Merge https://github.com/abetlen/llama-cpp-python

.github/workflows/build-and-release.yaml

@@ -29,7 +29,7 @@ jobs:
           python -m pip install -e .[all]
 
       - name: Build wheels
-        uses: pypa/cibuildwheel@v2.18.1
+        uses: pypa/cibuildwheel@v2.19.0
         env:
           # disable repair
           CIBW_REPAIR_WHEEL_COMMAND: ""
@@ -56,7 +56,7 @@ jobs:
           platforms: linux/arm64
 
       - name: Build wheels
-        uses: pypa/cibuildwheel@v2.18.1
+        uses: pypa/cibuildwheel@v2.19.0
         env:
           CIBW_SKIP: "*musllinux* pp*"
           CIBW_REPAIR_WHEEL_COMMAND: ""

.github/workflows/build-wheels-cuda.yaml

@@ -20,8 +20,8 @@ jobs:
         id: set-matrix
         run: |
           $matrix = @{
-              'os' = @('ubuntu-20.04', 'windows-latest')
+              'os' = @('ubuntu-latest', 'windows-latest')
-              'pyver' = @("3.10", "3.11", "3.12")
+              'pyver' = @("3.9", "3.10", "3.11", "3.12")
               'cuda' = @("12.1.1", "12.2.2", "12.3.2", "12.4.1")
               'releasetag' = @("basic")
           }
@@ -50,6 +50,7 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.pyver }}
+          cache: 'pip'
 
       - name: Setup Mamba
         uses: conda-incubator/setup-miniconda@v3.0.4
@@ -109,15 +110,15 @@ jobs:
           $env:VERBOSE = '1'
           $env:CMAKE_ARGS = '-DLLAMA_CUBLAS=on -DCMAKE_CUDA_ARCHITECTURES=all'
           $env:CMAKE_ARGS = "-DLLAMA_CUDA_FORCE_MMQ=ON $env:CMAKE_ARGS"
-          if ($env:AVXVER -eq 'AVX') {
+          # if ($env:AVXVER -eq 'AVX') {
           $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off'
-          }
+          # }
-          if ($env:AVXVER -eq 'AVX512') {
+          # if ($env:AVXVER -eq 'AVX512') {
-            $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX512=on'
+          #  $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX512=on'
-          }
+          # }
-          if ($env:AVXVER -eq 'basic') {
+          # if ($env:AVXVER -eq 'basic') {
-            $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off'
+          #  $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off'
-          }
+          # }
           python -m build --wheel
           # write the build tag to the output
           Write-Output "CUDA_VERSION=$cudaVersion" >> $env:GITHUB_ENV

.github/workflows/build-wheels-metal.yaml

@@ -6,81 +6,60 @@ permissions:
   contents: write
 
 jobs:
-  define_matrix:
-    name: Define Build Matrix
-    runs-on: ubuntu-latest
-    outputs:
-      matrix: ${{ steps.set-matrix.outputs.matrix }}
-    defaults:
-      run:
-        shell: pwsh
-
-    steps:
-      - name: Define Job Output
-        id: set-matrix
-        run: |
-          $matrix = @{
-              'os' = @('macos-11', 'macos-12', 'macos-13')
-              'pyver' = @('3.10', '3.11', '3.12')
-          }
-
-          $matrixOut = ConvertTo-Json $matrix -Compress
-          Write-Output ('matrix=' + $matrixOut) >> $env:GITHUB_OUTPUT
-
   build_wheels:
-    name: ${{ matrix.os }} Python ${{ matrix.pyver }}
+    name: Build wheels on ${{ matrix.os }}
-    needs: define_matrix
     runs-on: ${{ matrix.os }}
     strategy:
-      matrix: ${{ fromJSON(needs.define_matrix.outputs.matrix) }}
+      matrix:
-    env:
+        os: [macos-12, macos-13, macos-14]
-      OSVER: ${{ matrix.os }}
 
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: "recursive"
 
+      # Used to host cibuildwheel
       - uses: actions/setup-python@v5
         with:
-          python-version: ${{ matrix.pyver }}
+          python-version: "3.12"
+          cache: 'pip'
 
