From c23856b38b29c7eb46f7eb835e1767695fef5ed9 Mon Sep 17 00:00:00 2001
From: "riseproject-dev[bot]"
 <330740410+riseproject-dev[bot]@users.noreply.github.com>
Date: Thu, 22 Oct 2026 12:00:25 +0000
Subject: [PATCH 1/5] executorch: Add versions 1.5.0, 1.5.1

Signed-off-by: riseproject-dev[bot] <330740410+riseproject-dev[bot]@users.noreply.github.com>
---
 docs/packages/executorch.yaml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/docs/packages/executorch.yaml b/docs/packages/executorch.yaml
index fb148b98c445..635e147d4a2a 100644
--- a/docs/packages/executorch.yaml
+++ b/docs/packages/executorch.yaml
@@ -16,3 +16,5 @@ versions:
   - filename: executorch-1.4.1-cp314-cp314-manylinux_2_39_riscv64.whl
     sha256: 6cc15f5a4b2948289a4c0b4d86dd0202f903e8ac93e008c344a235312abf0bdc
     requires-python: <3.15,>=3.10
+- version: 1.5.0
+- version: 1.5.1

From 099fa63c560dbb4979631d56b0d2b4d980266a6b Mon Sep 17 00:00:00 2001
From: Ludovic Henry <git@ludovic.dev>
Date: Fri, 23 Oct 2026 02:20:45 +0800
Subject: [PATCH 2/5] executorch: Carry patches over to versions 1.5.0, 1.5.1

Rebased 0003 and 0004 (root CMakeLists.txt context moved upstream); 0001, 0002, 0005, 0006, 0007 are unchanged.
---
 ...de-vendored-third-party-licences-in-.patch |  34 ++++++
 ...rce-directory-name-check-for-cibuild.patch |  65 ++++++++++
 ...bind-extensions-as-MODULE-not-SHARED.patch | 115 ++++++++++++++++++
 ...able_lib-SHARED-for-llm-runner-linka.patch |  79 ++++++++++++
 ...-simd-arch-fatal-error-portable-fall.patch |  73 +++++++++++
 ...ry-cpuinfo-uarchs-once-cpuinfo-is-in.patch |  73 +++++++++++
 ...back-to-the-online-CPU-count-when-cp.patch |  63 ++++++++++
 7 files changed, 502 insertions(+)
 create mode 100644 patches/executorch/1.5.0/0001-packaging-Include-vendored-third-party-licences-in-.patch
 create mode 100644 patches/executorch/1.5.0/0002-cmake-relax-source-directory-name-check-for-cibuild.patch
 create mode 100644 patches/executorch/1.5.0/0003-cmake-build-pybind-extensions-as-MODULE-not-SHARED.patch
 create mode 100644 patches/executorch/1.5.0/0004-cmake-keep-portable_lib-SHARED-for-llm-runner-linka.patch
 create mode 100644 patches/executorch/1.5.0/0005-cmake-drop-FFHT-simd-arch-fatal-error-portable-fall.patch
 create mode 100644 patches/executorch/1.5.0/0006-xnnpack-only-query-cpuinfo-uarchs-once-cpuinfo-is-in.patch
 create mode 100644 patches/executorch/1.5.0/0007-threadpool-fall-back-to-the-online-CPU-count-when-cp.patch

diff --git a/patches/executorch/1.5.0/0001-packaging-Include-vendored-third-party-licences-in-.patch b/patches/executorch/1.5.0/0001-packaging-Include-vendored-third-party-licences-in-.patch
new file mode 100644
index 000000000000..7e65c08e10ed
--- /dev/null
+++ b/patches/executorch/1.5.0/0001-packaging-Include-vendored-third-party-licences-in-.patch
@@ -0,0 +1,34 @@
+From 0000000000000000000000000000000000000001 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Thu, 10 Jul 2026 00:00:00 +0000
+Subject: [PATCH] packaging: Include vendored third-party licences in the
+ wheel
+
+The compiled extensions in this wheel statically link FP16, FXdiv,
+XNNPACK, cpuinfo, pthreadpool, flatbuffers, flatcc, Eigen, pocketfft,
+nlohmann/json and pybind11, but `license-files` only reaches the
+project's own LICENSE. RISE distributes these wheels and must also
+carry the licence of every statically linked dependency, so widen the
+glob to pick up the `LICENSE.<dep>` files the riscv64 build stages at
+the project root before `pip wheel` runs.
+
+Upstream-Status: Inappropriate [only relevant to a distributor bundling vendored third-party licences; reproduces identically on any architecture]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ pyproject.toml | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/pyproject.toml b/pyproject.toml
+index 0000000000..0000000000 100644
+--- a/pyproject.toml
++++ b/pyproject.toml
+@@ -100,7 +100,7 @@
+ # package, and package_data to restrict the set up non-python files we
+ # include. See also setuptools/discovery.py for custom finders.
+ [tool.setuptools]
+-license-files = ["LICENSE"]
++license-files = ["LICENSE", "LICENSE.*"]
+
+ [tool.setuptools.package-dir]
+ # Tell setuptools to follow the symlink: src/executorch/* -> * for all first level
diff --git a/patches/executorch/1.5.0/0002-cmake-relax-source-directory-name-check-for-cibuild.patch b/patches/executorch/1.5.0/0002-cmake-relax-source-directory-name-check-for-cibuild.patch
new file mode 100644
index 000000000000..41e36734ac70
--- /dev/null
+++ b/patches/executorch/1.5.0/0002-cmake-relax-source-directory-name-check-for-cibuild.patch
@@ -0,0 +1,65 @@
+From 0000000000000000000000000000000000000002 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Thu, 10 Jul 2026 00:00:00 +0000
+Subject: [PATCH] cmake: stop requiring the checkout dir be named `executorch`
+
+CMakeLists.txt hard-fails configure unless the checkout's directory is
+named exactly `executorch`, because `_common_include_directories` adds
+the checkout's *parent* directory so in-tree code can say
+`#include <executorch/path/to/header.h>` (upstream tracks lifting this
+restriction at https://github.com/pytorch/executorch/issues/6475).
+
+cibuildwheel always mounts/copies the source into its manylinux
+container at the fixed path `/project`, so this check always trips
+there, regardless of what directory this repo is checked out to on the
+runner. Provide the same `<executorch/...>` include path through a
+configure-time symlink instead of requiring a literally named parent
+directory, so the build works unmodified under cibuildwheel.
+
+Upstream-Status: Inappropriate [only relevant to cibuildwheel, which always mounts the source at a fixed /project path inside its build container; see pytorch/executorch#6475]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ CMakeLists.txt | 20 +++++++++-----------
+ 1 file changed, 9 insertions(+), 11 deletions(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index be0cec2..4668d0a 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -447,22 +447,20 @@ if(NOT DEFINED CMAKE_POSITION_INDEPENDENT_CODE
+   list(APPEND _common_compile_options $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fPIC>)
+ endif()
+
+-# Let files say "include <executorch/path/to/header.h>".
+-# TODO(#6475): This requires/assumes that the repo lives in a directory named
+-# exactly `executorch`. Check the assumption first. Remove this check once we
+-# stop relying on the assumption.
+-cmake_path(GET CMAKE_CURRENT_SOURCE_DIR FILENAME _repo_dir_name)
+-if(NOT "${_repo_dir_name}" STREQUAL "executorch")
+-  message(
+-    FATAL_ERROR
+-      "The ExecuTorch repo must be cloned into a directory named exactly "
+-      "`executorch`; found `${_repo_dir_name}`. See "
+-      "https://github.com/pytorch/executorch/issues/6475 for progress on a "
+-      "fix for this restriction."
++# Let files say "include <executorch/path/to/header.h>", regardless of what
++# the checkout directory is actually named (see
++# https://github.com/pytorch/executorch/issues/6475). Provide that include
++# path unconditionally through a configure-time symlink, rather than
++# requiring the checkout itself be named `executorch`.
