llama-cpp-python 0.2.38__tar.gz → 0.2.39__tar.gz

llama_cpp_python-0.2.39/.git/FETCH_HEAD

@@ -0,0 +1 @@
+34f31040f610925552a66b3a033e31320b6f6ad8		'34f31040f610925552a66b3a033e31320b6f6ad8' of https://github.com/abetlen/llama-cpp-python

llama_cpp_python-0.2.39/.git/HEAD

@@ -0,0 +1 @@
+34f31040f610925552a66b3a033e31320b6f6ad8

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/.git/config

@@ -9,7 +9,7 @@
 [gc]
 	auto = 0
 [http "https://github.com/"]
-	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX2ptaUFjRVlmS1RIZmJGRE9hdnltMDZJQ0p2MGVoTjFxOGFWNQ==
+	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX3J2Wk5WNnkzdklaWVhQWDVqc2F4NXZUUTJQdXpUcDRCUXVHSQ==
 [submodule "vendor/llama.cpp"]
 	active = true
 	url = https://github.com/ggerganov/llama.cpp.git

llama_cpp_python-0.2.39/.git/logs/HEAD

@@ -0,0 +1 @@
+0000000000000000000000000000000000000000 34f31040f610925552a66b3a033e31320b6f6ad8 runner <runner@fv-az1019-910.m5ch4o5xfz2e5czaqtqza4205h.dx.internal.cloudapp.net> 1707241738 +0000	checkout: moving from master to refs/tags/v0.2.39

llama_cpp_python-0.2.39/.git/modules/vendor/llama.cpp/FETCH_HEAD

@@ -0,0 +1 @@
+b08f22c882a1443e6b97081f3ce718a4d1a741f8		'b08f22c882a1443e6b97081f3ce718a4d1a741f8' of https://github.com/ggerganov/llama.cpp

llama_cpp_python-0.2.39/.git/modules/vendor/llama.cpp/HEAD

@@ -0,0 +1 @@
+b08f22c882a1443e6b97081f3ce718a4d1a741f8

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/.git/modules/vendor/llama.cpp/config

@@ -13,7 +13,7 @@
 [gc]
 	auto = 0
 [http "https://github.com/"]
-	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX2ptaUFjRVlmS1RIZmJGRE9hdnltMDZJQ0p2MGVoTjFxOGFWNQ==
+	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX3J2Wk5WNnkzdklaWVhQWDVqc2F4NXZUUTJQdXpUcDRCUXVHSQ==
 [url "https://github.com/"]
 	insteadOf = git@github.com:
 	insteadOf = org-6826477@github.com:

llama_cpp_python-0.2.39/.git/modules/vendor/llama.cpp/logs/HEAD

@@ -0,0 +1,2 @@
+0000000000000000000000000000000000000000 17c97fb0620448b37516a3f53fea6c482b0a30a4 runner <runner@fv-az1019-910.m5ch4o5xfz2e5czaqtqza4205h.dx.internal.cloudapp.net> 1707241739 +0000	clone: from https://github.com/ggerganov/llama.cpp.git
+17c97fb0620448b37516a3f53fea6c482b0a30a4 b08f22c882a1443e6b97081f3ce718a4d1a741f8 runner <runner@fv-az1019-910.m5ch4o5xfz2e5czaqtqza4205h.dx.internal.cloudapp.net> 1707241741 +0000	checkout: moving from master to b08f22c882a1443e6b97081f3ce718a4d1a741f8

llama_cpp_python-0.2.39/.git/modules/vendor/llama.cpp/logs/refs/heads/master

@@ -0,0 +1 @@
+0000000000000000000000000000000000000000 17c97fb0620448b37516a3f53fea6c482b0a30a4 runner <runner@fv-az1019-910.m5ch4o5xfz2e5czaqtqza4205h.dx.internal.cloudapp.net> 1707241739 +0000	clone: from https://github.com/ggerganov/llama.cpp.git

llama_cpp_python-0.2.39/.git/modules/vendor/llama.cpp/logs/refs/remotes/origin/HEAD

