llama : fix hparams shadow (#1367)

fixes #1363

.github/workflows/build.yml

@@ -120,7 +120,7 @@ jobs:
           make
 
   macOS-latest-cmake:
-    runs-on: macOS-latest
+    runs-on: macos-latest
 
     steps:
       - name: Clone
@@ -148,22 +148,64 @@ jobs:
 
   windows-latest-cmake:
     runs-on: windows-latest
+    env:
+      OPENBLAS_VERSION: 0.3.23
+      OPENCL_VERSION: 2023.04.17
+      CLBLAST_VERSION: 1.5.3
 
     strategy:
       matrix:
         include:
-         - build: 'avx2'
-           defines: ''
-         - build: 'avx'
-           defines: '-DLLAMA_AVX2=OFF'
-         - build: 'avx512'
-           defines: '-DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'avx2'
+            defines: ''
+          - build: 'avx'
+            defines: '-DLLAMA_AVX2=OFF'
+          - build: 'avx512'
+            defines: '-DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'clblast'
+            defines: '-DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
+          - build: 'openblas'
+            defines: '-DLLAMA_OPENBLAS=ON -DBLAS_LIBRARIES="/LIBPATH:$env:RUNNER_TEMP/openblas/lib" -DOPENBLAS_INC="$env:RUNNER_TEMP/openblas/include"'
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v1
 
+      - name: Download OpenCL SDK
+        id: get_opencl
+        if: ${{ matrix.build == 'clblast' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/opencl.zip -L "https://github.com/KhronosGroup/OpenCL-SDK/releases/download/v${env:OPENCL_VERSION}/OpenCL-SDK-v${env:OPENCL_VERSION}-Win-x64.zip"
+          mkdir $env:RUNNER_TEMP/opencl
+          tar.exe -xvf $env:RUNNER_TEMP/opencl.zip --strip-components=1 -C $env:RUNNER_TEMP/opencl
+
+      - name: Download CLBlast
+        id: get_clblast
+        if: ${{ matrix.build == 'clblast' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/clblast.zip -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-Windows-x64.zip"
+          curl.exe -o $env:RUNNER_TEMP/CLBlast.LICENSE.txt -L "https://github.com/CNugteren/CLBlast/raw/${env:CLBLAST_VERSION}/LICENSE"
+          mkdir $env:RUNNER_TEMP/clblast
+          tar.exe -xvf $env:RUNNER_TEMP/clblast.zip -C $env:RUNNER_TEMP/clblast
+          foreach ($f in (gci -Recurse -Path "$env:RUNNER_TEMP/clblast" -Filter '*.cmake')) {
+            $txt = Get-Content -Path $f -Raw
+            $txt.Replace('C:/dependencies/opencl/', "$($env:RUNNER_TEMP.Replace('\','/'))/opencl/") | Set-Content -Path $f -Encoding UTF8
+          }
+
+      - name: Download OpenBLAS
+        id: get_openblas
+        if: ${{ matrix.build == 'openblas' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
+          curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
+          mkdir $env:RUNNER_TEMP/openblas
+          tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas
+          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
+          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
+          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
+          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
+
       - name: Build
         id: cmake_build
         run: |
@@ -171,6 +213,21 @@ jobs:
           cd build
           cmake .. ${{ matrix.defines }}
           cmake --build . --config Release
+          cp ../LICENSE ./bin/Release/llama.cpp.txt
+
+      - name: Add clblast.dll
+        id: add_clblast_dll
+        if: ${{ matrix.build == 'clblast' }}
+        run: |
+          cp $env:RUNNER_TEMP/clblast/lib/clblast.dll ./build/bin/Release
+          cp $env:RUNNER_TEMP/CLBlast.LICENSE.txt ./build/bin/Release/CLBlast-${env:CLBLAST_VERSION}.txt
+
+      - name: Add libopenblas.dll
+        id: add_libopenblas_dll
+        if: ${{ matrix.build == 'openblas' }}
+        run: |
+          cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
+          cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
 
       - name: Check AVX512F support
         id: check_avx512f
@@ -187,7 +244,7 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'avx512' || env.HAS_AVX512F == '1' }} # Test AVX-512 only when possible
+        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # Test AVX-512 only when possible
         run: |
           cd build
           ctest -C Release --verbose
@@ -210,6 +267,82 @@ jobs:
           path: |
             llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
 