++set(_executorch_include_shim "${CMAKE_BINARY_DIR}/_executorch_include")
++file(MAKE_DIRECTORY "${_executorch_include_shim}")
++if(NOT EXISTS "${_executorch_include_shim}/executorch")
++  file(CREATE_LINK "${CMAKE_CURRENT_SOURCE_DIR}"
++       "${_executorch_include_shim}/executorch" SYMBOLIC
+   )
+ endif()
+ set(_common_include_directories
+-    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/..>
++    $<BUILD_INTERFACE:${_executorch_include_shim}>
+     $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/runtime/core/portable_type/c10>
+     $<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR}/include>
+     $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
+--
+2.34.1
diff --git a/patches/executorch/1.5.0/0003-cmake-build-pybind-extensions-as-MODULE-not-SHARED.patch b/patches/executorch/1.5.0/0003-cmake-build-pybind-extensions-as-MODULE-not-SHARED.patch
new file mode 100644
index 000000000000..6819d082c2df
--- /dev/null
+++ b/patches/executorch/1.5.0/0003-cmake-build-pybind-extensions-as-MODULE-not-SHARED.patch
@@ -0,0 +1,115 @@
+From 0000000000000000000000000000000000000003 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Mon, 28 Sep 2026 00:00:00 +0000
+Subject: [PATCH] cmake: build pybind extensions as MODULE, not SHARED
+
+pybind11_add_module(<name> SHARED ...) makes CMake's Python_add_library()
+require the Python::Python target (COMPONENT Development.Embed) to exist
+at configure time, even though strip_python_lib() immediately removes the
+resulting Python::Python/pybind11::embed link dependency again -- these
+are ordinary dlopen()'d extension modules that resolve Python symbols
+from the running interpreter, never anything that embeds Python.
+
+Vanilla manylinux images (unlike the custom manylinux-builder images
+PyTorch's own CI uses) ship only a static libpythonX.Y.a and never
+satisfy Development.Embed, so find_package(Python) never defines
+Python::Python and every pybind11_add_module(... SHARED ...) call fails
+configure outright:
+
+  CMake Error: Python_ADD_LIBRARY: dependent target 'Python::Python' is
+  not defined. Did you miss to request COMPONENT 'Development.Embed'?
+
+Dropping SHARED (the pybind11 default, MODULE) only requires
+Development.Module, which these images do provide, and produces the
+exact same importable extension: strip_python_lib()'s job of avoiding a
+runtime libpython link was already redundant with MODULE all along.
+
+Upstream-Status: To upstream [not yet submitted; a real portability bug on any vanilla manylinux image lacking an embeddable libpython (Development.Embed), independent of riscv64 -- pytorch/executorch's own CI runs a custom manylinux-builder image that does provide one, so upstream has likely never hit this]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ CMakeLists.txt                       | 4 ++--
+ backends/apple/coreml/CMakeLists.txt | 2 +-
+ codegen/tools/CMakeLists.txt         | 2 +-
+ extension/llm/runner/CMakeLists.txt  | 2 +-
+ extension/training/CMakeLists.txt    | 2 +-
+ 5 files changed, 6 insertions(+), 6 deletions(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 93e02c5..ac73707 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -1397,7 +1397,7 @@ if(EXECUTORCH_BUILD_PYBIND)
+   endif()
+ 
+   # pybind portable_lib
+-  pybind11_add_module(portable_lib SHARED extension/pybindings/pybindings.cpp)
++  pybind11_add_module(portable_lib extension/pybindings/pybindings.cpp)
+   strip_python_lib(portable_lib)
+   # The actual output file needs a leading underscore so it can coexist with
+   # portable_lib.py in the same python package. PyTorch requires C++20, so
+@@ -1488,7 +1488,7 @@ if(EXECUTORCH_BUILD_PYBIND)
+   # pybind data_loader module - provides PyDataLoader type for external
+   # pybinding extensions to create custom data loaders
+   pybind11_add_module(
+-    data_loader SHARED extension/pybindings/pybindings_data_loader.cpp
++    data_loader extension/pybindings/pybindings_data_loader.cpp
+   )
+   strip_python_lib(data_loader)
+   target_include_directories(data_loader PRIVATE ${_common_include_directories})
+diff --git a/backends/apple/coreml/CMakeLists.txt b/backends/apple/coreml/CMakeLists.txt
+index 8df0cfd..fb07e6e 100644
+--- a/backends/apple/coreml/CMakeLists.txt
++++ b/backends/apple/coreml/CMakeLists.txt
+@@ -181,7 +181,7 @@ install(
+ 
+ if(EXECUTORCH_BUILD_PYBIND)
+   pybind11_add_module(
+-    executorchcoreml SHARED runtime/inmemoryfs/inmemory_filesystem_py.cpp
++    executorchcoreml runtime/inmemoryfs/inmemory_filesystem_py.cpp
+     runtime/inmemoryfs/inmemory_filesystem_utils.cpp
+   )
+   strip_python_lib(executorchcoreml)
+diff --git a/codegen/tools/CMakeLists.txt b/codegen/tools/CMakeLists.txt
+index 143d2c1..d95716d 100644
+--- a/codegen/tools/CMakeLists.txt
++++ b/codegen/tools/CMakeLists.txt
+@@ -8,7 +8,7 @@
+ # Check if pybind11 is available
+ 
+ # Create the selective_build pybind11 module
+-pybind11_add_module(selective_build SHARED selective_build.cpp)
++pybind11_add_module(selective_build selective_build.cpp)
+ strip_python_lib(selective_build)
+ 
+ # Set the output name to match the module name
+diff --git a/extension/llm/runner/CMakeLists.txt b/extension/llm/runner/CMakeLists.txt
+index 17016b9..110e252 100644
+--- a/extension/llm/runner/CMakeLists.txt
++++ b/extension/llm/runner/CMakeLists.txt
+@@ -113,7 +113,7 @@ endif()
+ if(EXECUTORCH_BUILD_PYBIND)
+   # Create the Python extension module for LLM runners
+   pybind11_add_module(
+-    _llm_runner SHARED ${CMAKE_CURRENT_SOURCE_DIR}/pybindings.cpp
++    _llm_runner ${CMAKE_CURRENT_SOURCE_DIR}/pybindings.cpp
+   )
+   strip_python_lib(_llm_runner)
+ 
+diff --git a/extension/training/CMakeLists.txt b/extension/training/CMakeLists.txt
+index 7b76723..cabf478 100644
+--- a/extension/training/CMakeLists.txt
++++ b/extension/training/CMakeLists.txt
+@@ -77,7 +77,7 @@ if(EXECUTORCH_BUILD_PYBIND)
+   endif()
+ 
+   pybind11_add_module(
+-    _training_lib SHARED
++    _training_lib
+     ${CMAKE_CURRENT_SOURCE_DIR}/pybindings/_training_lib.cpp
+   )
+   strip_python_lib(_training_lib)
+
+-- 
+2.34.1
+
diff --git a/patches/executorch/1.5.0/0004-cmake-keep-portable_lib-SHARED-for-llm-runner-linka.patch b/patches/executorch/1.5.0/0004-cmake-keep-portable_lib-SHARED-for-llm-runner-linka.patch
new file mode 100644
index 000000000000..370015c5cd67
--- /dev/null
+++ b/patches/executorch/1.5.0/0004-cmake-keep-portable_lib-SHARED-for-llm-runner-linka.patch
@@ -0,0 +1,79 @@
+From 0000000000000000000000000000000000000004 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Mon, 28 Sep 2026 00:00:00 +0000
+Subject: [PATCH] cmake: keep portable_lib SHARED for llm/runner linkage
+
+0003 converted all six pybind11_add_module(<name> SHARED ...) call sites to
+plain pybind11_add_module(<name> ...) (MODULE), since vanilla manylinux
+images never satisfy Python_add_library(SHARED)'s Python::Python
+(Development.Embed) requirement. portable_lib is not like the other five,
+though: extension/llm/runner/CMakeLists.txt does
+target_link_libraries(_llm_runner PRIVATE ... portable_lib ...), and CMake
+refuses to link a MODULE_LIBRARY into another target at all:
+
+  CMake Error: Target "portable_lib" of type MODULE_LIBRARY may not be
+  linked into another target. One may link only to INTERFACE, OBJECT,
+  STATIC or SHARED libraries, or to executables with the ENABLE_EXPORTS
+  property set.
