#define GGML_SYCL_COMMON_HPP
+
+#include <fstream>
+#include <iostream>
+
+#include "dpct/helper.hpp"
+#include "presets.hpp"
+
+#define GGML_COMMON_DECL_SYCL
+#define GGML_COMMON_IMPL_SYCL
+#include "ggml-common.h"
+
+void* ggml_sycl_host_malloc(size_t size);
+void ggml_sycl_host_free(void* ptr);
+
+static int g_ggml_sycl_debug = 0;
+#define GGML_SYCL_DEBUG(...)        \
+  do {                              \
+    if (g_ggml_sycl_debug)          \
+      fprintf(stderr, __VA_ARGS__); \
+  } while (0)
+
+#define CHECK_TRY_ERROR(expr)                                            \
+  [&]() {                                                                \
+    try {                                                                \
+      expr;                                                              \
+      return dpct::success;                                              \
+    } catch (std::exception const& e) {                                  \
+      std::cerr << e.what() << "\nException caught at file:" << __FILE__ \
+                << ", line:" << __LINE__ << ", func:" << __func__        \
+                << std::endl;                                            \
+      return dpct::default_error;                                        \
+    }                                                                    \
+  }()
+
+// #define DEBUG_SYCL_MALLOC
+
+static int g_work_group_size = 0;
+// typedef sycl::half ggml_fp16_t;
+
+#define __SYCL_ARCH__ DPCT_COMPATIBILITY_TEMP
+#define VER_4VEC 610 // todo for hardward optimize.
+#define VER_GEN9 700 // todo for hardward optimize.
+#define VER_GEN12 1000000 // todo for hardward optimize.
+#define VER_GEN13 (VER_GEN12 + 1030) // todo for hardward optimize.
+
+#define GGML_SYCL_MAX_NODES 8192 // TODO: adapt to hardwares
+
+// define for XMX in Intel GPU
+// TODO: currently, it's not used for XMX really.
+#if !defined(GGML_SYCL_FORCE_MMQ)
+    #define SYCL_USE_XMX
+#endif
+
+// max batch size to use MMQ kernels when tensor cores are available
+#define MMQ_MAX_BATCH_SIZE 32
+
+#if defined(_MSC_VER)
+#pragma warning(disable : 4244 4267) // possible loss of data
+#endif
+
+// dmmv = dequantize_mul_mat_vec
+#ifndef GGML_SYCL_DMMV_X
+#define GGML_SYCL_DMMV_X 32
+#endif
+#ifndef GGML_SYCL_MMV_Y
+#define GGML_SYCL_MMV_Y 1
+#endif
+
+typedef sycl::queue *queue_ptr;
+
+enum ggml_sycl_backend_gpu_mode {
+  SYCL_UNSET_GPU_MODE = -1,
+  SYCL_SINGLE_GPU_MODE = 0,
+  SYCL_MUL_GPU_MODE
+};
+
+static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size");
+
+static void crash() {
+  int* ptr = NULL;
+  *ptr = 0;
+}
+
+[[noreturn]] static void ggml_sycl_error(
+    const char* stmt,
+    const char* func,
+    const char* file,
+    const int line,
+    const char* msg) {
+  fprintf(stderr, "SYCL error: %s: %s\n", stmt, msg);
+  fprintf(stderr, "  in function %s at %s:%d\n", func, file, line);
+  GGML_ASSERT(!"SYCL error");
+}
+
+#define SYCL_CHECK(err)                     \
+  do {                                      \
+    auto err_ = (err);                      \
+    if (err_ != 0)                          \
+      ggml_sycl_error(                      \
+          #err,                             \
+          __func__,                         \
+          __FILE__,                         \
+          __LINE__,                         \
+          "Meet error in this line code!"); \
+  } while (0)
+
+#if DPCT_COMPAT_RT_VERSION >= 11100
+#define GGML_SYCL_ASSUME(x) __builtin_assume(x)
+#else
+#define GGML_SYCL_ASSUME(x)
+#endif // DPCT_COMPAT_RT_VERSION >= 11100
+
+#ifdef GGML_SYCL_F16
+typedef sycl::half dfloat; // dequantize float
+typedef sycl::half2 dfloat2;
+#else
+typedef float dfloat; // dequantize float
+typedef sycl::float2 dfloat2;
+#endif // GGML_SYCL_F16
+
+#define MMVQ_MAX_BATCH_SIZE  8
+
+static const int8_t kvalues_iq4nl[16]={-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+
+static int g_all_sycl_device_count = -1;
+static bool g_ggml_backend_sycl_buffer_type_initialized = false;
+
+static ggml_sycl_backend_gpu_mode g_ggml_sycl_backend_gpu_mode =
+    SYCL_UNSET_GPU_MODE;
+
+static void* g_scratch_buffer = nullptr;
+static size_t g_scratch_size = 0; // disabled by default
+static size_t g_scratch_offset = 0;
+
+[[noreturn]] static inline void bad_arch(const sycl::stream& stream_ct1) {
+  stream_ct1 << "ERROR: ggml-sycl was compiled without support for the "
+                "current GPU architecture.\n";
+  // __trap();
+  std::exit(1);
+
+  (void)bad_arch; // suppress unused function warning
+}
+
+int get_current_device_id();
+
+inline dpct::err0 ggml_sycl_set_device(const int device) try {
+
+  int current_device_id;
+  SYCL_CHECK(CHECK_TRY_ERROR(current_device_id = get_current_device_id()));
+
+  // GGML_SYCL_DEBUG("ggml_sycl_set_device device_id=%d,
+  // current_device_id=%d\n", device, current_device);
+  if (device == current_device_id) {
+    return 0;
+  }
+
+  return CHECK_TRY_ERROR(dpct::select_device(device));
+} catch (sycl::exception const& exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  crash();
+  std::exit(1);
+}
+
+//////////////////////
+
+struct ggml_sycl_device_info {
+    int device_count;
+
+    struct sycl_device_info {
+        int     cc;                 // compute capability
+        // int     nsm;                // number of streaming multiprocessors
+        // size_t  smpb;               // max. shared memory per block
+        bool    vmm;                // virtual memory support
+        size_t  total_vram;
+    };
+
+    sycl_device_info devices[GGML_SYCL_MAX_DEVICES] = {};
+
+    std::array<float, GGML_SYCL_MAX_DEVICES> default_tensor_split = {};
+};
+
+const ggml_sycl_device_info & ggml_sycl_info();
+
+struct ggml_sycl_pool {
+    virtual ~ggml_sycl_pool() = default;
+
+    virtual void * alloc(size_t size, size_t * actual_size) = 0;
+    virtual void free(void * ptr, size_t size) = 0;
+};
+
+template<typename T>
+struct ggml_sycl_pool_alloc {
+    ggml_sycl_pool * pool = nullptr;
+    T * ptr = nullptr;
+    size_t actual_size = 0;
+
+    explicit ggml_sycl_pool_alloc(ggml_sycl_pool & pool) : pool(&pool) {
+    }
+
+    ggml_sycl_pool_alloc(ggml_sycl_pool & pool, size_t size) : pool(&pool) {
+        alloc(size);
+    }
+
+    ~ggml_sycl_pool_alloc() {
+        if (ptr != nullptr) {
+            pool->free(ptr, actual_size);
+        }
+    }
+
+    // size is in number of elements
+    T * alloc(size_t size) {
+        GGML_ASSERT(pool != nullptr);
+        GGML_ASSERT(ptr == nullptr);
+        ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size);
+        return ptr;
+    }
+
+    T * alloc(ggml_sycl_pool & pool, size_t size) {
+        this->pool = &pool;
+        return alloc(size);
+    }
+
+    T * get() {
+        return ptr;
+    }
+
+    ggml_sycl_pool_alloc() = default;
+    ggml_sycl_pool_alloc(const ggml_sycl_pool_alloc &) = delete;
+    ggml_sycl_pool_alloc(ggml_sycl_pool_alloc &&) = delete;
+    ggml_sycl_pool_alloc& operator=(const ggml_sycl_pool_alloc &) = delete;
+    ggml_sycl_pool_alloc& operator=(ggml_sycl_pool_alloc &&) = delete;
+};
+
+// backend interface
+
+struct ggml_tensor_extra_gpu {
+  void* data_device[GGML_SYCL_MAX_DEVICES]; // 1 pointer for each device for split
+                                       // tensors
+  dpct::event_ptr events[GGML_SYCL_MAX_DEVICES]
+                        [GGML_SYCL_MAX_STREAMS]; // events for synchronizing multiple GPUs
+};
+
+struct ggml_backend_sycl_context {
+    int device;
+    std::string name;
+
+    queue_ptr qptrs[GGML_SYCL_MAX_DEVICES][GGML_SYCL_MAX_STREAMS] = { { nullptr } };
+
+    explicit ggml_backend_sycl_context(int device) :
+        device(device),
+        name(GGML_SYCL_NAME + std::to_string(device)) {
+    }
+
+    queue_ptr stream(int device, int stream) {
+        if (qptrs[device][stream] == nullptr) {
+            qptrs[device][stream] = &(dpct::get_current_device().default_queue());
+        }
+        return qptrs[device][stream];
+    }
+
+    queue_ptr stream() {
+        return stream(device, 0);
+    }
+
+    // pool
+    std::unique_ptr<ggml_sycl_pool> pools[GGML_SYCL_MAX_DEVICES];
+
+    static std::unique_ptr<ggml_sycl_pool> new_pool_for_device(queue_ptr qptr, int device);
+
+    ggml_sycl_pool & pool(int device) {
+        if (pools[device] == nullptr) {
+            pools[device] = new_pool_for_device(stream(device,0), device);
+        }
+        return *pools[device];
+    }
+
+    ggml_sycl_pool & pool() {
+        return pool(device);
+    }
+};
+
+
+#endif // GGML_SYCL_COMMON_HPP

ggml-sycl/convert.cpp

@@ -0,0 +1,544 @@
+#include "convert.hpp"
+#include "dequantize.hpp"
+#include "presets.hpp"
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k,
+                             const sycl::nd_item<3> &item_ct1) {
+    const int i = 2 * (item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                       item_ct1.get_local_id(2));
+
+    if (i >= k) {
+        return;
+    }
+
+    const int ib = i/qk; // block index
+    const int iqs = (i%qk)/qr; // quant index
+    const int iybs = i - i%qk; // y block start index
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(vx, ib, iqs, v);
+
+    y[iybs + iqs + 0] = v.x();
+    y[iybs + iqs + y_offset] = v.y();
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_sycl(const void *__restrict__ vx,
+                                  dst_t *__restrict__ y, const int k,
+                                  dpct::queue_ptr stream) {
+    const int num_blocks = (k + 2*SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / (2*SYCL_DEQUANTIZE_BLOCK_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+        stream->parallel_for(
+            sycl::nd_range<3>(
+                sycl::range<3>(1, 1, num_blocks) *
+                    sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE),
+                sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                dequantize_block<qk, qr, dequantize_kernel>(vx, y, k, item_ct1);
+            });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k,
+                                     dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+#if QK_K == 256
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q2_K(vx, y, item_ct1);
+                             });
+    }
+#else
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q2_K(vx, y, item_ct1);
+                             });
+    }
+
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k,
+                                     dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+#if QK_K == 256
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q3_K(vx, y, item_ct1);
+                             });
+    }
+#else
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q3_K(vx, y, item_ct1);
+                             });
+    }
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int k,
+                                     dpct::queue_ptr stream) {
+    const int nb32 = k / 32;
+    const int nb = (k + 255) / 256;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q4_0(vx, y, nb32, item_ct1);
+                             });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int k,
+                                     dpct::queue_ptr stream) {
+    const int nb32 = k / 32;
+    const int nb = (k + 255) / 256;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q4_1(vx, y, nb32, item_ct1);
+                             });
+    }
+}
+
+
+template <typename dst_t>
+static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int k,
+                                     dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q4_K(vx, y, item_ct1);
+                             });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k,
+                                     dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+#if QK_K == 256
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q5_K(vx, y, item_ct1);
+                             });
+    }
+#else
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q5_K(vx, y, item_ct1);
+                             });
+    }
+
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k,
+                                     dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+#if QK_K == 256
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q6_K(vx, y, item_ct1);
+                             });
+    }
+#else
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q6_K(vx, y, item_ct1);
+                             });
+    }
+
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int k,
+                                        dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq1_s(
+                                     vx, y, item_ct1, iq1s_grid_gpu
+                                     );
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq1_m_sycl(const void *vx, dst_t *y, const int k,
+                                        dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq1_m(
+                                     vx, y, item_ct1, iq1s_grid_gpu
+                                     );
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int k,
+                                        dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq2_xxs(
+                                     vx, y, item_ct1, iq2xxs_grid,
+                                     ksigns_iq2xs, kmask_iq2xs);
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int k,
+                                       dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq2_xs(
+                                     vx, y, item_ct1, iq2xs_grid,
+                                     ksigns_iq2xs, kmask_iq2xs);
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq2_s_sycl(const void *vx, dst_t *y, const int k,
+                                      dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq2_s(vx, y, item_ct1);
+                             });
+        });
+    }
+}
+
+
+template <typename dst_t>
+static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int k,
+                                        dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq3_xxs(
+                                     vx, y, item_ct1, iq3xxs_grid,
+                                     ksigns_iq2xs, kmask_iq2xs);
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int k,
+                                        dpct::queue_ptr stream) {
+    const int nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq3_s(
+                                     vx, y, item_ct1, kmask_iq2xs, iq3s_grid);
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int k,
+                                       dpct::queue_ptr stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+#if QK_K == 64
+    dequantize_row_iq4_nl_sycl(vx, y, k, stream);
+#else
+      {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                  cgh.parallel_for(
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                            sycl::range<3>(1, 1, 32),
+                                        sycl::range<3>(1, 1, 32)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                            dequantize_block_iq4_xs(vx, y, item_ct1);
+                      });
+            });
+      }
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int k,
+                                       dpct::queue_ptr stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+      {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                  cgh.parallel_for(
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                            sycl::range<3>(1, 1, 32),
+                                        sycl::range<3>(1, 1, 32)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                            dequantize_block_iq4_nl(vx, y, item_ct1);
+                      });
+            });
+      }
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int k,
+                          const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+
+    const src_t * x = (src_t *) vx;
+
+    y[i] = x[i];
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary_sycl(const void *__restrict__ vx,
+                               dst_t *__restrict__ y, const int k,
+                               dpct::queue_ptr stream) {
+    const int num_blocks = (k + SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / SYCL_DEQUANTIZE_BLOCK_SIZE;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(
+                sycl::range<3>(1, 1, num_blocks) *
+                    sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE),
+                sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                convert_unary<src_t>(vx, y, k, item_ct1);
+            });
+    }
+}
+
+to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_block_sycl<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_sycl<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_sycl<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_sycl<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_sycl;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_sycl;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_sycl;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_sycl;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_sycl;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_sycl;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_sycl;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_sycl;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_sycl;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_sycl;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_sycl;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_sycl;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_sycl;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_sycl;
+        case GGML_TYPE_F32:
+            return convert_unary_sycl<float>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_row_q4_0_sycl;
+        case GGML_TYPE_Q4_1:
+            return dequantize_row_q4_1_sycl;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_sycl<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_sycl<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_sycl;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_sycl;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_sycl;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_sycl;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_sycl;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_sycl;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_sycl;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_sycl;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_sycl;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_sycl;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_sycl;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_sycl;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_sycl;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_sycl;
+        case GGML_TYPE_F16:
+            return convert_unary_sycl<sycl::half>;
+        default:
+            return nullptr;
+    }
+}

ggml-sycl/convert.hpp

@@ -0,0 +1,27 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_CONVERT_HPP
+#define GGML_SYCL_CONVERT_HPP
+
+#include "common.hpp"
+
+template <typename T>
+using to_t_sycl_t = void (*)(const void *__restrict__ x, T *__restrict__ y,
+                             int k, dpct::queue_ptr stream);
+typedef to_t_sycl_t<float> to_fp32_sycl_t;
+typedef to_t_sycl_t<sycl::half> to_fp16_sycl_t;
+
+to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type);
+to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type);
+
+#endif // GGML_SYCL_CONVERT_HPP

