Add Q4_3 support to cuBLAS (#1086)

Broke it during conflict resolution in last PR

.github/workflows/build.yml

@@ -81,7 +81,6 @@ jobs:
       matrix:
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
         build_type: [Debug, Release]
-        accelerate: [ON, OFF]
 
     steps:
       - name: Clone
@@ -99,7 +98,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DLLAMA_ACCELERATE=${{ matrix.accelerate }}
+          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
           cmake --build . --config ${{ matrix.build_type }}
 
       - name: Test

.gitignore

@@ -1,11 +1,15 @@
 *.o
 *.a
+.DS_Store
+.build/
 .cache/
+.direnv/
+.envrc
+.swiftpm
+.venv
 .vs/
 .vscode/
-.DS_Store
 
-.build/
 build/
 build-em/
 build-debug/
@@ -30,12 +34,9 @@ models/*
 arm_neon.h
 compile_commands.json
 
-.envrc
-.direnv/
-
-.venv
 __pycache__
-.swiftpm
 
 zig-out/
 zig-cache/
+
+ppl-*.txt

CMakeLists.txt

@@ -110,6 +110,7 @@ if (APPLE AND LLAMA_ACCELERATE)
         message(WARNING "Accelerate framework not found")
     endif()
 endif()
+
 if (LLAMA_OPENBLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
@@ -150,6 +151,10 @@ if (LLAMA_CUBLAS)
     if (CUDAToolkit_FOUND)
         message(STATUS "cuBLAS found")
 
+        enable_language(CUDA)
+
+        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+
         add_compile_definitions(GGML_USE_CUBLAS)
 
         if (LLAMA_STATIC)
@@ -174,7 +179,6 @@ if (LLAMA_ALL_WARNINGS)
             -Wshadow
             -Wstrict-prototypes
             -Wpointer-arith
-            -Wno-unused-function
         )
         set(cxx_flags
             -Wall
@@ -242,21 +246,26 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
     message(STATUS "x86 detected")
     if (MSVC)
         if (LLAMA_AVX512)
-            add_compile_options(/arch:AVX512)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX512>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)
             # MSVC has no compile-time flags enabling specific
             # AVX512 extensions, neither it defines the
             # macros corresponding to the extensions.
             # Do it manually.
             if (LLAMA_AVX512_VBMI)
-                add_compile_definitions(__AVX512VBMI__)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
             endif()
             if (LLAMA_AVX512_VNNI)
-                add_compile_definitions(__AVX512VNNI__)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
             endif()
         elseif (LLAMA_AVX2)
-            add_compile_options(/arch:AVX2)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX2>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX2>)
         elseif (LLAMA_AVX)
-            add_compile_options(/arch:AVX)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX>)
         endif()
     else()
         if (LLAMA_F16C)
@@ -293,7 +302,8 @@ endif()
 
 add_library(ggml OBJECT
             ggml.c
-            ggml.h)
+            ggml.h
+            ${GGML_CUDA_SOURCES})
 
 target_include_directories(ggml PUBLIC .)
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
@@ -315,6 +325,14 @@ if (BUILD_SHARED_LIBS)
     target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
 endif()
 
+if (GGML_CUDA_SOURCES)
+    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
+    set_property(TARGET ggml PROPERTY CUDA_ARCHITECTURES OFF)
+    set_property(TARGET ggml PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
+    set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES OFF)
+endif()
+
+
 #
 # programs, examples and tests
 #

Makefile

@@ -1,3 +1,6 @@
+# Define the default target now so that it is always the first target
+default: main quantize quantize-stats perplexity embedding vdot
+
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
@@ -36,7 +39,7 @@ CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
 # warnings
-CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wno-unused-function
+CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith
 CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
 
 # OS specific
@@ -99,7 +102,10 @@ ifdef LLAMA_OPENBLAS
 endif
 ifdef LLAMA_CUBLAS
 	CFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include
-	LDFLAGS += -lcublas_static -lculibos -lcudart_static -lcublasLt_static -lpthread -ldl -L/usr/local/cuda/lib64
+	LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64
+	OBJS	+= ggml-cuda.o
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	nvcc -arch=native -c -o $@ $<
 endif
 ifdef LLAMA_GPROF
 	CFLAGS   += -pg
@@ -137,8 +143,6 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main quantize quantize-stats perplexity embedding vdot
-
 #
 # Build library
 #
@@ -155,35 +159,35 @@ common.o: examples/common.cpp examples/common.h
 clean:
 	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-q4_0-matmult
 