-      - name: Install Dependencies
+      - name: Install dependencies
         run: |
-          python -m pip install build wheel cmake
+          python -m pip install --upgrade pip
+          python -m pip install -e .[all]
 
-      - name: Build Wheel
+      - name: Build wheels
-        run: |
+        uses: pypa/cibuildwheel@v2.18.1
-          XCODE15PATH="/Applications/Xcode_15.0.app/Contents/Developer"
+        env:
-          XCODE15BINPATH="${XCODE15PATH}/Toolchains/XcodeDefault.xctoolchain/usr/bin"
+          # disable repair
-          export CMAKE_ARGS="-DLLAMA_NATIVE=off -DLLAMA_METAL=on"
+          CIBW_REPAIR_WHEEL_COMMAND: ""
-          [[ "$OSVER" == "macos-13" ]] && export CC="${XCODE15BINPATH}/cc" && export CXX="${XCODE15BINPATH}/c++" && export MACOSX_DEPLOYMENT_TARGET="13.0"
+          CIBW_ARCHS: "arm64"
-          [[ "$OSVER" == "macos-12" ]] && export MACOSX_DEPLOYMENT_TARGET="12.0"
+          CIBW_ENVIRONMENT: CMAKE_ARGS="-DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_APPLE_SILICON_PROCESSOR=arm64 -DLLAMA_METAL=on"
-          [[ "$OSVER" == "macos-11" ]] && export MACOSX_DEPLOYMENT_TARGET="11.0"
+          CIBW_BUILD: "cp39-* cp310-* cp311-* cp312-*"
+        with:
+          package-dir: .
+          output-dir: wheelhouse2
 
-          export CMAKE_OSX_ARCHITECTURES="arm64" && export ARCHFLAGS="-arch arm64"
+      - uses: actions/upload-artifact@v4
-          VERBOSE=1 python -m build --wheel
+        with:
+          name: wheels-mac_${{ matrix.os }}
+          path: ./wheelhouse2/*.whl
 
-          if [[ "$OSVER" == "macos-13" ]]; then
+  release:
-            export SDKROOT="${XCODE15PATH}/Platforms/MacOSX.platform/Developer/SDKs/MacOSX14.0.sdk"
+    name: Release
-            export MACOSX_DEPLOYMENT_TARGET="14.0"
+    needs: [build_wheels]
-            VERBOSE=1 python -m build --wheel
+    runs-on: ubuntu-latest
-          fi
 
-          for file in ./dist/*.whl; do cp "$file" "${file/arm64.whl/aarch64.whl}"; done
+    steps:
-
+      - uses: actions/download-artifact@v4
-          export CMAKE_OSX_ARCHITECTURES="x86_64" && export CMAKE_ARGS="-DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_METAL=on" && export ARCHFLAGS="-arch x86_64"
+        with:
-          VERBOSE=1 python -m build --wheel
+          merge-multiple: true
-
+          path: dist2
-          if [[ "$OSVER" == "macos-13" ]]; then
-            export SDKROOT="${XCODE15PATH}/Platforms/MacOSX.platform/Developer/SDKs/MacOSX14.0.sdk"
-            export MACOSX_DEPLOYMENT_TARGET="14.0"
-            VERBOSE=1 python -m build --wheel
-          fi
           
       - uses: softprops/action-gh-release@v2
         with:
-          files: dist/*
+          files: dist2/*
           # set release name to <tag>-metal
           tag_name: ${{ github.ref_name }}-metal
         env:

.github/workflows/publish-to-test.yaml

@@ -22,7 +22,8 @@ jobs:
     - name: Set up Python
       uses: actions/setup-python@v5
       with:
-        python-version: "3.8"
+        python-version: "3.11"
+        cache: 'pip'
     - name: Append Dev Version to __version__
       run: |
         DEV_VERSION=${{ github.event.inputs.dev_version }}
@@ -31,11 +32,11 @@ jobs:
         sed -i 's/__version__ = \".*\"/__version__ = \"'"${NEW_VERSION}"'\"/' llama_cpp/__init__.py
     - name: Install dependencies
       run: |
-        python3 -m pip install --upgrade pip build
+        python -m pip install --upgrade pip build
-        python3 -m pip install -e .[all]
+        python -m pip install -e .[all]
     - name: Build source distribution
       run: |
-        python3 -m build --sdist
+        python -m build --sdist
     - name: Publish to Test PyPI
       uses: pypa/gh-action-pypi-publish@release/v1
       with:

.github/workflows/publish.yaml

@@ -16,14 +16,14 @@ jobs:
     - name: Set up Python
       uses: actions/setup-python@v5
       with:
-        python-version: "3.8"
+        python-version: "3.9"
     - name: Install dependencies
       run: |
-        python3 -m pip install --upgrade pip build
+        python -m pip install --upgrade pip build
-        python3 -m pip install -e .[all]
+        python -m pip install -e .[all]
     - name: Build source distribution
       run: |
-        python3 -m build --sdist
+        python -m build --sdist
     - name: Publish distribution to PyPI
       # TODO: move to tag based releases
       # if: startsWith(github.ref, 'refs/tags')

.github/workflows/test-pypi.yaml

@@ -8,57 +8,60 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"]
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
 
     steps:
       - name: Set up Python ${{ matrix.python-version }}
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
+          cache: 'pip'
       - name: Install dependencies
         run: |
-          python3 -m pip install --upgrade pip
+          python -m pip install --upgrade pip
-          python3 -m pip install --verbose llama-cpp-python[all]
+          python -m pip install --verbose llama-cpp-python[all]
       - name: Test with pytest
         run: |
-          python3 -c "import llama_cpp"
+          python -c "import llama_cpp"
 
   build-windows:
 
     runs-on: windows-latest
     strategy:
       matrix:
-        python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"]
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
 
     steps:
       - name: Set up Python ${{ matrix.python-version }}
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
+          cache: 'pip'
       - name: Install dependencies
         run: |
-          python3 -m pip install --upgrade pip
+          python -m pip install --upgrade pip
-          python3 -m pip install --verbose llama-cpp-python[all]
+          python -m pip install --verbose llama-cpp-python[all]
       - name: Test with pytest
         run: |
-          python3 -c "import llama_cpp"
+          python -c "import llama_cpp"
 
   build-macos:
 
     runs-on: macos-latest
     strategy:
       matrix:
-        python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"]
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
 
     steps:
       - name: Set up Python ${{ matrix.python-version }}
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
+          cache: 'pip'
       - name: Install dependencies
         run: |
-          python3 -m pip install --upgrade pip
+          python -m pip install --upgrade pip
-          python3 -m pip install --verbose llama-cpp-python[all]
+          python -m pip install --verbose llama-cpp-python[all]
       - name: Test with pytest
         run: |
-          python3 -c "import llama_cpp"
+          python -c "import llama_cpp"

.github/workflows/test.yaml

@@ -14,7 +14,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
 
     steps:
       - uses: actions/checkout@v4
@@ -24,20 +24,21 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
+          cache: 'pip'
       - name: Install dependencies
         run: |
-          python3 -m pip install --upgrade pip
+          python -m pip install --upgrade pip
-          python3 -m pip install .[all] -v
+          python -m pip install .[all] -v
       - name: Test with pytest
         run: |
-          python3 -m pytest
+          python -m pytest
 
   build-windows:
 
     runs-on: windows-latest
     strategy:
       matrix:
-        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
 
     steps:
       - uses: actions/checkout@v4
@@ -47,20 +48,21 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
+          cache: 'pip'
       - name: Install dependencies
         run: |
-          python3 -m pip install --upgrade pip
+          python -m pip install --upgrade pip
-          python3 -m pip install .[all] -v
+          python -m pip install .[all] -v
       - name: Test with pytest
         run: |
-          python3 -m pytest
+          python -m pytest
 
   build-macos:
 