ggml-sycl/dequantize.hpp

@@ -0,0 +1,690 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_DEQUANTIZE_HPP
+#define GGML_SYCL_DEQUANTIZE_HPP
+
+#include "common.hpp"
+
+typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
+
+static __dpct_inline__ void dequantize_q4_0(const void *vx, const int ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x() = vui & 0xF;
+    v.y() = vui >> 4;
+
+#ifdef GGML_SYCL_F16
+    // v = v - {8.0f, 8.0f};
+    // v = v * {d, d};
+    v.s0() = (v.s0() - 8.0f) * d;
+    v.s1() = (v.s1() - 8.0f) * d;
+
+#else
+    v.x() = (v.x() - 8.0f) * d;
+    v.y() = (v.y() - 8.0f) * d;
+#endif // GGML_SYCL_F16
+}
+
+static __dpct_inline__ void dequantize_q4_1(const void *vx, const int ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const dfloat d = x[ib].dm[0];
+    const dfloat m = x[ib].dm[1];
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x() = vui & 0xF;
+    v.y() = vui >> 4;
+
+#ifdef GGML_SYCL_F16
+    // v = v * {d, d};
+    // v = v + {m, m};
+    v.s0() = (v.s0() * d) + m;
+    v.s1() = (v.s1() * d) + m;
+
+#else
+    v.x() = (v.x() * d) + m;
+    v.y() = (v.y() * d) + m;
+#endif // GGML_SYCL_F16
+}
+
+static __dpct_inline__ void dequantize_q5_0(const void *vx, const int ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q5_0 * x = (const block_q5_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y() = ((x[ib].qs[iqs] >> 4) | xh_1);
+
+#ifdef GGML_SYCL_F16
+    // v = v - {16.0f, 16.0f};
+    // v = v * {d, d};
+    v.s0() = (v.s0() - 16.0f) * d;
+    v.s1() = (v.s1() - 16.0f) * d;
+
+#else
+    v.x() = (v.x() - 16.0f) * d;
+    v.y() = (v.y() - 16.0f) * d;
+#endif // GGML_SYCL_F16
+}
+
+static __dpct_inline__ void dequantize_q5_1(const void *vx, const int ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q5_1 * x = (const block_q5_1 *) vx;
+
+    const dfloat d = x[ib].dm[0];
+    const dfloat m = x[ib].dm[1];
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y() = ((x[ib].qs[iqs] >> 4) | xh_1);
+
+#ifdef GGML_SYCL_F16
+    // v = v * {d, d};
+    // v = v + {m, m};
+    v.s0() = (v.s0() * d) + m;
+    v.s1() = (v.s1() * d) + m;
+#else
+    v.x() = (v.x() * d) + m;
+    v.y() = (v.y() * d) + m;
+#endif // GGML_SYCL_F16
+}
+
+static __dpct_inline__ void dequantize_q8_0(const void *vx, const int ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q8_0 * x = (const block_q8_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    v.x() = x[ib].qs[iqs + 0];
+    v.y() = x[ib].qs[iqs + 1];
+
+#ifdef GGML_SYCL_F16
+    // v = v * {d, d};
+    v.s0() *= d;
+    v.s1() *= d;
+#else
+    v.x() *= d;
+    v.y() *= d;
+#endif // GGML_SYCL_F16
+}
+
+template<typename dst_t>
+static void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int i = item_ct1.get_group(2);
+
+    // assume 32 threads
+    const int tid = item_ct1.get_local_id(2);
+    const int il  = tid/8;
+    const int ir  = tid%8;
+    const int ib = 8*i + ir;
+    if (ib >= nb32) {
+        return;
+    }
+
+    dst_t * y = yy + 256*i + 32*ir + 4*il;
+
+    const block_q4_0 * x = (const block_q4_0 *)vx + ib;
+    const float d = sycl::vec<sycl::half, 1>(x->d)
+                        .convert<float, sycl::rounding_mode::automatic>()[0];
+    const float dm = -8*d;
+
+    const uint8_t * q = x->qs + 4*il;
+
+    for (int l = 0; l < 4; ++l) {
+        y[l+ 0] = d * (q[l] & 0xF) + dm;
+        y[l+16] = d * (q[l] >>  4) + dm;
+    }
+}
+
+template<typename dst_t>
+static void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int i = item_ct1.get_group(2);
+
+    // assume 32 threads
+    const int tid = item_ct1.get_local_id(2);
+    const int il  = tid/8;
+    const int ir  = tid%8;
+    const int ib = 8*i + ir;
+    if (ib >= nb32) {
+        return;
+    }
+
+    dst_t * y = yy + 256*i + 32*ir + 4*il;
+
+    const block_q4_1 * x = (const block_q4_1 *)vx + ib;
+    const sycl::float2 d =
+        x->dm.convert<float, sycl::rounding_mode::automatic>();
+
+    const uint8_t * q = x->qs + 4*il;
+
+    for (int l = 0; l < 4; ++l) {
+        y[l + 0] = d.x() * (q[l] & 0xF) + d.y();
+        y[l + 16] = d.x() * (q[l] >> 4) + d.y();
+    }
+}
+
+
+//================================== k-quants
+
+template<typename dst_t>
+static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int i = item_ct1.get_group(2);
+    const block_q2_K * x = (const block_q2_K *) vx;
+
+    const int tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int n   = tid/32;
+    const int l   = tid - 32*n;
+    const int is  = 8*n + l/16;
+
+    const uint8_t q = x[i].qs[32*n + l];
+    dst_t * y = yy + i*QK_K + 128*n;
+
+    float dall = x[i].dm[0];
+    float dmin = x[i].dm[1];
+    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
+    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
+    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
+    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
+#else
+    const int is = tid/16;  // 0 or 1
+    const int il = tid%16;  // 0...15
+    const uint8_t q = x[i].qs[il] >> (2*is);
+    dst_t * y = yy + i*QK_K + 16*is + il;
+
+    float dall = x[i].dm[0];
+    float dmin = x[i].dm[1];
+    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
+    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
+#endif
+
+}
+
+template<typename dst_t>
+static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int i = item_ct1.get_group(2);
+    const block_q3_K * x = (const block_q3_K *) vx;
+
+#if QK_K == 256
+    const int r = item_ct1.get_local_id(2) / 4;
+    const int tid = r/2;
+    const int is0 = r%2;
+    const int l0 = 16 * is0 + 4 * (item_ct1.get_local_id(2) % 4);
+    const int n = tid / 4;
+    const int j = tid - 4*n;
+
+    uint8_t m = 1 << (4*n + j);
+    int is = 8*n + 2*j + is0;
+    int shift = 2*j;
+
+    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
+                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
+                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
+                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
+    float d_all = x[i].d;
+    float dl = d_all * (us - 32);
+
+    dst_t * y = yy + i*QK_K + 128*n + 32*j;
+    const uint8_t * q = x[i].qs + 32*n;
+    const uint8_t * hm = x[i].hmask;
+
+    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
+#else
+    const int tid = item_ct1.get_local_id(2);
+    const int is  = tid/16;  // 0 or 1
+    const int il  = tid%16;  // 0...15
+    const int im  = il/8;    // 0...1
+    const int in  = il%8;    // 0...7
+
+    dst_t * y = yy + i*QK_K + 16*is + il;
+
+    const uint8_t q = x[i].qs[il] >> (2*is);
+    const uint8_t h = x[i].hmask[in] >> (2*is + im);
+    const float   d = (float)x[i].d;
+
+    if (is == 0) {
+        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
+        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
+    } else {
+        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
+        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
+    }
+#endif
+
+}
+
+#if QK_K == 256
+static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
+    if (j < 4) {
+        d = q[j] & 63; m = q[j + 4] & 63;
+    } else {
+        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+    }
+}
+#endif
+
+template<typename dst_t>
+static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  const sycl::nd_item<3> &item_ct1) {
+    const block_q4_K * x = (const block_q4_K *) vx;
+
+    const int i = item_ct1.get_group(2);
+
+#if QK_K == 256
+    // assume 32 threads
+    const int tid = item_ct1.get_local_id(2);
+    const int il  = tid/8;
+    const int ir  = tid%8;
+    const int is  = 2*il;
+    const int n   = 4;
+
+    dst_t * y = yy + i*QK_K + 64*il + n*ir;
+
+    const float dall = x[i].dm[0];
+    const float dmin = x[i].dm[1];
+
+    const uint8_t * q = x[i].qs + 32*il + n*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+    for (int l = 0; l < n; ++l) {
+        y[l + 0] = d1 * (q[l] & 0xF) - m1;
+        y[l +32] = d2 * (q[l] >>  4) - m2;
+    }
+#else
+    const int tid = item_ct1.get_local_id(2);
+    const uint8_t * q = x[i].qs;
+    dst_t * y = yy + i*QK_K;
+    const float d = (float)x[i].dm[0];
+    const float m = (float)x[i].dm[1];
+    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
+    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  const sycl::nd_item<3> &item_ct1) {
+    const block_q5_K * x = (const block_q5_K *) vx;
+
+    const int i = item_ct1.get_group(2);
+
+#if QK_K == 256
+    // assume 64 threads - this is very slightly better than the one below
+    const int tid = item_ct1.get_local_id(2);
+    const int il  = tid/16;   // il is in 0...3
+    const int ir  = tid%16;   // ir is in 0...15
+    const int is  = 2*il;     // is is in 0...6
+
+    dst_t * y = yy + i*QK_K + 64*il + 2*ir;
+
+    const float dall = x[i].dm[0];
+    const float dmin = x[i].dm[1];
+
+    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
+    const uint8_t * qh = x[i].qh + 2*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+
+    uint8_t   hm  = 1 << (2*il);
+    y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
+    y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
+    hm <<= 1;
+    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
+    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
+#else
+    const int tid = item_ct1.get_local_id(2);
+    const uint8_t q = x[i].qs[tid];
+    const int im = tid/8;  // 0...3
+    const int in = tid%8;  // 0...7
+    const int is = tid/16; // 0 or 1
+    const uint8_t h = x[i].qh[in] >> im;
+    const float d = x[i].d;
+    dst_t * y = yy + i*QK_K + tid;
+    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
+    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  const sycl::nd_item<3> &item_ct1) {
+    const block_q6_K * x = (const block_q6_K *) vx;
+
+    const int i = item_ct1.get_group(2);
+#if QK_K == 256
+
+    // assume 64 threads - this is very slightly better than the one below
+    const int tid = item_ct1.get_local_id(2);
+    const int ip  = tid/32;   // ip is 0 or 1
+    const int il  = tid - 32*ip; // 0...32
+    const int is  = 8*ip + il/16;
+
+    dst_t * y = yy + i*QK_K + 128*ip + il;
+
+    const float d = x[i].d;
+
+    const uint8_t * ql = x[i].ql + 64*ip + il;
+    const uint8_t   qh = x[i].qh[32*ip + il];
+    const int8_t  * sc = x[i].scales + is;
+
+    y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+#else
+
+    // assume 32 threads
+    const int tid = item_ct1.get_local_id(2);
+    const int ip  = tid/16;         // 0 or 1
+    const int il  = tid - 16*ip;    // 0...15
+
+    dst_t * y = yy + i*QK_K + 16*ip + il;
+
+    const float d = x[i].d;
+
+    const uint8_t   ql = x[i].ql[16*ip + il];
+    const uint8_t   qh = x[i].qh[il] >> (2*ip);
+    const int8_t  * sc = x[i].scales;
+
+    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                     const sycl::nd_item<3> &item_ct1,
+                                     const uint64_t *iq2xxs_grid_ptr,
+                                     const uint8_t *ksigns_iq2xs_ptr,
+                                     const uint8_t *kmask_iq2xs_ptr) {
+
+    const int i = item_ct1.get_group(2);
+    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
+
+    const int tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const uint8_t  * grid = (const uint8_t *)(iq2xxs_grid_ptr + aux8[il]);
+    const uint32_t aux32 = q2[2] | (q2[3] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs_ptr[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs_ptr[j] ? -1.f : 1.f);
+#else
+    assert(false);
+#endif
+
+}
+
+template<typename dst_t>
+static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                    const sycl::nd_item<3> &item_ct1,
+                                    const uint64_t *iq2xs_grid,
+                                    const uint8_t *ksigns_iq2xs,
+                                    const uint8_t *kmask_iq2xs) {
+
+    const int i = item_ct1.get_group(2);
+    const block_iq2_xs * x = (const block_iq2_xs *) vx;
+
+    const int tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                       const sycl::nd_item<3> &item_ct1) {
+
+    const int i = item_ct1.get_group(2);
+    const block_iq2_s * x = (const block_iq2_s *) vx;
+
+    const int tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
+#pragma unroll
+    for (int j = 0; j < 8; ++j)
+        y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    assert(false);
+
+#endif
+
+}
+
+template<typename dst_t>
+static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                     const sycl::nd_item<3> &item_ct1,
+                                     const uint32_t *iq3xxs_grid,
+                                     const uint8_t *ksigns_iq2xs,
+                                     const uint8_t *kmask_iq2xs) {
+
+    const int i = item_ct1.get_group(2);
+    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
+
+    const int tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t  * q3 = x[i].qs + 8*ib;
+    const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
+    const uint8_t  * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                       const sycl::nd_item<3> &item_ct1,
+                       const uint8_t *kmask_iq2xs, const uint32_t *iq3s_grid) {
+
+    const int i = item_ct1.get_group(2);
+    const block_iq3_s * x = (const block_iq3_s *) vx;
+
+    const int tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * qs = x[i].qs + 8*ib;
+    const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
+    const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
+    const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
+    const uint8_t signs = x[i].signs[4*ib + il];
+#pragma unroll
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                       const sycl::nd_item<3> &item_ct1,
+                       const uint32_t *iq1s_grid_gpu) {
+
+    const int i = item_ct1.get_group(2);
+    const block_iq1_s * x = (const block_iq1_s  *) vx;
+
+    const int tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
+    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1);
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                       const sycl::nd_item<3> &item_ct1,
+                       const uint32_t *iq1s_grid_gpu) {
+
+    const int i = item_ct1.get_group(2);
+    const block_iq1_m * x = (const block_iq1_m  *) vx;
+
+    const int tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * sc = (const uint16_t *)x[i].scales;
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const int ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
+    const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
+    const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq4_nl(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                        const sycl::nd_item<3> &item_ct1) {
+
+    const int i = item_ct1.get_group(2);
+    const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
+
+    const int tid = item_ct1.get_local_id(2);
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = (float)x[ib].d;
+#pragma unroll
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+
+}
+
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq4_xs(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                        const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_group(2);
+    const block_iq4_xs * x = (const block_iq4_xs *)vx;
+
+    const int tid = item_ct1.get_local_id(2);
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
+    const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
+#pragma unroll
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+}
+
+
+#endif // GGML_SYCL_DEQUANTIZE_HPP

ggml-sycl/dmmv.hpp

@@ -0,0 +1,27 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_DMMV_HPP
+#define GGML_SYCL_DMMV_HPP
+
+#include "common.hpp"
+
+
+void ggml_sycl_op_dequantize_mul_mat_vec(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
+    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
+    const int64_t src1_ncols, const int64_t src1_padded_row_size,
+    const dpct::queue_ptr &stream);
+
+#endif // GGML_SYCL_DMMV_HPP

ggml-sycl/mmq.hpp

@@ -0,0 +1,33 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_MMQ_HPP
+#define GGML_SYCL_MMQ_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_mul_mat_q(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor* src0,
+    const ggml_tensor* src1,
+    ggml_tensor* dst,
+    const char* src0_dd_i,
+    const float* src1_ddf_i,
+    const char* src1_ddq_i,
+    float* dst_dd_i,
+    const int64_t row_low,
+    const int64_t row_high,
+    const int64_t src1_ncols,
+    const int64_t src1_padded_row_size,
+    const dpct::queue_ptr& stream);
+
+#endif // GGML_SYCL_MMQ_HPP

ggml-sycl/mmvq.hpp

@@ -0,0 +1,27 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_MMVQ_HPP
+#define GGML_SYCL_MMVQ_HPP
+
+#include "common.hpp"
+
+
+void ggml_sycl_op_mul_mat_vec_q(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
+    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
+    const int64_t src1_ncols, const int64_t src1_padded_row_size,
+    const dpct::queue_ptr &stream);
+
+#endif // GGML_SYCL_MMVQ_HPP

ggml-sycl/presets.hpp

@@ -0,0 +1,67 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_PRESETS_HPP
+#define GGML_SYCL_PRESETS_HPP
+
+#define GGML_SYCL_MAX_STREAMS       8
+#define GGML_SYCL_MAX_BUFFERS       256
+#define GGML_SYCL_MAX_DEVICES       48
+#define GGML_SYCL_NAME "SYCL"
+
+#define WARP_SIZE 32
+#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
+
+#define SYCL_GELU_BLOCK_SIZE 256
+#define SYCL_SILU_BLOCK_SIZE 256
+#define SYCL_TANH_BLOCK_SIZE 256
+#define SYCL_RELU_BLOCK_SIZE 256
+#define SYCL_HARDSIGMOID_BLOCK_SIZE 256
+#define SYCL_HARDSWISH_BLOCK_SIZE 256
+#define SYCL_SQR_BLOCK_SIZE 256
+#define SYCL_CPY_BLOCK_SIZE 32
+#define SYCL_SCALE_BLOCK_SIZE 256
+#define SYCL_CLAMP_BLOCK_SIZE 256
+#define SYCL_ROPE_BLOCK_SIZE 256
+#define SYCL_ALIBI_BLOCK_SIZE 32
+#define SYCL_DIAG_MASK_INF_BLOCK_SIZE 32
+#define SYCL_QUANTIZE_BLOCK_SIZE 256
+#define SYCL_DEQUANTIZE_BLOCK_SIZE 256
+#define SYCL_GET_ROWS_BLOCK_SIZE 256
+#define SYCL_UPSCALE_BLOCK_SIZE 256
+#define SYCL_CONCAT_BLOCK_SIZE 256
+#define SYCL_PAD_BLOCK_SIZE 256
+#define SYCL_ACC_BLOCK_SIZE 256
+#define SYCL_IM2COL_BLOCK_SIZE 256
+#define SYCL_POOL2D_BLOCK_SIZE 256
+
+// dmmv = dequantize_mul_mat_vec
+#ifndef GGML_SYCL_DMMV_X
+#define GGML_SYCL_DMMV_X 32
+#endif
+#ifndef GGML_SYCL_MMV_Y
+#define GGML_SYCL_MMV_Y 1
+#endif
+
+#ifndef K_QUANTS_PER_ITERATION
+#define K_QUANTS_PER_ITERATION 2
+#else
+static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
+#endif
+
+#ifndef GGML_SYCL_PEER_MAX_BATCH_SIZE
+#define GGML_SYCL_PEER_MAX_BATCH_SIZE 128
+#endif // GGML_SYCL_PEER_MAX_BATCH_SIZE
+
+#define MUL_MAT_SRC1_COL_STRIDE 128
+
+#endif // GGML_SYCL_PRESETS_HPP

gguf-py/gguf/constants.py

@@ -40,6 +40,7 @@ class Keys:
         LEADING_DENSE_BLOCK_COUNT         = "{arch}.leading_dense_block_count"
         FEED_FORWARD_LENGTH               = "{arch}.feed_forward_length"
         EXPERT_FEED_FORWARD_LENGTH        = "{arch}.expert_feed_forward_length"
+        EXPERT_SHARED_FEED_FORWARD_LENGTH = "{arch}.expert_shared_feed_forward_length"
         USE_PARALLEL_RESIDUAL             = "{arch}.use_parallel_residual"
         TENSOR_DATA_LAYOUT                = "{arch}.tensor_data_layout"
         EXPERT_COUNT                      = "{arch}.expert_count"

gguf-py/gguf/gguf_writer.py

@@ -394,6 +394,9 @@ class GGUFWriter:
     def add_expert_feed_forward_length(self, length: int) -> None:
         self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
 
+    def add_expert_shared_feed_forward_length(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_SHARED_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+
     def add_parallel_residual(self, use: bool) -> None:
         self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
 

gguf-py/scripts/gguf-dump.py

@@ -14,7 +14,7 @@ import numpy as np
 if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
     sys.path.insert(0, str(Path(__file__).parent.parent))
 
-from gguf import GGUFReader, GGUFValueType  # noqa: E402
+from gguf import GGUFReader, GGUFValueType, ReaderTensor  # noqa: E402
 
 logger = logging.getLogger("gguf-dump")
 
@@ -101,25 +101,285 @@ def dump_metadata_json(reader: GGUFReader, args: argparse.Namespace) -> None:
     json.dump(result, sys.stdout)
 