+  windows-latest-cmake-cublas:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        cuda: ['12.1.0', '11.7.1']
+        build: ['cublas']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - uses: Jimver/cuda-toolkit@v0.2.10
+        id: cuda-toolkit
+        with:
+          cuda: ${{ matrix.cuda }}
+          # TODO(green-sky): _dev seems to fail, and non dev are not enought
+          #sub-packages: '["nvcc", "cudart", "cublas", "cudart_dev", "cublas_dev"]'
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DLLAMA_CUBLAS=ON
+          cmake --build . --config Release
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
+
+      - name: Copy and pack Cuda runtime
+        if: ${{ matrix.cuda == '12.1.0' }}
+        # TODO(green-sky): paths are cuda 12 specific
+        run: |
+          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
+          mkdir '.\build\bin\cudart\'
+          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_12.dll" '.\build\bin\cudart\'
+          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_12.dll" '.\build\bin\cudart\'
+          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_12.dll" '.\build\bin\cudart\'
+          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\*
+
+      - name: Copy and pack Cuda runtime
+        if: ${{ matrix.cuda == '11.7.1' }}
+        # TODO(green-sky): paths are cuda 11 specific
+        run: |
+          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
+          mkdir '.\build\bin\cudart\'
+          ls "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin"
+          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_110.dll" '.\build\bin\cudart\'
+          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_11.dll" '.\build\bin\cudart\'
+          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_11.dll" '.\build\bin\cudart\'
+          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\*
+
+      - name: Upload Cuda runtime
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -221,6 +354,7 @@ jobs:
       - macOS-latest-make
       - macOS-latest-cmake
       - windows-latest-cmake
+      - windows-latest-cmake-cublas
 
     steps:
       - name: Download artifacts

.gitignore

@@ -21,6 +21,7 @@ build-sanitize-addr/
 build-sanitize-thread/
 
 models/*
+*.bin
 
 /main
 /quantize
@@ -42,5 +43,6 @@ zig-out/
 zig-cache/
 
 ppl-*.txt
+qnt-*.txt
 
 examples/jeopardy/results.txt

Makefile

@@ -107,7 +107,11 @@ ifndef LLAMA_NO_ACCELERATE
 endif
 ifdef LLAMA_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
-	LDFLAGS += -lopenblas
+	ifneq ($(shell grep -e "Arch Linux" -e "ID_LIKE=arch" /etc/os-release 2>/dev/null),)
+		LDFLAGS += -lopenblas -lcblas
+	else
+		LDFLAGS += -lopenblas
+	endif
 endif
 ifdef LLAMA_CUBLAS
 	CFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
@@ -121,7 +125,12 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 endif
 ifdef LLAMA_CLBLAST
 	CFLAGS  += -DGGML_USE_CLBLAST
-	LDFLAGS += -lclblast -lOpenCL
+	# Mac provides OpenCL as a framework
+	ifeq ($(UNAME_S),Darwin)
+		LDFLAGS += -lclblast -framework OpenCL
+	else
+		LDFLAGS += -lclblast -lOpenCL
+	endif
 	OBJS    += ggml-opencl.o
 ggml-opencl.o: ggml-opencl.c ggml-opencl.h
 	$(CC) $(CFLAGS) -c $< -o $@

README.md

@@ -18,10 +18,12 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant
 
 - Plain C/C++ implementation without dependencies
 - Apple silicon first-class citizen - optimized via ARM NEON and Accelerate framework
-- AVX2 support for x86 architectures
+- AVX, AVX2 and AVX512 support for x86 architectures
 - Mixed F16 / F32 precision
-- 4-bit integer quantization support
+- 4-bit, 5-bit and 8-bit integer quantization support
 - Runs on the CPU
+- OpenBLAS support
+- cuBLAS and CLBlast support
 
 The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022).
 Since then, the project has improved significantly thanks to many contributions. This project is for educational purposes and serves
@@ -43,6 +45,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
 - [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
+- [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
 
 **Bindings:**
 
@@ -213,7 +216,6 @@ Building the program with BLAS support may lead to some performance improvements
       ```bash
       make LLAMA_OPENBLAS=1
       ```
-      Note: In order to build on Arch Linux with OpenBLAS support enabled you must edit the Makefile adding at the end of the line 105: `-lcblas`
 
     - On Windows:
 
@@ -255,6 +257,8 @@ Building the program with BLAS support may lead to some performance improvements
     cmake --build . --config Release
     ```
 