-    runs-on: macos-13
+    runs-on: macos-latest
     strategy:
       matrix:
-        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
 
     steps:
       - uses: actions/checkout@v4
@@ -70,13 +72,14 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
+          cache: 'pip'
       - name: Install dependencies
         run: |
-          python3 -m pip install --upgrade pip
+          python -m pip install --upgrade pip
-          python3 -m pip install .[all] --verbose
+          python -m pip install .[all] --verbose
       - name: Test with pytest
         run: |
-          python3 -m pytest
+          python -m pytest
 
   # build-linux-opencl:
 
@@ -98,29 +101,29 @@ jobs:
   #         sudo apt-get install -y --no-install-recommends llvm intel-oneapi-runtime-opencl intel-oneapi-runtime-compilers libclblast-dev
   #     - name: Install dependencies
   #       run: |
-  #         python3 -m pip install --upgrade pip
+  #         python -m pip install --upgrade pip
-  #         CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install .[all] --verbose
+  #         CMAKE_ARGS="-DLLAMA_CLBLAST=on" python -m pip install .[all] --verbose
   #     - name: Test with pytest
   #       run: |
-  #         python3 -m pytest
+  #         python -m pytest
 
 
   build-macos-metal:
 
-    runs-on: macos-13
+    runs-on: macos-latest
 
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: "recursive"
-      - name: Set up Python 3.8
+      - name: Set up Python 3.9
         uses: actions/setup-python@v5
         with:
-          python-version: "3.8"
+          python-version: "3.9"
       - name: Install dependencies
         run: |
-          python3 -m pip install --upgrade pip
+          python -m pip install --upgrade pip
-          CMAKE_ARGS="-DLLAMA_METAL=on" python3 -m pip install .[all] --verbose
+          CMAKE_ARGS="-DLLAMA_METAL=on" python -m pip install .[all] --verbose
       - name: Test with pytest
         run: |
-          python3 -m pytest
+          python -m pytest

examples/high_level_api/high_level_api_infill.py

@@ -0,0 +1,33 @@
+import argparse
+
+from llama_cpp import Llama
+
+parser = argparse.ArgumentParser()
+parser.add_argument("-m", "--model", type=str, default="../models/7B/ggml-models.bin")
+parser.add_argument("-p", "--prompt", type=str, default="def add(")
+parser.add_argument("-s", "--suffix", type=str, default="\n    return sum\n\n")
+parser.add_argument("-i", "--spm-infill", action='store_true')
+args = parser.parse_args()
+
+llm = Llama(model_path=args.model, n_gpu_layers=-1, spm_infill=args.spm_infill)
+
+output = llm.create_completion(
+      temperature = 0.0,
+      repeat_penalty = 1.0,
+      prompt = args.prompt,
+      suffix = args.suffix,
+)
+
+# Models sometimes repeat suffix in response, attempt to filter that
+response = output["choices"][0]["text"]
+response_stripped = response.rstrip()
+unwanted_response_suffix = args.suffix.rstrip()
+unwanted_response_length = len(unwanted_response_suffix)
+
+filtered = False
+if unwanted_response_suffix and response_stripped[-unwanted_response_length:] == unwanted_response_suffix:
+    response = response_stripped[:-unwanted_response_length]
+    filtered = True
+
+print(f"Fill-in-Middle completion{' (filtered)' if filtered else ''}:\n\n{args.prompt}\033[32m{response}\033[{'33' if filtered else '0'}m{args.suffix}\033[0m")
+

llama_cpp/_internals.py

@@ -9,6 +9,7 @@ from typing import (
     Sequence,
 )
 from dataclasses import dataclass, field
+from contextlib import ExitStack
 
 import numpy as np
 import numpy.typing as npt
@@ -27,9 +28,6 @@ class _LlamaModel:
     """Intermediate Python wrapper for a llama.cpp llama_model.
     NOTE: For stability it's recommended you use the Llama class instead."""
 