@@ -0,0 +1 @@
+0000000000000000000000000000000000000000 17c97fb0620448b37516a3f53fea6c482b0a30a4 runner <runner@fv-az1019-910.m5ch4o5xfz2e5czaqtqza4205h.dx.internal.cloudapp.net> 1707241739 +0000	clone: from https://github.com/ggerganov/llama.cpp.git

llama_cpp_python-0.2.39/.git/modules/vendor/llama.cpp/packed-refs

@@ -0,0 +1,2 @@
+# pack-refs with: peeled fully-peeled sorted 
+17c97fb0620448b37516a3f53fea6c482b0a30a4 refs/remotes/origin/master

llama_cpp_python-0.2.39/.git/modules/vendor/llama.cpp/refs/heads/master

@@ -0,0 +1 @@
+17c97fb0620448b37516a3f53fea6c482b0a30a4

llama_cpp_python-0.2.39/.git/modules/vendor/llama.cpp/shallow

@@ -0,0 +1,2 @@
+17c97fb0620448b37516a3f53fea6c482b0a30a4
+b08f22c882a1443e6b97081f3ce718a4d1a741f8

llama_cpp_python-0.2.39/.git/refs/tags/v0.2.39

@@ -0,0 +1 @@
+34f31040f610925552a66b3a033e31320b6f6ad8

llama_cpp_python-0.2.39/.git/shallow

@@ -0,0 +1 @@
+34f31040f610925552a66b3a033e31320b6f6ad8

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/CHANGELOG.md

@@ -7,11 +7,16 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.2.39]
+
+- feat: Update llama.cpp to ggerganov/llama.cpp@b08f22c882a1443e6b97081f3ce718a4d1a741f8
+- fix: Fix destructor logging bugs by using llama_log_callback to avoid suppress_stdout_stderr by @abetlen in 59760c85eddc72dfcc1839f43760ef72c23d6874
+
 ## [0.2.38]
 
 - feat: Update llama.cpp to ggerganov/llama.cpp@1cfb5372cf5707c8ec6dde7c874f4a44a6c4c915
 - feat: Add speculative decoding by @abetlen in #1120
-- fix: Pass raise_exception and add_generation_prompt to jinja2 chat template 078cca0361bf5a94d2cf52ed04980d20e32d6f95
+- fix: Pass raise_exception and add_generation_prompt to jinja2 chat template by @abetlen in 078cca0361bf5a94d2cf52ed04980d20e32d6f95
 
 ## [0.2.37]
 

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: llama_cpp_python
-Version: 0.2.38
+Version: 0.2.39
 Summary: Python bindings for the llama.cpp library
 Author-Email: Andrei Betlen <abetlen@gmail.com>
 License: MIT

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/llama_cpp/__init__.py

@@ -1,4 +1,4 @@
 from .llama_cpp import *
 from .llama import *
 
-__version__ = "0.2.38"
+__version__ = "0.2.39"

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/llama_cpp/_internals.py

@@ -18,8 +18,6 @@ from .llama_grammar import LlamaGrammar
 
 import llama_cpp.llama_cpp as llama_cpp
 
-from ._utils import suppress_stdout_stderr
-
 
 # Python wrappers over llama.h structs
 
@@ -30,7 +28,6 @@ class _LlamaModel:
 
     _llama_free_model = None
     # NOTE: this must be "saved" here to avoid exceptions when calling __del__
-    _suppress_stdout_stderr = suppress_stdout_stderr
 
     def __init__(
         self,
@@ -48,16 +45,14 @@ class _LlamaModel:
         if not os.path.exists(path_model):
             raise ValueError(f"Model path does not exist: {path_model}")
 
-        with self._suppress_stdout_stderr(disable=self.verbose):
-            self.model = llama_cpp.llama_load_model_from_file(
-                self.path_model.encode("utf-8"), self.params
-            )
+        self.model = llama_cpp.llama_load_model_from_file(
+            self.path_model.encode("utf-8"), self.params
+        )
 
     def __del__(self):
-        with self._suppress_stdout_stderr(disable=self.verbose):
-            if self.model is not None and self._llama_free_model is not None:
-                self._llama_free_model(self.model)
-                self.model = None
+        if self.model is not None and self._llama_free_model is not None:
+            self._llama_free_model(self.model)
+            self.model = None
 
     def vocab_type(self) -> int:
         assert self.model is not None
@@ -240,8 +235,6 @@ class _LlamaContext:
     NOTE: For stability it's recommended you use the Llama class instead."""
 