+
+This is a CMake generator restriction on the target type itself (checked in
+every CMake version's link-dependency analysis), not something any
+pybind11_add_module() argument can opt out of -- so portable_lib has to stay
+SHARED, and the Development.Embed gap has to be closed some other way.
+
+FindPython's __Python_ADD_LIBRARY() (Modules/FindPython/Support.cmake)
+requires the Python::Python target to already exist before it will even
+add_library(... SHARED ...), and then links the new target against it:
+
+  if (NOT type STREQUAL "MODULE" AND NOT TARGET ${prefix}::Python)
+    message (SEND_ERROR "...dependent target '${prefix}::Python' is not
+    defined. Did you miss to request COMPONENT 'Development.Embed'?")
+  ...
+  target_link_libraries (${name} PRIVATE ${prefix}::Python)
+
+Nothing requires that target to carry a real embeddable libpython, though:
+defining Python::Python ourselves as a link-free INTERFACE IMPORTED target
+when Development.Embed didn't find one satisfies the existence check, and
+strip_python_lib() (already applied to portable_lib right below) then
+removes Python::Python/pybind11::embed from portable_lib's real link line
+exactly as it already does for the other five, now-MODULE targets -- so no
+embeddable libpython is ever actually linked, and portable_lib comes out
+identical to what it built as on upstream's own manylinux-builder CI, which
+does have Development.Embed and never exercises this fallback.
+
+Upstream-Status: To upstream [not yet submitted; a real portability bug on any vanilla manylinux image lacking an embeddable libpython (Development.Embed), independent of riscv64, same as 0003]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ CMakeLists.txt | 15 ++++++++++++++-
+ 1 file changed, 14 insertions(+), 1 deletion(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index ac73707..c2bcca2 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -1397,7 +1397,20 @@ if(EXECUTORCH_BUILD_PYBIND)
+   endif()
+ 
+   # pybind portable_lib
+-  pybind11_add_module(portable_lib extension/pybindings/pybindings.cpp)
++  # Unlike the other five pybind11_add_module() call sites, portable_lib is
++  # itself target_link_libraries()'d into extension/llm/runner's _llm_runner,
++  # and CMake refuses to link a MODULE_LIBRARY into another target -- so this
++  # one has to stay SHARED. Python_add_library(SHARED) unconditionally
++  # requires a Python::Python target to already exist (Development.Embed),
++  # which vanilla manylinux images never provide; define it as a link-free
++  # stand-in when missing so the SHARED add_library() call is allowed to
++  # proceed. strip_python_lib() below removes it (and pybind11::embed) from
++  # portable_lib's real link line either way, so no embeddable libpython is
++  # ever actually required.
++  if(NOT TARGET Python::Python)
++    add_library(Python::Python INTERFACE IMPORTED)
++  endif()
++  pybind11_add_module(portable_lib SHARED extension/pybindings/pybindings.cpp)
+   strip_python_lib(portable_lib)
+   # The actual output file needs a leading underscore so it can coexist with
+   # portable_lib.py in the same python package. PyTorch requires C++20, so
+
+--
+2.34.1
diff --git a/patches/executorch/1.5.0/0005-cmake-drop-FFHT-simd-arch-fatal-error-portable-fall.patch b/patches/executorch/1.5.0/0005-cmake-drop-FFHT-simd-arch-fatal-error-portable-fall.patch
new file mode 100644
index 000000000000..5ec70df461bd
--- /dev/null
+++ b/patches/executorch/1.5.0/0005-cmake-drop-FFHT-simd-arch-fatal-error-portable-fall.patch
@@ -0,0 +1,73 @@
+From 0000000000000000000000000000000000000005 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Mon, 28 Sep 2026 00:00:00 +0000
+Subject: [PATCH] cmake: drop FFHT SIMD arch fatal error, portable fallback exists
+
+extension/llm/custom_ops/CMakeLists.txt picks a vendored FFHT (Fast
+Hadamard Transform) SIMD kernel source file to compile based on
+CMAKE_SYSTEM_PROCESSOR -- fht_neon.c for aarch64/arm64/armv7, fht_avx.c for
+x86_64/AMD64 -- and hard-errors for every other architecture, riscv64
+included:
+
+  CMake Error: Unsupported CMAKE_SYSTEM_PROCESSOR riscv64. (If 32-bit x86,
+  try using fht_avx.c and send a PR if it works!)
+
+But nothing on such an architecture actually needs an FFHT SIMD kernel.
+extension/llm/custom_ops/spinquant/fast_hadamard_transform.h's
+fast_hadamard_transform_ffht_impl() -- the only caller of fht_float()/
+fht_double() outside the vendored FFHT sources themselves -- already
+branches on the compiler-defined architecture macros, not
+CMAKE_SYSTEM_PROCESSOR:
+
+  #if defined(__aarch64__) || defined(__x86_64__)
+    fht_float(vec, log2_vec_size);
+    normalize_after_fht(vec, log2_vec_size);
+  #else
+    fast_hadamard_transform_simple_impl(vec, log2_vec_size);
+  #endif
+
+so on any architecture other than __aarch64__/__x86_64__ -- riscv64
+included -- the portable, scalar fast_hadamard_transform_simple_impl() is
+already used at compile time, and the FFHT SIMD kernel's fht_float()/
+fht_double() symbols are never referenced, let alone linked. The
+CMakeLists.txt FATAL_ERROR was refusing to configure over a source file
+that would never have been needed. Replace it with a STATUS message and
+add no FFHT source at all for these architectures; the vendored fht_avx.c/
+fht_neon.c/fht_sse.c/dumb_fht.c files use a different, narrower API
+(void dumb_fht(float*, int) vs. fht_float()'s int fht_float(float*, int),
+and no fht_double()/_oop equivalents) and were never a drop-in match
+regardless of architecture.
+
+Upstream-Status: To upstream [not yet submitted; the FATAL_ERROR blocks every architecture outside x86_64/aarch64/armv7 even though fast_hadamard_transform.h already has a portable fallback that needs none of them, independent of riscv64]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ extension/llm/custom_ops/CMakeLists.txt | 12 +++++++++---
+ 1 file changed, 9 insertions(+), 3 deletions(-)
+
+diff --git a/extension/llm/custom_ops/CMakeLists.txt b/extension/llm/custom_ops/CMakeLists.txt
+index fa847cb..d9f6e4f 100644
+--- a/extension/llm/custom_ops/CMakeLists.txt
++++ b/extension/llm/custom_ops/CMakeLists.txt
+@@ -58,10 +58,16 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|AMD64)")
+        "extension/llm/custom_ops/spinquant/third-party/FFHT/fht_avx.c"
+   )
+ else()
++  # No FFHT SIMD kernel exists for this architecture, but none is required:
++  # fast_hadamard_transform_ffht_impl() (fast_hadamard_transform.h) only
++  # calls fht_float() when __aarch64__ or __x86_64__ is defined, and already
++  # falls back at compile time to the portable
++  # fast_hadamard_transform_simple_impl() everywhere else, so nothing here
++  # links against fht_float/fht_double.
+   message(
+-    FATAL_ERROR
+-      "Unsupported CMAKE_SYSTEM_PROCESSOR ${CMAKE_SYSTEM_PROCESSOR}. (If \
+-32-bit x86, try using fht_avx.c and send a PR if it works!)"
++    STATUS
++      "No FFHT SIMD kernel for CMAKE_SYSTEM_PROCESSOR ${CMAKE_SYSTEM_PROCESSOR}; "
++      "using the portable fast_hadamard_transform_simple_impl fallback."