 
+def markdown_table_with_alignment_support(header_map: list[dict[str, str]], data: list[dict[str, Any]]):
+    # JSON to Markdown table formatting: https://stackoverflow.com/a/72983854/2850957
+
+    # Alignment Utility Function
+    def strAlign(padding: int, alignMode: str | None, strVal: str):
+        if alignMode == 'center':
+            return strVal.center(padding)
+        elif alignMode == 'right':
+            return strVal.rjust(padding - 1) + ' '
+        elif alignMode == 'left':
+            return ' ' + strVal.ljust(padding - 1)
+        else: # default left
+            return ' ' + strVal.ljust(padding - 1)
+
+    def dashAlign(padding: int, alignMode: str | None):
+        if alignMode == 'center':
+            return ':' + '-' * (padding - 2) + ':'
+        elif alignMode == 'right':
+            return '-' * (padding - 1) + ':'
+        elif alignMode == 'left':
+            return ':' + '-' * (padding - 1)
+        else: # default left
+            return '-' * (padding)
+
+    # Calculate Padding For Each Column Based On Header and Data Length
+    rowsPadding = {}
+    for index, columnEntry in enumerate(header_map):
+        padCount = max([len(str(v)) for d in data for k, v in d.items() if k == columnEntry['key_name']], default=0) + 2
+        headerPadCount = len(columnEntry['header_name']) + 2
+        rowsPadding[index] = headerPadCount if padCount <= headerPadCount else padCount
+
+    # Render Markdown Header
+    rows = []
+    rows.append('|'.join(strAlign(rowsPadding[index], columnEntry.get('align'), str(columnEntry['header_name'])) for index, columnEntry in enumerate(header_map)))
+    rows.append('|'.join(dashAlign(rowsPadding[index], columnEntry.get('align')) for index, columnEntry in enumerate(header_map)))
+
+    # Render Tabular Data
+    for item in data:
+        rows.append('|'.join(strAlign(rowsPadding[index], columnEntry.get('align'), str(item[columnEntry['key_name']])) for index, columnEntry in enumerate(header_map)))
+
+    # Convert Tabular String Rows Into String
+    tableString = ""
+    for row in rows:
+        tableString += f'|{row}|\n'
+
+    return tableString
+
+
+def element_count_rounded_notation(count: int) -> str:
+    if count > 1e15 :
+        # Quadrillion
+        scaled_amount = count * 1e-15
+        scale_suffix = "Q"
+    elif count > 1e12 :
+        # Trillions
+        scaled_amount = count * 1e-12
+        scale_suffix = "T"
+    elif count > 1e9 :
+        # Billions
+        scaled_amount = count * 1e-9
+        scale_suffix = "B"
+    elif count > 1e6 :
+        # Millions
+        scaled_amount = count * 1e-6
+        scale_suffix = "M"
+    elif count > 1e3 :
+        # Thousands
+        scaled_amount = count * 1e-3
+        scale_suffix = "K"
+    else:
+        # Under Thousands
+        scaled_amount = count
+        scale_suffix = ""
+    return f"{'~' if count > 1e3 else ''}{round(scaled_amount)}{scale_suffix}"
+
+
+def translate_tensor_name(name):
+    words = name.split(".")
+
+    # Source: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#standardized-tensor-names
+    abbreviation_dictionary = {
+        'token_embd': 'Token embedding',
+        'pos_embd': 'Position embedding',
+        'output_norm': 'Output normalization',
+        'output': 'Output',
+        'attn_norm': 'Attention normalization',
+        'attn_norm_2': 'Attention normalization',
+        'attn_qkv': 'Attention query-key-value',
+        'attn_q': 'Attention query',
+        'attn_k': 'Attention key',
+        'attn_v': 'Attention value',
+        'attn_output': 'Attention output',
+        'ffn_norm': 'Feed-forward network normalization',
+        'ffn_up': 'Feed-forward network "up"',
+        'ffn_gate': 'Feed-forward network "gate"',
+        'ffn_down': 'Feed-forward network "down"',
+        'ffn_gate_inp': 'Expert-routing layer for the Feed-forward network in Mixture of Expert models',
+        'ffn_gate_exp': 'Feed-forward network "gate" layer per expert in Mixture of Expert models',
+        'ffn_down_exp': 'Feed-forward network "down" layer per expert in Mixture of Expert models',
+        'ffn_up_exp': 'Feed-forward network "up" layer per expert in Mixture of Expert models',
+        'ssm_in': 'State space model input projections',
+        'ssm_conv1d': 'State space model rolling/shift',
+        'ssm_x': 'State space model selective parametrization',
+        'ssm_a': 'State space model state compression',
+        'ssm_d': 'State space model skip connection',
+        'ssm_dt': 'State space model time step',
+        'ssm_out': 'State space model output projection',
+        'blk': 'Block'
+    }
+
+    expanded_words = []
+    for word in words:
+        word_norm = word.strip().lower()
+        if word_norm in abbreviation_dictionary:
+            expanded_words.append(abbreviation_dictionary[word_norm].title())
+        else:
+            expanded_words.append(word.title())
+
+    return ' '.join(expanded_words)
+
+
+def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None:
+    host_endian, file_endian = get_file_host_endian(reader)
+    markdown_content = ""
+    markdown_content += f'# {args.model} - GGUF Internal File Dump\n\n'
+    markdown_content += f'- Endian: {file_endian} endian\n'
+    markdown_content += '\n'
+    markdown_content += '## Key Value Metadata Store\n\n'
+    markdown_content += f'There are {len(reader.fields)} key-value pairs in this file\n'
+    markdown_content += '\n'
+
+    kv_dump_table: list[dict[str, str | int]] = []
+    for n, field in enumerate(reader.fields.values(), 1):
+        if not field.types:
+            pretty_type = 'N/A'
+        elif field.types[0] == GGUFValueType.ARRAY:
+            nest_count = len(field.types) - 1
+            pretty_type = '[' * nest_count + str(field.types[-1].name) + ']' * nest_count
+        else:
+            pretty_type = str(field.types[-1].name)
+
+        total_elements = len(field.data)
+        value = ""
+        if len(field.types) == 1:
+            curr_type = field.types[0]
+            if curr_type == GGUFValueType.STRING:
+                value = repr(str(bytes(field.parts[-1]), encoding='utf-8')[:60])
+            elif curr_type in reader.gguf_scalar_to_np:
+                value = str(field.parts[-1][0])
+        else:
+            if field.types[0] == GGUFValueType.ARRAY:
+                curr_type = field.types[1]
+                if curr_type == GGUFValueType.STRING:
+                    render_element = min(5, total_elements)
+                    for element_pos in range(render_element):
+                        value += repr(str(bytes(field.parts[-1 - element_pos]), encoding='utf-8')[:5]) + (", " if total_elements > 1 else "")
+                elif curr_type in reader.gguf_scalar_to_np:
+                    render_element = min(7, total_elements)
+                    for element_pos in range(render_element):
+                        value += str(field.parts[-1 - element_pos][0]) + (", " if total_elements > 1 else "")
+                value = f'[ {value}{" ..." if total_elements > 1 else ""} ]'
+        kv_dump_table.append({"n":n, "pretty_type":pretty_type, "total_elements":total_elements, "field_name":field.name, "value":value})
+
+    kv_dump_table_header_map = [
+        {'key_name':'n',                'header_name':'POS',      'align':'right'},
+        {'key_name':'pretty_type',      'header_name':'TYPE',     'align':'left'},
+        {'key_name':'total_elements',   'header_name':'Count',    'align':'right'},
+        {'key_name':'field_name',       'header_name':'Key',      'align':'left'},
+        {'key_name':'value',            'header_name':'Value',    'align':'left'},
+    ]
+
+    markdown_content += markdown_table_with_alignment_support(kv_dump_table_header_map, kv_dump_table)
+
+    markdown_content += "\n"
+
+    if not args.no_tensors:
+        # Group tensors by their prefix and maintain order
+        tensor_prefix_order: list[str] = []
+        tensor_name_to_key: dict[str, int] = {}
+        tensor_groups: dict[str, list[ReaderTensor]] = {}
+        total_elements = sum(tensor.n_elements for tensor in reader.tensors)
+
+        # Parsing Tensors Record
+        for key, tensor in enumerate(reader.tensors):
+            tensor_components = tensor.name.split('.')
+
+            # Classify Tensor Group
+            tensor_group_name = "base"
+            if tensor_components[0] == 'blk':
+                tensor_group_name = f"{tensor_components[0]}.{tensor_components[1]}"
+
+            # Check if new Tensor Group
+            if tensor_group_name not in tensor_groups:
+                tensor_groups[tensor_group_name] = []
+                tensor_prefix_order.append(tensor_group_name)
+
+            # Record Tensor and Tensor Position
+            tensor_groups[tensor_group_name].append(tensor)
+            tensor_name_to_key[tensor.name] = key
+
+        # Tensors Mapping Dump
+        markdown_content += f'## Tensors Overview {element_count_rounded_notation(total_elements)} Elements\n\n'
+        markdown_content += f'Total number of elements in all tensors: {total_elements} Elements\n'
+        markdown_content += '\n'
+
+        for group in tensor_prefix_order:
+            tensors = tensor_groups[group]
+            group_elements = sum(tensor.n_elements for tensor in tensors)
+            markdown_content += f"- [{translate_tensor_name(group)} Tensor Group - {element_count_rounded_notation(group_elements)} Elements](#{group.replace('.', '_')})\n"
+
+        markdown_content += "\n"
+
+        for group in tensor_prefix_order:
+            tensors = tensor_groups[group]
+            group_elements = sum(tensor.n_elements for tensor in tensors)
+            group_percentage = group_elements / total_elements * 100
+            markdown_content += f"### <a name=\"{group.replace('.', '_')}\">{translate_tensor_name(group)} Tensor Group : {element_count_rounded_notation(group_elements)} Elements</a>\n\n"
+
+            # Precalculate column sizing for visual consistency
+            prettify_element_est_count_size: int = 1
+            prettify_element_count_size: int = 1
+            prettify_dimension_max_widths: dict[int, int] = {}
+            for tensor in tensors:
+                prettify_element_est_count_size = max(prettify_element_est_count_size, len(str(element_count_rounded_notation(tensor.n_elements))))
+                prettify_element_count_size = max(prettify_element_count_size, len(str(tensor.n_elements)))
+                for i, dimension_size in enumerate(list(tensor.shape) + [1] * (4 - len(tensor.shape))):
+                    prettify_dimension_max_widths[i] = max(prettify_dimension_max_widths.get(i,1), len(str(dimension_size)))
+
+            # Generate Tensor Layer Table Content
+            tensor_dump_table: list[dict[str, str | int]] = []
+            for tensor in tensors:
+                human_friendly_name = translate_tensor_name(tensor.name.replace(".weight", ".(W)").replace(".bias", ".(B)"))
+                pretty_dimension = ' x '.join(f'{str(d):>{prettify_dimension_max_widths[i]}}' for i, d in enumerate(list(tensor.shape) + [1] * (4 - len(tensor.shape))))
+                element_count_est = f"({element_count_rounded_notation(tensor.n_elements):>{prettify_element_est_count_size}})"
+                element_count_string = f"{element_count_est} {tensor.n_elements:>{prettify_element_count_size}}"
+                type_name_string = f"{tensor.tensor_type.name}"
+                tensor_dump_table.append({"t_id":tensor_name_to_key[tensor.name], "layer_name":tensor.name, "human_layer_name":human_friendly_name, "element_count":element_count_string, "pretty_dimension":pretty_dimension, "tensor_type":type_name_string})
+
+            tensor_dump_table_header_map = [
+                {'key_name':'t_id',             'header_name':'T_ID',                             'align':'right'},
+                {'key_name':'layer_name',       'header_name':'Tensor Layer Name',                'align':'left'},
+                {'key_name':'human_layer_name', 'header_name':'Human Friendly Tensor Layer Name', 'align':'left'},
+                {'key_name':'element_count',    'header_name':'Elements',                         'align':'left'},
+                {'key_name':'pretty_dimension', 'header_name':'Shape',                            'align':'left'},
+                {'key_name':'tensor_type',      'header_name':'Type',                             'align':'left'},
+            ]
+
+            markdown_content += markdown_table_with_alignment_support(tensor_dump_table_header_map, tensor_dump_table)
+
+            markdown_content += "\n"
+            markdown_content += f"- Total elements in {group}: ({element_count_rounded_notation(group_elements):>4}) {group_elements}\n"
+            markdown_content += f"- Percentage of total elements: {group_percentage:.2f}%\n"
+            markdown_content += "\n\n"
+
+        print(markdown_content)  # noqa: NP100
+
+
 def main() -> None:
     parser = argparse.ArgumentParser(description="Dump GGUF file metadata")
     parser.add_argument("model",           type=str,            help="GGUF format model filename")
     parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata")
     parser.add_argument("--json",       action="store_true", help="Produce JSON output")
     parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)")
+    parser.add_argument("--markdown",   action="store_true", help="Produce markdown output")
     parser.add_argument("--verbose",    action="store_true", help="increase output verbosity")
 
     args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
 
     logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
 
-    if not args.json:
+    if not args.json and not args.markdown:
         logger.info(f'* Loading: {args.model}')
 
     reader = GGUFReader(args.model, 'r')
 
     if args.json:
         dump_metadata_json(reader, args)
+    elif args.markdown:
+        dump_markdown_metadata(reader, args)
     else:
         dump_metadata(reader, args)
 

llama.cpp

@@ -286,6 +286,7 @@ enum llm_kv {
     LLM_KV_LEADING_DENSE_BLOCK_COUNT,
     LLM_KV_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
     LLM_KV_USE_PARALLEL_RESIDUAL,
     LLM_KV_TENSOR_DATA_LAYOUT,
     LLM_KV_EXPERT_COUNT,
@@ -371,6 +372,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_LEADING_DENSE_BLOCK_COUNT,         "%s.leading_dense_block_count"         },
     { LLM_KV_FEED_FORWARD_LENGTH,               "%s.feed_forward_length"               },
     { LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        "%s.expert_feed_forward_length"        },
+    { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
     { LLM_KV_USE_PARALLEL_RESIDUAL,             "%s.use_parallel_residual"             },
     { LLM_KV_TENSOR_DATA_LAYOUT,                "%s.tensor_data_layout"                },
     { LLM_KV_EXPERT_COUNT,                      "%s.expert_count"                      },
@@ -1278,6 +1280,126 @@ struct no_init {
 };
 
 struct llama_file {
+
+#if defined(_WIN32)
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    HANDLE fp_win32;
+    size_t size;
+
+private:
+    std::string GetErrorMessageWin32(DWORD error_code) const {
+        std::string ret;
+        LPSTR lpMsgBuf = NULL;
+        DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+                                    NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
+        if (!bufLen) {
+            ret = format("Win32 error code: %s", error_code);
+        } else {
+            ret = lpMsgBuf;
+            LocalFree(lpMsgBuf);
+        }
+
+        return ret;
+    }
+
+public:
+
+    llama_file(const char * fname, const char * mode) {
+        fp = ggml_fopen(fname, mode);
+        if (fp == NULL) {
+            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
+        }
+        fp_win32 = (HANDLE) _get_osfhandle(_fileno(fp));
+        seek(0, SEEK_END);
+        size = tell();
+        seek(0, SEEK_SET);
+    }
+
+    size_t tell() const {
+        // SetFilePointerEx returns the current position when seeking relative 0 bytes
+        LARGE_INTEGER li;
+        li.QuadPart = 0;
+        BOOL ret = SetFilePointerEx(fp_win32, li, &li, FILE_CURRENT);
+        if (!ret) {
+            throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
+        }
+
+        return li.QuadPart;
+    }
+
+    void seek(size_t offset, int whence) const {
+        // no need to convert SEEK_* to FILE_*. The enums are the same.
+        // Still, keep static asserts to avoid failures in the future.
+        static_assert(SEEK_SET == FILE_BEGIN, "SEEK_SET != FILE_BEGIN");
+        static_assert(SEEK_CUR == FILE_CURRENT, "SEEK_CUR != FILE_CURRENT");
+        static_assert(SEEK_END == FILE_END, "SEEK_END != FILE_END");
+
+        LARGE_INTEGER li;
+        li.QuadPart = offset;
+        BOOL ret = SetFilePointerEx(fp_win32, li, NULL, whence);
+        if (!ret) {
+            throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
+        }
+    }
+
+    void read_raw(void * ptr, size_t len) const {
+        // On Win32 ReadFile is significant faster than fread which is again significant faster than std::fstream. Thus
+        // use the Win32 API to do file io instead of the C/C++ library functions.
+
+        // There are conditions under which ReadFile cannot read chunks >64MB.
+        // Thus split the operation into smaller chunks if len exceeds this limit.
+        size_t bytes_read = 0;
+        while (bytes_read < len) {
+            size_t chunk_size = std::min<size_t>(len - bytes_read, 64*1024*1024);
+            DWORD chunk_read = 0;
+            BOOL result = ReadFile(fp_win32, reinterpret_cast<char*>(ptr) + bytes_read, chunk_size, &chunk_read, NULL);
+            if (!result) {
+                throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
+            }
+            if (chunk_read < chunk_size || chunk_read == 0) {
+                throw std::runtime_error("unexpectedly reached end of file");
+            }
+
+            bytes_read += chunk_read;
+        } ;
+    }
+
+    uint32_t read_u32() const {
+        uint32_t val;
+        read_raw(&val, sizeof(val));
+        return val;
+    }
+
+    void write_raw(const void * ptr, size_t len) const {
+        // There are conditions under which WriteFile cannot write chunks >64MB.
+        // Thus split the operation into smaller chunks if len exceeds this limit.
+        size_t bytes_written = 0;
+        while (bytes_written < len) {
+            size_t chunk_size = std::min<size_t>(len - bytes_written, 64*1024*1024);
+            DWORD chunk_written = 0;
+            BOOL result = WriteFile(fp_win32, reinterpret_cast<char const*>(ptr) + bytes_written, chunk_size, &chunk_written, NULL);
+            if (!result) {
+                throw std::runtime_error(format("write error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
+            }
+            if (chunk_written < chunk_size || chunk_written == 0) {
+                throw std::runtime_error("unexpectedly failed to write bytes");
+            }
+
+            bytes_written += chunk_written;
+        }
+    }
+
+    void write_u32(std::uint32_t val) const {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+#else
     // use FILE * so we don't have to re-open the file to mmap
     FILE * fp;
     size_t size;
@@ -1298,7 +1420,10 @@ struct llama_file {
 #else
         long ret = std::ftell(fp);
 #endif
-        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        if (ret == -1) {
+            throw std::runtime_error(format("ftell error: %s", strerror(errno)));
+        }
+
         return (size_t) ret;
     }
 
@@ -1308,7 +1433,9 @@ struct llama_file {
 #else
         int ret = std::fseek(fp, (long) offset, whence);
 #endif
-        GGML_ASSERT(ret == 0); // same
+        if (ret != 0) {
+            throw std::runtime_error(format("seek error: %s", strerror(errno)));
+        }
     }
 
     void read_raw(void * ptr, size_t len) const {
@@ -1351,6 +1478,7 @@ struct llama_file {
             std::fclose(fp);
         }
     }
+#endif
 };
 using llama_files = std::vector<std::unique_ptr<llama_file>>;
 
@@ -1844,6 +1972,7 @@ struct llama_hparams {
     uint32_t n_lora_q = 0;
     uint32_t n_lora_kv = 0;
     uint32_t n_ff_exp = 0;
+    uint32_t n_ff_shexp = 0;
     uint32_t n_expert_shared = 0;
     float    expert_weights_scale = 0.0;
 