-main: examples/main/main.cpp ggml.o llama.o common.o
+main: examples/main/main.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-quantize: examples/quantize/quantize.cpp ggml.o llama.o
+quantize: examples/quantize/quantize.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o
+quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
+perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
+embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-vdot: pocs/vdot/vdot.cpp ggml.o
+vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-libllama.so: llama.o ggml.o
+libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 #
 # Tests
 #
 
-benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o
+benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o benchmark-q4_0-matmult $(LDFLAGS)
 	./benchmark-q4_0-matmult
 

README.md

@@ -7,6 +7,10 @@
 
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
+**Warnings**
+
+- `Q4_2` and `Q4_3` are still in development. Do not expect any kind of backward compatibility until they are finalized
+
 **Hot topics:**
 
 - [Added LoRA support](https://github.com/ggerganov/llama.cpp/pull/820)
@@ -15,7 +19,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ## Description
 
-The main goal is to run the model using 4-bit quantization on a MacBook
+The main goal of llama.cpp is to run the llama model using 4-bit quantization on a MacBook.
 
 - Plain C/C++ implementation without dependencies
 - Apple silicon first-class citizen - optimized via ARM NEON and Accelerate framework
@@ -152,7 +156,7 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
 
 ## Usage
 
-Here are the step for the LLaMA-7B model.
+Here are the steps for the LLaMA-7B model.
 
 ### Get the Code
 
@@ -210,8 +214,7 @@ When running the larger models, make sure you have enough disk space to store al
 
 ### Memory/Disk Requirements
 
-As the models are currently fully loaded into memory, you will need adequate disk space to save them
-and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
+As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
 
 | model | original size | quantized size (4-bit) |
 |-------|---------------|------------------------|
@@ -223,18 +226,18 @@ and sufficient RAM to load them. At the moment, memory and disk requirements are
 ### Interactive mode
 
 If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
-In this mode, you can always interrupt generation by pressing Ctrl+C and enter one or more lines of text which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt which makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
+In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
 
-Here is an example few-shot interaction, invoked with the command
+Here is an example of a few-shot interaction, invoked with the command
 
 ```bash
-# default arguments using 7B model
+# default arguments using a 7B model
 ./examples/chat.sh
 
-# advanced chat with 13B model
+# advanced chat with a 13B model
 ./examples/chat-13B.sh
 
-# custom arguments using 13B model
+# custom arguments using a 13B model
 ./main -m ./models/13B/ggml-model-q4_0.bin -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
 ```
 
@@ -273,7 +276,7 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
 ### Using [GPT4All](https://github.com/nomic-ai/gpt4all)
 
 - Obtain the `gpt4all-lora-quantized.bin` model
-- It is distributed in the old `ggml` format which is now obsoleted
+- It is distributed in the old `ggml` format, which is now obsoleted
 - You have to convert it to the new format using [./convert-gpt4all-to-ggml.py](./convert-gpt4all-to-ggml.py). You may also need to
 convert the model from the old format to the new format with [./migrate-ggml-2023-03-30-pr613.py](./migrate-ggml-2023-03-30-pr613.py):
 
@@ -287,7 +290,7 @@ convert the model from the old format to the new format with [./migrate-ggml-202
 
 ### Obtaining and verifying the Facebook LLaMA original model and Stanford Alpaca model data
 
-- **Under no circumstances share IPFS, magnet links, or any other links to model downloads anywhere in this respository, including in issues, discussions or pull requests. They will be immediately deleted.**
+- **Under no circumstances should IPFS, magnet links, or any other links to model downloads be shared anywhere in this repository, including in issues, discussions, or pull requests. They will be immediately deleted.**
 - The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository.
 - Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data.
 - Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.
@@ -299,29 +302,27 @@ convert the model from the old format to the new format with [./migrate-ggml-202
 
   `shasum -a 256 --ignore-missing -c SHA256SUMS` on macOS
 
-- If your issue is with model generation quality then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
-  - LLaMA:
-    - [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/)
-    - [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)
-  - GPT-3
-    - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165)
-  - GPT-3.5 / InstructGPT / ChatGPT:
-    - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
-    - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
+- If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
+- LLaMA:
+- [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/)
+- [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)
+- GPT-3
+- [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165)
+- GPT-3.5 / InstructGPT / ChatGPT:
+- [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
+- [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
 
-### Perplexity (Measuring model quality)
+### Perplexity (measuring model quality)
 
-You can use the `perplexity` example to measure perplexity over the given prompt.  For more background,
-see https://huggingface.co/docs/transformers/perplexity.  However, in general, lower perplexity is better for LLMs.
+You can use the `perplexity` example to measure perplexity over the given prompt. For more background, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity). However, in general, lower perplexity is better for LLMs.
 
