Update llama cpp python 0.3.16

recipe/0001-Adapt-shared-library-relocation.patch

@@ -0,0 +1,91 @@
+From 8156a3728b89cbb944abf5af8376100da8832965 Mon Sep 17 00:00:00 2001
+From: Julien Jerphanion <[email protected]>
+Date: Fri, 22 Aug 2025 10:22:47 +0200
+Subject: [PATCH] Adapt shared library relocation
+
+Signed-off-by: Julien Jerphanion <[email protected]>
+---
+ llama_cpp/_ctypes_extensions.py | 11 +++++++++--
+ llama_cpp/llama_cpp.py          | 13 +++++++++++++
+ 2 files changed, 22 insertions(+), 2 deletions(-)
+
+diff --git a/llama_cpp/_ctypes_extensions.py b/llama_cpp/_ctypes_extensions.py
+index e88ed38..0acd159 100644
+--- a/llama_cpp/_ctypes_extensions.py
++++ b/llama_cpp/_ctypes_extensions.py
+@@ -29,16 +29,21 @@ def load_shared_library(lib_base_name: str, base_path: pathlib.Path):
+     if sys.platform.startswith("linux") or sys.platform.startswith("freebsd"):
+         lib_paths += [
+             base_path / f"lib{lib_base_name}.so",
++            f"lib{lib_base_name}.so",
+         ]
+     elif sys.platform == "darwin":
+         lib_paths += [
+             base_path / f"lib{lib_base_name}.so",
+             base_path / f"lib{lib_base_name}.dylib",
++            f"{lib_base_name}.so",
++            f"lib{lib_base_name}.dylib",
+         ]
+     elif sys.platform == "win32":
+         lib_paths += [
+             base_path / f"{lib_base_name}.dll",
+             base_path / f"lib{lib_base_name}.dll",
++            f"{lib_base_name}.dll",
++            f"lib{lib_base_name}.dll",
+         ]
+     else:
+         raise RuntimeError("Unsupported platform")
+@@ -62,14 +67,16 @@ def load_shared_library(lib_base_name: str, base_path: pathlib.Path):
+
+     # Try to load the shared library, handling potential errors
+     for lib_path in lib_paths:
+-        if lib_path.exists():
++        if isinstance(lib_path, str) or lib_path.exists():
+             try:
+                 return ctypes.CDLL(str(lib_path), **cdll_args)  # type: ignore
++            except OSError:
++                pass
+             except Exception as e:
+                 raise RuntimeError(f"Failed to load shared library '{lib_path}': {e}")
+
+     raise FileNotFoundError(
+-        f"Shared library with base name '{lib_base_name}' not found"
++        f"Shared library with base name '{lib_base_name}' not found in {lib_paths}."
+     )
+
+
+diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py
+index 711d42a..a23c778 100644
+--- a/llama_cpp/llama_cpp.py
++++ b/llama_cpp/llama_cpp.py
+@@ -3,6 +3,7 @@ from __future__ import annotations
+ import os
+ import ctypes
+ import pathlib
++import sys
+
+ from typing import (
+     Callable,
+@@ -32,7 +33,19 @@ if TYPE_CHECKING:
+
+ # Specify the base name of the shared library to load
+ _lib_base_name = "llama"
++
+ _override_base_path = os.environ.get("LLAMA_CPP_LIB_PATH")
++if sys.platform.startswith("win") and _override_base_path is None:
++    # llama.cpp windows' builds' DLL are stored in: `$CONDA_PREFIX/Library/bin/`
++    # We cannot assume that `$CONDA_PREFIX` is set, so we will use this
++    # file position to determine the prefix directory.
++
++    # This file directory in the prefix: `$CONDA_PREFIX/lib/site-packages/llama_cpp`
++    __this_file_dir = pathlib.Path(os.path.abspath(os.path.dirname(__file__)))
++    # Prefix directory: `$CONDA_PREFIX`
++    __prefix_dir = __this_file_dir.parent.parent.parent
++    _override_base_path = __prefix_dir / "Library" / "bin"
++
+ _base_path = pathlib.Path(os.path.abspath(os.path.dirname(__file__))) / "lib" if _override_base_path is None else pathlib.Path(_override_base_path)
+ # Load the library
+ _lib = load_shared_library(_lib_base_name, _base_path)
+-- 
+2.50.1
+

recipe/0002-Fix-ARM-SVE-detection-for-old-kernel-headers.patch

@@ -0,0 +1,32 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Anaconda <[email protected]>
+Date: Mon, 11 Nov 2025 00:00:00 +0000
+Subject: [PATCH] Fix ARM SVE detection for old kernel headers
+
+PR_SVE_VL_LEN_MASK and PR_SVE_GET_VL are not available in older kernel
+headers (< Linux 4.15). Add guards to disable SVE vector length detection
+when these constants are not defined.
+
+This fixes compilation on aarch64 with older glibc/kernel headers.
+---
+ vendor/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c | 4 ++++
+ 1 file changed, 4 insertions(+)
+
+diff --git a/vendor/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c b/vendor/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c
+index 1111111..2222222 100644
+--- a/vendor/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c
++++ b/vendor/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c
+@@ -687,7 +687,11 @@ static int ggml_get_n_threads(struct ggml_compute_params * params, int n_tasks)
+
+ static void ggml_init_arm_arch_features(void) {
+ #if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
++#if defined(PR_SVE_VL_LEN_MASK) && defined(PR_SVE_GET_VL)
+     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
++#else
++    ggml_arm_arch_features.sve_cnt = 0;
++#endif
+ #endif
+ }
+
+--
+2.39.5