@@ -1892,6 +2021,7 @@ struct llama_hparams {
         if (this->n_lora_q           != other.n_lora_q)           return true;
         if (this->n_lora_kv          != other.n_lora_kv)          return true;
         if (this->n_ff_exp           != other.n_ff_exp)           return true;
+        if (this->n_ff_shexp         != other.n_ff_shexp)         return true;
         if (this->n_expert_shared    != other.n_expert_shared)    return true;
 
         if (this->rope_finetuned  != other.rope_finetuned)  return true;
@@ -2180,16 +2310,17 @@ struct llama_vocab {
     id special_cls_id  = -1;
     id special_mask_id = -1;
 
-    int special_add_bos = -1; // -1 unknown, 1 add, 0 don't add.
-    int special_add_eos = -1; // -1 unknown, 1 add, 0 don't add.
-
     id linefeed_id       = 13;
     id special_prefix_id = -1;
     id special_suffix_id = -1;
     id special_middle_id = -1;
     id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
 
-    bool add_space_prefix = true;
+    // tokenizer flags
+    bool tokenizer_add_space_prefix = true;
+    bool tokenizer_add_bos          = false;
+    bool tokenizer_add_eos          = false;
+    bool tokenizer_ignore_merges    = false;
 
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const {
         GGML_ASSERT(token_left.find(' ') == std::string::npos);
@@ -3721,6 +3852,44 @@ struct llama_model_loader {
         std::vector<no_init<uint8_t>> read_buf;
         std::vector<std::future<std::pair<ggml_tensor *, bool>>> validation_result;
 
+#if defined(GGML_USE_CUDA)
+        // 4 staging buffers for async uploads, each sized 1MB seems to be a good default for single NVMe drives.
+        // NVMe raid configurations might require more / larger buffers.
+        constexpr size_t num_buffers = 4;
+        constexpr size_t buffer_size = 1 * 1024 * 1024; // 1MB
+
+        std::vector<ggml_backend_buffer_t> host_buffers;
+        std::vector<void*> host_ptrs;
+        std::vector<ggml_backend_event_t> events;
+        size_t buffer_idx = 0; // buffer to use for async loads
+
+        ggml_backend_t cuda_backend = nullptr;
+        if (!use_mmap && !check_tensors) {
+            // When not using mmaped io use async uploads from pinned memory to GPU memory.
+            // First determine if the CUDA backend is active, and if so, determine the device ID.
+            ggml_backend_buffer_t buf = bufs_mmap.count(0) ? bufs_mmap.at(0) : nullptr;
+            if (buf) {
+                ggml_backend_buffer_type_t buffer_type = ggml_backend_buffer_get_type(buf);
+                for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) {
+                    auto * cuda_buffer_type = ggml_backend_cuda_buffer_type(i);
+                    if (buffer_type == cuda_buffer_type) {
+                        cuda_backend = ggml_backend_cuda_init(i);
+                        break;
+                    }
+                }
+            }
+
+            // If the cuda backend is active create pinned memory buffers and events for synchronisation.
+            if (cuda_backend) {
+                for (size_t idx = 0; idx < num_buffers; ++idx) {
+                    host_buffers.emplace_back(ggml_backend_buft_alloc_buffer(llama_default_buffer_type_cpu(true), buffer_size));
+                    host_ptrs.emplace_back(ggml_backend_buffer_get_base(host_buffers[idx]));
+                    events.emplace_back(ggml_backend_event_new(cuda_backend));
+                }
+            }
+        }
+#endif
+
         for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) {
             const auto * weight = get_weight(ggml_get_name(cur));
             if (weight == nullptr) {
@@ -3776,6 +3945,29 @@ struct llama_model_loader {
                         }));
                     }
                 } else {
+#if defined(GGML_USE_CUDA)
+                    // If cuda_backend is valid load the tensor in chunks to pinned memory and upload the buffers asynchronously to the GPU.
+                    if (cuda_backend) {
+                        file->seek(weight->offs, SEEK_SET);
+
+                        size_t bytes_read = 0;
+
+                        while (bytes_read < n_size) {
+                            size_t read_iteration = std::min<size_t>(buffer_size, n_size - bytes_read);
+
+                            ggml_backend_event_synchronize(events[buffer_idx]);
+                            file->read_raw(host_ptrs[buffer_idx], read_iteration);
+                            ggml_backend_tensor_set_async(cuda_backend, cur, host_ptrs[buffer_idx], bytes_read, read_iteration);
+                            ggml_backend_event_record(events[buffer_idx]);
+
+                            bytes_read += read_iteration;
+                            ++buffer_idx;
+                            buffer_idx %= num_buffers;
+                        }
+                    }
+                    else
+#endif
+                    {
                         read_buf.resize(n_size);
                         file->seek(weight->offs, SEEK_SET);
                         file->read_raw(read_buf.data(), n_size);
@@ -3785,10 +3977,23 @@ struct llama_model_loader {
                         }
                     }
                 }
+            }
 
             size_done += n_size;
         }
 
+#if defined(GGML_USE_CUDA)
+        // free temporary resources used for async cuda uploads
+        if (cuda_backend) {
+            for (size_t idx = 0; idx < num_buffers;++idx) {
+                ggml_backend_event_synchronize(events[idx]);
+                ggml_backend_event_free(events[idx]);
+                ggml_backend_buffer_free(host_buffers[idx]);
+            }
+            ggml_backend_free(cuda_backend);
+        }
+#endif
+
         // check validation results
         bool validation_failed = false;
         for (auto & future : validation_result) {
@@ -4255,6 +4460,9 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_QWEN2MOE:
             {
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
+
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 24: model.type = e_model::MODEL_A2_7B; break;
@@ -4563,7 +4771,7 @@ static void llm_load_vocab(
 
             const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str());
             if (add_space_prefix_keyidx != -1) {
-                vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx);
+                vocab.tokenizer_add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx);
             } // The default value of add_space_prefix is true.
         } else if (tokenizer_model == "bert") {
             vocab.type = LLAMA_VOCAB_TYPE_WPM;
@@ -4576,13 +4784,13 @@ static void llm_load_vocab(
             vocab.special_pad_id  = 0;
             vocab.special_cls_id  = 101;
             vocab.special_mask_id = 103;
-            vocab.add_space_prefix = false;
+            vocab.tokenizer_add_space_prefix = false;
         } else if (tokenizer_model == "gpt2") {
             vocab.type = LLAMA_VOCAB_TYPE_BPE;
 
             const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str());
             if (add_space_prefix_keyidx != -1) {
-                vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx);
+                vocab.tokenizer_add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx);
             }
 
             // read bpe merges and populate bpe ranks
@@ -4640,6 +4848,8 @@ static void llm_load_vocab(
                     tokenizer_pre == "llama-v3" ||
                     tokenizer_pre == "llama-bpe") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
+                vocab.tokenizer_ignore_merges = true;
+                vocab.tokenizer_add_bos = true;
             } else if (
                     tokenizer_pre == "deepseek-llm") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM;
@@ -4690,6 +4900,14 @@ static void llm_load_vocab(
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
+        } else if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
+            vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+            vocab.tokenizer_add_bos = true;
+            vocab.tokenizer_add_eos = false;
+        } else if (vocab.type == LLAMA_VOCAB_TYPE_WPM) {
+            vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+            vocab.tokenizer_add_bos = true;
+            vocab.tokenizer_add_eos = false;
         } else {
             vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
         }
@@ -4834,10 +5052,10 @@ static void llm_load_vocab(
             bool temp = true;
 
             if (ml.get_key(LLM_KV_TOKENIZER_ADD_BOS, temp, false)) {
-                vocab.special_add_bos = int(temp);
+                vocab.tokenizer_add_bos = temp;
             }
             if (ml.get_key(LLM_KV_TOKENIZER_ADD_EOS, temp, false)) {
-                vocab.special_add_eos = int(temp);
+                vocab.tokenizer_add_eos = temp;
             }
         }
 
@@ -4937,7 +5155,7 @@ static void llm_load_vocab(
         );
 
         // set attributes by model/tokenizer name
-        if (_contains_any(tokenizer_pre, {"jina-v2-es", "jina-v2-de"})) {
+        if (_contains_any(tokenizer_pre, {"jina-v2-de", "jina-v2-es", "jina-v2-code"})) {
             _set_token_attr("<mask>", LLAMA_TOKEN_ATTR_LSTRIP, true);
         } else if (_contains_any(model_name, {"phi-3", "phi3"})) {
             for (auto id : vocab.cache_special_tokens) {
@@ -5040,6 +5258,11 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
         LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
         LLAMA_LOG_INFO("%s: rope_yarn_log_mul    = %.4f\n",   __func__, hparams.rope_yarn_log_mul);
     }
+
+    if (model.arch == LLM_ARCH_QWEN2MOE) {
+        LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
+    }
 }
 
 // Returns false if cancelled by progress_callback
@@ -5183,7 +5406,7 @@ static bool llm_load_tensors(
     // create tensors for the weights
     {
         const int64_t n_embd       = hparams.n_embd;
-        const int64_t n_embd_head  = n_embd / hparams.n_head;
+        const int64_t n_embd_head  = (hparams.n_head == 0) ? 0 : n_embd / hparams.n_head;
         const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
         const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa();
         const int64_t n_embd_gqa   = n_embd_v_gqa;
@@ -5826,16 +6049,17 @@ static bool llm_load_tensors(
                         GGML_ASSERT(hparams.n_expert_used > 0);
 
                         // MoE branch
-                        auto n_ff_exp = n_ff / hparams.n_expert_used;
+                        auto n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / hparams.n_expert_used;
                         layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert});
                         layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert});
                         layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert});
 
                         // Shared expert branch
+                        auto n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
                         layer.ffn_gate_inp_shexp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), {n_embd});
-                        layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd,   n_ff});
-                        layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {  n_ff, n_embd});
-                        layer.ffn_up_shexp   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd,   n_ff});
+                        layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp});
+                        layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd});
+                        layer.ffn_up_shexp   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp});
                     }
                 } break;
             case LLM_ARCH_PHI2:
@@ -6625,16 +6849,6 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
         }
 #endif
 
-#ifdef GGML_USE_SYCL
-        if (params.split_mode == LLAMA_SPLIT_MODE_NONE) {
-            ggml_backend_sycl_set_single_device_mode(params.main_gpu);
-            //SYCL use device index (0, 1, 2) directly, uer input device id, then convert to device index.
-            params.main_gpu = ggml_backend_sycl_get_device_index(params.main_gpu);
-        } else {
-            ggml_backend_sycl_set_mul_device_mode();
-        }
-#endif
-
         if (!llm_load_tensors(
             ml, model, params.n_gpu_layers, params.split_mode,  params.main_gpu, params.tensor_split, params.use_mlock,
             params.progress_callback, params.progress_callback_user_data
@@ -12955,113 +13169,143 @@ struct llm_bigram_bpe {
 };
 
 struct llm_tokenizer_bpe {
-    llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {}
-
-    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
-        int final_prev_index = -1;
-        bool ignore_merges = false;
-
-        std::vector<std::string> word_collection;
-        switch (vocab.type) {
-            case LLAMA_VOCAB_TYPE_BPE:
+    llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {
+        GGML_ASSERT(vocab.type == LLAMA_VOCAB_TYPE_BPE);
         switch (vocab.type_pre) {
             case LLAMA_VOCAB_PRE_TYPE_LLAMA3:
-                        ignore_merges = true;
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     // original regex from tokenizer.json
                     //"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
 
                     // adapted: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989
                     "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
-                        });
+                };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_DBRX:
             case LLAMA_VOCAB_PRE_TYPE_SMAUG:
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     // same as llama3
                     "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
-                        });
+                };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM:
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     "[\r\n]",
                     "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
                     "\\s?[!-/:-~！-／：-～‘-‟　-。]+",
                     "\\s+$",
                     "[一-龥ࠀ-一가-퟿]+",
                     "\\p{N}+",
-                        });
+                };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER:
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     "[\r\n]",
                     "\\s?\\p{L}+",
                     "\\s?\\p{P}+",
                     "[一-龥ࠀ-一가-퟿]+",
                     "\\p{N}",
-                        });
+                };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_FALCON:
-                        word_collection = unicode_regex_split(text, {
-                            "[\\p{P}\\$\\+<=>\\^~\\|]+",
+                regex_exprs = {
+                    "[\\p{P}\\$\\+<=>\\^~\\|`]+",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                     "[0-9][0-9][0-9]",
-                        });
+                };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_MPT:
                 // TODO: MPT pre-tokenization regexes are unknown
                 //       the following are close, but not exact. run the following:
                 //       ./bin/test-tokenizer-0 ../models/ggml-vocab-mpt.gguf
                 GGML_ASSERT("MPT pre-tokenization regexes are unknown - fixes needed");
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     "\\s?\\p{L}+",
                     "\\s?\\p{P}+",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
-                        });
+                };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_STARCODER:
             case LLAMA_VOCAB_PRE_TYPE_REFACT:
             case LLAMA_VOCAB_PRE_TYPE_COMMAND_R:
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     "\\p{N}",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
-                        });
+                };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_GPT2:
             case LLAMA_VOCAB_PRE_TYPE_OLMO:
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
-                        });
+                };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_STABLELM2:
             case LLAMA_VOCAB_PRE_TYPE_QWEN2:
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     // original regex from tokenizer.json
                     // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
                     "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
-                        });
+                };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_PORO:
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     " ?[^(\\s|.,!?…。，、।۔،)]+",
-                        });
+                };
                 break;
             default:
                 // default regex for BPE tokenization pre-processing
-                        word_collection = unicode_regex_split(text, {
+                regex_exprs = {
                     "[\\p{P}\\$\\+<=>\\^~\\|]+",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                     "\\p{N}+",
                     "[0-9][0-9][0-9]",
-                        });
+                };
                 break;
         }
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
     }
 
+    void append(const llama_vocab::id token_id, std::vector<llama_vocab::id> & output) const {
+        output.push_back(token_id);
+    }
+
+    bool append_bos(std::vector<llama_vocab::id> & output) const {
+        if (vocab.tokenizer_add_bos) {
+            GGML_ASSERT(vocab.special_bos_id != -1);
+            output.push_back(vocab.special_bos_id);
+            return true;
+        }
+        return false;
+    }
+
+    bool append_eos(std::vector<llama_vocab::id> & output) const {
+        if (vocab.tokenizer_add_eos) {
+            GGML_ASSERT(vocab.special_eos_id != -1);
+            output.push_back(vocab.special_eos_id);
+            return true;
+        }
+        return false;
+    }
+
+    void check_double_bos_eos(const std::vector<llama_vocab::id> & output) const {
+        if (vocab.tokenizer_add_bos && output.size() >= 2 && output[1] == vocab.special_bos_id) {
+            LLAMA_LOG_WARN(
+                "%s: Added a BOS token to the prompt as specified by the model but the prompt "
+                "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
+                "Are you sure this is what you want?\n", __FUNCTION__);
+        }
+        if (vocab.tokenizer_add_eos && output.size() >= 2 && *(output.end()-2) == vocab.special_eos_id) {
+            LLAMA_LOG_WARN(
+                "%s: Added a EOS token to the prompt as specified by the model but the prompt "
+                "also ends with a EOS token. So now the final prompt ends with 2 EOS tokens. "
+                "Are you sure this is what you want?\n", __FUNCTION__);
+        }
+    }
+
+    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+        int final_prev_index = -1;
+
+        const auto word_collection = unicode_regex_split(text, regex_exprs);
+
         symbols_final.clear();
 
         for (auto & word : word_collection) {
@@ -13071,7 +13315,7 @@ struct llm_tokenizer_bpe {
             int index = 0;
             size_t offset = 0;
 
-            if (ignore_merges && vocab.token_to_id.find(word) != vocab.token_to_id.end()) {
+            if (vocab.tokenizer_ignore_merges && vocab.token_to_id.find(word) != vocab.token_to_id.end()) {
                 symbols.emplace_back(llm_symbol{-1, -1, word.c_str(), word.size()});
                 offset = word.size();
             }
@@ -13152,10 +13396,9 @@ struct llm_tokenizer_bpe {
                     for (auto j = str.begin(); j != str.end(); ++j) {
                         std::string byte_str(1, *j);
                         auto token_multibyte = vocab.token_to_id.find(byte_str);
-                        if (token_multibyte == vocab.token_to_id.end()) {
-                            throw std::runtime_error("ERROR: byte not found in vocab");
+                        if (token_multibyte != vocab.token_to_id.end()) {
+                            output.push_back(token_multibyte->second);
                         }
-                        output.push_back((*token_multibyte).second);
                     }
                 } else {
                     output.push_back((*token).second);
@@ -13194,6 +13437,8 @@ private:
 
     const llama_vocab & vocab;
 
+    std::vector<std::string> regex_exprs;
+
     std::vector<llm_symbol> symbols;
     std::vector<llm_symbol> symbols_final;
 
@@ -13256,7 +13501,7 @@ struct llm_tokenizer_wpm {
         const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
         std::vector<std::string> words(1, "");
 
-        for (const char32_t cpt : cpts_nfd) {
+        for (const uint32_t cpt : cpts_nfd) {
             const auto flags = unicode_cpt_flags(cpt);
 
             if (flags.is_whitespace) {
@@ -13474,7 +13719,7 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
 
                 bool is_prev_special = false;
 
-                if (add_special && vocab.special_add_bos != 0) {
+                if (add_special && vocab.tokenizer_add_bos) {
                     GGML_ASSERT(vocab.special_bos_id != -1);
                     output.push_back(vocab.special_bos_id);
                     is_prev_special = true;
@@ -13484,7 +13729,7 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
                         auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
 
-                        if (vocab.add_space_prefix) {
+                        if (vocab.tokenizer_add_space_prefix) {
                             if (!output.size() || is_prev_special) {  // prefix with space if first token
                                 raw_text = " " + raw_text;
                             }
@@ -13502,23 +13747,24 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                     }
                 }
 
-                if (add_special && vocab.special_add_bos != 0 && output.size() >= 2 && output[1] == vocab.special_bos_id) {
+                if (add_special && vocab.tokenizer_add_bos && output.size() >= 2 && output[1] == vocab.special_bos_id) {
                     LLAMA_LOG_WARN(
                         "%s: Added a BOS token to the prompt as specified by the model but the prompt "
                         "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
                         "Are you sure this is what you want?\n", __FUNCTION__);
                 }
 
-                if (add_special && vocab.special_add_eos == 1) {
+                if (add_special && vocab.tokenizer_add_eos) {
                     GGML_ASSERT(vocab.special_eos_id != -1);
                     output.push_back(vocab.special_eos_id);
                 }
             } break;
         case LLAMA_VOCAB_TYPE_BPE:
             {
-                if (add_special && vocab.special_add_bos != 0) {
-                    GGML_ASSERT(vocab.special_bos_id != -1);
-                    output.push_back(vocab.special_bos_id);
+                llm_tokenizer_bpe tokenizer(vocab);
+
+                if (add_special) {
+                    tokenizer.append_bos(output);
                 }
 
                 for (const auto & fragment : fragment_buffer) {
@@ -13528,23 +13774,15 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        llm_tokenizer_bpe tokenizer(vocab);
                         tokenizer.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
-                        output.push_back(fragment.token);
+                        tokenizer.append(fragment.token, output);
                     }
                 }
 