 #### Latest measurements
 
-The latest perplexity scores for the various model sizes and quantizations are being tracked in [discussion #406](https://github.com/ggerganov/llama.cpp/discussions/406).  `llama.cpp` is measuring very well
-compared to the baseline implementations.  Quantization has a small negative impact to quality, but, as you can see, running
+The latest perplexity scores for the various model sizes and quantizations are being tracked in [discussion #406](https://github.com/ggerganov/llama.cpp/discussions/406). `llama.cpp` is measuring very well compared to the baseline implementations. Quantization has a small negative impact on quality, but, as you can see, running
 13B at q4_0 beats the 7B f16 model by a significant amount.
 
-All measurements are done against wikitext2 test dataset (https://paperswithcode.com/dataset/wikitext-2), with default options (512 length context).
-Note that the changing the context length will have a significant impact on perplexity (longer context = better perplexity).
+All measurements are done against the wikitext2 test dataset (https://paperswithcode.com/dataset/wikitext-2), with default options (512 length context).
+Note that changing the context length will have a significant impact on perplexity (longer context = better perplexity).
 ```
 Perplexity - model options
 5.5985 - 13B, q4_0
@@ -363,7 +364,7 @@ https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b0
 
 #### Prerequisites
 * Docker must be installed and running on your system.
-* Create a folder to store big models & intermediate files (in ex. im using /llama/models)
+* Create a folder to store big models & intermediate files (ex. /llama/models)
 
 #### Images
 We have two Docker images available for this project:
@@ -377,17 +378,17 @@ The easiest way to download the models, convert them to ggml and optimize them i
 
 Replace `/path/to/models` below with the actual path where you downloaded the models.
 
- ```bash
+```bash
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
 ```
 
-On complete, you are ready to play!
+On completion, you are ready to play!
 