recipe/bld.bat

@@ -0,0 +1,22 @@
+:: Set CMake arguments for vendored llama.cpp build
+set CMAKE_ARGS=%CMAKE_ARGS% -DLLAMA_BUILD=ON
+set CMAKE_ARGS=%CMAKE_ARGS% -DLLAVA_BUILD=OFF
+
+:: Install the package
+%PYTHON% -m pip install . -vv --no-deps --no-build-isolation
+if errorlevel 1 exit 1
+
+:: Move DLLs from site-packages/bin to Library/bin (standard conda location)
+:: This matches the expectation in the patched llama_cpp.py
+if not exist %LIBRARY_BIN% mkdir %LIBRARY_BIN%
+
+if exist %SP_DIR%\bin\*.dll (
+    move %SP_DIR%\bin\*.dll %LIBRARY_BIN%\
+    if errorlevel 1 exit 1
+)
+
+:: Also move DLLs from site-packages/llama_cpp/lib to Library/bin
+if exist %SP_DIR%\llama_cpp\lib\*.dll (
+    move %SP_DIR%\llama_cpp\lib\*.dll %LIBRARY_BIN%\
+    if errorlevel 1 exit 1
+)

recipe/meta.yaml

@@ -1,61 +1,68 @@
 {% set name = "llama-cpp-python" %}
-# NOTE: VERIFY llama_cpp_version before merging!
-{% set version = "0.2.24" %}
-{% set llama_cpp_version = "0.0.1660" %}
+{% set version = "0.3.16" %}
+# Building vendored llama.cpp (commit 4227c9b) instead of using external dependency
+# because llama-cpp-python 0.3.16 requires llama.cpp >= b6173, but the available
+# llama.cpp b6872 has breaking API changes (removed llama_get_kv_self function)
 
 package:
   name: {{ name|lower }}
   version: {{ version }}
 
 source:
   url: https://pypi.io/packages/source/{{ name[0] }}/{{ name }}/llama_cpp_python-{{ version }}.tar.gz
-  sha256: 85f8fd110b4b90599d5ff427bd4a1a4db6e70817c60ba8aa609fa5c645761ec1
+  sha256: 34ed0f9bd9431af045bb63d9324ae620ad0536653740e9bb163a2e1fcb973be6
   patches:
-    # Asks cdll to look for the library in the path as well.
-    - try-lib-in-path.patch
+    # Adapt shared library relocation for conda environments
+    - 0001-Adapt-shared-library-relocation.patch
+    # Fix ARM SVE detection for old kernel headers on aarch64
+    - 0002-Fix-ARM-SVE-detection-for-old-kernel-headers.patch  # [linux and aarch64]
 
 build:
   number: 0
-  script:
-    {% macro cmake_args(key, value) -%}
-    - export CMAKE_ARGS="${CMAKE_ARGS} {{ key }}={{ value }}"    # [unix]
-    - set CMAKE_ARGS=%CMAKE_ARGS% {{ key }}={{ value }}          # [win]
-    {%- endmacro %}
-
-    {{ cmake_args("-DLLAMA_BUILD", "OFF") }}
-    {{ cmake_args("-DLLAVA_BUILD", "OFF") }}
-
-    - {{ PYTHON }} -m pip install . -vv
+  missing_dso_whitelist:  # [win]
+    - $RPATH/ggml-base.dll  # [win]
+    - $RPATH/ggml-cpu.dll   # [win]
+    - $RPATH/ggml.dll       # [win]
+  script:  # [unix]
+    - export CMAKE_ARGS="${CMAKE_ARGS} -DLLAMA_BUILD=ON"    # [unix]
+    - export CMAKE_ARGS="${CMAKE_ARGS} -DLLAVA_BUILD=OFF"   # [unix]
+    - {{ PYTHON }} -m pip install . -vv --no-deps --no-build-isolation  # [unix]
 requirements:
   build:
     - python                                 # [build_platform != target_platform]
     - cross-python_{{ target_platform }}     # [build_platform != target_platform]
-
+    - {{ stdlib('c') }}
     - {{ compiler('c') }}
     - {{ compiler('cxx') }}
     - cmake
-    - make
+    - make  # [not win]
     - pkgconfig
 
   host:
     - python
     - scikit-build-core >=0.5.1
     - pip
+    - llvm-openmp    # [osx]
 
   run:
     - python
     - typing-extensions >=4.5.0
     - numpy >=1.20.0
     - diskcache >=5.6.1
+    - jinja2 >=2.11.3
 
-    - llama.cpp {{ llama_cpp_version }}
+    # Note: No external llama.cpp dependency - using vendored version (commit 4227c9b)
+    # Vendored llama.cpp requires OpenMP
+    - _openmp_mutex  # [linux]
+    - llvm-openmp    # [osx]
 
     # Split into llama-cpp-python-server
     - uvicorn >=0.22.0
     - fastapi >=0.100.0
     - pydantic-settings >=2.0.1
     - sse-starlette >=1.6.1
-    - starlette-context >=0.3.6,<0.4
+    - starlette-context >=0.3.6
+    - pyyaml >=5.1
 test:
   imports:
     - llama_cpp
@@ -67,6 +74,12 @@ test:
 about:
   home: https://github.com/abetlen/llama-cpp-python
   summary: Python bindings for the llama.cpp library
+  description: |
+    Python bindings for llama.cpp, providing a simple Python interface for
+    inference with Large Language Models (LLMs) using the llama.cpp backend.
+    Supports CPU and GPU acceleration with vendored llama.cpp library.
+  dev_url: https://github.com/abetlen/llama-cpp-python
+  doc_url: https://llama-cpp-python.readthedocs.io
   license: MIT
   license_file:
     - LICENSE.md