+Note: Because llama.cpp uses multiple CUDA streams for matrix multiplication results [are not guaranteed to be reproducible](https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility). If you need reproducibility, set `GGML_CUDA_MAX_STREAMS` in the file `ggml-cuda.cu` to 1.
+
 ### Prepare Data & Run
 
 ```bash
@@ -294,17 +298,25 @@ Several quantization methods are supported. They differ in the resulting model d
 
 | Model | Measure      | F16    | Q4_0   | Q4_1   | Q4_2   | Q5_0   | Q5_1   | Q8_0   |
 |------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|-------:|
-|    7B | perplexity   | 5.9565 | 6.2103 | 6.1286 | 6.1698 | 6.0139 | 5.9934 | 5.9571 |
+|    7B | perplexity   | 5.9066 | 6.1620 | 6.0910 | 6.1466 | 5.9862 | 5.9481 | 5.9069 |
 |    7B | file size    |  13.0G |   4.0G |   4.8G |   4.0G |   4.4G |   4.8G |   7.1G |
 |    7B | ms/tok @ 4th |    128 |     56 |     61 |     84 |     91 |     95 |     75 |
 |    7B | ms/tok @ 8th |    128 |     47 |     55 |     48 |     53 |     59 |     75 |
 |    7B | bits/weight  |   16.0 |    5.0 |    6.0 |    5.0 |    5.5 |    6.0 |    9.0 |
-|   13B | perplexity   | 5.2455 | 5.3748 | 5.3471 | 5.3433 | 5.2768 | 5.2582 | 5.2458 |
+|   13B | perplexity   | 5.2543 | 5.3863 | 5.3607 | 5.3513 | 5.2856 | 5.2706 | 5.2548 |
 |   13B | file size    |  25.0G |   7.6G |   9.1G |   7.6G |   8.4G |   9.1G |    14G |
 |   13B | ms/tok @ 4th |    239 |    104 |    113 |    160 |    176 |    185 |    141 |
 |   13B | ms/tok @ 8th |    240 |     85 |     99 |     97 |    108 |    117 |    147 |
 |   13B | bits/weight  |   16.0 |    5.0 |    6.0 |    5.0 |    5.5 |    6.0 |    9.0 |
 
+### Perplexity (measuring model quality)
+
+You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
+For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
+
+The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
+The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads.
+
 ### Interactive mode
 
 If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
@@ -403,26 +415,6 @@ If your issue is with model generation quality, then please at least scan the fo
     - [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)
-
-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 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 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
-5.9565 - 7B, f16
-6.3001 - 7B, q4_1
-6.5949 - 7B, q4_0
-6.5995 - 7B, q4_0, --memory_f16
-```
-
 #### How to run
 
 1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research

convert.py

@@ -67,6 +67,7 @@ FTYPE_TO_DATA_TYPE: Dict[int, DataType] = \
     {ftype: dtype for (dtype, ftype) in DATA_TYPE_TO_FTYPE.items()}
 
 DATA_TYPE_TO_NUMPY: Dict[DataType, 'np.dtype[Any]'] = {
+    DT_BF16: np.dtype(np.uint16),
     DT_F16: np.dtype(np.float16),
     DT_F32: np.dtype(np.float32),
     DT_I32: np.dtype(np.int32),
@@ -276,6 +277,12 @@ class Tensor(metaclass=ABCMeta):
     def to_ggml(self) -> 'GGMLCompatibleTensor': ...
 
 
+def bf16_to_fp32(bf16_arr: np.ndarray) -> np.ndarray:
+    assert bf16_arr.dtype == np.uint16, f"Input array should be of dtype uint16, but got {bf16_arr.dtype}"
+    fp32_arr = bf16_arr.astype(np.uint32) << 16
+    return fp32_arr.view(np.float32)
+
+
 class UnquantizedTensor(Tensor):
     def __init__(self, ndarray: NDArray) -> None:
         assert isinstance(ndarray, np.ndarray)
@@ -284,6 +291,8 @@ class UnquantizedTensor(Tensor):
 
     def astype(self, data_type: DataType) -> Tensor:
         dtype = DATA_TYPE_TO_NUMPY[data_type]
+        if self.data_type == DT_BF16:
+            self.ndarray = bf16_to_fp32(self.ndarray)
         return UnquantizedTensor(self.ndarray.astype(dtype))
 
     def to_ggml(self) -> 'UnquantizedTensor':
@@ -686,6 +695,7 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
         return LazyStorage(load=load, kind=pid[1], description=description)
 
+   # @staticmethod
     def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,  # pyright: ignore[reportSelfClsParameterName]
                                requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
         assert isinstance(storage, LazyStorage)
@@ -696,12 +706,18 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'pickled storage_offset={storage_offset} in {storage.description}'
         return LazyTensor(load, list(size), storage.kind.data_type, description)
 
+    # @staticmethod
+    def rebuild_from_type_v2(func, new_type, args, state):
+        return func(*args)
+
     CLASSES: Dict[Any, Any] = {
+        ('torch._tensor', '_rebuild_from_type_v2'): rebuild_from_type_v2,
         ('torch._utils', '_rebuild_tensor_v2'): lazy_rebuild_tensor_v2,
         ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16),