 ```bash
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
-or with light image:
+or with a light image:
 
 ```bash
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512
@@ -408,7 +409,7 @@ docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /mode
 - Always consider cross-compatibility with other operating systems and architectures
 - Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
 - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
-- Clean-up any trailing whitespaces, use 4 spaces indentation, brackets on same line, `void * ptr`, `int & a`
+- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 
 ### Docs

examples/quantize-stats/quantize-stats.cpp

@@ -15,6 +15,8 @@
 #include <string>
 #include <unordered_map>
 #include <vector>
+#include <thread>
+#include <mutex>
 
 struct quantize_stats_params {
     std::string model = "models/7B/ggml-model-f16.bin";
@@ -27,7 +29,6 @@ struct quantize_stats_params {
     std::vector<enum ggml_type> include_types;
 };
 
-const int64_t SCRATCH_ELEMENTS = 32*32;
 const size_t HISTOGRAM_BUCKETS = 150;
 const double HISTOGRAM_RANGE = 0.03;
 
@@ -90,6 +91,13 @@ void update_error_stats(int64_t nelements, const float * input, const float * ou
     stats.num_samples += nelements;
 }
 
+void combine_error_stats(error_stats & into, const error_stats & from) {
+    into.num_samples += from.num_samples;
+    into.total_error += from.total_error;
+    if (from.max_error > into.max_error) into.max_error = from.max_error;
+    for (size_t i=0; i<HISTOGRAM_BUCKETS; ++i) into.error_histogram[i] += from.error_histogram[i];
+}
+
 double find_quantile(const error_stats & stats, double quantile) {
     double sum = std::accumulate(std::begin(stats.error_histogram), std::end(stats.error_histogram), 0.0);
 
@@ -130,6 +138,36 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
 
+void test_roundtrip_on_chunk(
+        const ggml_tensor * layer,
+        int64_t offset,
+        int64_t chunk_size,
+        const quantize_fns_t & qfns,
+        bool use_reference,
+        float * input_scratch,
+        char * quantized_scratch,
+        float * output_scratch,
+        error_stats & stats) {
+
+    if (layer->type == GGML_TYPE_F16) {
+        for (int i = 0; i < chunk_size; i++) {
+            input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
+        }
+    } else {
+        input_scratch = ggml_get_data_f32(layer) + offset;
+    }
+
+    if (use_reference) {
+        qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size);
+    } else {
+        qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size);
+    }
+    qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size);
+
+    update_error_stats(chunk_size, input_scratch, output_scratch, stats);
+}
+
+
 // Run quantization function for a single layer and update error stats
 void test_roundtrip_on_layer(
         std::string & name,
@@ -137,40 +175,61 @@ void test_roundtrip_on_layer(
         const quantize_fns_t & qfns,
         bool use_reference,
         const ggml_tensor * layer,
-        float * input_scratch,
-        char *quantized_scratch,
-        float * output_scratch,
-        error_stats & total_error) {
+        std::vector<float> & input_scratch,
+        std::vector<char> & quantized_scratch,
+        std::vector<float> & output_scratch,
+        error_stats & total_error,
+        int max_thread = 0) {
 
     assert(tensor_is_contiguous(layer));
     error_stats layer_error {};
-    int64_t nelements = ggml_nelements(layer);
+    uint64_t nelements = ggml_nelements(layer);
 
-    for (int64_t offset = 0; offset < nelements; offset += SCRATCH_ELEMENTS) {
-        int64_t chunk_size = std::min(SCRATCH_ELEMENTS, nelements - offset);
-
-        if (layer->type == GGML_TYPE_F16) {
-            for (int i = 0; i < chunk_size; i++) {
-                input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
-            }
-        } else {
-            input_scratch = ggml_get_data_f32(layer) + offset;
-        }
-
-        if (use_reference) {
-            qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size);
-        } else {
-            qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size);
-        }
-        qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size);
-
-        update_error_stats(chunk_size, input_scratch, output_scratch, total_error);
-        if (print_layer_stats) {
-            update_error_stats(chunk_size, input_scratch, output_scratch, layer_error);
-        }
+    float* input_scratch_ptr = nullptr;
+    if (layer->type == GGML_TYPE_F16) {
+        if (input_scratch.size() < nelements) input_scratch.resize(nelements);
+        input_scratch_ptr = input_scratch.data();
     }
+    if (quantized_scratch.size() < 4*nelements) quantized_scratch.resize(4*nelements);
+    if (output_scratch.size() < nelements) output_scratch.resize(nelements);
+
+    if (max_thread < 1) max_thread = std::thread::hardware_concurrency();
+    int chunk_size = 32*512;
+    int num_chunks = (nelements + chunk_size - 1)/chunk_size;
+
+    if (num_chunks < 2 || max_thread < 2) {
+        test_roundtrip_on_chunk(layer, 0, nelements, qfns, use_reference, input_scratch_ptr, quantized_scratch.data(),
+                output_scratch.data(), print_layer_stats ? layer_error : total_error);