+   )
+ endif()
+
+--
+2.34.1
diff --git a/patches/executorch/1.5.0/0006-xnnpack-only-query-cpuinfo-uarchs-once-cpuinfo-is-in.patch b/patches/executorch/1.5.0/0006-xnnpack-only-query-cpuinfo-uarchs-once-cpuinfo-is-in.patch
new file mode 100644
index 000000000000..8632bcb6d9ac
--- /dev/null
+++ b/patches/executorch/1.5.0/0006-xnnpack-only-query-cpuinfo-uarchs-once-cpuinfo-is-in.patch
@@ -0,0 +1,73 @@
+From 0000000000000000000000000000000000000006 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Fri, 9 Oct 2026 00:00:00 +0000
+Subject: [PATCH 6/7] xnnpack: only query cpuinfo uarchs once cpuinfo is
+ initialized
+
+init_hardware_config() checks cpuinfo_initialize() before reading the L1/L2
+data cache sizes, but the loop that fills hardware_config.uarch[] right
+after it sits outside that check and calls cpuinfo_get_uarch()
+unconditionally. When cpuinfo cannot initialize, cpuinfo_get_uarch() is
+fatal:
+
+  Fatal error in cpuinfo: cpuinfo_get_uarch called before cpuinfo is initialized
+
+and the process aborts. XnnpackBackend's static registration calls
+xnn_initialize() while _portable_lib is being loaded, so this kills
+`import executorch.extension.pybindings.portable_lib` outright.
+
+xnn_init_hardware_config() deliberately skips its cpuinfo_initialize()
+gate on XNN_ARCH_RISCV and XNN_ARCH_PPC64 (both detect their vector
+extensions through getauxval() instead), so on those architectures
+init_hardware_config() is reached with cpuinfo uninitialized whenever
+cpuinfo cannot parse the host's topology - e.g. a riscv64 kernel that
+reports /sys/devices/system/cpu/cpuN/topology/core_id as -1. Only query
+the uarchs when cpuinfo_initialize() succeeds and otherwise record
+xnn_uarch_unknown, which is what the !XNN_ENABLE_CPUINFO branch already
+does.
+
+Upstream-Status: To upstream [not yet submitted to google/XNNPACK; unchanged on XNNPACK master, affects every architecture whose xnn_init_hardware_config() skips the cpuinfo_initialize() gate (riscv64, ppc64) on hosts where cpuinfo cannot initialize]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ .../XNNPACK/src/configs/hardware-config.c     | 22 ++++++++++++-------
+ 1 file changed, 14 insertions(+), 8 deletions(-)
+
+diff --git a/backends/xnnpack/third-party/XNNPACK/src/configs/hardware-config.c b/backends/xnnpack/third-party/XNNPACK/src/configs/hardware-config.c
+index da518a1..620c202 100644
+--- a/backends/xnnpack/third-party/XNNPACK/src/configs/hardware-config.c
++++ b/backends/xnnpack/third-party/XNNPACK/src/configs/hardware-config.c
+@@ -448,16 +448,22 @@ static void init_hardware_config(void) {
+       }
+     }
+ 
++    if (cpuinfo_initialize()) {
+ #if XNN_MAX_UARCH_TYPES > 1
+-    // Print what we think we know about the microarchs.
+-    xnn_log_info("cpuinfo_get_uarchs_count: %u.", cpuinfo_get_uarchs_count());
+-    for (int i = 0; i < cpuinfo_get_uarchs_count(); i++) {
+-      xnn_log_info("cpu_get_uarch(%i): 0x%x", i, cpuinfo_get_uarch(i)->uarch);
+-    }
++      // Print what we think we know about the microarchs.
++      xnn_log_info("cpuinfo_get_uarchs_count: %u.", cpuinfo_get_uarchs_count());
++      for (int i = 0; i < cpuinfo_get_uarchs_count(); i++) {
++        xnn_log_info("cpu_get_uarch(%i): 0x%x", i, cpuinfo_get_uarch(i)->uarch);
++      }
+ #endif  // XNN_MAX_UARCH_TYPES > 1
+-    for (size_t i = 0; i < XNN_MAX_UARCH_TYPES; ++i) {
+-      const struct cpuinfo_uarch_info* uarch = cpuinfo_get_uarch(i);
+-      hardware_config.uarch[i] = uarch ? cpuinfo_to_xnn_uarch(uarch->uarch) : xnn_uarch_unknown;
++      for (size_t i = 0; i < XNN_MAX_UARCH_TYPES; ++i) {
++        const struct cpuinfo_uarch_info* uarch = cpuinfo_get_uarch(i);
++        hardware_config.uarch[i] = uarch ? cpuinfo_to_xnn_uarch(uarch->uarch) : xnn_uarch_unknown;
++      }
++    } else {
++      for (size_t i = 0; i < XNN_MAX_UARCH_TYPES; ++i) {
++        hardware_config.uarch[i] = xnn_uarch_unknown;
++      }
+     }
+ #else
+   xnn_log_warning("Unable to determine L1/L2 data cache properties.");
+-- 
+2.43.0
+
diff --git a/patches/executorch/1.5.0/0007-threadpool-fall-back-to-the-online-CPU-count-when-cp.patch b/patches/executorch/1.5.0/0007-threadpool-fall-back-to-the-online-CPU-count-when-cp.patch
new file mode 100644
index 000000000000..d74a1e05fb51
--- /dev/null
+++ b/patches/executorch/1.5.0/0007-threadpool-fall-back-to-the-online-CPU-count-when-cp.patch
@@ -0,0 +1,63 @@
+From 0000000000000000000000000000000000000007 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Fri, 9 Oct 2026 00:00:00 +0000
+Subject: [PATCH 7/7] threadpool: fall back to the online CPU count when
+ cpuinfo fails
+
+get_threadpool() returns nullptr when cpuinfo_initialize() fails, but
+none of its callers handle that: get_pthreadpool() aborts with "Failed to
+acquire an instance of ThreadPool!" (reached from XNNCompiler.cpp when
+creating every XNNPACK runtime), and parallel_for(),
+_threadpool_get_thread_count() and the LLM sdpa op dereference the
+result directly. So on any host where cpuinfo cannot initialize - e.g. a
+riscv64 kernel that reports /sys/devices/system/cpu/cpuN/topology/core_id
+as -1 - running an XNNPACK-delegated or parallelised program crashes.
+
+cpuinfo is only used here to choose a thread count: pthreadpool itself is
+built with PTHREADPOOL_USE_CPUINFO=0 and starts its threads without it.
+Size the pool from std::thread::hardware_concurrency() (capped at the
+same tsan limit of 63) when cpuinfo is unavailable, instead of leaving
+every caller without a threadpool.
+
+Upstream-Status: To upstream [not yet submitted; unchanged on pytorch/executorch main, affects any host where cpuinfo_initialize() fails, independent of riscv64]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ extension/threadpool/threadpool.cpp | 13 ++++++++-----
+ 1 file changed, 8 insertions(+), 5 deletions(-)
+
+diff --git a/extension/threadpool/threadpool.cpp b/extension/threadpool/threadpool.cpp
+index 2845b4a..a3a8c60 100644
+--- a/extension/threadpool/threadpool.cpp
++++ b/extension/threadpool/threadpool.cpp
+@@ -11,6 +11,7 @@
+ 
+ #include <algorithm>
+ #include <memory>
++#include <thread>
+ 
+ #include <executorch/extension/threadpool/threadpool_guard.h>
+ #include <executorch/runtime/platform/assert.h>
+@@ -127,12 +128,14 @@ void ThreadPool::run(
+ ThreadPool* get_threadpool() {
+   executorch::runtime::runtime_init();
+ 
+-  if (!cpuinfo_initialize()) {
+-    ET_LOG(Error, "cpuinfo initialization failed");
+-    return nullptr; // NOLINT(facebook-hte-NullableReturn)
+-  }
+-
+   static const int num_threads = ([]() {
++    if (!cpuinfo_initialize()) {
++      // pthreadpool does not need cpuinfo, so size the pool from the online
++      // CPU count instead of leaving every caller without a threadpool.