-    _llama_free_model = None
-    # NOTE: this must be "saved" here to avoid exceptions when calling __del__
-
     def __init__(
         self,
         *,
@@ -40,8 +38,7 @@ class _LlamaModel:
         self.path_model = path_model
         self.params = params
         self.verbose = verbose
-
+        self._exit_stack = ExitStack()
-        self._llama_free_model = llama_cpp._lib.llama_free_model  # type: ignore
 
         self.model = None
 
@@ -56,11 +53,17 @@ class _LlamaModel:
         if self.model is None:
             raise ValueError(f"Failed to load model from file: {path_model}")
 
-    def __del__(self):
+        def free_model():
-        if self.model is not None and self._llama_free_model is not None:
+            if self.model is None:
-            self._llama_free_model(self.model)
+                return
+            llama_cpp.llama_free_model(self.model)
             self.model = None
 
+        self._exit_stack.callback(free_model)
+
+    def close(self):
+        self._exit_stack.close()
+
     def vocab_type(self) -> int:
         assert self.model is not None
         return llama_cpp.llama_vocab_type(self.model)
@@ -170,6 +173,14 @@ class _LlamaModel:
         assert self.model is not None
         return llama_cpp.llama_token_eot(self.model)
 
+    def add_bos_token(self) -> int:
+        assert self.model is not None
+        return llama_cpp.llama_add_bos_token(self.model)
+
+    def add_eos_token(self) -> int:
+        assert self.model is not None
+        return llama_cpp.llama_add_eos_token(self.model)
+
     # Tokenization
 
     def tokenize(self, text: bytes, add_bos: bool, special: bool):
@@ -249,8 +260,6 @@ class _LlamaContext:
     """Intermediate Python wrapper for a llama.cpp llama_context.
     NOTE: For stability it's recommended you use the Llama class instead."""
 
-    _llama_free = None
-
     def __init__(
         self,
         *,
@@ -261,24 +270,28 @@ class _LlamaContext:
         self.model = model
         self.params = params
         self.verbose = verbose
+        self._exit_stack = ExitStack()
 
-        self._llama_free = llama_cpp._lib.llama_free  # type: ignore
         self.ctx = None
 
         assert self.model.model is not None
 
-        self.ctx = llama_cpp.llama_new_context_with_model(
+        self.ctx = llama_cpp.llama_new_context_with_model(self.model.model, self.params)
-            self.model.model, self.params
-        )
 
         if self.ctx is None:
             raise ValueError("Failed to create llama_context")
 
-    def __del__(self):
+        def free_ctx():
-        if self.ctx is not None and self._llama_free is not None:
+            if self.ctx is None:
-            self._llama_free(self.ctx)
+                return
+            llama_cpp.llama_free(self.ctx)
             self.ctx = None
 
+        self._exit_stack.callback(free_ctx)
+
+    def close(self):
+        self._exit_stack.close()
+
     def n_ctx(self) -> int:
         assert self.ctx is not None
         return llama_cpp.llama_n_ctx(self.ctx)
@@ -493,8 +506,6 @@ class _LlamaContext:
 
 
 class _LlamaBatch:
-    _llama_batch_free = None
-
     def __init__(
         self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True
     ):
@@ -502,19 +513,24 @@ class _LlamaBatch:
         self.embd = embd
         self.n_seq_max = n_seq_max
         self.verbose = verbose
-
+        self._exit_stack = ExitStack()
-        self._llama_batch_free = llama_cpp._lib.llama_batch_free  # type: ignore
 
         self.batch = None
         self.batch = llama_cpp.llama_batch_init(
             self._n_tokens, self.embd, self.n_seq_max
         )
 
-    def __del__(self):
+        def free_batch():
-        if self.batch is not None and self._llama_batch_free is not None:
+            if self.batch is None:
-            self._llama_batch_free(self.batch)
+                return
+            llama_cpp.llama_batch_free(self.batch)
             self.batch = None
 