+    } else {
+        auto & stats = print_layer_stats ? layer_error : total_error;
+        std::mutex mutex;
+        uint64_t counter = 0;
+        auto compute = [&mutex, &counter, &stats, &qfns, nelements, layer, use_reference, input_scratch_ptr,
+             &quantized_scratch, &output_scratch, chunk_size] () {
+            error_stats local_stats {};
+            while (true) {
+                std::unique_lock<std::mutex> lock(mutex);
+                uint64_t offset = counter; counter += chunk_size;
+                if (offset >= nelements) {
+                    combine_error_stats(stats, local_stats);
+                    break;
+                }
+                lock.unlock();
+                uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset;
+                test_roundtrip_on_chunk(layer, offset, chunk, qfns, use_reference, input_scratch_ptr + offset,
+                        quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats);
+            }
+        };
+        int nthread = std::min(num_chunks, max_thread);
+        std::vector<std::thread> workers(nthread-1);
+        for (auto& w : workers) w = std::thread(compute);
+        compute();
+        for (auto& w : workers) w.join();
+    }
+
     if (print_layer_stats) {
         print_error_stats(name, layer_error, false);
+        combine_error_stats(total_error, layer_error);
     }
 }
 
@@ -181,6 +240,7 @@ int main(int argc, char ** argv) {
 
     // read command line
 
+    int max_thread = 0;
     bool invalid_param = false;
     std::string arg;
     for (int i = 1; i < argc; i++) {
@@ -230,6 +290,12 @@ int main(int argc, char ** argv) {
                 fprintf(stderr, "error: %s not in list of types\n", argv[i]);
                 invalid_param = true;
             }
+        } else if (arg == "-n" || arg == "--num-threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            max_thread = atoi(argv[i]);
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             quantize_stats_print_usage(argc, argv);
@@ -295,9 +361,9 @@ int main(int argc, char ** argv) {
     }
     printf("testing %d layers with max size %" PRId64 "\n", included_layers, max_nelements);
     // allocate scratch space
-    std::vector<float> input_scratch(SCRATCH_ELEMENTS);
-    std::vector<char> quantized_scratch(SCRATCH_ELEMENTS*4);
-    std::vector<float> output_scratch(SCRATCH_ELEMENTS);
+    std::vector<float> input_scratch;
+    std::vector<char> quantized_scratch;
+    std::vector<float> output_scratch;
 
     // loop throught quantization types
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
@@ -328,10 +394,11 @@ int main(int argc, char ** argv) {
                         qfns,
                         params.reference,
                         kv_tensor.second,
-                        input_scratch.data(),
-                        quantized_scratch.data(),
-                        output_scratch.data(),
-                        global_stats
+                        input_scratch,
+                        quantized_scratch,
+                        output_scratch,
+                        global_stats,
+                        max_thread
                 );
             }
 

examples/quantize/quantize.cpp

@@ -10,11 +10,12 @@
 int main(int argc, char ** argv) {
     ggml_time_init();
 
-    if (argc != 4) {
-        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+    if (argc < 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type [nthread]\n", argv[0]);
         fprintf(stderr, "  type = %d - q4_0\n", LLAMA_FTYPE_MOSTLY_Q4_0);
         fprintf(stderr, "  type = %d - q4_1\n", LLAMA_FTYPE_MOSTLY_Q4_1);
         fprintf(stderr, "  type = %d - q4_2\n", LLAMA_FTYPE_MOSTLY_Q4_2);
+        fprintf(stderr, "  type = %d - q4_3\n", LLAMA_FTYPE_MOSTLY_Q4_3);
         return 1;
     }
 
@@ -29,6 +30,7 @@ int main(int argc, char ** argv) {
     const std::string fname_out = argv[2];
 
     const enum llama_ftype ftype = (enum llama_ftype)atoi(argv[3]);
+    int nthread = argc > 4 ? atoi(argv[4]) : 0;
 
     const int64_t t_main_start_us = ggml_time_us();
 
@@ -38,7 +40,7 @@ int main(int argc, char ** argv) {
     {
         const int64_t t_start_us = ggml_time_us();
 
-        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype)) {
+        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype, nthread)) {
             fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
             return 1;
         }

ggml-cuda.cu

@@ -0,0 +1,154 @@
+#include <stdint.h>
+#include <cuda_fp16.h>
+#include "ggml-cuda.h"
+
+typedef uint16_t ggml_fp16_t;
+static_assert(sizeof(__half) == sizeof(ggml_fp16_t), "wrong fp16 size");
+
+#define QK4_0 32
+typedef struct {
+    float   d;              // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    float   d;              // delta
+    float   m;              // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK4_2 16
+typedef struct {
+    __half  d;              // delta
+    uint8_t qs[QK4_2 / 2];  // nibbles / quants
+} block_q4_2;
+static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding");
+
+#define QK4_3 16
+typedef struct {
+    __half  d;         // delta
+    __half  m;         // min
+    uint8_t qs[QK4_3 / 2]; // nibbles / quants
+} block_q4_3;