++      ET_LOG(Info, "cpuinfo initialization failed, using the online CPU count");
++      const uint32_t online = std::thread::hardware_concurrency();
++      return std::min(std::max(online, 1u), 63u);
++    }
+ #if defined(EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES)
+     // Use threads=cores.
+     auto result = cpuinfo_get_processors_count();
+-- 
+2.43.0
+

From d29541a80b5d38f31771c137a883c36211b1ebca Mon Sep 17 00:00:00 2001
From: Ludovic Henry <git@ludovic.dev>
Date: Fri, 23 Oct 2026 02:21:50 +0800
Subject: [PATCH 3/5] executorch: Carry patches over to version 1.5.1

Same set as 1.5.0 (rebased 0003 and 0004; the rest unchanged).
---
 ...de-vendored-third-party-licences-in-.patch |  34 ++++++
 ...rce-directory-name-check-for-cibuild.patch |  65 ++++++++++
 ...bind-extensions-as-MODULE-not-SHARED.patch | 115 ++++++++++++++++++
 ...able_lib-SHARED-for-llm-runner-linka.patch |  79 ++++++++++++
 ...-simd-arch-fatal-error-portable-fall.patch |  73 +++++++++++
 ...ry-cpuinfo-uarchs-once-cpuinfo-is-in.patch |  73 +++++++++++
 ...back-to-the-online-CPU-count-when-cp.patch |  63 ++++++++++
 7 files changed, 502 insertions(+)
 create mode 100644 patches/executorch/1.5.1/0001-packaging-Include-vendored-third-party-licences-in-.patch
 create mode 100644 patches/executorch/1.5.1/0002-cmake-relax-source-directory-name-check-for-cibuild.patch
 create mode 100644 patches/executorch/1.5.1/0003-cmake-build-pybind-extensions-as-MODULE-not-SHARED.patch
 create mode 100644 patches/executorch/1.5.1/0004-cmake-keep-portable_lib-SHARED-for-llm-runner-linka.patch
 create mode 100644 patches/executorch/1.5.1/0005-cmake-drop-FFHT-simd-arch-fatal-error-portable-fall.patch
 create mode 100644 patches/executorch/1.5.1/0006-xnnpack-only-query-cpuinfo-uarchs-once-cpuinfo-is-in.patch
 create mode 100644 patches/executorch/1.5.1/0007-threadpool-fall-back-to-the-online-CPU-count-when-cp.patch

diff --git a/patches/executorch/1.5.1/0001-packaging-Include-vendored-third-party-licences-in-.patch b/patches/executorch/1.5.1/0001-packaging-Include-vendored-third-party-licences-in-.patch
new file mode 100644
index 000000000000..7e65c08e10ed
--- /dev/null
+++ b/patches/executorch/1.5.1/0001-packaging-Include-vendored-third-party-licences-in-.patch
@@ -0,0 +1,34 @@
+From 0000000000000000000000000000000000000001 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Thu, 10 Jul 2026 00:00:00 +0000
+Subject: [PATCH] packaging: Include vendored third-party licences in the
+ wheel
+
+The compiled extensions in this wheel statically link FP16, FXdiv,
+XNNPACK, cpuinfo, pthreadpool, flatbuffers, flatcc, Eigen, pocketfft,
+nlohmann/json and pybind11, but `license-files` only reaches the
+project's own LICENSE. RISE distributes these wheels and must also
+carry the licence of every statically linked dependency, so widen the
+glob to pick up the `LICENSE.<dep>` files the riscv64 build stages at
+the project root before `pip wheel` runs.
+
+Upstream-Status: Inappropriate [only relevant to a distributor bundling vendored third-party licences; reproduces identically on any architecture]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ pyproject.toml | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/pyproject.toml b/pyproject.toml
+index 0000000000..0000000000 100644
+--- a/pyproject.toml
++++ b/pyproject.toml
+@@ -100,7 +100,7 @@
+ # package, and package_data to restrict the set up non-python files we
+ # include. See also setuptools/discovery.py for custom finders.
+ [tool.setuptools]
+-license-files = ["LICENSE"]
++license-files = ["LICENSE", "LICENSE.*"]
+
+ [tool.setuptools.package-dir]
+ # Tell setuptools to follow the symlink: src/executorch/* -> * for all first level
diff --git a/patches/executorch/1.5.1/0002-cmake-relax-source-directory-name-check-for-cibuild.patch b/patches/executorch/1.5.1/0002-cmake-relax-source-directory-name-check-for-cibuild.patch
new file mode 100644
index 000000000000..41e36734ac70
--- /dev/null
+++ b/patches/executorch/1.5.1/0002-cmake-relax-source-directory-name-check-for-cibuild.patch
@@ -0,0 +1,65 @@
+From 0000000000000000000000000000000000000002 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Thu, 10 Jul 2026 00:00:00 +0000
+Subject: [PATCH] cmake: stop requiring the checkout dir be named `executorch`
+
+CMakeLists.txt hard-fails configure unless the checkout's directory is
+named exactly `executorch`, because `_common_include_directories` adds
+the checkout's *parent* directory so in-tree code can say
+`#include <executorch/path/to/header.h>` (upstream tracks lifting this
+restriction at https://github.com/pytorch/executorch/issues/6475).
+
+cibuildwheel always mounts/copies the source into its manylinux
+container at the fixed path `/project`, so this check always trips
+there, regardless of what directory this repo is checked out to on the
+runner. Provide the same `<executorch/...>` include path through a
+configure-time symlink instead of requiring a literally named parent
+directory, so the build works unmodified under cibuildwheel.
+
+Upstream-Status: Inappropriate [only relevant to cibuildwheel, which always mounts the source at a fixed /project path inside its build container; see pytorch/executorch#6475]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ CMakeLists.txt | 20 +++++++++-----------
+ 1 file changed, 9 insertions(+), 11 deletions(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index be0cec2..4668d0a 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -447,22 +447,20 @@ if(NOT DEFINED CMAKE_POSITION_INDEPENDENT_CODE
+   list(APPEND _common_compile_options $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fPIC>)
+ endif()
+
+-# Let files say "include <executorch/path/to/header.h>".
+-# TODO(#6475): This requires/assumes that the repo lives in a directory named
+-# exactly `executorch`. Check the assumption first. Remove this check once we
+-# stop relying on the assumption.
+-cmake_path(GET CMAKE_CURRENT_SOURCE_DIR FILENAME _repo_dir_name)
+-if(NOT "${_repo_dir_name}" STREQUAL "executorch")
+-  message(
+-    FATAL_ERROR
+-      "The ExecuTorch repo must be cloned into a directory named exactly "
+-      "`executorch`; found `${_repo_dir_name}`. See "
+-      "https://github.com/pytorch/executorch/issues/6475 for progress on a "
+-      "fix for this restriction."
++# Let files say "include <executorch/path/to/header.h>", regardless of what
++# the checkout directory is actually named (see
++# https://github.com/pytorch/executorch/issues/6475). Provide that include
++# path unconditionally through a configure-time symlink, rather than
++# requiring the checkout itself be named `executorch`.