         ('torch', 'HalfStorage'): LazyStorageKind(DT_F16),
         ('torch', 'FloatStorage'): LazyStorageKind(DT_F32),
         ('torch', 'IntStorage'): LazyStorageKind(DT_I32),
+        ('torch', 'Tensor'): LazyTensor,
     }
 
     def find_class(self, module: str, name: str) -> Any:
@@ -750,7 +766,7 @@ def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
             return UnquantizedTensor(np.frombuffer(buf, dtype=numpy_dtype).reshape(shape))
         description = f'safetensors begin={begin} end={end} type={data_type} path={path}'
         return LazyTensor(load, shape, data_type, description)
-    model = {name: convert(info) for (name, info) in header.items()}
+    model = {name: convert(info) for (name, info) in header.items() if name != '__metadata__'}
     return ModelPlus(model=model, paths=[path], format='safetensors', vocab=None)
 
 
@@ -961,7 +977,7 @@ class OutputFile:
 
 def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFileType:
     wq_type = model["layers.0.attention.wq.weight"].data_type
-    if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
+    if output_type_str == "f32" or (output_type_str is None and wq_type in (DT_F32, DT_BF16)):
         return GGMLFileType.AllF32
     if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16):
         return GGMLFileType.MostlyF16
@@ -1035,8 +1051,12 @@ def load_some_model(path: Path) -> ModelPlus:
     '''Load a model of any supported format.'''
     # Be extra-friendly and accept either a file or a directory:
     if path.is_dir():
-        globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt"]
-        files = [file for glob in globs for file in path.glob(glob)]
+        # Check if it's a set of safetensors files first
+        files = list(path.glob("model-00001-of-*.safetensors"))
+        if not files:
+            # Try the PyTorch patterns too, with lower priority
+            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt"]
+            files = [file for glob in globs for file in path.glob(glob)]
         if not files:
             # Try GGML too, but with lower priority, since if both a non-GGML
             # model and a GGML model exist in the same directory, we assume the

examples/common.cpp

@@ -100,6 +100,9 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         arg = argv[i];
 
         if (arg == "-s" || arg == "--seed") {
+#if defined(GGML_USE_CUBLAS)
+            fprintf(stderr, "WARNING: when using cuBLAS generation results are NOT guaranteed to be reproducible.\n");
+#endif
             if (++i >= argc) {
                 invalid_param = true;
                 break;
@@ -324,6 +327,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.input_prefix = argv[i];
+        } else if (arg == "--in-suffix") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.input_suffix = argv[i];
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             gpt_print_usage(argc, argv, default_params);
@@ -362,6 +371,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --session FNAME       file to cache model state in (may be large!) (default: none)\n");
     fprintf(stderr, "  --random-prompt       start with a randomized prompt.\n");
     fprintf(stderr, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
+    fprintf(stderr, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
     fprintf(stderr, "  -f FNAME, --file FNAME\n");
     fprintf(stderr, "                        prompt file to start generation.\n");
     fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict);
@@ -428,8 +438,8 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
 // TODO: not great allocating this every time
 std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
     // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
-    std::vector<llama_token> res(text.size() + (int)add_bos);
-    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
+    std::vector<llama_token> res(text.size() + (int) add_bos);
+    const int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
     assert(n >= 0);
     res.resize(n);
 

examples/common.h

@@ -43,6 +43,7 @@ struct gpt_params {
     std::string prompt = "";
     std::string path_session = "";       // path to file for saving/loading model eval state
     std::string input_prefix = "";       // string to prefix user inputs with
+    std::string input_suffix = "";       // string to suffix user inputs with
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
 
     std::string lora_adapter = "";  // lora adapter path

examples/main/README.md

@@ -112,6 +112,14 @@ The `--in-prefix` flag is used to add a prefix to your input, primarily, this is
 ./main -r "User:" --in-prefix " "
 ```
 