+static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding");
+
+
+
+static __global__ void dequantize_block_q4_0(const void * vx, float * y) {
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_0; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = (vi0 - 8)*d;
+        const float v1 = (vi1 - 8)*d;
+
+        y[i*QK4_0 + l + 0] = v0;
+        y[i*QK4_0 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q4_1(const void * vx, float * y) {
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+    const float m = x[i].m;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_1; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = vi0*d + m;
+        const float v1 = vi1*d + m;
+
+        y[i*QK4_1 + l + 0] = v0;
+        y[i*QK4_1 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q4_2(const void * vx, float * y) {
+    const block_q4_2 * x = (const block_q4_2 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_2; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = (vi0 - 8)*d;
+        const float v1 = (vi1 - 8)*d;
+
+        y[i*QK4_2 + l + 0] = v0;
+        y[i*QK4_2 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q4_3(const void * vx, float * y) {
+    const block_q4_3 * x = (const block_q4_3 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+    const float m = x[i].m;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_3; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = vi0*d + m;
+        const float v1 = vi1*d + m;
+
+        y[i*QK4_3 + l + 0] = v0;
+        y[i*QK4_3 + l + 1] = v1;
+    }
+}
+
+extern "C" {
+    __host__ void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_0;
+        dequantize_block_q4_0<<<nb, 1, 0, stream>>>(vx, y);
+    }
+
+    __host__ void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_1;
+        dequantize_block_q4_1<<<nb, 1, 0, stream>>>(vx, y);
+    }
+
+    __host__ void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_2;
+        dequantize_block_q4_2<<<nb, 1, 0, stream>>>(vx, y);
+    }
+
+    __host__ void dequantize_row_q4_3_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_3;
+        dequantize_block_q4_3<<<nb, 1, 0, stream>>>(vx, y);
+    }
+}

ggml-cuda.h

@@ -0,0 +1,12 @@
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+void dequantize_row_q4_3_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+
+#ifdef  __cplusplus
+}
+#endif

ggml.h

@@ -205,7 +205,8 @@ enum ggml_type {
     GGML_TYPE_Q4_0 = 2,
     GGML_TYPE_Q4_1 = 3,
     GGML_TYPE_Q4_2 = 4,
-    GGML_TYPE_Q8_0 = 5,
+    GGML_TYPE_Q4_3 = 5,
+    GGML_TYPE_Q8_0 = 6,
     GGML_TYPE_I8,
     GGML_TYPE_I16,
     GGML_TYPE_I32,
@@ -360,6 +361,8 @@ const char * ggml_type_name(enum ggml_type type);
 
 size_t ggml_element_size(const struct ggml_tensor * tensor);
 
+bool ggml_is_quantized(enum ggml_type type);
+
 struct ggml_context * ggml_init(struct ggml_init_params params);
 void ggml_free(struct ggml_context * ctx);
 
@@ -808,6 +811,9 @@ enum ggml_opt_result ggml_opt(
 size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
 size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
 size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_3(const float * src, void * dst, int n, int k, int64_t * hist);
+
+size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
 
 //
 // system info

llama.cpp

@@ -24,6 +24,9 @@
 #include <memory>
 #include <algorithm>
 #include <initializer_list>
+#include <thread>
+#include <atomic>
+#include <mutex>
 
 #define LLAMA_USE_SCRATCH
 #define LLAMA_MAX_SCRATCH_BUFFERS 16
@@ -479,6 +482,7 @@ struct llama_file_loader {
                 case GGML_TYPE_Q4_0:
                 case GGML_TYPE_Q4_1:
                 case GGML_TYPE_Q4_2:
+                case GGML_TYPE_Q4_3:
                     break;
                 default: {
                     throw format("unrecognized tensor type %u\n", shard.type);
@@ -552,6 +556,7 @@ struct llama_file_saver {
             case GGML_TYPE_Q4_0:
             case GGML_TYPE_Q4_1:
             case GGML_TYPE_Q4_2:
+            case GGML_TYPE_Q4_3:
                 break;
             default: LLAMA_ASSERT(false);
         }
@@ -841,6 +846,7 @@ static const char *llama_ftype_name(enum llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16:
                                       return "mostly Q4_1, some F16";
         case LLAMA_FTYPE_MOSTLY_Q4_2: return "mostly Q4_2";
+        case LLAMA_FTYPE_MOSTLY_Q4_3: return "mostly Q4_3";
         default:                      return "unknown, may not work";
     }
 }
@@ -1569,15 +1575,20 @@ static llama_vocab::id llama_sample_top_p_top_k(
 // quantization
 //
 
-static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, enum llama_ftype ftype) {
+static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, enum llama_ftype ftype, int nthread) {
     ggml_type quantized_type;
     switch (ftype) {
         case LLAMA_FTYPE_MOSTLY_Q4_0: quantized_type = GGML_TYPE_Q4_0; break;
         case LLAMA_FTYPE_MOSTLY_Q4_1: quantized_type = GGML_TYPE_Q4_1; break;
         case LLAMA_FTYPE_MOSTLY_Q4_2: quantized_type = GGML_TYPE_Q4_2; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_3: quantized_type = GGML_TYPE_Q4_3; break;
         default: throw format("invalid output file type %d\n", ftype);
     };
 
+    if (nthread <= 0) {