     _llama_free = None
-    # NOTE: this must be "saved" here to avoid exceptions when calling __del__
-    _suppress_stdout_stderr = suppress_stdout_stderr
 
     def __init__(
         self,
@@ -256,16 +249,16 @@ class _LlamaContext:
 
         self._llama_free = llama_cpp._lib.llama_free  # type: ignore
 
-        with self._suppress_stdout_stderr(disable=self.verbose):
-            self.ctx = llama_cpp.llama_new_context_with_model(
-                self.model.model, self.params
-            )
+        assert self.model.model is not None
+
+        self.ctx = llama_cpp.llama_new_context_with_model(
+            self.model.model, self.params
+        )
 
     def __del__(self):
-        with self._suppress_stdout_stderr(disable=self.verbose):
-            if self.ctx is not None and self._llama_free is not None:
-                self._llama_free(self.ctx)
-                self.ctx = None
+        if self.ctx is not None and self._llama_free is not None:
+            self._llama_free(self.ctx)
+            self.ctx = None
 
     def n_ctx(self) -> int:
         assert self.ctx is not None
@@ -493,8 +486,6 @@ class _LlamaContext:
 
 class _LlamaBatch:
     _llama_batch_free = None
-    # NOTE: this must be "saved" here to avoid exceptions when calling __del__
-    _suppress_stdout_stderr = suppress_stdout_stderr
 
     def __init__(
         self, *, n_tokens: int, embd: int, n_seq_max: int, verbose: bool = True
@@ -506,16 +497,14 @@ class _LlamaBatch:
 
         self._llama_batch_free = llama_cpp._lib.llama_batch_free  # type: ignore
 
-        with self._suppress_stdout_stderr(disable=self.verbose):
-            self.batch = llama_cpp.llama_batch_init(
-                self.n_tokens, self.embd, self.n_seq_max
-            )
+        self.batch = llama_cpp.llama_batch_init(
+            self.n_tokens, self.embd, self.n_seq_max
+        )
 
     def __del__(self):
-        with self._suppress_stdout_stderr(disable=self.verbose):
-            if self.batch is not None and self._llama_batch_free is not None:
-                self._llama_batch_free(self.batch)
-                self.batch = None
+        if self.batch is not None and self._llama_batch_free is not None:
+            self._llama_batch_free(self.batch)
+            self.batch = None
 
     def set_batch(self, batch: Sequence[int], n_past: int, logits_all: bool):
         assert self.batch is not None

llama_cpp_python-0.2.39/llama_cpp/_logger.py

@@ -0,0 +1,37 @@
+import sys
+import ctypes
+import logging
+
+import llama_cpp
+
+# enum ggml_log_level {
+#     GGML_LOG_LEVEL_ERROR = 2,
+#     GGML_LOG_LEVEL_WARN = 3,
+#     GGML_LOG_LEVEL_INFO = 4,
+#     GGML_LOG_LEVEL_DEBUG = 5
+# };
+GGML_LOG_LEVEL_TO_LOGGING_LEVEL = {
+    2: logging.ERROR,
+    3: logging.WARNING,
+    4: logging.INFO,
+    5: logging.DEBUG,
+}
+
+logger = logging.getLogger("llama-cpp-python")
+
+
+@llama_cpp.llama_log_callback
+def llama_log_callback(
+    level: int,
+    text: bytes,
+    user_data: ctypes.c_void_p,
+):
+    if logger.level <= GGML_LOG_LEVEL_TO_LOGGING_LEVEL[level]:
+        print(text.decode("utf-8"), end="", flush=True, file=sys.stderr)
+
+
+llama_cpp.llama_log_set(llama_log_callback, ctypes.c_void_p(0))
+
+
+def set_verbose(verbose: bool):
+    logger.setLevel(logging.DEBUG if verbose else logging.ERROR)

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/llama_cpp/llama.py

@@ -35,7 +35,6 @@ from llama_cpp.llama_speculative import LlamaDraftModel
 import numpy as np
 import numpy.typing as npt
 