+### In-Suffix
+
+The `--in-suffix` flag is used to add a suffix after your input. This is useful for adding an "Assistant:" prompt after the user's input. It's added after the new-line character (`\n`) that's automatically added to the end of the user's input. Here's an example of how to use the `--in-suffix` flag in conjunction with the `--reverse-prompt` flag:
+
+```sh
+./main -r "User:" --in-prefix " " --in-suffix "Assistant:"
+```
+
 ### Instruction Mode
 
 Instruction mode is particularly useful when working with Alpaca models, which are designed to follow user instructions for specific tasks:

examples/main/main.cpp

@@ -260,6 +260,10 @@ int main(int argc, char ** argv) {
         if (!params.input_prefix.empty()) {
             fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str());
         }
+
+        if (!params.input_suffix.empty()) {
+            fprintf(stderr, "Input suffix: '%s'\n", params.input_suffix.c_str());
+        }
     }
     fprintf(stderr, "sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
             params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
@@ -309,7 +313,8 @@ int main(int argc, char ** argv) {
             if (n_past + (int) embd.size() > n_ctx) {
                 const int n_left = n_past - params.n_keep;
 
-                n_past = params.n_keep;
+                // always keep the first token - BOS
+                n_past = std::max(1, params.n_keep);
 
                 // insert n_left/2 tokens at the start of embd from last_n_tokens
                 embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size());
@@ -327,7 +332,6 @@ int main(int argc, char ** argv) {
             }
 
             // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
-            // REVIEW
             if (n_session_consumed < (int) session_tokens.size()) {
                 size_t i = 0;
                 for ( ; i < embd.size(); i++) {
@@ -440,10 +444,10 @@ int main(int argc, char ** argv) {
                         id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
                     } else {
                         // Temperature sampling
-                        llama_sample_top_k(ctx, &candidates_p, top_k);
-                        llama_sample_tail_free(ctx, &candidates_p, tfs_z);
-                        llama_sample_typical(ctx, &candidates_p, typical_p);
-                        llama_sample_top_p(ctx, &candidates_p, top_p);
+                        llama_sample_top_k(ctx, &candidates_p, top_k, 1);
+                        llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
+                        llama_sample_typical(ctx, &candidates_p, typical_p, 1);
+                        llama_sample_top_p(ctx, &candidates_p, top_p, 1);
                         llama_sample_temperature(ctx, &candidates_p, temp);
                         id = llama_sample_token(ctx, &candidates_p);
                     }
@@ -567,6 +571,11 @@ int main(int argc, char ** argv) {
                 // Add tokens to embd only if the input buffer is non-empty
                 // Entering a empty line lets the user pass control back
                 if (buffer.length() > 1) {
+                    // append input suffix if any
+                    if (!params.input_suffix.empty()) {
+                        buffer += params.input_suffix;
+                        printf("%s", params.input_suffix.c_str());
+                    }
 
                     // instruct mode: insert instruction prefix
                     if (params.instruct && !is_antiprompt) {

examples/perplexity/perplexity.cpp

@@ -25,46 +25,68 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
+    // BOS tokens will be added for each chunk before eval
     auto tokens = ::llama_tokenize(ctx, params.prompt, true);
 
-    int count = 0;
-    int seq_count = tokens.size() / params.n_ctx;
-    int n_vocab = llama_n_vocab(ctx);
+    int count   = 0;
+
+    const int n_chunk = tokens.size() / params.n_ctx;
+    const int n_vocab = llama_n_vocab(ctx);
+    const int n_batch = params.n_batch;
 
     double nll = 0.0;
-    fprintf(stderr, "%s : calculating perplexity over %d chunks, batch_size=%d\n", __func__, seq_count, params.n_batch);
+    fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
 
-    for (int i = 0; i < seq_count; ++i) {
-        int start = i * params.n_ctx;
-        int end = start + params.n_ctx;
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * params.n_ctx;
+        const int end   = start + params.n_ctx;
+
+        const int num_batches = (params.n_ctx + n_batch - 1) / n_batch;
 
         std::vector<float> logits;
-        int num_batches = (params.n_ctx + params.n_batch - 1) / params.n_batch;
-        auto start_t = std::chrono::high_resolution_clock::now();
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
         for (int j = 0; j < num_batches; ++j) {
-            int batch_start = start + j * params.n_batch;
-            int batch_size = std::min(end - batch_start, params.n_batch);
-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * params.n_batch, params.n_threads)) {
+            const int batch_start = start + j * n_batch;
+            const int batch_size  = std::min(end - batch_start, n_batch);
+
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
+
+            // add BOS token for the first batch of each chunk
+            if (j == 0) {
+                tokens[batch_start] = llama_token_bos();
+            }
+
+            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return;
             }
-            auto batch_logits = llama_get_logits(ctx);
+
+            // restore the original token in case it was set to BOS
+            tokens[batch_start] = token_org;
+
+            const auto batch_logits = llama_get_logits(ctx);
             logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
         }
-        auto end_t = std::chrono::high_resolution_clock::now();
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
         if (i == 0) {
-            const float seconds = std::chrono::duration<float>(end_t - start_t).count();
-            printf("%.2f seconds per pass - ETA ", seconds);
-            int total_seconds = (int)(seconds * seq_count);
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
             if (total_seconds >= 60*60) {
-                printf("%d hours ", total_seconds / (60*60));
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
             }
-            printf("%d minutes\n", total_seconds / 60);
+            fprintf(stderr, "%d minutes\n", total_seconds / 60);
         }
+
         // We get the logits for all the tokens in the context window (params.n_ctx)
         // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
-        // calculate the perplexity over the last half the window (so the model always has
+        // calculate the perplexity over the last half of the window (so the model always has
         // some context to predict the token).
         //
         // We rely on the fact that attention in the forward pass only looks at previous
@@ -76,10 +98,12 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         // process the entire prompt.
         for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) {
             // Calculate probability of next token, given the previous ones.
-            std::vector<float> tok_logits(
-                logits.begin() + j * n_vocab,
+            const std::vector<float> tok_logits(
+                logits.begin() + (j + 0) * n_vocab,
                 logits.begin() + (j + 1) * n_vocab);
-            float prob = softmax(tok_logits)[tokens[start + j + 1]];
+
+            const float prob = softmax(tok_logits)[tokens[start + j + 1]];
+
             nll += -std::log(prob);
             ++count;
         }