+        nthread = std::thread::hardware_concurrency();
+    }
+
     std::unique_ptr<llama_model_loader> model_loader(new llama_model_loader(fname_inp.c_str(), /*use_mmap*/ false,
                                                                             /*vocab_only*/ false));
     llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), ftype);
@@ -1586,6 +1597,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     size_t total_size_new = 0;
     std::vector<int64_t> hist_all(1 << 4, 0);
 
+    std::vector<std::thread> workers;
+    std::mutex mutex;
+
     size_t idx = 0;
     for (llama_load_tensor & tensor : model_loader->tensors_map.tensors) {
         llama_buffer read_data;
@@ -1639,21 +1653,37 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             new_data = work.addr;
             std::vector<int64_t> hist_cur(1 << 4, 0);
 
-            switch (new_type) {
-                case GGML_TYPE_Q4_0:
-                    {
-                        new_size = ggml_quantize_q4_0(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
-                    } break;
-                case GGML_TYPE_Q4_1:
-                    {
-                        new_size = ggml_quantize_q4_1(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
-                    } break;
-                case GGML_TYPE_Q4_2:
-                    {
-                        new_size = ggml_quantize_q4_2(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
-                    } break;
-                default:
-                    LLAMA_ASSERT(false);
+            int chunk_size = 32 * 512;
+            const int nchunk = (nelements + chunk_size - 1)/chunk_size;
+            const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1;
+            if (nthread_use < 2) {
+                new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nelements, hist_cur.data());
+            } else {
+                size_t counter = 0;
+                new_size = 0;
+                auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, nelements, chunk_size] () {
+                    std::vector<int64_t> local_hist;
+                    size_t local_size = 0;
+                    while (true) {
+                        std::unique_lock<std::mutex> lock(mutex);
+                        size_t first = counter; counter += chunk_size;
+                        if (first >= nelements) {
+                            if (!local_hist.empty()) {
+                                for (int j=0; j<int(local_hist.size()); ++j) hist_cur[j] += local_hist[j];
+                                new_size += local_size;
+                            }
+                            break;
+                        }
+                        lock.unlock();
+                        size_t last = std::min(nelements, first + chunk_size);
+                        if (local_hist.empty()) local_hist.resize(hist_cur.size(), 0);
+                        local_size += ggml_quantize_chunk(new_type, f32_data, new_data, first, last - first, local_hist.data());
+                    }
+                };
+                if (int(workers.size()) < nthread_use - 1) workers.resize(nthread_use - 1);
+                for (int it = 0; it < nthread_use - 1; ++it) workers[it] = std::thread(compute);
+                compute();
+                for (int it = 0; it < nthread_use - 1; ++it) workers[it].join();
             }
 
             printf("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0);
@@ -1775,9 +1805,10 @@ void llama_free(struct llama_context * ctx) {
 int llama_model_quantize(
         const char * fname_inp,
         const char * fname_out,
-  enum llama_ftype   ftype) {
+  enum llama_ftype   ftype,
+        int          nthread) {
     try {
-        llama_model_quantize_internal(fname_inp, fname_out, ftype);
+        llama_model_quantize_internal(fname_inp, fname_out, ftype, nthread);
         return 0;
     } catch (const std::string & err) {
         fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.c_str());
@@ -1963,7 +1994,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
                 base_t = dest_t;
             }
 
-            if (base_t->type == GGML_TYPE_Q4_0 || base_t->type == GGML_TYPE_Q4_1 || base_t->type == GGML_TYPE_Q4_2) {
+            if (ggml_is_quantized(base_t->type)) {
                 if (!warned) {
                     fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, "
                                     "use a f16 or f32 base model with --lora-base\n", __func__);

llama.h

@@ -73,6 +73,7 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
         LLAMA_FTYPE_MOSTLY_Q4_2 = 5,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_3 = 6,  // except 1d tensors
     };
 
     LLAMA_API struct llama_context_params llama_context_default_params();
@@ -92,10 +93,12 @@ extern "C" {
 
     // TODO: not great API - very likely to change
     // Returns 0 on success
+    // nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
     LLAMA_API int llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
-      enum llama_ftype   ftype);
+      enum llama_ftype   ftype,
+            int          nthread);
 
     // Apply a LoRA adapter to a loaded model
     // path_base_model is the path to a higher quality model to use as a base for