-from ._utils import suppress_stdout_stderr
 from ._internals import (
     _LlamaModel,  # type: ignore
     _LlamaContext,  # type: ignore
@@ -44,6 +43,7 @@ from ._internals import (
     _LlamaSamplingParams,  # type: ignore
     _LlamaSamplingContext,  # type: ignore
 )
+from ._logger import set_verbose
 
 
 class Llama:
@@ -169,10 +169,11 @@ class Llama:
         """
         self.verbose = verbose
 
+        set_verbose(verbose)
+
         self.numa = numa
         if not Llama.__backend_initialized:
-            with suppress_stdout_stderr(disable=self.verbose):
-                llama_cpp.llama_backend_init(self.numa)
+            llama_cpp.llama_backend_init(self.numa)
             Llama.__backend_initialized = True
 
         self.model_path = model_path

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/llama_cpp/llama_cpp.py

@@ -445,7 +445,7 @@ class llama_model_params(Structure):
 #     uint32_t n_batch;           // prompt processing maximum batch size
 #     uint32_t n_threads;         // number of threads to use for generation
 #     uint32_t n_threads_batch;   // number of threads to use for batch processing
-#     int8_t   rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
+#     int32_t  rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
 
 #     // ref: https://github.com/ggerganov/llama.cpp/pull/2054
 #     float    rope_freq_base;   // RoPE base frequency, 0 = from model
@@ -502,7 +502,7 @@ class llama_context_params(Structure):
         ("n_batch", c_uint32),
         ("n_threads", c_uint32),
         ("n_threads_batch", c_uint32),
-        ("rope_scaling_type", c_int8),
+        ("rope_scaling_type", c_int32),
         ("rope_freq_base", c_float),
         ("rope_freq_scale", c_float),
         ("yarn_ext_factor", c_float),

llama_cpp_python-0.2.39/vendor/llama.cpp/.devops/main-intel.Dockerfile

@@ -0,0 +1,28 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
+
+ARG LLAMA_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+RUN mkdir build && \
+    cd build && \
+    if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
+        echo "LLAMA_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
+    fi && \
+    cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake --build . --config Release --target main
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
+
+COPY --from=build /app/build/bin/main /main
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/main" ]

llama_cpp_python-0.2.39/vendor/llama.cpp/.devops/main-vulkan.Dockerfile

@@ -0,0 +1,29 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget
+
+# Install Vulkan SDK
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DLLAMA_VULKAN=1 && \
+    cmake --build . --config Release --target main
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/main /main && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/main" ]

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/vendor/llama.cpp/.devops/nix/package.nix

@@ -13,18 +13,22 @@
   cudaPackages,
   darwin,
   rocmPackages,
+  vulkan-headers,
+  vulkan-loader,
   clblast,
   useBlas ? builtins.all (x: !x) [
     useCuda
     useMetalKit
     useOpenCL
     useRocm
+    useVulkan
   ],
   useCuda ? config.cudaSupport,
   useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
   useMpi ? false, # Increases the runtime closure size by ~700M
   useOpenCL ? false,
   useRocm ? config.rocmSupport,
+  useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
 }@inputs:
 
@@ -48,7 +52,8 @@ let
     ++ lib.optionals useMetalKit [ "MetalKit" ]
     ++ lib.optionals useMpi [ "MPI" ]
     ++ lib.optionals useOpenCL [ "OpenCL" ]
-    ++ lib.optionals useRocm [ "ROCm" ];
+    ++ lib.optionals useRocm [ "ROCm" ]
+    ++ lib.optionals useVulkan [ "Vulkan" ];
 
   pnameSuffix =
     strings.optionalString (suffices != [ ])
@@ -108,6 +113,11 @@ let
     hipblas
     rocblas
   ];
+
+  vulkanBuildInputs = [
+    vulkan-headers
+    vulkan-loader
+  ];
 in
 
 effectiveStdenv.mkDerivation (
@@ -164,7 +174,8 @@ effectiveStdenv.mkDerivation (
       ++ optionals useCuda cudaBuildInputs
       ++ optionals useMpi [ mpi ]
       ++ optionals useOpenCL [ clblast ]
-      ++ optionals useRocm rocmBuildInputs;
+      ++ optionals useRocm rocmBuildInputs
+      ++ optionals useVulkan vulkanBuildInputs;
 
