diff --git a/misc/py-llama-cpp-python/Makefile b/misc/py-llama-cpp-python/Makefile
index 23c50dd516d5..4e6f3573f767 100644
--- a/misc/py-llama-cpp-python/Makefile
+++ b/misc/py-llama-cpp-python/Makefile
@@ -1,42 +1,41 @@
 PORTNAME=	llama-cpp-python
 DISTVERSIONPREFIX=	v
-DISTVERSION=	0.3.16
-PORTREVISION=	1
+DISTVERSION=	0.3.20
 CATEGORIES=	misc # machine-learning
 PKGNAMEPREFIX=	${PYTHON_PKGNAMEPREFIX}
 
 MAINTAINER=	yuri@FreeBSD.org
 COMMENT=	Python bindings for the llama.cpp library
 WWW=		https://llama-cpp-python.readthedocs.io/en/latest/ \
 		https://github.com/abetlen/llama-cpp-python
 
 LICENSE=	MIT
 LICENSE_FILE=	${WRKSRC}/LICENSE.md
 
 BUILD_DEPENDS=	${PYTHON_PKGNAMEPREFIX}scikit-build-core>=0.9.2:devel/py-scikit-build-core@${PY_FLAVOR} \
 		cmake:devel/cmake-core
 LIB_DEPENDS=	libvulkan.so:graphics/vulkan-loader
 RUN_DEPENDS=	${PYTHON_PKGNAMEPREFIX}diskcache>=5.6.1:devel/py-diskcache@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}Jinja2>=2.11.3:devel/py-Jinja2@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}numpy1>=1.16:math/py-numpy1@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}typing-extensions>=4.5.0:devel/py-typing-extensions@${PY_FLAVOR}
 RUN_DEPENDS+=	${PYTHON_PKGNAMEPREFIX}fastapi>=0.100.0:www/py-fastapi@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}pydantic-settings>=2.0.1:devel/py-pydantic-settings@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}sse-starlette>=1.6.1:www/py-sse-starlette@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}starlette-context>=0.3.6:www/py-starlette-context@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}uvicorn>=0.22.0:www/py-uvicorn@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}pyyaml>=5.1:devel/py-pyyaml@${PY_FLAVOR}
 TEST_DEPENDS=	${PYTHON_PKGNAMEPREFIX}httpx>=0.24.1:www/py-httpx@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}huggingface-hub>=0.23.0:misc/py-huggingface-hub@${PY_FLAVOR} \
 		${PYTHON_PKGNAMEPREFIX}scipy>=1.10:science/py-scipy@${PY_FLAVOR}
 
 USES=		python shebangfix
 USE_PYTHON=	pep517 autoplist pytest
 
 USE_GITHUB=	yes
 GH_ACCOUNT=	abetlen
-GH_TUPLE=	ggerganov:llama.cpp:b6598:cpp/vendor/llama.cpp
+GH_TUPLE=	ggerganov:llama.cpp:f49e917:cpp/vendor/llama.cpp
 
 SHEBANG_GLOB=	*.py
 
 .include <bsd.port.mk>
diff --git a/misc/py-llama-cpp-python/distinfo b/misc/py-llama-cpp-python/distinfo
index 7f4849c553a3..3dc559d95177 100644
--- a/misc/py-llama-cpp-python/distinfo
+++ b/misc/py-llama-cpp-python/distinfo
@@ -1,5 +1,5 @@
-TIMESTAMP = 1758915707
-SHA256 (abetlen-llama-cpp-python-v0.3.16_GH0.tar.gz) = d6ae5a6ac40dda4d14c6bb8f5e9504d28f442cf810263661d457c948c386f2a4
-SIZE (abetlen-llama-cpp-python-v0.3.16_GH0.tar.gz) = 279565
-SHA256 (ggerganov-llama.cpp-b6598_GH0.tar.gz) = cd296792f49695bd44e885a5c1fcefe4ef72f3da7f8933be1378f944116515a3
-SIZE (ggerganov-llama.cpp-b6598_GH0.tar.gz) = 25833773
+TIMESTAMP = 1775551031
+SHA256 (abetlen-llama-cpp-python-v0.3.20_GH0.tar.gz) = 2e4a39603059cca4eb8630136253ab668ca245ed51d486969e778b69f45603ee
+SIZE (abetlen-llama-cpp-python-v0.3.20_GH0.tar.gz) = 286693
+SHA256 (ggerganov-llama.cpp-f49e917_GH0.tar.gz) = 9b683589792de24a79042b5b232451c5697ef64dfe0f8ca0bc77347538223877
+SIZE (ggerganov-llama.cpp-f49e917_GH0.tar.gz) = 29723031
diff --git a/misc/py-llama-cpp-python/files/patch-vendor_llama.cpp_ggml_src_ggml-cpu_CMakeLists.txt b/misc/py-llama-cpp-python/files/patch-vendor_llama.cpp_ggml_src_ggml-cpu_CMakeLists.txt
index e39877694888..15b68b26763e 100644
--- a/misc/py-llama-cpp-python/files/patch-vendor_llama.cpp_ggml_src_ggml-cpu_CMakeLists.txt
+++ b/misc/py-llama-cpp-python/files/patch-vendor_llama.cpp_ggml_src_ggml-cpu_CMakeLists.txt
@@ -1,18 +1,19 @@
---- vendor/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt.orig	2025-12-30 12:29:09 UTC
+--- vendor/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt.orig	2026-04-03 02:35:15 UTC
 +++ vendor/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt
-@@ -387,15 +387,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
+@@ -396,16 +396,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
+             string(TOUPPER "${POWER10_M}" POWER10_M_UPPER)
              string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M_UPPER}")
              string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
- 
+-
 -            if (EXTRACTED_NUMBER GREATER_EQUAL 10)
--                list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
+-                list(APPEND ARCH_FLAGS -mcpu=power10)
 -            elseif (EXTRACTED_NUMBER EQUAL 9)
--                list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
+-                list(APPEND ARCH_FLAGS -mcpu=power9)
 -            elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
 -                list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
 -            else()
 -                list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
 -            endif()
          elseif(GGML_CPU_ALL_VARIANTS)
              # Begin with the lowest baseline
              set(ARCH_DEFINITIONS "")