++set(_executorch_include_shim "${CMAKE_BINARY_DIR}/_executorch_include")
++file(MAKE_DIRECTORY "${_executorch_include_shim}")
++if(NOT EXISTS "${_executorch_include_shim}/executorch")
++  file(CREATE_LINK "${CMAKE_CURRENT_SOURCE_DIR}"
++       "${_executorch_include_shim}/executorch" SYMBOLIC
+   )
+ endif()
+ set(_common_include_directories
+-    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/..>
++    $<BUILD_INTERFACE:${_executorch_include_shim}>
+     $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/runtime/core/portable_type/c10>
+     $<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR}/include>
+     $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
+--
+2.34.1
diff --git a/patches/executorch/1.5.1/0003-cmake-build-pybind-extensions-as-MODULE-not-SHARED.patch b/patches/executorch/1.5.1/0003-cmake-build-pybind-extensions-as-MODULE-not-SHARED.patch
new file mode 100644
index 000000000000..6819d082c2df
--- /dev/null
+++ b/patches/executorch/1.5.1/0003-cmake-build-pybind-extensions-as-MODULE-not-SHARED.patch
@@ -0,0 +1,115 @@
+From 0000000000000000000000000000000000000003 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Mon, 28 Sep 2026 00:00:00 +0000
+Subject: [PATCH] cmake: build pybind extensions as MODULE, not SHARED
+
+pybind11_add_module(<name> SHARED ...) makes CMake's Python_add_library()
+require the Python::Python target (COMPONENT Development.Embed) to exist
+at configure time, even though strip_python_lib() immediately removes the
+resulting Python::Python/pybind11::embed link dependency again -- these
+are ordinary dlopen()'d extension modules that resolve Python symbols
+from the running interpreter, never anything that embeds Python.
+
+Vanilla manylinux images (unlike the custom manylinux-builder images
+PyTorch's own CI uses) ship only a static libpythonX.Y.a and never
+satisfy Development.Embed, so find_package(Python) never defines
+Python::Python and every pybind11_add_module(... SHARED ...) call fails
+configure outright:
+
+  CMake Error: Python_ADD_LIBRARY: dependent target 'Python::Python' is
+  not defined. Did you miss to request COMPONENT 'Development.Embed'?
+
+Dropping SHARED (the pybind11 default, MODULE) only requires
+Development.Module, which these images do provide, and produces the
+exact same importable extension: strip_python_lib()'s job of avoiding a
+runtime libpython link was already redundant with MODULE all along.
+
+Upstream-Status: To upstream [not yet submitted; a real portability bug on any vanilla manylinux image lacking an embeddable libpython (Development.Embed), independent of riscv64 -- pytorch/executorch's own CI runs a custom manylinux-builder image that does provide one, so upstream has likely never hit this]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ CMakeLists.txt                       | 4 ++--
+ backends/apple/coreml/CMakeLists.txt | 2 +-
+ codegen/tools/CMakeLists.txt         | 2 +-
+ extension/llm/runner/CMakeLists.txt  | 2 +-
+ extension/training/CMakeLists.txt    | 2 +-
+ 5 files changed, 6 insertions(+), 6 deletions(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 93e02c5..ac73707 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -1397,7 +1397,7 @@ if(EXECUTORCH_BUILD_PYBIND)
+   endif()
+ 
+   # pybind portable_lib
+-  pybind11_add_module(portable_lib SHARED extension/pybindings/pybindings.cpp)
++  pybind11_add_module(portable_lib extension/pybindings/pybindings.cpp)
+   strip_python_lib(portable_lib)
+   # The actual output file needs a leading underscore so it can coexist with
+   # portable_lib.py in the same python package. PyTorch requires C++20, so
+@@ -1488,7 +1488,7 @@ if(EXECUTORCH_BUILD_PYBIND)
+   # pybind data_loader module - provides PyDataLoader type for external
+   # pybinding extensions to create custom data loaders
+   pybind11_add_module(
+-    data_loader SHARED extension/pybindings/pybindings_data_loader.cpp
++    data_loader extension/pybindings/pybindings_data_loader.cpp
+   )
+   strip_python_lib(data_loader)
+   target_include_directories(data_loader PRIVATE ${_common_include_directories})
+diff --git a/backends/apple/coreml/CMakeLists.txt b/backends/apple/coreml/CMakeLists.txt
+index 8df0cfd..fb07e6e 100644
+--- a/backends/apple/coreml/CMakeLists.txt
++++ b/backends/apple/coreml/CMakeLists.txt
+@@ -181,7 +181,7 @@ install(
+ 
+ if(EXECUTORCH_BUILD_PYBIND)
+   pybind11_add_module(
+-    executorchcoreml SHARED runtime/inmemoryfs/inmemory_filesystem_py.cpp
++    executorchcoreml runtime/inmemoryfs/inmemory_filesystem_py.cpp
+     runtime/inmemoryfs/inmemory_filesystem_utils.cpp
+   )
+   strip_python_lib(executorchcoreml)
+diff --git a/codegen/tools/CMakeLists.txt b/codegen/tools/CMakeLists.txt
+index 143d2c1..d95716d 100644
+--- a/codegen/tools/CMakeLists.txt
++++ b/codegen/tools/CMakeLists.txt
+@@ -8,7 +8,7 @@
+ # Check if pybind11 is available
+ 
+ # Create the selective_build pybind11 module
+-pybind11_add_module(selective_build SHARED selective_build.cpp)
++pybind11_add_module(selective_build selective_build.cpp)
+ strip_python_lib(selective_build)
+ 
+ # Set the output name to match the module name
+diff --git a/extension/llm/runner/CMakeLists.txt b/extension/llm/runner/CMakeLists.txt
+index 17016b9..110e252 100644
+--- a/extension/llm/runner/CMakeLists.txt
++++ b/extension/llm/runner/CMakeLists.txt
+@@ -113,7 +113,7 @@ endif()
+ if(EXECUTORCH_BUILD_PYBIND)
+   # Create the Python extension module for LLM runners
+   pybind11_add_module(
+-    _llm_runner SHARED ${CMAKE_CURRENT_SOURCE_DIR}/pybindings.cpp
++    _llm_runner ${CMAKE_CURRENT_SOURCE_DIR}/pybindings.cpp
+   )
+   strip_python_lib(_llm_runner)
+ 
+diff --git a/extension/training/CMakeLists.txt b/extension/training/CMakeLists.txt
+index 7b76723..cabf478 100644
+--- a/extension/training/CMakeLists.txt
++++ b/extension/training/CMakeLists.txt
+@@ -77,7 +77,7 @@ if(EXECUTORCH_BUILD_PYBIND)
+   endif()
+ 
+   pybind11_add_module(
+-    _training_lib SHARED
++    _training_lib
+     ${CMAKE_CURRENT_SOURCE_DIR}/pybindings/_training_lib.cpp
+   )
+   strip_python_lib(_training_lib)
+
+-- 
+2.34.1
+
diff --git a/patches/executorch/1.5.1/0004-cmake-keep-portable_lib-SHARED-for-llm-runner-linka.patch b/patches/executorch/1.5.1/0004-cmake-keep-portable_lib-SHARED-for-llm-runner-linka.patch
new file mode 100644
index 000000000000..370015c5cd67
--- /dev/null
+++ b/patches/executorch/1.5.1/0004-cmake-keep-portable_lib-SHARED-for-llm-runner-linka.patch
@@ -0,0 +1,79 @@
+From 0000000000000000000000000000000000000004 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Mon, 28 Sep 2026 00:00:00 +0000
+Subject: [PATCH] cmake: keep portable_lib SHARED for llm/runner linkage
+
+0003 converted all six pybind11_add_module(<name> SHARED ...) call sites to
+plain pybind11_add_module(<name> ...) (MODULE), since vanilla manylinux
+images never satisfy Python_add_library(SHARED)'s Python::Python
+(Development.Embed) requirement. portable_lib is not like the other five,
+though: extension/llm/runner/CMakeLists.txt does
+target_link_libraries(_llm_runner PRIVATE ... portable_lib ...), and CMake
+refuses to link a MODULE_LIBRARY into another target at all:
+
+  CMake Error: Target "portable_lib" of type MODULE_LIBRARY may not be
+  linked into another target. One may link only to INTERFACE, OBJECT,
+  STATIC or SHARED libraries, or to executables with the ENABLE_EXPORTS
+  property set.
+
+This is a CMake generator restriction on the target type itself (checked in
+every CMake version's link-dependency analysis), not something any
+pybind11_add_module() argument can opt out of -- so portable_lib has to stay
+SHARED, and the Development.Embed gap has to be closed some other way.