examples/quantize/quantize.cpp

@@ -6,23 +6,47 @@
 #include <map>
 #include <string>
 
-static const std::map<std::string, enum llama_ftype> LLAMA_FTYPE_MAP = {
-  {"q4_0", LLAMA_FTYPE_MOSTLY_Q4_0},
-  {"q4_1", LLAMA_FTYPE_MOSTLY_Q4_1},
-  {"q4_2", LLAMA_FTYPE_MOSTLY_Q4_2},
-  {"q5_0", LLAMA_FTYPE_MOSTLY_Q5_0},
-  {"q5_1", LLAMA_FTYPE_MOSTLY_Q5_1},
-  {"q8_0", LLAMA_FTYPE_MOSTLY_Q8_0},
+static const std::map<std::string, llama_ftype> LLAMA_FTYPE_MAP = {
+    {"q4_0", LLAMA_FTYPE_MOSTLY_Q4_0},
+    {"q4_1", LLAMA_FTYPE_MOSTLY_Q4_1},
+    {"q4_2", LLAMA_FTYPE_MOSTLY_Q4_2},
+    {"q5_0", LLAMA_FTYPE_MOSTLY_Q5_0},
+    {"q5_1", LLAMA_FTYPE_MOSTLY_Q5_1},
+    {"q8_0", LLAMA_FTYPE_MOSTLY_Q8_0},
 };
 
+bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::string & ftype_str_out) {
+    auto it = LLAMA_FTYPE_MAP.find(ftype_str);
+    if (it != LLAMA_FTYPE_MAP.end()) {
+        ftype = it->second;
+        ftype_str_out = it->first;
+        return true;
+    }
+    // try to parse as an integer
+    try {
+        int ftype_int = std::stoi(ftype_str);
+        for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
+            if (it->second == ftype_int) {
+                ftype = it->second;
+                ftype_str_out = it->first;
+                return true;
+            }
+        }
+    }
+    catch (...) {
+        // stoi failed
+    }
+    return false;
+}
+
 // usage:
-//  ./quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type
+//  ./quantize models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
 //
 int main(int argc, char ** argv) {
     ggml_time_init();
 
-    if (argc < 4) {
-        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type [nthread]\n", argv[0]);
+    if (argc < 3) {
+        fprintf(stderr, "usage: %s model-f32.bin [model-quant.bin] type [nthreads]\n", argv[0]);
         for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
             fprintf(stderr, "  type = \"%s\" or %d\n", it->first.c_str(), it->second);
         }
@@ -36,24 +60,62 @@ int main(int argc, char ** argv) {
         ggml_free(ctx);
     }
 
+    // parse command line arguments
     const std::string fname_inp = argv[1];
-    const std::string fname_out = argv[2];
+    std::string fname_out;
+    int nthread;
+    llama_ftype ftype;
 
-    enum llama_ftype ftype;
-    if (argv[3][0] == 'q') {
-        auto it = LLAMA_FTYPE_MAP.find(argv[3]);
-        if (it == LLAMA_FTYPE_MAP.end()) {
-            fprintf(stderr, "%s: unknown ftype '%s'\n", __func__, argv[3]);
+    int arg_idx = 2;
+    std::string ftype_str;
+    if (try_parse_ftype(argv[arg_idx], ftype, ftype_str)) {
+        // argv[2] is the ftype
+        std::string fpath;
+        const size_t pos = fname_inp.find_last_of('/');
+        if (pos != std::string::npos) {
+            fpath = fname_inp.substr(0, pos + 1);
+        }
+        // export as [inp path]/ggml-model-[ftype].bin
+        fname_out = fpath + "ggml-model-" + ftype_str + ".bin";
+        arg_idx++;
+    }
+    else {
+        // argv[2] is the output path
+        fname_out = argv[arg_idx];
+        arg_idx++;
+
+        if (argc <= arg_idx) {
+            fprintf(stderr, "%s: missing ftype\n", __func__);
+            return 1;
+        }
+        // argv[3] is the ftype
+        if (!try_parse_ftype(argv[arg_idx], ftype, ftype_str)) {
+            fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]);
+            return 1;
+        }
+        arg_idx++;
+    }
+
+    // parse nthreads
+    if (argc > arg_idx) {
+        try {
+            nthread = std::stoi(argv[arg_idx]);
+        }
+        catch (const std::exception & e) {
+            fprintf(stderr, "%s: invalid nthread '%s' (%s)\n", __func__, argv[arg_idx], e.what());
             return 1;
         }
-        ftype = it->second;
     } else {
-        ftype = (enum llama_ftype)atoi(argv[3]);
+        nthread = 0;
     }
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
-    int nthread = argc > 4 ? atoi(argv[4]) : 0;
+    fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
+    if (nthread > 0) {
+        fprintf(stderr, " using %d threads", nthread);
+    }
+    fprintf(stderr, "\n");
 