+        self._exit_stack.callback(free_batch)
+
+    def close(self):
+        self._exit_stack.close()
+
     def n_tokens(self) -> int:
         assert self.batch is not None
         return self.batch.n_tokens

llama_cpp/llama.py

@@ -9,7 +9,9 @@ import ctypes
 import typing
 import fnmatch
 import warnings
+import contextlib
 import multiprocessing
+from types import TracebackType
 
 from typing import (
     List,
@@ -21,6 +23,7 @@ from typing import (
     Deque,
     Callable,
     Dict,
+    Type,
 )
 from collections import deque
 from pathlib import Path
@@ -115,6 +118,7 @@ class Llama:
         type_k: Optional[int] = None,
         type_v: Optional[int] = None,
         # Misc
+        spm_infill: bool = False,
         verbose: bool = True,
         # Extra Params
         **kwargs,  # type: ignore
@@ -185,6 +189,7 @@ class Llama:
             verbose: Print verbose output to stderr.
             type_k: KV cache data type for K (default: f16)
             type_v: KV cache data type for V (default: f16)
+            spm_infill: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this.
 
         Raises:
             ValueError: If the model path does not exist.
@@ -343,12 +348,16 @@ class Llama:
         self.lora_scale = lora_scale
         self.lora_path = lora_path
 
+        self.spm_infill = spm_infill
+
         if not os.path.exists(model_path):
             raise ValueError(f"Model path does not exist: {model_path}")
 
-        self._model = _LlamaModel(
+        self._stack = contextlib.ExitStack()
+
+        self._model = self._stack.enter_context(contextlib.closing(_LlamaModel(
             path_model=self.model_path, params=self.model_params, verbose=self.verbose
-        )
+        )))
 
         # Override tokenizer
         self.tokenizer_ = tokenizer or LlamaTokenizer(self)
@@ -360,18 +369,18 @@ class Llama:
             self.context_params.n_ctx = self._model.n_ctx_train()
             self.context_params.n_batch = self.n_batch
 
-        self._ctx = _LlamaContext(
+        self._ctx = self._stack.enter_context(contextlib.closing(_LlamaContext(
             model=self._model,
             params=self.context_params,
             verbose=self.verbose,
-        )
+        )))
 
-        self._batch = _LlamaBatch(
+        self._batch = self._stack.enter_context(contextlib.closing(_LlamaBatch(
             n_tokens=self.n_batch,
             embd=0,
             n_seq_max=self.context_params.n_ctx,
             verbose=self.verbose,
-        )
+        )))
 
         if self.lora_path:
             if self._model.apply_lora_from_file(
@@ -972,14 +981,33 @@ class Llama:
 