     cmakeFlags =
       [
@@ -178,6 +189,7 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "LLAMA_HIPBLAS" useRocm)
         (cmakeBool "LLAMA_METAL" useMetalKit)
         (cmakeBool "LLAMA_MPI" useMpi)
+        (cmakeBool "LLAMA_VULKAN" useVulkan)
       ]
       ++ optionals useCuda [
         (
@@ -218,6 +230,7 @@ effectiveStdenv.mkDerivation (
         useMpi
         useOpenCL
         useRocm
+        useVulkan
         ;
 
       shell = mkShell {
@@ -242,11 +255,11 @@ effectiveStdenv.mkDerivation (
       # Configurations we don't want even the CI to evaluate. Results in the
       # "unsupported platform" messages. This is mostly a no-op, because
       # cudaPackages would've refused to evaluate anyway.
-      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
+      badPlatforms = optionals (useCuda || useOpenCL || useVulkan) lib.platforms.darwin;
 
       # Configurations that are known to result in build failures. Can be
       # overridden by importing Nixpkgs with `allowBroken = true`.
-      broken = (useMetalKit && !effectiveStdenv.isDarwin);
+      broken = (useMetalKit && !effectiveStdenv.isDarwin) || (useVulkan && effectiveStdenv.isDarwin);
 
       description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
       homepage = "https://github.com/ggerganov/llama.cpp/";

llama_cpp_python-0.2.39/vendor/llama.cpp/.devops/server-intel.Dockerfile

@@ -0,0 +1,28 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
+
+ARG LLAMA_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+RUN mkdir build && \
+    cd build && \
+    if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
+        echo "LLAMA_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
+    fi && \
+    cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake --build . --config Release --target server
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
+
+COPY --from=build /app/build/bin/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

llama_cpp_python-0.2.39/vendor/llama.cpp/.devops/server-vulkan.Dockerfile

@@ -0,0 +1,29 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget
+
+# Install Vulkan SDK
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DLLAMA_VULKAN=1 && \
+    cmake --build . --config Release --target server
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/server /server && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

{llama_cpp_python-0.2.38 → llama_cpp_python-0.2.39}/vendor/llama.cpp/CMakeLists.txt

@@ -79,7 +79,7 @@ if (NOT MSVC)
 endif()
 
 if (WIN32)
-    option(LLAMA_WIN_VER                     "llama: Windows Version"                           0x602)
+    set(LLAMA_WIN_VER "0x602" CACHE STRING "llama: Windows Version")
 endif()
 
 # 3rd party libs
@@ -100,6 +100,10 @@ option(LLAMA_HIPBLAS                         "llama: use hipBLAS"
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_VULKAN                          "llama: use Vulkan"                                OFF)
+option(LLAMA_VULKAN_CHECK_RESULTS            "llama: run Vulkan op checks"                      OFF)
+option(LLAMA_VULKAN_DEBUG                    "llama: enable Vulkan debug output"                OFF)
+option(LLAMA_VULKAN_VALIDATE                 "llama: enable Vulkan validation"                  OFF)
+option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"                          OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
@@ -431,6 +435,22 @@ if (LLAMA_VULKAN)
 
         add_compile_definitions(GGML_USE_VULKAN)
 
+        if (LLAMA_VULKAN_CHECK_RESULTS)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_CHECK_RESULTS)
+        endif()
+
+        if (LLAMA_VULKAN_DEBUG)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_DEBUG)
+        endif()
+
+        if (LLAMA_VULKAN_VALIDATE)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_VALIDATE)
+        endif()
+
+        if (LLAMA_VULKAN_RUN_TESTS)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_RUN_TESTS)
+        endif()
+
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
     else()
         message(WARNING "Vulkan not found")
@@ -789,9 +809,9 @@ if (LLAMA_CCACHE)
     if (LLAMA_CCACHE_FOUND)
         set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache)
         set(ENV{CCACHE_SLOPPINESS} time_macros)
-        message(STATUS "Using ccache")
+        message(STATUS "ccache found, compilation results will be cached. Disable with LLAMA_CCACHE=OFF.")
     else()
-        message(STATUS "Warning: ccache not found - consider installing it or use LLAMA_CCACHE=OFF")
+        message(STATUS "Warning: ccache not found - consider installing it for faster compilation or disable this warning with LLAMA_CCACHE=OFF")
     endif ()
 endif()