     const int64_t t_main_start_us = ggml_time_us();
 

ggml-cuda.cu

@@ -348,7 +348,7 @@ static void ggml_cuda_pool_free(void * ptr, size_t size) {
     CUDA_CHECK(cudaFree(ptr));
 }
 
-#define GGML_CUDA_MAX_STREAMS 8
+#define GGML_CUDA_MAX_STREAMS 8 // Set this to 1 for reproducible matrix multiplication.
 #define GGML_CUDA_MAX_EVENTS 64
 static cublasHandle_t g_cublasH = nullptr;
 static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_STREAMS] = { nullptr };

ggml.c

@@ -137,6 +137,9 @@ inline static void* ggml_aligned_malloc(size_t size) {
 
 #if defined(GGML_USE_ACCELERATE)
 #include <Accelerate/Accelerate.h>
+#if defined(GGML_USE_CLBLAST) // allow usage of CLBlast alongside Accelerate functions
+#include "ggml-opencl.h"
+#endif
 #elif defined(GGML_USE_OPENBLAS)
 #include <cblas.h>
 #elif defined(GGML_USE_CUBLAS)
@@ -180,9 +183,13 @@ typedef double ggml_float;
 #undef bool
 #define bool _Bool
 #else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
 #include <immintrin.h>
 #endif
 #endif
+#endif
 
 #ifdef __F16C__
 

llama-util.h

@@ -14,6 +14,7 @@
 
 #include <string>
 #include <vector>
+#include <stdexcept>
 
 #ifdef __has_include
     #if __has_include(<unistd.h>)
@@ -74,7 +75,7 @@ struct llama_file {
     llama_file(const char * fname, const char * mode) {
         fp = std::fopen(fname, mode);
         if (fp == NULL) {
-            throw format("failed to open %s: %s", fname, std::strerror(errno));
+            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
         }
         seek(0, SEEK_END);
         size = tell();
@@ -107,10 +108,10 @@ struct llama_file {
         errno = 0;
         std::size_t ret = std::fread(ptr, size, 1, fp);
         if (ferror(fp)) {
-            throw format("read error: %s", strerror(errno));
+            throw std::runtime_error(format("read error: %s", strerror(errno)));
         }
         if (ret != 1) {
-            throw std::string("unexpectedly reached end of file");
+            throw std::runtime_error(std::string("unexpectedly reached end of file"));
         }
     }
 
@@ -133,7 +134,7 @@ struct llama_file {
         errno = 0;
         size_t ret = std::fwrite(ptr, size, 1, fp);
         if (ret != 1) {
-            throw format("write error: %s", strerror(errno));
+            throw std::runtime_error(format("write error: %s", strerror(errno)));
         }
     }
 
@@ -180,7 +181,7 @@ struct llama_mmap {
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
         if (addr == MAP_FAILED) {
-            throw format("mmap failed: %s", strerror(errno));
+            throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
         }
 
         if (prefetch) {
@@ -207,7 +208,7 @@ struct llama_mmap {
         DWORD error = GetLastError();
 
         if (hMapping == NULL) {
-            throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());
+            throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
         }
 
         addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
@@ -215,7 +216,7 @@ struct llama_mmap {
         CloseHandle(hMapping);
 
         if (addr == NULL) {
-            throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str());
+            throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
         }
 