-                if (add_special && vocab.special_add_bos != 0 && output.size() >= 2 && output[1] == vocab.special_bos_id) {
-                    LLAMA_LOG_WARN(
-                        "%s: Added a BOS token to the prompt as specified by the model but the prompt "
-                        "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
-                        "Are you sure this is what you want?\n", __FUNCTION__);
-                }
-
-                if (add_special && vocab.special_add_eos == 1) {
-                    GGML_ASSERT(vocab.special_add_eos != -1);
-                    output.push_back(vocab.special_eos_id);
+                if (add_special) {
+                    tokenizer.append_eos(output);
+                    tokenizer.check_double_bos_eos(output);
                 }
             } break;
         case LLAMA_VOCAB_TYPE_WPM:
@@ -16070,6 +16308,11 @@ struct llama_context * llama_new_context_with_model(
         params.flash_attn = false;
     }
 
+    if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) {
+        LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
     if (params.type_v != GGML_TYPE_F16 && !params.flash_attn) {
         LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
         return nullptr;
@@ -16241,8 +16484,7 @@ struct llama_context * llama_new_context_with_model(
         if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
             ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu);
             if (backend == nullptr) {
-                int main_gpu_id = ggml_backend_sycl_get_device_id(model->main_gpu);
-                LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, main_gpu_id, model->main_gpu);
+                LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, model->main_gpu);
                 llama_free(ctx);
                 return nullptr;
             }
@@ -18113,11 +18355,11 @@ llama_token llama_token_nl(const struct llama_model * model) {
 }
 
 int32_t llama_add_bos_token(const struct llama_model * model) {
-    return model->vocab.special_add_bos;
+    return model->vocab.tokenizer_add_bos;
 }
 
 int32_t llama_add_eos_token(const struct llama_model * model) {
-    return model->vocab.special_add_eos;
+    return model->vocab.tokenizer_add_eos;
 }
 
 llama_token llama_token_prefix(const struct llama_model * model) {

scripts/gen-unicode-data.py

@@ -1,83 +1,143 @@
-import regex
-import ctypes
+import array
 import unicodedata
-
-
-class CoodepointFlags (ctypes.Structure):
-    _fields_ = [  # see definition in unicode.h
-        ("is_undefined",   ctypes.c_uint16, 1),
-        ("is_number",      ctypes.c_uint16, 1),  # regex: \p{N}
-        ("is_letter",      ctypes.c_uint16, 1),  # regex: \p{L}
-        ("is_separator",   ctypes.c_uint16, 1),  # regex: \p{Z}
-        ("is_accent_mark", ctypes.c_uint16, 1),  # regex: \p{M}
-        ("is_punctuation", ctypes.c_uint16, 1),  # regex: \p{P}
-        ("is_symbol",      ctypes.c_uint16, 1),  # regex: \p{S}
-        ("is_control",     ctypes.c_uint16, 1),  # regex: \p{C}
-    ]
-
-
-assert (ctypes.sizeof(CoodepointFlags) == 2)
+import requests
 
 
 MAX_CODEPOINTS = 0x110000
 
-regex_number      = regex.compile(r'\p{N}')
-regex_letter      = regex.compile(r'\p{L}')
-regex_separator   = regex.compile(r'\p{Z}')
-regex_accent_mark = regex.compile(r'\p{M}')
-regex_punctuation = regex.compile(r'\p{P}')
-regex_symbol      = regex.compile(r'\p{S}')
-regex_control     = regex.compile(r'\p{C}')
-regex_whitespace  = regex.compile(r'\s')
+UNICODE_DATA_URL = "https://www.unicode.org/Public/UCD/latest/ucd/UnicodeData.txt"
 
-codepoint_flags = (CoodepointFlags * MAX_CODEPOINTS)()
+
+# see https://www.unicode.org/L2/L1999/UnicodeData.html
+def unicode_data_iter():
+    res = requests.get(UNICODE_DATA_URL)
+    res.raise_for_status()
+    data = res.content.decode()
+
+    prev = []
+
+    for line in data.splitlines():
+        # ej: 0000;<control>;Cc;0;BN;;;;;N;NULL;;;;
+        line = line.split(";")
+
+        cpt = int(line[0], base=16)
+        assert cpt < MAX_CODEPOINTS
+
+        cpt_lower = int(line[-2] or "0", base=16)
+        assert cpt_lower < MAX_CODEPOINTS
+
+        cpt_upper = int(line[-3] or "0", base=16)
+        assert cpt_upper < MAX_CODEPOINTS
+
+        categ = line[2].strip()
+        assert len(categ) == 2
+
+        bidir = line[4].strip()
+        assert len(categ) == 2
+
+        name = line[1]
+        if name.endswith(", First>"):
+            prev = (cpt, cpt_lower, cpt_upper, categ, bidir)
+            continue
+        if name.endswith(", Last>"):
+            assert prev[1:] == (0, 0, categ, bidir)
+            for c in range(prev[0], cpt):
+                yield (c, cpt_lower, cpt_upper, categ, bidir)
+
+        yield (cpt, cpt_lower, cpt_upper, categ, bidir)
+
+
+# see definition in unicode.h
+CODEPOINT_FLAG_UNDEFINED   = 0x0001  #
+CODEPOINT_FLAG_NUMBER      = 0x0002  # \p{N}
+CODEPOINT_FLAG_LETTER      = 0x0004  # \p{L}
+CODEPOINT_FLAG_SEPARATOR   = 0x0008  # \p{Z}
+CODEPOINT_FLAG_MARK        = 0x0010  # \p{M}
+CODEPOINT_FLAG_PUNCTUATION = 0x0020  # \p{P}
+CODEPOINT_FLAG_SYMBOL      = 0x0040  # \p{S}
+CODEPOINT_FLAG_CONTROL     = 0x0080  # \p{C}
+
+UNICODE_CATEGORY_TO_FLAG = {
+    "Cn": CODEPOINT_FLAG_UNDEFINED,    # Undefined
+    "Cc": CODEPOINT_FLAG_CONTROL,      # Control
+    "Cf": CODEPOINT_FLAG_CONTROL,      # Format
+    "Co": CODEPOINT_FLAG_CONTROL,      # Private Use
+    "Cs": CODEPOINT_FLAG_CONTROL,      # Surrrogate
+    "Ll": CODEPOINT_FLAG_LETTER,       # Lowercase Letter
+    "Lm": CODEPOINT_FLAG_LETTER,       # Modifier Letter
+    "Lo": CODEPOINT_FLAG_LETTER,       # Other Letter
+    "Lt": CODEPOINT_FLAG_LETTER,       # Titlecase Letter
+    "Lu": CODEPOINT_FLAG_LETTER,       # Uppercase Letter
+    "L&": CODEPOINT_FLAG_LETTER,       # Cased Letter
+    "Mc": CODEPOINT_FLAG_MARK,         # Spacing Mark
+    "Me": CODEPOINT_FLAG_MARK,         # Enclosing Mark
+    "Mn": CODEPOINT_FLAG_MARK,         # Nonspacing Mark
+    "Nd": CODEPOINT_FLAG_NUMBER,       # Decimal Number
+    "Nl": CODEPOINT_FLAG_NUMBER,       # Letter Number
+    "No": CODEPOINT_FLAG_NUMBER,       # Other Number
+    "Pc": CODEPOINT_FLAG_PUNCTUATION,  # Connector Punctuation
+    "Pd": CODEPOINT_FLAG_PUNCTUATION,  # Dash Punctuation
+    "Pe": CODEPOINT_FLAG_PUNCTUATION,  # Close Punctuation
+    "Pf": CODEPOINT_FLAG_PUNCTUATION,  # Final Punctuation
+    "Pi": CODEPOINT_FLAG_PUNCTUATION,  # Initial Punctuation
+    "Po": CODEPOINT_FLAG_PUNCTUATION,  # Other Punctuation
+    "Ps": CODEPOINT_FLAG_PUNCTUATION,  # Open Punctuation
+    "Sc": CODEPOINT_FLAG_SYMBOL,       # Currency Symbol
+    "Sk": CODEPOINT_FLAG_SYMBOL,       # Modifier Symbol
+    "Sm": CODEPOINT_FLAG_SYMBOL,       # Math Symbol
+    "So": CODEPOINT_FLAG_SYMBOL,       # Other Symbol
+    "Zl": CODEPOINT_FLAG_SEPARATOR,    # Line Separator
+    "Zp": CODEPOINT_FLAG_SEPARATOR,    # Paragraph Separator
+    "Zs": CODEPOINT_FLAG_SEPARATOR,    # Space Separator
+}
+
+
+codepoint_flags = array.array('H', [CODEPOINT_FLAG_UNDEFINED]) * MAX_CODEPOINTS
 table_whitespace = []
 table_lowercase = []
 table_uppercase = []
 table_nfd = []
 
-for codepoint in range(MAX_CODEPOINTS):
+for (cpt, cpt_lower, cpt_upper, categ, bidir) in unicode_data_iter():
     # convert codepoint to unicode character
-    char = chr(codepoint)
+    char = chr(cpt)
 
-    # regex categories
-    flags = codepoint_flags[codepoint]
-    flags.is_number      = bool(regex_number.match(char))
-    flags.is_letter      = bool(regex_letter.match(char))
-    flags.is_separator   = bool(regex_separator.match(char))
-    flags.is_accent_mark = bool(regex_accent_mark.match(char))
-    flags.is_punctuation = bool(regex_punctuation.match(char))
-    flags.is_symbol      = bool(regex_symbol.match(char))
-    flags.is_control     = bool(regex_control.match(char))
-    flags.is_undefined   = bytes(flags)[0] == 0
-    assert (not flags.is_undefined)
-
-    # whitespaces
-    if bool(regex_whitespace.match(char)):
-        table_whitespace.append(codepoint)
+    # codepoint category flags
+    codepoint_flags[cpt] = UNICODE_CATEGORY_TO_FLAG[categ]
 
     # lowercase conversion
-    lower = ord(char.lower()[0])
-    if codepoint != lower:
-        table_lowercase.append((codepoint, lower))
+    if cpt_lower:
+        table_lowercase.append((cpt, cpt_lower))
 
     # uppercase conversion
-    upper = ord(char.upper()[0])
-    if codepoint != upper:
-        table_uppercase.append((codepoint, upper))
+    if cpt_upper:
+        table_uppercase.append((cpt, cpt_upper))
 
     # NFD normalization
     norm = ord(unicodedata.normalize('NFD', char)[0])
-    if codepoint != norm:
-        table_nfd.append((codepoint, norm))
+    if cpt != norm:
+        table_nfd.append((cpt, norm))
+
+
+# whitespaces, see "<White_Space>" https://www.unicode.org/Public/UCD/latest/ucd/PropList.txt
+table_whitespace.extend(range(0x0009, 0x000D + 1))
+table_whitespace.extend(range(0x2000, 0x200A + 1))
+table_whitespace.extend([0x0020, 0x0085, 0x00A0, 0x1680, 0x2028, 0x2029, 0x202F, 0x205F, 0x3000])
+
+
+# sort by codepoint
+table_whitespace.sort()
+table_lowercase.sort()
+table_uppercase.sort()
+table_nfd.sort()
 
 
 # group ranges with same flags
 ranges_flags = [(0, codepoint_flags[0])]  # start, flags
 for codepoint, flags in enumerate(codepoint_flags):
-    if bytes(flags) != bytes(ranges_flags[-1][1]):
+    if flags != ranges_flags[-1][1]:
         ranges_flags.append((codepoint, flags))
-ranges_flags.append((MAX_CODEPOINTS, CoodepointFlags()))
+ranges_flags.append((MAX_CODEPOINTS, 0x0000))
 
 
 # group ranges with same nfd
@@ -90,8 +150,8 @@ for codepoint, norm in table_nfd:
     ranges_nfd[-1] = (start, codepoint, norm)
 
 
-# Generate 'unicode-data.cpp'
-
+# Generate 'unicode-data.cpp':
+#   python ./scripts//gen-unicode-data.py > unicode-data.cpp
 
 def out(line=""):
     print(line, end='\n')  # noqa
@@ -110,12 +170,12 @@ out("""\
 
 out("const std::vector<std::pair<uint32_t, uint16_t>> unicode_ranges_flags = {  // start, flags // last=next_start-1")
 for codepoint, flags in ranges_flags:
-    flags = int.from_bytes(bytes(flags), "little")
     out("{0x%06X, 0x%04X}," % (codepoint, flags))
 out("};\n")
 
 out("const std::unordered_set<uint32_t> unicode_set_whitespace = {")
-out(", ".join("0x%06X" % cpt for cpt in table_whitespace))
+for codepoint in table_whitespace:
+    out("0x%06X," % codepoint)
 out("};\n")
 
 out("const std::unordered_map<uint32_t, uint32_t> unicode_map_lowercase = {")

scripts/sync-ggml.last

@@ -1 +1 @@
-2aae01fd9b8f9399f343cf18f46f38996ef52e2c
+5653a195935ea3ac54652644c9daf154dbc1571b

sgemm.cpp

@@ -43,8 +43,10 @@
 // [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
 //     Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
 
+#if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Wpedantic"
 #pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
 
 #include "sgemm.h"
 #include "ggml-impl.h"

tests/test-backend-ops.cpp

@@ -1063,6 +1063,33 @@ struct test_sqr : public test_case {
     }
 };
 
+// GGML_OP_SQRT
+struct test_sqrt : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_sqrt(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_sqrt(ctx, a);
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        // fill with positive values
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            init_tensor_uniform(t, 0.0f, 100.0f);
+        }
+    }
+};
+
 // GGML_OP_CLAMP
 struct test_clamp : public test_case {
     const ggml_type type;
@@ -2200,6 +2227,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     }
 
     test_cases.emplace_back(new test_sqr());
+    test_cases.emplace_back(new test_sqrt());
     test_cases.emplace_back(new test_clamp());
 
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10,  1,  1}, 5));

tests/test-tokenizer-random.py

@@ -11,13 +11,15 @@ import logging
 import argparse
 import subprocess
 import random
+import unicodedata
 
 from typing import Callable, Iterator
 
 import cffi
 from transformers import AutoTokenizer
 
-logger = logging.getLogger("test-tokenizer-random-bpe")
+
+logger = logging.getLogger("test-tokenizer-random")
 
 
 class LibLlama:
@@ -155,9 +157,14 @@ def generator_custom_text_edge_cases() -> Iterator[str]:
         'Cửa Việt',   # llama-3, ignore_merges = true
         '<s>a',       # Phi-3 fail
         '<unk><|endoftext|><s>',  # Phi-3 fail
-        'a\na',       # TODO: Bert fail
+        'a\na',            # bert fail
+        '"`',              # falcon
+        ' \u2e4e',         # falcon
+        'a\xa0\xa0\x00b',  # jina-v2-es
+        'one <mask>',      # jina-v2-es  <mask> lstrip=true
         'a </s> b',        # rstrip phi-3
         'a <mask> b',      # lstrip jina-v2
+        '\xa0aC',          # deepseek
     ]
 
 
@@ -189,17 +196,23 @@ def generator_random_added_tokens(tokenizer, iterations=100) -> Iterator[str]:
     for m in range(iterations):
         rand.seed(m)
         words = rand.choices(all_tokens, k=500)
-        if words[0] == tokenizer.bos_token:  # skip spam warning of double BOS
+        if words and words[0] == tokenizer.bos_token:  # skip spam warning of double BOS
             while len(words) > 1 and words[1] == tokenizer.bos_token:  # leave one starting BOS
                 words.pop(0)
             if tokenizer.add_bos_token:  # drop all starting BOS
                 words.pop(0)
+        if words and words[-1] == tokenizer.eos_token:  # skip spam warning of double EOS
+            while len(words) > 1 and words[-2] == tokenizer.eos_token:  # leave one trailing EOS
+                words.pop(-1)
+            if tokenizer.add_bos_token:  # drop all trailing EOS
+                words.pop(-1)
         yield "".join(words)
 
 
 def generator_random_chars(iterations=100) -> Iterator[str]:
     """Brute force random text with simple characters"""
 
+    NUM_WORDS = 400
     WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
     CHARS = list(sorted(set("""
         ABCDEFGHIJKLMNOPQRSTUVWXYZ
@@ -213,12 +226,50 @@ def generator_random_chars(iterations=100) -> Iterator[str]:
     for m in range(iterations):
         rand.seed(m)
         text = []
-        num_words = rand.randint(300, 400)
-        for i in range(num_words):
+        for _ in range(NUM_WORDS):
             k = rand.randint(1, 7)
             word = rand.choices(CHARS, k=k)
-            space = rand.choice(WHITESPACES)
-            text.append("".join(word) + space)
+            word.append(rand.choice(WHITESPACES))
+            text.append("".join(word))
+        yield "".join(text)
+
+
+def generator_unicodes() -> Iterator[str]:
+    """Iterate unicode characters"""
+
+    MAX_CODEPOINTS = 0x30000  # 0x110000
+
+    def _valid(cpt):
+        if cpt >= 0x30000:  # unassigned and supplement­ary
+            return False
+        if 0x00D800 <= cpt <= 0x00F8FF:  # Surrogates
+            return False
+        if unicodedata.category(chr(cpt)) == "Cn":
+            return False
+        return True
+
+    characters = [chr(cpt) for cpt in range(1, MAX_CODEPOINTS) if _valid(cpt)]
+
+    yield from characters
+
+
+def generator_random_unicodes(iterations=100) -> Iterator[str]:
+    """Brute force random text with unicode characters"""
+
+    NUM_WORDS = 200
+    WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
+
+    characters = list(generator_unicodes())
+
+    rand = random.Random()
+    for m in range(iterations):
+        rand.seed(m)
+        text = []
+        for _ in range(NUM_WORDS):
+            k = rand.randint(1, 7)
+            word = rand.choices(characters, k=k)
+            word.append(rand.choice(WHITESPACES))
+            text.append("".join(word))
         yield "".join(text)
 
 
@@ -256,25 +307,7 @@ def generator_random_vocab_words(vocab: list[str], iterations=100) -> Iterator[s
         yield "".join(text)
 
 
-def generator_random_bytes(iterations=100) -> Iterator[str]:
-    """Brute force random bytes"""
-
-    WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
-
-    rand = random.Random()
-    for m in range(iterations):
-        rand.seed(m)
-        text = []
-        num_words = rand.randint(300, 400)
-        for i in range(num_words):
-            k = rand.randint(1, 8)
-            word = [chr(r) for r in rand.randbytes(k) if r]
-            word.append(rand.choice(WHITESPACES))
-            text.append("".join(word))
-        yield "".join(text)
-
-
-def test_compare_tokenizer(func_tokenize1: Callable, func_tokenize2: Callable, generator: Iterator[str]):
+def compare_tokenizers(func_tokenize1: Callable, func_tokenize2: Callable, generator: Iterator[str]):
 
     def find_first_mismatch(ids1: list[int], ids2: list[int]):
         for i, (a, b) in enumerate(zip(ids1, ids2)):
@@ -284,20 +317,34 @@ def test_compare_tokenizer(func_tokenize1: Callable, func_tokenize2: Callable, g
             return -1
         return min(len(ids1), len(ids2))
 