+
+FindPython's __Python_ADD_LIBRARY() (Modules/FindPython/Support.cmake)
+requires the Python::Python target to already exist before it will even
+add_library(... SHARED ...), and then links the new target against it:
+
+  if (NOT type STREQUAL "MODULE" AND NOT TARGET ${prefix}::Python)
+    message (SEND_ERROR "...dependent target '${prefix}::Python' is not
+    defined. Did you miss to request COMPONENT 'Development.Embed'?")
+  ...
+  target_link_libraries (${name} PRIVATE ${prefix}::Python)
+
+Nothing requires that target to carry a real embeddable libpython, though:
+defining Python::Python ourselves as a link-free INTERFACE IMPORTED target
+when Development.Embed didn't find one satisfies the existence check, and
+strip_python_lib() (already applied to portable_lib right below) then
+removes Python::Python/pybind11::embed from portable_lib's real link line
+exactly as it already does for the other five, now-MODULE targets -- so no
+embeddable libpython is ever actually linked, and portable_lib comes out
+identical to what it built as on upstream's own manylinux-builder CI, which
+does have Development.Embed and never exercises this fallback.
+
+Upstream-Status: To upstream [not yet submitted; a real portability bug on any vanilla manylinux image lacking an embeddable libpython (Development.Embed), independent of riscv64, same as 0003]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ CMakeLists.txt | 15 ++++++++++++++-
+ 1 file changed, 14 insertions(+), 1 deletion(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index ac73707..c2bcca2 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -1397,7 +1397,20 @@ if(EXECUTORCH_BUILD_PYBIND)
+   endif()
+ 
+   # pybind portable_lib
+-  pybind11_add_module(portable_lib extension/pybindings/pybindings.cpp)
++  # Unlike the other five pybind11_add_module() call sites, portable_lib is
++  # itself target_link_libraries()'d into extension/llm/runner's _llm_runner,
++  # and CMake refuses to link a MODULE_LIBRARY into another target -- so this
++  # one has to stay SHARED. Python_add_library(SHARED) unconditionally
++  # requires a Python::Python target to already exist (Development.Embed),
++  # which vanilla manylinux images never provide; define it as a link-free
++  # stand-in when missing so the SHARED add_library() call is allowed to
++  # proceed. strip_python_lib() below removes it (and pybind11::embed) from
++  # portable_lib's real link line either way, so no embeddable libpython is
++  # ever actually required.
++  if(NOT TARGET Python::Python)
++    add_library(Python::Python INTERFACE IMPORTED)
++  endif()
++  pybind11_add_module(portable_lib SHARED extension/pybindings/pybindings.cpp)
+   strip_python_lib(portable_lib)
+   # The actual output file needs a leading underscore so it can coexist with
+   # portable_lib.py in the same python package. PyTorch requires C++20, so
+
+--
+2.34.1
diff --git a/patches/executorch/1.5.1/0005-cmake-drop-FFHT-simd-arch-fatal-error-portable-fall.patch b/patches/executorch/1.5.1/0005-cmake-drop-FFHT-simd-arch-fatal-error-portable-fall.patch
new file mode 100644
index 000000000000..5ec70df461bd
--- /dev/null
+++ b/patches/executorch/1.5.1/0005-cmake-drop-FFHT-simd-arch-fatal-error-portable-fall.patch
@@ -0,0 +1,73 @@
+From 0000000000000000000000000000000000000005 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Mon, 28 Sep 2026 00:00:00 +0000
+Subject: [PATCH] cmake: drop FFHT SIMD arch fatal error, portable fallback exists
+
+extension/llm/custom_ops/CMakeLists.txt picks a vendored FFHT (Fast
+Hadamard Transform) SIMD kernel source file to compile based on
+CMAKE_SYSTEM_PROCESSOR -- fht_neon.c for aarch64/arm64/armv7, fht_avx.c for
+x86_64/AMD64 -- and hard-errors for every other architecture, riscv64
+included:
+
+  CMake Error: Unsupported CMAKE_SYSTEM_PROCESSOR riscv64. (If 32-bit x86,
+  try using fht_avx.c and send a PR if it works!)
+
+But nothing on such an architecture actually needs an FFHT SIMD kernel.
+extension/llm/custom_ops/spinquant/fast_hadamard_transform.h's
+fast_hadamard_transform_ffht_impl() -- the only caller of fht_float()/
+fht_double() outside the vendored FFHT sources themselves -- already
+branches on the compiler-defined architecture macros, not
+CMAKE_SYSTEM_PROCESSOR:
+
+  #if defined(__aarch64__) || defined(__x86_64__)
+    fht_float(vec, log2_vec_size);
+    normalize_after_fht(vec, log2_vec_size);
+  #else
+    fast_hadamard_transform_simple_impl(vec, log2_vec_size);
+  #endif
+
+so on any architecture other than __aarch64__/__x86_64__ -- riscv64
+included -- the portable, scalar fast_hadamard_transform_simple_impl() is
+already used at compile time, and the FFHT SIMD kernel's fht_float()/
+fht_double() symbols are never referenced, let alone linked. The
+CMakeLists.txt FATAL_ERROR was refusing to configure over a source file
+that would never have been needed. Replace it with a STATUS message and
+add no FFHT source at all for these architectures; the vendored fht_avx.c/
+fht_neon.c/fht_sse.c/dumb_fht.c files use a different, narrower API
+(void dumb_fht(float*, int) vs. fht_float()'s int fht_float(float*, int),
+and no fht_double()/_oop equivalents) and were never a drop-in match
+regardless of architecture.
+
+Upstream-Status: To upstream [not yet submitted; the FATAL_ERROR blocks every architecture outside x86_64/aarch64/armv7 even though fast_hadamard_transform.h already has a portable fallback that needs none of them, independent of riscv64]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ extension/llm/custom_ops/CMakeLists.txt | 12 +++++++++---
+ 1 file changed, 9 insertions(+), 3 deletions(-)
+
+diff --git a/extension/llm/custom_ops/CMakeLists.txt b/extension/llm/custom_ops/CMakeLists.txt
+index fa847cb..d9f6e4f 100644
+--- a/extension/llm/custom_ops/CMakeLists.txt
++++ b/extension/llm/custom_ops/CMakeLists.txt
+@@ -58,10 +58,16 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|AMD64)")
+        "extension/llm/custom_ops/spinquant/third-party/FFHT/fht_avx.c"
+   )
+ else()
++  # No FFHT SIMD kernel exists for this architecture, but none is required:
++  # fast_hadamard_transform_ffht_impl() (fast_hadamard_transform.h) only
++  # calls fht_float() when __aarch64__ or __x86_64__ is defined, and already
++  # falls back at compile time to the portable
++  # fast_hadamard_transform_simple_impl() everywhere else, so nothing here
++  # links against fht_float/fht_double.
+   message(
+-    FATAL_ERROR
+-      "Unsupported CMAKE_SYSTEM_PROCESSOR ${CMAKE_SYSTEM_PROCESSOR}. (If \
+-32-bit x86, try using fht_avx.c and send a PR if it works!)"
++    STATUS
++      "No FFHT SIMD kernel for CMAKE_SYSTEM_PROCESSOR ${CMAKE_SYSTEM_PROCESSOR}; "
++      "using the portable fast_hadamard_transform_simple_impl fallback."