         completion_id: str = f"cmpl-{str(uuid.uuid4())}"
         created: int = int(time.time())
+        bos_token_id: int = self.token_bos()
+        cls_token_id: int = self._model.token_cls()
+        sep_token_id: int = self._model.token_sep()
         prefix_token_id: int = self._model.token_prefix()
         middle_token_id: int = self._model.token_middle()
         suffix_token_id: int = self._model.token_suffix()
+        add_space_prefix: bool = self.metadata.get("tokenizer.ggml.add_space_prefix", "true") == "true"
+        bos_tokens: List[int] = [cls_token_id if cls_token_id != -1 else bos_token_id]
+        eos_tokens: List[int] = [sep_token_id if sep_token_id != -1 else self.token_eos()]
+
+        if (isinstance(prompt, list) and suffix is None) or self._model.add_bos_token() == 0 or bos_tokens[:1] == [-1]:
+            bos_tokens = []
+
+        if (isinstance(prompt, list) and suffix is None) or (self._model.add_eos_token() != 1 and sep_token_id == -1):
+            eos_tokens = []
+
+        suffix_space_prefix: int = 0
+        # Tokenizer hack to remove leading space
+        if add_space_prefix and suffix_token_id >= 0 and suffix:
+            suffix = "☺" + suffix
+            suffix_space_prefix = 2
+
         # If prompt is empty, initialize completion with BOS token to avoid
         # detokenization including a space at the beginning of the completion
-        completion_tokens: List[int] = [] if len(prompt) > 0 else [self.token_bos()]
+        completion_tokens: List[int] = [] if len(prompt) > 0 else [bos_token_id]
         # Add blank space to start of prompt to match OG llama tokenizer
-        prompt_tokens: List[int] = (
+        prefix_tokens: List[int] = (
             (
                 [prefix_token_id]
                 if prefix_token_id >= 0 and suffix is not None
@@ -988,24 +1016,20 @@ class Llama:
             +
             (
                 (
-                    self.tokenize(prompt.encode("utf-8"), add_bos=(prefix_token_id < 0 or suffix is None), special=(prefix_token_id < 0 or suffix is None))
+                    self.tokenize(prompt.encode("utf-8"), add_bos=False, special=(prefix_token_id < 0 or suffix is None))
                     if prompt != ""
-                    else (
+                    else []
-                        []
-                        if prefix_token_id >= 0 and suffix is not None
-                        else [self.token_bos()]
-                    )
                 )
                 if isinstance(prompt, str)
                 else prompt
             )
-            +
+        )
-            (
+        suffix_tokens: List[int] = (
             (
                 [suffix_token_id]
                 +
                 (
-                        self.tokenize(suffix.encode("utf-8"), add_bos=False, special=False)
+                    self.tokenize(suffix.encode("utf-8"), add_bos=False, special=False)[suffix_space_prefix:]
                     if suffix
                     else []
                 )
@@ -1013,13 +1037,12 @@ class Llama:
             if suffix_token_id >= 0 and suffix is not None
             else []
         )
-            +
+        middle_tokens: List[int] = (
-            (
             [middle_token_id]
             if middle_token_id >= 0 and suffix is not None
             else []
         )
-        )
+        prompt_tokens: List[int] = bos_tokens + ((suffix_tokens + prefix_tokens + middle_tokens) if self.spm_infill else (prefix_tokens + suffix_tokens + middle_tokens)) + eos_tokens
         text: bytes = b""
         returned_tokens: int = 0
         stop = (
@@ -1176,7 +1199,7 @@ class Llama:
                     # not sure how to handle this branch when dealing
                     # with CJK output, so keep it unchanged
                     for token in remaining_tokens:
-                        if token == self.token_bos():
+                        if token == bos_token_id:
                             continue
                         token_end_position += len(self.detokenize([token], prev_tokens=prompt_tokens + completion_tokens[:returned_tokens]))
                         # Check if stop sequence is in the token
@@ -1303,7 +1326,7 @@ class Llama:
 
                 logprobs_or_none: Optional[CompletionLogprobs] = None
                 if logprobs is not None:
-                    if token == self.token_bos():
+                    if token == bos_token_id:
                         continue
                     token_str = self.detokenize([token]).decode(
                         "utf-8", errors="ignore"
@@ -1431,7 +1454,7 @@ class Llama:
             for idx, (token, token_str, logprobs_token) in enumerate(
                 zip(all_tokens, all_token_strs, all_logprobs)
             ):
-                if token == self.token_bos():
+                if token == bos_token_id:
                     continue
                 text_offsets.append(
                     text_offset
@@ -1858,6 +1881,7 @@ class Llama:
             type_k=self.context_params.type_k,
             type_v=self.context_params.type_v,
             # Misc
+            spm_infill=self.spm_infill,
             verbose=self.verbose,
         )
 
@@ -1940,6 +1964,10 @@ class Llama:
         """Return the pooling type."""
         return self._ctx.pooling_type()
 
+    def close(self) -> None:
+        """Explicitly free the model from memory."""
+        self._stack.close()
+
     @staticmethod
     def logits_to_logprobs(
         logits: Union[npt.NDArray[np.single], List], axis: int = -1

llama_cpp/server/model.py

@@ -44,6 +44,8 @@ class LlamaProxy:
             if self._current_model is not None:
                 return self._current_model
 
+        if self._current_model:
+            self._current_model.close()
         self._current_model = None
 
         settings = self._model_settings_dict[model]
@@ -65,6 +67,7 @@ class LlamaProxy:
 
     def free(self):
         if self._current_model:
+            self._current_model.close()
             del self._current_model
 
     @staticmethod

vendor/llama.cpp

@@ -1 +1 @@
-Subproject commit fd5ea0f897ecb3659d6c269ef6f3d833e865ead7
+Subproject commit 172c8256840ffd882ab9992ecedbb587d9b21f15