-    t0 = time.perf_counter()
+    t_tokenizer1 = 0
+    t_tokenizer2 = 0
+    t_start = time.perf_counter()
+    num_errors = 10
+
     logger.info("%s: %s" % (generator.__name__, "ini"))
     for text in generator:
+        # print(repr(text), hex(ord(text[0])), text.encode())
+        t0 = time.perf_counter()
         ids1 = func_tokenize1(text)
+        t1 = time.perf_counter()
         ids2 = func_tokenize2(text)
+        t2 = time.perf_counter()
+        t_tokenizer1 += t1 - t0
+        t_tokenizer2 += t2 - t1
         if ids1 != ids2:
             i = find_first_mismatch(ids1, ids2)
             ids1 = list(ids1)[max(0, i - 2) : i + 5 + 1]
             ids2 = list(ids2)[max(0, i - 2) : i + 5 + 1]
-            logger.info(" TokenIDs: " + str(ids1))
-            logger.info(" Expected: " + str(ids2))
-            raise Exception()
-    t1 = time.perf_counter()
-    logger.info("%s: end, time: %.3f secs" % (generator.__name__, t1 - t0))
+            logger.error(" TokenIDs: " + str(ids1))
+            logger.error(" Expected: " + str(ids2))
+            # raise Exception()
+            num_errors += 1
+            if num_errors > 10:
+                break
+
+    t_total = time.perf_counter() - t_start
+    logger.info("%s: end,  tok1: %.3f  tok2: %.3f  total: %.3f" % (generator.__name__, t_tokenizer1, t_tokenizer2, t_total))
 
 
 def main(argv: list[str] = None):
@@ -307,7 +354,8 @@ def main(argv: list[str] = None):
     parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
     args = parser.parse_args(argv)
 
-    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
+    logging.basicConfig(level = logging.DEBUG if args.verbose else logging.INFO)
+    logger.info(f"VOCABFILE: '{args.vocab_file}'")
 
     model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=4096))
     tokenizer = AutoTokenizer.from_pretrained(args.dir_tokenizer)
@@ -321,18 +369,22 @@ def main(argv: list[str] = None):
     ids = func_tokenize2("a")
     assert 1 <= len(ids) <= 3
     add_bos_token = len(ids) > 1 and tokenizer.bos_token_id == ids[0]
+    add_eos_token = len(ids) > 1 and tokenizer.eos_token_id == ids[-1]
     tokenizer.add_bos_token = getattr(tokenizer, "add_bos_token", add_bos_token)
+    tokenizer.add_eos_token = getattr(tokenizer, "add_eos_token", add_eos_token)
 
     vocab = list(sorted(tokenizer.batch_decode(list(tokenizer.get_vocab().values()), skip_special_tokens=True)))
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text())
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases())
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_vocab_words(vocab))
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_added_lr_strip(tokenizer))
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_added_tokens(tokenizer, 10_000))
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_chars(10_000))
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_chars(vocab, 10_000))
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_words(vocab, 5_000))
-    # test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_bytes(10_000)) # FAIL
+
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_custom_text())
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases())
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_unicodes())
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_vocab_words(vocab))
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_added_lr_strip(tokenizer))
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_added_tokens(tokenizer, 10_000))
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_chars(10_000))
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_unicodes(10_000))
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_vocab_chars(vocab, 10_000))
+    compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_vocab_words(vocab, 5_000))
 
     model.free()
 
@@ -340,20 +392,40 @@ def main(argv: list[str] = None):
 if __name__ == "__main__":
     # main()
 
+    logging.basicConfig(
+        level    = logging.DEBUG,
+        format   = "%(asctime)s.%(msecs)03d %(name)s %(levelname)s %(message)s",
+        datefmt  = "%Y-%m-%d %H:%M:%S",
+        filename = logger.name + ".log",
+        filemode = "a"
+    )
+
     path_tokenizers   = "./models/tokenizers/"
     path_vocab_format = "./models/ggml-vocab-%s.gguf"
 
     # import os
     # tokenizers = os.listdir(path_tokenizers)
     tokenizers = [
-        "llama-spm",   # SPM
-        "phi-3",       # SPM
-        "jina-v2-en",  # WPM
-        "bert-bge",    # WPM
+        # "llama-spm",   # SPM
+        # "phi-3",       # SPM
+        # "bert-bge",    # WPM
+        # "jina-v2-en",  # WPM
+        "gpt-2",          # BPE
+        "llama-bpe",      # BPE
+        "falcon",         # BPE
+        "starcoder",      # BPE
+        "jina-v2-es",     # BPE
+        "jina-v2-de",     # BPE
+        "jina-v2-code",   # BPE
+        "smaug-bpe",      # BPE
+        "phi-2",          # BPE
+        "deepseek-coder", # BPE
+        "deepseek-llm",   # BPE
     ]
 
     for tokenizer in tokenizers:
-        print("\n" + "=" * 50 + "\n" + tokenizer + "\n")  # noqa
+        logger.info("=" * 50)
+        logger.info(f"TOKENIZER: '{tokenizer}'")
         vocab_file = path_vocab_format % tokenizer
         dir_tokenizer = path_tokenizers + "/" + tokenizer
         main([vocab_file, dir_tokenizer, "--verbose"])

unicode.cpp

@@ -226,8 +226,9 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
         assert(offset_end <= cpts.size());
         start = offset_end;
 
-        auto _get_cpt = [&] (const size_t pos) -> char32_t {
-            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
+        static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
+        auto _get_cpt = [&] (const size_t pos) -> uint32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
         auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
@@ -253,18 +254,18 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
         };
 
         for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
-            const char32_t cpt = _get_cpt(pos);
+            const uint32_t cpt = _get_cpt(pos);
             const auto flags = _get_flags(pos);
 
             // regex: 's|'t|'re|'ve|'m|'ll|'d
             if (cpt == '\'' && pos+1 < offset_end) {
-                char32_t cpt_next = _get_cpt(pos+1);
+                uint32_t cpt_next = _get_cpt(pos+1);
                 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
                     pos += _add_token(pos+2);
                     continue;
                 }
                 if (pos+2 < offset_end) {
-                    char32_t cpt_next_next = _get_cpt(pos+2);
+                    uint32_t cpt_next_next = _get_cpt(pos+2);
                     if ((cpt_next == 'r' && cpt_next_next == 'e') ||
                         (cpt_next == 'v' && cpt_next_next == 'e') ||
                         (cpt_next == 'l' && cpt_next_next == 'l')) {
@@ -309,7 +310,7 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
             }
 
             // regex: \s+(?!\S)
-            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) {
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
                 pos += num_whitespaces - 1;
                 _add_token(pos);
                 continue;
@@ -344,8 +345,9 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
         assert(offset_end <= cpts.size());
         start = offset_end;
 
-        auto _get_cpt = [&] (const size_t pos) -> char32_t {
-            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
+        static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
+        auto _get_cpt = [&] (const size_t pos) -> uint32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
         auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
@@ -371,18 +373,18 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
         };
 
         for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
-            const char32_t cpt = _get_cpt(pos);
+            const uint32_t cpt = _get_cpt(pos);
             const auto flags = _get_flags(pos);
 
             // regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
             if (cpt == '\'' && pos+1 < offset_end) {
-                char32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
+                uint32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
                 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
                     pos += _add_token(pos+2);
                     continue;
                 }
                 if (pos+2 < offset_end) {
-                    char32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
+                    uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
                     if ((cpt_next == 'r' && cpt_next_next == 'e') ||
                         (cpt_next == 'v' && cpt_next_next == 'e') ||
                         (cpt_next == 'l' && cpt_next_next == 'l')) {
@@ -424,7 +426,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
                 while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
                     flags2 = _get_flags(++pos);
                 }
-                char32_t cpt2 = _get_cpt(pos);
+                uint32_t cpt2 = _get_cpt(pos);
                 while (cpt2 == '\r' || cpt2 == '\n') {
                     cpt2 = _get_cpt(++pos);
                 }
@@ -435,7 +437,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             size_t num_whitespaces = 0;
             size_t last_end_r_or_n = 0;
             while (_get_flags(pos+num_whitespaces).is_whitespace) {
-                char32_t cpt2 = _get_cpt(pos+num_whitespaces);
+                uint32_t cpt2 = _get_cpt(pos+num_whitespaces);
                 if (cpt2 == '\r' || cpt2 == '\n') {
                     last_end_r_or_n = pos + num_whitespaces + 1;
                 }
@@ -450,7 +452,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             }
 
             // regex: \s+(?!\S)
-            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) {
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
                 pos += num_whitespaces - 1;
                 _add_token(pos);
                 continue;
@@ -626,7 +628,7 @@ uint8_t unicode_utf8_to_byte(const std::string & utf8) {
     return map.at(utf8);
 }
 
-char32_t unicode_tolower(char32_t cp) {
+uint32_t unicode_tolower(uint32_t cp) {
     auto it = unicode_map_lowercase.find(cp);
     return it == unicode_map_lowercase.end() ? cp : it->second;
 }
@@ -679,10 +681,14 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 continue;
             }
 
-            const int cpt_flag = unicode_cpt_flags(cpts[i]).category_flag();
+            const auto flags = unicode_cpt_flags(cpts[i]);
 
-            if (k_ucat_cpt.find(cpt_flag) != k_ucat_cpt.end()) {
-                text_collapsed[i] = k_ucat_cpt.at(cpt_flag);
+            if (flags.is_whitespace) {
+                //NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does.
+                //text_collapsed[i] = (char) 0x85;  // <Next Line> as whitespace fallback
+                text_collapsed[i] = (char) 0x0B;    // <vertical tab> as whitespace fallback
+            } else if (k_ucat_cpt.find(flags.category_flag()) != k_ucat_cpt.end()) {
+                text_collapsed[i] = k_ucat_cpt.at(flags.category_flag());
             } else {
                 text_collapsed[i] = (char) 0xD0; // fallback
             }
@@ -766,9 +772,16 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets);
             } else {
                 // no unicode category used, we can use std::wregex directly
-                const std::wstring wtext       = unicode_wstring_from_utf8(text);
                 const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr);
 
+                // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
+                std::wstring wtext(cpts.begin(), cpts.end());
+                for (size_t i = 0; i < wtext.size(); ++i) {
+                    if (wtext[i] > 0x7F && unicode_cpt_flags(wtext[i]).is_whitespace) {
+                        wtext[i] = 0x0B;
+                    }
+                }
+
                 //printf("text: %s\n", text.c_str());
                 //printf("regex_expr: %s\n", regex_expr.c_str());
                 bpe_offsets = unicode_regex_split_stl(wtext, wregex_expr, bpe_offsets);

unicode.h

@@ -58,6 +58,6 @@ codepoint_flags unicode_cpt_flags(const std::string & utf8);
 std::string unicode_byte_to_utf8(uint8_t byte);
 uint8_t unicode_utf8_to_byte(const std::string & utf8);
 
-char32_t unicode_tolower(char32_t cp);
+uint32_t unicode_tolower(uint32_t cp);
 
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);

vulkan-shaders/add.comp

@@ -0,0 +1,12 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    if (gl_GlobalInvocationID.x >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) + FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+}

vulkan-shaders/argsort.comp

@@ -0,0 +1,71 @@
+#version 450
+
+#include "types.comp"
+
+#define BLOCK_SIZE 1024
+#define ASC 0
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1)          buffer D {int data_d[];};
+
+layout (push_constant) uniform parameter {
+    uint ncols;
+    uint ncols_pad;
+    uint order;
+} p;
+
+shared int dst_row[BLOCK_SIZE];
+
+void swap(uint idx0, uint idx1) {
+    int tmp = dst_row[idx0];
+    dst_row[idx0] = dst_row[idx1];
+    dst_row[idx1] = tmp;
+}
+
+void main() {
+    // bitonic sort
+    const int col = int(gl_LocalInvocationID.x);
+    const uint row = gl_WorkGroupID.y;
+
+    if (col >= p.ncols_pad) {
+        return;
+    }
+
+    const uint row_offset = row * p.ncols;
+
+    // initialize indices
+    dst_row[col] = col;
+    barrier();
+
+    for (uint k = 2; k <= p.ncols_pad; k *= 2) {
+        for (uint j = k / 2; j > 0; j /= 2) {
+            const uint ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (dst_row[col] >= p.ncols ||
+                        (dst_row[ixj] < p.ncols && (p.order == ASC ?
+                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
+                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
+                    ) {
+                        swap(col, ixj);
+                    }
+                } else {
+                    if (dst_row[ixj] >= p.ncols ||
+                        (dst_row[col] < p.ncols && (p.order == ASC ?
+                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
+                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
+                    ) {
+                        swap(col, ixj);
+                    }
+                }
+            }
+            barrier();
+        }
+    }
+
+    if (col < p.ncols) {
+        data_d[row_offset + col] = dst_row[col];
+    }
+}

vulkan-shaders/clamp.comp

@@ -0,0 +1,13 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    if (gl_GlobalInvocationID.x >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
+    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
+}

vulkan-shaders/copy.comp

@@ -0,0 +1,16 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    if (gl_GlobalInvocationID.x >= p.ne) {
+        return;
+    }
+
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
+#else
+    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = data_a[src0_idx(gl_GlobalInvocationID.x)];
+#endif
+}

vulkan-shaders/dequant_f32.comp

@@ -0,0 +1,20 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {float data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x * 16;
+
+    if (i >= p.nel) {
+        return;
+    }
+
+    [[unroll]] for (uint l = 0; l < 16; l++) {
+        data_b[i + l] = D_TYPE(data_a[i + l]);
+    }
+}

vulkan-shaders/dequant_funcs.comp

@@ -0,0 +1,60 @@
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+#endif
+
+#if defined(DATA_A_F32)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
+}
+#endif
+
+#if defined(DATA_A_F16)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
+}
+#endif
+
+#if defined(DATA_A_Q4_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
+}
+#endif
+
+#if defined(DATA_A_Q4_1)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const float m = float(data_a[a_offset + ib].m);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2(vui & 0xF, vui >> 4) * d + m;
+}
+#endif
+
+#if defined(DATA_A_Q5_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0];
+    const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
+}
+#endif
+
+#if defined(DATA_A_Q5_1)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const float m = float(data_a[a_offset + ib].m);
+    const uint uint_qh = data_a[a_offset + ib].qh;
+    const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
+}
+#endif

vulkan-shaders/dequant_head.comp

@@ -0,0 +1,13 @@
+#extension GL_EXT_control_flow_attributes : require
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint M;
+    uint K;
+    uint stride_a;
+    uint stride_b;
+    uint nel;
+} p;
+
+#include "types.comp"

vulkan-shaders/dequant_q2_k.comp

@@ -0,0 +1,34 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint tid = gl_LocalInvocationID.x;
+        const uint ip = tid / 32;
+        const uint il = tid - 32 * ip;
+        const uint is = 8 * ip + il / 16;
+
+        const uint y_idx = i * QUANT_K + 128 * ip + il;
+
+        const uint ql_idx = 32 * ip + il;
+        const uint8_t qs = data_a[i].qs[32 * ip + il];
+
+        FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+        data_b[y_idx +  0] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+0] & 0xF) * ((qs >> 0) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+0] >> 4));
+        data_b[y_idx + 32] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+2] & 0xF) * ((qs >> 2) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+2] >> 4));
+        data_b[y_idx + 64] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+4] & 0xF) * ((qs >> 4) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+4] >> 4));
+        data_b[y_idx + 96] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+6] & 0xF) * ((qs >> 6) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+6] >> 4));
+    }
+}

vulkan-shaders/dequant_q3_k.comp

@@ -0,0 +1,42 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = uint(gl_WorkGroupID.x * 256 + wgy);
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint r = gl_LocalInvocationID.x / 4;
+        const uint tid = r / 2;
+        const uint is0 = r % 2;
+        const uint l0 = 16 * is0 + 4 * (gl_LocalInvocationID.x % 4);
+        const uint n = tid / 4;
+        const uint j = tid - 4*n;
+
+        const uint8_t m = uint8_t(1 << (4*n + j));
+        const uint is = 8*n + 2*j + is0;
+        const uint shift = 2*j;
+
+        const int8_t us = int8_t(is <  4 ? (data_a[i].scales[is-0] & 0xF) | (((data_a[i].scales[is+8] >> 0) & 3) << 4) :
+                                 is <  8 ? (data_a[i].scales[is-0] & 0xF) | (((data_a[i].scales[is+4] >> 2) & 3) << 4) :
+                                 is < 12 ? (data_a[i].scales[is-8] >>  4) | (((data_a[i].scales[is+0] >> 4) & 3) << 4) :
+                                           (data_a[i].scales[is-8] >>  4) | (((data_a[i].scales[is-4] >> 6) & 3) << 4));
+        const FLOAT_TYPE d_all = FLOAT_TYPE(data_a[i].d);
+        const FLOAT_TYPE dl    = d_all * FLOAT_TYPE(us - 32);
+
+        const uint y_idx = i * QUANT_K + 128 * n + 32 * j;
+        const uint qs_idx = 32*n;
+
+        for (uint l = l0; l < l0 + 4; ++l) {
+            data_b[y_idx + l] = D_TYPE(dl * FLOAT_TYPE(int8_t((data_a[i].qs[qs_idx + l] >> shift) & 3) - (((data_a[i].hmask[l] & m) != 0) ? 0 : 4)));
+        }
+    }
+}

vulkan-shaders/dequant_q4_0.comp

@@ -0,0 +1,32 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q4_0 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint b_idx = 1024*i + 32*ir + 8*il;
+
+    const float d = float(data_a[ib].d);
+    const float dm = -8.0f * d;
+
+    const uint q_idx = 8*il;
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * (data_a[ib].qs[q_idx + l] & 0xF) + dm);
+        data_b[b_idx + l + 16] = D_TYPE(d * (data_a[ib].qs[q_idx + l] >>  4) + dm);
+    }
+}

vulkan-shaders/dequant_q4_1.comp

@@ -0,0 +1,32 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q4_1 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint b_idx = 1024*i + 32*ir + 8*il;
+
+    const float d = float(data_a[ib].d);
+    const float m = float(data_a[ib].m);
+
+    const uint q_idx = 8*il;
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * (data_a[ib].qs[q_idx + l] & 0xF) + m);
+        data_b[b_idx + l + 16] = D_TYPE(d * (data_a[ib].qs[q_idx + l] >>  4) + m);
+    }
+}