         #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
@@ -245,7 +246,7 @@ struct llama_mmap {
 
     llama_mmap(struct llama_file *, bool prefetch = true) {
         (void)prefetch;
-        throw std::string("mmap not supported");
+        throw std::runtime_error(std::string("mmap not supported"));
     }
 #endif
 };

llama.cpp

@@ -970,8 +970,6 @@ static void llama_model_load_internal(
 
     // prepare memory for the weights
     {
-        const auto & hparams = model.hparams;
-
         const uint32_t n_embd  = hparams.n_embd;
         const uint32_t n_layer = hparams.n_layer;
         const uint32_t n_vocab = hparams.n_vocab;
@@ -1052,6 +1050,13 @@ static bool llama_eval_internal(
             const int   n_tokens,
             const int   n_past,
             const int   n_threads) {
+
+    // enforce that the first token is BOS
+    if (n_past == 0 && tokens[0] != llama_token_bos()) {
+        fprintf(stderr, "%s: first token must be BOS\n", __func__);
+        return false;
+    }
+
     const int64_t t_start_us = ggml_time_us();
 
     const int N = n_tokens;
@@ -1482,7 +1487,7 @@ static std::vector<llama_vocab::id> llama_tokenize(const llama_vocab & vocab, co
     }
 
     if (bos) {
-        output.push_back(1);
+        output.push_back(llama_token_bos());
     }
 
     tokenizer.tokenize(text, output);
@@ -1791,7 +1796,7 @@ llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_
     float k = powf((epsilon_hat * powf(2, *mu)) / (1 - powf(N, -epsilon_hat)), 1 / s_hat);
 
     // Sample the next word X using top-k sampling
-    llama_sample_top_k(nullptr, candidates, int(k));
+    llama_sample_top_k(nullptr, candidates, int(k), 1);
     if (ctx) {
         ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
     }
@@ -2727,11 +2732,14 @@ int llama_eval(
         fprintf(stderr, "%s: failed to eval\n", __func__);
         return 1;
     }
+
     // get a more accurate load time, upon first eval
+    // TODO: fix this
     if (!ctx->has_evaluated_once) {
         ctx->t_load_us = ggml_time_us() - ctx->t_start_us;
         ctx->has_evaluated_once = true;
     }
+
     return 0;
 }
 

llama.h

@@ -202,16 +202,16 @@ extern "C" {
     LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep = 1);
+    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep = 1);
+    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
 
     /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep = 1);
+    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
-    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep = 1);
+    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
     LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
 
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.

scripts/ppl-run-all.sh

@@ -0,0 +1,43 @@
+#!/bin/bash
+
+#
+# quantize
+#
+
+# 7B
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-7b-q4_0.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-7b-q4_1.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_2.bin q4_2 2>&1 | tee ../qnt-7b-q4_2.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-7b-q5_0.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-7b-q5_1.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-7b-q8_0.txt
+
+# 13B
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-13b-q4_0.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-13b-q4_1.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_2.bin q4_2 2>&1 | tee ../qnt-13b-q4_2.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-13b-q5_0.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-13b-q5_1.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-13b-q8_0.txt
+
+#
+# perplexity
+#
+
+# 7B
+time ./bin/perplexity -m ../models/7B/ggml-model-f16.bin  -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-f16.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_0.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_1.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q4_2.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_2.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_0.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_1.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q8_0.txt
+
+# 13B
+time ./bin/perplexity -m ../models/13B/ggml-model-f16.bin  -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-f16.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_0.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_1.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q4_2.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_2.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_0.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_1.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q8_0.txt

tests/test-sampling.cpp

@@ -32,7 +32,7 @@ void test_top_k(const std::vector<float> & probs,
     llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
     llama_sample_softmax(nullptr, &candidates_p);
     DUMP(&candidates_p);
-    llama_sample_top_k(nullptr, &candidates_p, k);
+    llama_sample_top_k(nullptr, &candidates_p, k, 1);
     DUMP(&candidates_p);
 
     assert(candidates_p.size == expected_probs.size());
@@ -57,7 +57,7 @@ void test_top_p(const std::vector<float> & probs,
     llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
     llama_sample_softmax(nullptr, &candidates_p);
     DUMP(&candidates_p);
-    llama_sample_top_p(nullptr, &candidates_p, p);
+    llama_sample_top_p(nullptr, &candidates_p, p, 1);
     DUMP(&candidates_p);
 
     assert(candidates_p.size == expected_probs.size());
@@ -80,7 +80,7 @@ void test_tfs(const std::vector<float> & probs,
 
     llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
     DUMP(&candidates_p);
-    llama_sample_tail_free(nullptr, &candidates_p, z);
+    llama_sample_tail_free(nullptr, &candidates_p, z, 1);
     DUMP(&candidates_p);
 
     assert(candidates_p.size == expected_probs.size());
@@ -103,7 +103,7 @@ void test_typical(const std::vector<float> & probs,
 
     llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
     DUMP(&candidates_p);
-    llama_sample_typical(nullptr, &candidates_p, p);
+    llama_sample_typical(nullptr, &candidates_p, p, 1);
     DUMP(&candidates_p);
 
     assert(candidates_p.size == expected_probs.size());