vulkan-shaders/dequant_q4_k.comp

@@ -0,0 +1,56 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint tid = gl_LocalInvocationID.x;
+        const uint il = tid / 8;
+        const uint ir = tid % 8;
+        const uint is = 2 * il;
+        const uint n = 4;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+
+        const uint y_idx = i * QUANT_K + 64 * il + n * ir;
+        const uint qs_idx = 32*il + n * ir;
+
+        uint8_t sc;
+        uint8_t m;
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is] & 63);
+            m  = uint8_t(data_a[i].scales[is + 4] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
+        }
+        const FLOAT_TYPE d1 = dall * sc;
+        const FLOAT_TYPE m1 = dmin * m;
+
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is + 1] & 63);
+            m  = uint8_t(data_a[i].scales[is + 5] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
+        }
+        const FLOAT_TYPE d2 = dall * sc;
+        const FLOAT_TYPE m2 = dmin * m;
+
+        [[unroll]] for (uint l = 0; l < n; ++l) {
+            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] & 0xF) - m1);
+            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] >>  4) - m2);
+        }
+    }
+}

vulkan-shaders/dequant_q5_0.comp

@@ -0,0 +1,34 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q5_0 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint b_idx = 1024*i + 32*ir + 8*il;
+
+    const float d = float(data_a[ib].d);
+    const uint qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0];
+
+    const uint q_idx = 8*il;
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        const uint iqs = q_idx + l;
+        const uint vui = uint(data_a[ib].qs[iqs]);
+        data_b[b_idx + l +  0] = D_TYPE(d * (((vui & 0xF) | (((qh >> iqs) << 4) & 0x10)) - 16.0f));
+        data_b[b_idx + l + 16] = D_TYPE(d * (((vui >>  4) | ((qh >> (iqs + 12)) & 0x10)) - 16.0f));
+    }
+}

vulkan-shaders/dequant_q5_1.comp

@@ -0,0 +1,35 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q5_1 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint b_idx = 1024*i + 32*ir + 8*il;
+
+    const float d = float(data_a[ib].d);
+    const float m = float(data_a[ib].m);
+    const uint qh = data_a[ib].qh;
+
+    const uint q_idx = 8*il;
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        const uint iqs = q_idx + l;
+        const uint vui = uint(data_a[ib].qs[iqs]);
+        data_b[b_idx + l +  0] = D_TYPE(d * (((vui & 0xF) | (((qh >> iqs) << 4) & 0x10))) + m);
+        data_b[b_idx + l + 16] = D_TYPE(d * (((vui >>  4) | ((qh >> (iqs + 12)) & 0x10))) + m);
+    }
+}

vulkan-shaders/dequant_q5_k.comp

@@ -0,0 +1,58 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint tid = gl_LocalInvocationID.x;
+        const uint il = tid / 16;
+        const uint ir = tid % 16;
+        const uint is = 2 * il;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+
+        const uint y_idx = i * QUANT_K + 64 * il + 2 * ir;
+        const uint qs_idx = 32*il + 2 * ir;
+        const uint qh_idx = 2 * ir;
+
+        uint8_t sc;
+        uint8_t m;
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is] & 63);
+            m  = uint8_t(data_a[i].scales[is + 4] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
+        }
+        const FLOAT_TYPE d1 = dall * sc;
+        const FLOAT_TYPE m1 = dmin * m;
+
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is + 1] & 63);
+            m  = uint8_t(data_a[i].scales[is + 5] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
+        }
+        const FLOAT_TYPE d2 = dall * sc;
+        const FLOAT_TYPE m2 = dmin * m;
+
+        const uint8_t hm1 = uint8_t(1 << (2 * il    ));
+        const uint8_t hm2 = uint8_t(1 << (2 * il + 1));
+        data_b[y_idx     ] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx    ] & 0xF) + (((data_a[i].qh[qh_idx    ] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx +  1] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1] & 0xF) + (((data_a[i].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx    ]  >> 4) + (((data_a[i].qh[qh_idx    ] & hm2) != 0) ? 16 : 0)) - m2);
+        data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1]  >> 4) + (((data_a[i].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2);
+    }
+}

vulkan-shaders/dequant_q6_k.comp

@@ -0,0 +1,33 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+        const uint tid = gl_LocalInvocationID.x;
+        const uint ip = tid / 32;
+        const uint il = tid - 32 * ip;
+        const uint is = 8 * ip + il / 16;
+
+        const uint y_idx = i * QUANT_K + 128 * ip + il;
+
+        const uint ql_idx = 64 * ip + il;
+        const uint8_t qh = data_a[i].qh[32 * ip + il];
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[i].d);
+
+        data_b[y_idx +  0] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 0] * (int8_t((data_a[i].ql[ql_idx +  0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32)));
+        data_b[y_idx + 32] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 2] * (int8_t((data_a[i].ql[ql_idx + 32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32)));
+        data_b[y_idx + 64] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 4] * (int8_t((data_a[i].ql[ql_idx +  0] >>  4) | (((qh >> 4) & 3) << 4)) - 32)));
+        data_b[y_idx + 96] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 6] * (int8_t((data_a[i].ql[ql_idx + 32] >>  4) | (((qh >> 6) & 3) << 4)) - 32)));
+    }
+}

vulkan-shaders/dequant_q8_0.comp

@@ -0,0 +1,31 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q8_0 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint b_idx = 1024*i + 32*ir + 16*il;
+
+    const float d = float(data_a[ib].d);
+
+    const uint q_idx = 16*il;
+
+    [[unroll]] for (uint l = 0; l < 16; l += 2) {
+        data_b[b_idx + l    ] = D_TYPE(d * data_a[ib].qs[q_idx + l    ]);
+        data_b[b_idx + l + 1] = D_TYPE(d * data_a[ib].qs[q_idx + l + 1]);
+    }
+}

vulkan-shaders/diag_mask_inf.comp

@@ -0,0 +1,34 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_control_flow_attributes : enable
+
+layout (push_constant) uniform parameter
+{
+    uint ncols;
+    uint rows_per_channel;
+    uint n_past;
+} p;
+
+#include "types.comp"
+
+layout(local_size_x = 512, 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 col = gl_GlobalInvocationID.y;
+    const uint row = gl_GlobalInvocationID.x;
+
+    if (col >= p.ncols) {
+        return;
+    }
+
+    const uint i = row*p.ncols + col;
+    if (col > p.n_past + row % p.rows_per_channel) {
+        data_d[i] = D_TYPE(uintBitsToFloat(0xFF800000));
+    } else {
+        data_d[i] = D_TYPE(data_a[i]);
+    }
+}

vulkan-shaders/div.comp

@@ -0,0 +1,12 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    if (gl_GlobalInvocationID.x >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) / FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+}

vulkan-shaders/gelu.comp

@@ -0,0 +1,25 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, 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 float GELU_COEF_A    = 0.044715f;
+    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+    const uint i = gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float xi = float(data_a[i]);
+    const float val = SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi);
+    data_d[i] = D_TYPE(0.5f*xi*(2.0f - 2.0f / (exp(2 * val) + 1)));
+}

vulkan-shaders/generic_binary_head.comp

@@ -0,0 +1,48 @@
+#extension GL_EXT_shader_16bit_storage : require
+
+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 ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
+    uint d_offset;
+    float param1; float param2;
+} p;
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+uint src0_idx(uint idx) {
+    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
+    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
+}
+
+uint src1_idx(uint idx) {
+    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
+    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+
+    return (i03 % p.ne13)*p.nb13 + (i02 % p.ne12)*p.nb12 + (i01 % p.ne11)*p.nb11 + (i00 % p.ne10)*p.nb10;
+}
+
+uint dst_idx(uint idx) {
+    const uint i23 = idx / (p.ne22*p.ne21*p.ne20);
+    const uint i23_offset = i23 * p.ne22*p.ne21*p.ne20;
+    const uint i22 = (idx - i23_offset) / (p.ne21*p.ne20);
+    const uint i22_offset = i22*p.ne21*p.ne20;
+    const uint i21 = (idx - i23_offset - i22_offset) / p.ne20;
+    const uint i20 = idx - i23_offset - i22_offset - i21*p.ne20;
+    return i23*p.nb23 + i22*p.nb22 + i21*p.nb21 + i20*p.nb20;
+}

vulkan-shaders/generic_head.comp

@@ -0,0 +1,9 @@
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint KX;
+    uint KY;
+    float param1;
+    float param2;
+} p;

vulkan-shaders/generic_unary_head.comp

@@ -0,0 +1,35 @@
+#extension GL_EXT_shader_16bit_storage : require
+
+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;
+    float param1; float param2;
+} p;
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+uint src0_idx(uint idx) {
+    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
+    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
+}
+
+uint dst_idx(uint idx) {
+    const uint i13 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
+    const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10);
+    const uint i12_offset = i12*p.ne11*p.ne10;
+    const uint i11 = (idx - i13_offset - i12_offset) / p.ne10;
+    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
+    return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
+}

vulkan-shaders/get_rows.comp

@@ -0,0 +1,26 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint i00 = gl_GlobalInvocationID.x;
+    const uint i10 = gl_GlobalInvocationID.y;
+    const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
+    const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
+
+    if (i00 >= p.ne00) {
+        return;
+    }
+
+    const uint i01 = data_b[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;
+
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);
+#else
+    data_d[d_offset + i00] = data_a[a_offset + i00];
+#endif
+}

vulkan-shaders/get_rows_quant.comp

@@ -0,0 +1,31 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+#include "dequant_funcs.comp"
+
+void main() {
+    const uint i00 = (gl_GlobalInvocationID.x)*2;
+    const uint i10 = gl_GlobalInvocationID.y;
+    const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
+    const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
+
+    if (i00 >= p.ne00) {
+        return;
+    }
+
+    const uint i01 = data_b[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 ib = a_offset + i00/QUANT_K; // block index
+    const uint iqs = (i00%QUANT_K)/QUANT_R; // quant index
+    const uint iybs = i00 - i00%QUANT_K; // dst block start index
+    const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
+
+    vec2 v = dequantize(ib, iqs, 0);
+
+    data_d[d_offset + iybs + iqs           ] = D_TYPE(v.x);
+    data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y);
+}

vulkan-shaders/mul.comp

@@ -0,0 +1,12 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    if (gl_GlobalInvocationID.x >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+}

vulkan-shaders/mul_mat_split_k_reduce.comp

@@ -0,0 +1,29 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {float data_a[];};
+layout (binding = 1) writeonly buffer D {float data_d[];};
+
+layout (push_constant) uniform parameter {
+    uint ne;
+    uint k_num;
+} p;
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    float result = 0.0f;
+
+    [[unroll]] for (uint i = 0; i < p.k_num; i++) {
+        result += data_a[i * p.ne + idx];
+    }
+
+    data_d[idx] = result;
+}

vulkan-shaders/mul_mat_vec.comp

@@ -0,0 +1,50 @@
+#version 450
+
+#ifdef FLOAT16
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#endif
+
+#include "mul_mat_vec_base.comp"
+
+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;
+    const uint tid = gl_LocalInvocationID.x;
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
+
+    tmp[tid] = FLOAT_TYPE(0.0f);
+
+    [[unroll]] for (uint i = 0; i < p.ncols/BLOCK_SIZE; i += 2) {
+        const uint col = i*BLOCK_SIZE + 2*tid;
+        const uint ib = (row*p.ncols + col)/QUANT_K; // block index
+        const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
+        const uint iybs = col - col%QUANT_K; // y block start index
+
+        vec2 v = dequantize(ib, iqs, a_offset / QUANT_K);
+
+        // matrix multiplication
+        tmp[tid] += FLOAT_TYPE(v.x) * FLOAT_TYPE(data_b[b_offset + iybs + iqs]) +
+                    FLOAT_TYPE(v.y) * FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}

vulkan-shaders/mul_mat_vec_base.comp

@@ -0,0 +1,81 @@
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_shader_8bit_storage : require
+
+#define K_QUANTS_PER_ITERATION 2
+
+#ifdef MUL_MAT_ID
+#define EXPERT_COUNT 8
+#endif
+
+#include "types.comp"
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+#ifdef MUL_MAT_ID
+layout (binding = 3) readonly buffer IDS {int data_ids[];};
+#endif
+
+#include "dequant_funcs.comp"
+
+layout (push_constant) uniform parameter
+{
+    uint ncols;
+    uint stride_a;
+    uint stride_b;
+    uint stride_d;
+
+    uint batch_stride_a;
+    uint batch_stride_b;
+    uint batch_stride_d;
+
+#ifdef MUL_MAT_ID
+    uint nei0;
+    uint ne11;
+#else
+    uint ne02;
+    uint ne12;
+    uint broadcast2;
+    uint broadcast3;
+#endif
+} p;
+
+void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.y;
+#else
+    const uint batch_idx = gl_GlobalInvocationID.y;
+#endif
+
+#ifndef MUL_MAT_ID
+    const uint i13 = batch_idx / p.ne12;
+    const uint i12 = batch_idx % p.ne12;
+
+    const uint i03 = i13 / p.broadcast3;
+    const uint i02 = i12 / p.broadcast2;
+
+    const uint batch_idx_a = i03 * p.ne02 + i02;
+#else
+    const uint expert_id = data_ids[expert_idx];
+#endif
+
+    a_offset =
+#ifdef MUL_MAT_ID
+            expert_id * p.batch_stride_a;
+#else
+            batch_idx_a * p.batch_stride_a;
+#endif
+    b_offset =
+#ifdef MUL_MAT_ID
+            (expert_idx % p.ne11) * p.stride_b;
+#else
+            batch_idx * p.batch_stride_b;
+#endif
+    d_offset =
+#ifdef MUL_MAT_ID
+            expert_idx * p.stride_d;
+#else
+            batch_idx * p.batch_stride_d;
+#endif
+}

vulkan-shaders/mul_mat_vec_nc.comp

@@ -0,0 +1,71 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#define BLOCK_SIZE 32
+#define FLOAT_TYPE float
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
+
+layout (push_constant) uniform parameter
+{
+    uint ncols_x;
+    uint nrows_x;
+    uint row_stride_x;
+    uint channel_stride_x;
+    uint channel_x_divisor;
+    uint b_offset;
+    uint d_offset;
+} p;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint tid       = gl_LocalInvocationID.x;
+    const uint row_x     = gl_GlobalInvocationID.y;
+    const uint channel   = gl_GlobalInvocationID.z;
+    const uint channel_x = channel / p.channel_x_divisor;
+
+    const uint nrows_y   = p.ncols_x;
+    const uint nrows_dst = p.nrows_x;
+    const uint row_dst   = row_x;
+
+    const uint idst = channel*nrows_dst + row_dst;
+
+    tmp[tid] = 0.0f;
+
+    for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
+        const uint col_x = col_x0 + tid;
+
+        if (col_x >= p.ncols_x) {
+            break;
+        }
+
+        const uint row_y = col_x;
+
+        const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+        const uint iy = channel*nrows_y + row_y;
+
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
+
+        tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    if (tid == 0) {
+        dst[idst] = tmp[0];
+    }
+}

vulkan-shaders/mul_mat_vec_p021.comp

@@ -0,0 +1,73 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#define BLOCK_SIZE 32
+#define FLOAT_TYPE float
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
+
+layout (push_constant) uniform parameter
+{
+    uint ncols_x;
+    uint nrows_x;
+    uint nchannels_x;
+    uint nchannels_y;
+    uint b_offset;
+    uint d_offset;
+} p;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint row_x = gl_GlobalInvocationID.y;
+    const uint channel = gl_GlobalInvocationID.z;
+    const uint channel_x = channel / (p.nchannels_y / p.nchannels_x);
+
+    const uint nrows_y = p.ncols_x;
+    const uint nrows_dst = p.nrows_x;
+    const uint row_dst = row_x;
+
+    tmp[tid] = FLOAT_TYPE(0.0f);
+
+    for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
+        const uint col_x = col_x0 + tid;
+
+        if (col_x >= p.ncols_x) {
+            break;
+        }
+
+        // x is transposed and permuted
+        const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x;
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
+
+        const uint row_y = col_x;
+
+        // y is not transposed but permuted
+        const uint iy = channel*nrows_y + row_y;
+
+        tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
+    }
+
+    // dst is not transposed and not permuted
+    const uint idst = channel*nrows_dst + row_dst;
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    if (tid == 0) {
+        dst[idst] = tmp[0];
+    }
+}

vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -0,0 +1,73 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x;
+
+    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;
+
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
+
+    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
+
+    const uint l0 = K_QUANTS_PER_ITERATION*v_in;            // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint s_offset = 8*v_im;
+    const uint y_offset = 128*v_im + l0;
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const uint y_idx = i * QUANT_K + y_offset;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
+
+        FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
+        FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum1 += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3);
+            sum2 += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF)
+                  + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF);
+        }
+        tmp[16 * ix + tid] += dall * sum1 - dmin * sum2;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}

vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -0,0 +1,66 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x;
+
+    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;
+
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
+
+    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
+
+    const uint8_t m = uint8_t(1 << (4 * v_im));
+
+    const uint l0 = K_QUANTS_PER_ITERATION*v_in;            // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 128*v_im + l0;
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    const uint s_shift = 4 * v_im;
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const uint y_idx = i * QUANT_K + y_offset;
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (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))
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (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))
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (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))
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (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))
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (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))
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (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))
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (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))
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (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));
+        }
+        tmp[16 * ix + tid] += d * sum;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}

vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -0,0 +1,115 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x;
+
+    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;
+
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const uint step = 8/K_QUANTS_PER_ITERATION;             // 8 or 4
+
+    const uint il = tid/step;                               // 0...3
+    const uint ir = tid - step*il;                          // 0...7 or 0...3
+    const uint n =  2 * K_QUANTS_PER_ITERATION;             // 2 or 4
+
+    const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const uint v_in = il % 2;
+
+    const uint l0 = n * (2 * ir + v_in);            // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 64*v_im + l0;
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const uint y1_idx = i * QUANT_K + y_offset;
+        const uint y2_idx = y1_idx + 128;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
+
+        const uint8_t sc0 = uint8_t(  data_a[ib0 + i].scales[v_im * 2    ]       & 0x3f);
+        const uint8_t sc1 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 1]       & 0x3f);
+        const uint8_t sc2 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 4]       & 0x3f);
+        const uint8_t sc3 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 5]       & 0x3f);
+        const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2    ] & 0xc0) >> 2));
+        const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
+        const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
+        const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
+
+#if K_QUANTS_PER_ITERATION == 2
+        const uint8_t q4_0  = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
+        const uint8_t q4_1  = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
+        const uint8_t q4_2  = uint8_t(data_a[ib0 + i].qs[q_offset +  2] & 0xf);
+        const uint8_t q4_3  = uint8_t(data_a[ib0 + i].qs[q_offset +  3] & 0xf);
+        const uint8_t q4_4  = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
+        const uint8_t q4_5  = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
+        const uint8_t q4_6  = uint8_t(data_a[ib0 + i].qs[q_offset +  2]  >> 4);
+        const uint8_t q4_7  = uint8_t(data_a[ib0 + i].qs[q_offset +  3]  >> 4);
+        const uint8_t q4_8  = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
+        const uint8_t q4_9  = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
+        const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] & 0xf);
+        const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] & 0xf);
+        const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
+        const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
+        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 66]  >> 4);
+        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67]  >> 4);
+
+        const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1 + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * q4_3);
+        const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_5 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7);
+        const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx]) * q4_8 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_9 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * q4_10 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * q4_11);
+        const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_12 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_13 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * q4_14 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15);
+        const FLOAT_TYPE smin = FLOAT_TYPE(
+            FLOAT_TYPE(data_b[b_offset + y1_idx    ]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx    ]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
+          + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
+          + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * sc7
+          + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7
+        );
+        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin);
+#else
+        const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
+        const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
+        const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
+        const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
+        const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
+        const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
+        const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
+        const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
+
+        const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx     ]) * q4_0  + FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * q4_1);
+        const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_2  + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
+        const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx     ]) * q4_4  + FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * q4_5);
+        const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
+        const FLOAT_TYPE smin = FLOAT_TYPE(
+            FLOAT_TYPE(data_b[b_offset + y1_idx]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
+          + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
+        );
+
+        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) + sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
+#endif
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}

vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -0,0 +1,111 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x;
+
+    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;
+
+    const uint tid = gl_LocalInvocationID.x/2;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%2;  // 0 or 0, 1
+
+    const uint il = tid/4;                           // 0...3
+    const uint ir = tid - 4*il;                      // 0...7 or 0...3
+
+    const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const uint v_in = il % 2;
+
+    const uint l0 = 4*ir + 2*v_in;                   // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 64*v_im + l0;
+
+    const uint8_t hm1 = uint8_t(1 << (2*v_im));
+    const uint8_t hm2 = uint8_t(hm1 << 4);
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) {
+        const uint y1_idx = i * QUANT_K + y_offset;
+        const uint y2_idx = y1_idx + 128;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
+
+        const uint8_t sc0 = uint8_t(  data_a[ib0 + i].scales[v_im * 2    ]       & 0x3f);
+        const uint8_t sc1 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 1]       & 0x3f);
+        const uint8_t sc2 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 4]       & 0x3f);
+        const uint8_t sc3 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 5]       & 0x3f);
+        const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2    ] & 0xc0) >> 2));
+        const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
+        const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
+        const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
+
+        const uint8_t q4_0  = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
+        const uint8_t q4_1  = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
+        const uint8_t q4_2  = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf);
+        const uint8_t q4_3  = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf);
+        const uint8_t q4_4  = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
+        const uint8_t q4_5  = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
+        const uint8_t q4_6  = uint8_t(data_a[ib0 + i].qs[q_offset + 16]  >> 4);
+        const uint8_t q4_7  = uint8_t(data_a[ib0 + i].qs[q_offset + 17]  >> 4);
+        const uint8_t q4_8  = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
+        const uint8_t q4_9  = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
+        const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf);
+        const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf);
+        const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
+        const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
+        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80]  >> 4);
+        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81]  >> 4);
+
+        const FLOAT_TYPE sx = FLOAT_TYPE(
+            FLOAT_TYPE(data_b[b_offset + y1_idx     ]) * (q4_0 + (((data_a[ib0 + i].qh[l0     ] & hm1) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * (q4_1 + (((data_a[ib0 + i].qh[l0 +  1] & hm1) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) * (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0))
+        );
+        const FLOAT_TYPE sy = FLOAT_TYPE(
+            FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * (q4_4 + (((data_a[ib0 + i].qh[l0     ] & (hm1 << 1)) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * (q4_5 + (((data_a[ib0 + i].qh[l0 +  1] & (hm1 << 1)) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) * (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0))
+        );
+        const FLOAT_TYPE sz = FLOAT_TYPE(
+            FLOAT_TYPE(data_b[b_offset + y2_idx     ]) * (q4_8  + (((data_a[ib0 + i].qh[l0     ] & hm2) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * (q4_9  + (((data_a[ib0 + i].qh[l0 +  1] & hm2) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) * (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0))
+        );
+        const FLOAT_TYPE sw = FLOAT_TYPE(
+            FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * (q4_12 + (((data_a[ib0 + i].qh[l0     ] & (hm2 << 1)) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * (q4_13 + (((data_a[ib0 + i].qh[l0 +  1] & (hm2 << 1)) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) * (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0))
+          + FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0))
+        );
+        const FLOAT_TYPE smin = FLOAT_TYPE(
+            (FLOAT_TYPE(data_b[b_offset + y1_idx]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17])) * sc2 + (FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49])) * sc3
+          + (FLOAT_TYPE(data_b[b_offset + y2_idx]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17])) * sc6 + (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7
+        );
+        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin);
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}

vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -0,0 +1,79 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x;
+
+    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;
+
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
+
+    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
+
+#if K_QUANTS_PER_ITERATION == 1
+    const uint l0 = v_in;                                   // 0...15
+    const uint is = 0;
+#else
+    const uint l0 = 4 * v_in;                               // 0, 4, 8, ..., 28
+    const uint is = v_in / 4;
+#endif
+
+    const uint ql_offset = 64*v_im + l0;
+    const uint qh_offset = 32*v_im + l0;
+    const uint s_offset  =  8*v_im + is;
+    const uint y_offset = 128*v_im + l0;
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const uint y_idx   = i * QUANT_K + y_offset;
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+#if K_QUANTS_PER_ITERATION == 1
+        FLOAT_TYPE sum = FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32)
+                       + FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32)
+                       + FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32)
+                       + FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32)
+                       + FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32)
+                       + FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32)
+                       + FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32)
+                       + FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32);
+        tmp[16 * ix + tid] += sum;
+#else
+        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+        [[unroll]] for (int l = 0; l < 4; ++l) {
+            sum += FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32)
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32)
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32)
+                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32);
+        }
+        tmp[16 * ix + tid] += sum;
+#endif
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+       }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}

vulkan-shaders/mul_mm.comp

@@ -0,0 +1,494 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#ifdef FLOAT16
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#endif
+
+#ifdef MUL_MAT_ID
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+#endif
+
+#include "types.comp"
+
+#ifndef LOAD_VEC_A
+#define LOAD_VEC_A 1
+#endif
+#ifndef LOAD_VEC_B
+#define LOAD_VEC_B 1
+#endif
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+#ifdef MUL_MAT_ID
+layout (binding = 3) readonly buffer IDS {int data_ids[];};
+#endif
+
+layout (push_constant) uniform parameter
+{
+    uint M;
+    uint N;
+    uint K;
+    uint stride_a;
+    uint stride_b;
+    uint stride_d;
+
+    uint batch_stride_a;
+    uint batch_stride_b;
+    uint batch_stride_d;
+
+#ifdef MUL_MAT_ID
+    uint nei0;
+    uint nei1;
+    uint nbi1;
+    uint ne11;
+#else
+    uint k_split;
+    uint ne02;
+    uint ne12;
+    uint broadcast2;
+    uint broadcast3;
+#endif
+} p;
+
+layout (constant_id = 1) const uint BM = 64;
+layout (constant_id = 2) const uint BN = 64;
+layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
+layout (constant_id = 4) const uint WM = 32;
+layout (constant_id = 5) const uint WN = 32;
+layout (constant_id = 6) const uint WMITER = 2;
+layout (constant_id = 7) const uint TM = 4;
+layout (constant_id = 8) const uint TN = 2;
+layout (constant_id = 9) const uint WARP = 32;
+
+shared FLOAT_TYPE buf_a[BM * (BK+1)];
+shared FLOAT_TYPE buf_b[BN * (BK+1)];
+
+#ifdef MUL_MAT_ID
+shared u16vec2 row_ids[2048];
+#endif
+
+void main() {
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.z;
+#else
+    const uint batch_idx = gl_GlobalInvocationID.z;
+
+    const uint i13 = batch_idx / p.ne12;
+    const uint i12 = batch_idx % p.ne12;
+
+    const uint i03 = i13 / p.broadcast3;
+    const uint i02 = i12 / p.broadcast2;
+
+    const uint batch_idx_a = i03 * p.ne02 + i02;
+#endif
+
+    const uint blocks_m = (p.M + BM - 1) / BM;
+    const uint ir = gl_WorkGroupID.x % blocks_m;
+    const uint ik = gl_WorkGroupID.x / blocks_m;
+    const uint ic = gl_WorkGroupID.y;
+
+    const uint warp_i = gl_LocalInvocationID.x / WARP;
+    const uint warp_r = warp_i % (BM / WM);
+    const uint warp_c = warp_i / (BM / WM);
+
+    const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
+    const uint WSUBM = WM / WMITER;
+    const uint WSUBN = WN / WNITER;
+
+    const uint tiw = gl_LocalInvocationID.x % WARP;
+    const uint tiwr = tiw % (WSUBM / TM);
+    const uint tiwc = tiw / (WSUBM / TM);
+
+    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
+    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
+    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B);
+    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B);
+
+    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A / BK;
+    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
+
+#ifdef MUL_MAT_ID
+    uint _ne1 = 0;
+    for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
+        for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
+            if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
+                row_ids[_ne1] = u16vec2(ii0, ii1);
+                _ne1++;
+            }
+        }
+    }
+
+    barrier();
+
+    // Workgroup has no work
+    if (ic * BN >= _ne1) return;
+#endif
+
+#ifdef MUL_MAT_ID
+    const uint start_k = 0;
+    const uint end_k = p.K;
+#else
+    const uint start_k = ik * p.k_split;
+    const uint end_k = min(p.K, (ik + 1) * p.k_split);
+#endif
+
+    uint pos_a = (
+#ifdef MUL_MAT_ID
+        expert_idx * p.batch_stride_a +
+#else
+        batch_idx_a * p.batch_stride_a +
+#endif
+        ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
+#ifdef MUL_MAT_ID
+    uint pos_b = 0;
+#else
+    uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B;
+#endif
+
+    float sums[WMITER * TM * WNITER * TN];
+    FLOAT_TYPE cache_a[WMITER * TM];
+    FLOAT_TYPE cache_b[WNITER * TN];
+
+    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    [[unroll]] for (uint block = start_k; block < end_k; block += BK) {
+        [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) {
+
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+#if LOAD_VEC_A == 8
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            buf_a[buf_idx    ] = FLOAT_TYPE(data_a[idx][0].x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx][0].y);
+            buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx][0].z);
+            buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx][0].w);
+            buf_a[buf_idx + 4] = FLOAT_TYPE(data_a[idx][1].x);
+            buf_a[buf_idx + 5] = FLOAT_TYPE(data_a[idx][1].y);
+            buf_a[buf_idx + 6] = FLOAT_TYPE(data_a[idx][1].z);
+            buf_a[buf_idx + 7] = FLOAT_TYPE(data_a[idx][1].w);
+#elif LOAD_VEC_A == 4
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            buf_a[buf_idx    ] = FLOAT_TYPE(data_a[idx].x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx].y);
+            buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx].z);
+            buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx].w);
+#else
+            if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) {
+                buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]);
+            } else {
+                buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(0.0f);
+            }
+#endif
+#elif defined(DATA_A_Q4_0)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q4_1)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const float m = float(data_a[ib].m);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = vec2(vui & 0xF, vui >> 4) * d + m;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q5_0)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const uint uint_qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0];
+            const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q5_1)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const float m = float(data_a[ib].m);
+            const uint uint_qh = data_a[ib].qh;
+            const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q8_0)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 16;
+            const uint iqs = (idx & 0xF) * 2;
+
+            const float d = float(data_a[ib].d);
+            const vec2 v = vec2(int(data_a[ib].qs[iqs]), int(data_a[ib].qs[iqs + 1])) * d;
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q2_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                         // 2 values per idx
+            const uint iqs = idx % 128;                        // 0..127
+
+            const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
+            const uint scalesi = iqs / 8;                      // 0..15
+            const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6
+
+            const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
+            const uint scales = data_a[ib].scales[scalesi];
+            const vec2 d = vec2(data_a[ib].d);
+
+            const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q3_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                   // 2 values per idx
+            const uint iqs = idx % 128;                  // 0..127
+
+            const uint n = iqs / 64;                     // 0,1
+            const uint qsi = n * 32 + (iqs % 16) * 2;    // 0,2,4..62
+            const uint hmi =          (iqs % 16) * 2;    // 0,2,4..30
+            const uint j = (iqs % 64) / 4;               // 0..3
+            const uint is = iqs / 8;                     // 0..15
+            const uint halfsplit = ((iqs % 64) / 16);    // 0,1,2,3
+            const uint qsshift = halfsplit * 2;          // 0,2,4,6
+            const uint m = 1 << (4 * n + halfsplit);     // 1,2,4,8,16,32,64,128
+
+            const int8_t us = int8_t(is <  4 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+8] >> 0) & 3) << 4) :
+                                    is <  8 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+4] >> 2) & 3) << 4) :
+                                    is < 12 ? (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is+0] >> 4) & 3) << 4) :
+                                            (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is-4] >> 6) & 3) << 4));
+            const float dl = float(data_a[ib].d) * float(us - 32);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi    ] & m) != 0) ? 0 : 4)));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
+#elif defined(DATA_A_Q4_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                 // 2 values per idx
+            const uint iqs = idx % 128;                // 0..127
+
+            const uint n = iqs / 32;                   // 0,1,2,3
+            const uint b = (iqs % 32) / 16;            // 0,1
+            const uint is = 2 * n + b;                 // 0..7
+            const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+
+            const vec2 loadd = vec2(data_a[ib].d);
+
+            uint8_t sc;
+            uint8_t mbyte;
+            if (is < 4) {
+                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
+                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
+            } else {
+                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
+                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
+            }
+            const float d = loadd.x * sc;
+            const float m = loadd.y * mbyte;
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) - m);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) - m);
+#elif defined(DATA_A_Q5_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                 // 2 values per idx
+            const uint iqs = idx % 128;                // 0..127
+
+            const uint n = iqs / 32;                   // 0,1,2,3
+            const uint b = (iqs % 32) / 16;            // 0,1
+            const uint is = 2 * n + b;                 // 0..7
+            const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+            const uint qhi = (iqs % 16) * 2;           // 0,2,4..30
+
+            const uint8_t hm = uint8_t(1 << (iqs / 16));
+
+            const vec2 loadd = vec2(data_a[ib].d);
+
+            uint8_t sc;
+            uint8_t mbyte;
+            if (is < 4) {
+                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
+                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
+            } else {
+                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
+                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
+            }
+            const float d = loadd.x * sc;
+            const float m = loadd.y * mbyte;
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0)) - m);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0)) - m);
+#elif defined(DATA_A_Q6_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                  // 2 values per idx
+            const uint iqs = idx % 128;                 // 0..127
+
+            const uint n = iqs / 64;                    // 0,1
+            const uint b = (iqs % 64) / 32;             // 0,1
+            const uint is_b = (iqs % 16) / 8;           // 0,1
+            const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
+            const uint is = 8 * n + qhshift + is_b;     // 0..15
+            const uint qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
+            const uint qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62
+
+            const float dscale = float(data_a[ib].d) * float(data_a[ib].scales[is]);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+#endif
+        }
+        [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
+#if LOAD_VEC_B == 8
+#ifdef MUL_MAT_ID
+            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
+            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
+#else
+            const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
+#endif
+            const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
+            buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx][0].x);
+            buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx][0].y);
+            buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx][0].z);
+            buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx][0].w);
+            buf_b[buf_idx + 4] = FLOAT_TYPE(data_b[idx][1].x);
+            buf_b[buf_idx + 5] = FLOAT_TYPE(data_b[idx][1].y);
+            buf_b[buf_idx + 6] = FLOAT_TYPE(data_b[idx][1].z);
+            buf_b[buf_idx + 7] = FLOAT_TYPE(data_b[idx][1].w);
+#elif LOAD_VEC_B == 4
+#ifdef MUL_MAT_ID
+            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
+            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
+#else
+            const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
+#endif
+            const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
+            buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x);
+            buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y);
+            buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z);
+            buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w);
+#elif !MUL_MAT_ID
+            if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) {
+                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
+            } else {
+                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f);
+            }
+#else
+            const uint row_i = ic * BN + loadc_b + l;
+            if (row_i < _ne1) {
+                const u16vec2 row_idx = row_ids[row_i];
+                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
+            } else {
+                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f);
+            }
+#endif
+        }
+
+        barrier();
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (uint i = 0; i < BK; i++) {
+            // Load from shared into cache
+            [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+                [[unroll]] for (uint j = 0; j < TM; j++) {
+                    cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * (BK+1) + i];
+                }
+            }
+            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+                [[unroll]] for (uint j = 0; j < TN; j++) {
+                    cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * (BK+1) + i];
+                }
+            }
+
+            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+                [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+                    [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+                        [[unroll]] for (uint cr = 0; cr < TM; cr++) {
+                            sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr] += float(cache_a[wsir * TM + cr]) * float(cache_b[wsic * TN + cc]);
+                        }
+                    }
+                }
+            }
+        }
+
+        barrier();
+    }
+
+    const uint dr = ir * BM + warp_r * WM;
+    const uint dc = ic * BN + warp_c * WN;
+
+#ifndef MUL_MAT_ID
+    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+#endif
+
+    [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+        [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+
+            const uint dr_warp = dr + wsir * WSUBM + tiwr * TM;
+            const uint dc_warp = dc + wsic * WSUBN + tiwc * TN;
+            [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+#ifdef MUL_MAT_ID
+                const uint row_i = dc_warp + cc;
+                if (row_i >= _ne1) break;
+
+                const u16vec2 row_idx = row_ids[row_i];
+#endif
+                [[unroll]] for (uint cr = 0; cr < TM; cr++) {
+#ifdef MUL_MAT_ID
+                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+#else
+                    if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
+                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    }
+#endif
+                }
+            }
+        }
+    }
+}

vulkan-shaders/norm.comp

@@ -0,0 +1,44 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, 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[];};
+
+shared vec2 sum[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    sum[tid] = vec2(0.0f, 0.0f);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        const float xi = float(data_a[row*p.KX + col]);
+        sum[tid].x += xi;
+        sum[tid].y += xi * xi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sum[tid] += sum[tid + s];
+        }
+        barrier();
+    }
+
+    const float mean = sum[0].x / p.KX;
+    const float var = sum[0].y / p.KX - mean * mean;
+    const float inv_std = inversesqrt(var + p.param1);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        data_d[row*p.KX + col] = D_TYPE((float(data_a[row*p.KX + col]) - mean) * inv_std);
+    }
+}

vulkan-shaders/relu.comp

@@ -0,0 +1,21 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, 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.KX) {
+        return;
+    }
+
+    data_d[i] = max(float(data_a[i]), 0);
+}