+   )
+ endif()
+
+--
+2.34.1
diff --git a/patches/executorch/1.5.1/0006-xnnpack-only-query-cpuinfo-uarchs-once-cpuinfo-is-in.patch b/patches/executorch/1.5.1/0006-xnnpack-only-query-cpuinfo-uarchs-once-cpuinfo-is-in.patch
new file mode 100644
index 000000000000..8632bcb6d9ac
--- /dev/null
+++ b/patches/executorch/1.5.1/0006-xnnpack-only-query-cpuinfo-uarchs-once-cpuinfo-is-in.patch
@@ -0,0 +1,73 @@
+From 0000000000000000000000000000000000000006 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Fri, 9 Oct 2026 00:00:00 +0000
+Subject: [PATCH 6/7] xnnpack: only query cpuinfo uarchs once cpuinfo is
+ initialized
+
+init_hardware_config() checks cpuinfo_initialize() before reading the L1/L2
+data cache sizes, but the loop that fills hardware_config.uarch[] right
+after it sits outside that check and calls cpuinfo_get_uarch()
+unconditionally. When cpuinfo cannot initialize, cpuinfo_get_uarch() is
+fatal:
+
+  Fatal error in cpuinfo: cpuinfo_get_uarch called before cpuinfo is initialized
+
+and the process aborts. XnnpackBackend's static registration calls
+xnn_initialize() while _portable_lib is being loaded, so this kills
+`import executorch.extension.pybindings.portable_lib` outright.
+
+xnn_init_hardware_config() deliberately skips its cpuinfo_initialize()
+gate on XNN_ARCH_RISCV and XNN_ARCH_PPC64 (both detect their vector
+extensions through getauxval() instead), so on those architectures
+init_hardware_config() is reached with cpuinfo uninitialized whenever
+cpuinfo cannot parse the host's topology - e.g. a riscv64 kernel that
+reports /sys/devices/system/cpu/cpuN/topology/core_id as -1. Only query
+the uarchs when cpuinfo_initialize() succeeds and otherwise record
+xnn_uarch_unknown, which is what the !XNN_ENABLE_CPUINFO branch already
+does.
+
+Upstream-Status: To upstream [not yet submitted to google/XNNPACK; unchanged on XNNPACK master, affects every architecture whose xnn_init_hardware_config() skips the cpuinfo_initialize() gate (riscv64, ppc64) on hosts where cpuinfo cannot initialize]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ .../XNNPACK/src/configs/hardware-config.c     | 22 ++++++++++++-------
+ 1 file changed, 14 insertions(+), 8 deletions(-)
+
+diff --git a/backends/xnnpack/third-party/XNNPACK/src/configs/hardware-config.c b/backends/xnnpack/third-party/XNNPACK/src/configs/hardware-config.c
+index da518a1..620c202 100644
+--- a/backends/xnnpack/third-party/XNNPACK/src/configs/hardware-config.c
++++ b/backends/xnnpack/third-party/XNNPACK/src/configs/hardware-config.c
+@@ -448,16 +448,22 @@ static void init_hardware_config(void) {
+       }
+     }
+ 
++    if (cpuinfo_initialize()) {
+ #if XNN_MAX_UARCH_TYPES > 1
+-    // Print what we think we know about the microarchs.
+-    xnn_log_info("cpuinfo_get_uarchs_count: %u.", cpuinfo_get_uarchs_count());
+-    for (int i = 0; i < cpuinfo_get_uarchs_count(); i++) {
+-      xnn_log_info("cpu_get_uarch(%i): 0x%x", i, cpuinfo_get_uarch(i)->uarch);
+-    }
++      // Print what we think we know about the microarchs.
++      xnn_log_info("cpuinfo_get_uarchs_count: %u.", cpuinfo_get_uarchs_count());
++      for (int i = 0; i < cpuinfo_get_uarchs_count(); i++) {
++        xnn_log_info("cpu_get_uarch(%i): 0x%x", i, cpuinfo_get_uarch(i)->uarch);
++      }
+ #endif  // XNN_MAX_UARCH_TYPES > 1
+-    for (size_t i = 0; i < XNN_MAX_UARCH_TYPES; ++i) {
+-      const struct cpuinfo_uarch_info* uarch = cpuinfo_get_uarch(i);
+-      hardware_config.uarch[i] = uarch ? cpuinfo_to_xnn_uarch(uarch->uarch) : xnn_uarch_unknown;
++      for (size_t i = 0; i < XNN_MAX_UARCH_TYPES; ++i) {
++        const struct cpuinfo_uarch_info* uarch = cpuinfo_get_uarch(i);
++        hardware_config.uarch[i] = uarch ? cpuinfo_to_xnn_uarch(uarch->uarch) : xnn_uarch_unknown;
++      }
++    } else {
++      for (size_t i = 0; i < XNN_MAX_UARCH_TYPES; ++i) {
++        hardware_config.uarch[i] = xnn_uarch_unknown;
++      }
+     }
+ #else
+   xnn_log_warning("Unable to determine L1/L2 data cache properties.");
+-- 
+2.43.0
+
diff --git a/patches/executorch/1.5.1/0007-threadpool-fall-back-to-the-online-CPU-count-when-cp.patch b/patches/executorch/1.5.1/0007-threadpool-fall-back-to-the-online-CPU-count-when-cp.patch
new file mode 100644
index 000000000000..d74a1e05fb51
--- /dev/null
+++ b/patches/executorch/1.5.1/0007-threadpool-fall-back-to-the-online-CPU-count-when-cp.patch
@@ -0,0 +1,63 @@
+From 0000000000000000000000000000000000000007 Mon Sep 17 00:00:00 2001
+From: RISE Project CI <ci@riseproject.dev>
+Date: Fri, 9 Oct 2026 00:00:00 +0000
+Subject: [PATCH 7/7] threadpool: fall back to the online CPU count when
+ cpuinfo fails
+
+get_threadpool() returns nullptr when cpuinfo_initialize() fails, but
+none of its callers handle that: get_pthreadpool() aborts with "Failed to
+acquire an instance of ThreadPool!" (reached from XNNCompiler.cpp when
+creating every XNNPACK runtime), and parallel_for(),
+_threadpool_get_thread_count() and the LLM sdpa op dereference the
+result directly. So on any host where cpuinfo cannot initialize - e.g. a
+riscv64 kernel that reports /sys/devices/system/cpu/cpuN/topology/core_id
+as -1 - running an XNNPACK-delegated or parallelised program crashes.
+
+cpuinfo is only used here to choose a thread count: pthreadpool itself is
+built with PTHREADPOOL_USE_CPUINFO=0 and starts its threads without it.
+Size the pool from std::thread::hardware_concurrency() (capped at the
+same tsan limit of 63) when cpuinfo is unavailable, instead of leaving
+every caller without a threadpool.
+
+Upstream-Status: To upstream [not yet submitted; unchanged on pytorch/executorch main, affects any host where cpuinfo_initialize() fails, independent of riscv64]
+
+Signed-off-by: RISE Project CI <ci@riseproject.dev>
+---
+ extension/threadpool/threadpool.cpp | 13 ++++++++-----
+ 1 file changed, 8 insertions(+), 5 deletions(-)
+
+diff --git a/extension/threadpool/threadpool.cpp b/extension/threadpool/threadpool.cpp
+index 2845b4a..a3a8c60 100644
+--- a/extension/threadpool/threadpool.cpp
++++ b/extension/threadpool/threadpool.cpp
+@@ -11,6 +11,7 @@
+ 
+ #include <algorithm>
+ #include <memory>
++#include <thread>
+ 
+ #include <executorch/extension/threadpool/threadpool_guard.h>
+ #include <executorch/runtime/platform/assert.h>
+@@ -127,12 +128,14 @@ void ThreadPool::run(
+ ThreadPool* get_threadpool() {
+   executorch::runtime::runtime_init();
+ 
+-  if (!cpuinfo_initialize()) {
+-    ET_LOG(Error, "cpuinfo initialization failed");
+-    return nullptr; // NOLINT(facebook-hte-NullableReturn)
+-  }
+-
+   static const int num_threads = ([]() {
++    if (!cpuinfo_initialize()) {
++      // pthreadpool does not need cpuinfo, so size the pool from the online
++      // CPU count instead of leaving every caller without a threadpool.
++      ET_LOG(Info, "cpuinfo initialization failed, using the online CPU count");
++      const uint32_t online = std::thread::hardware_concurrency();
++      return std::min(std::max(online, 1u), 63u);
++    }
+ #if defined(EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES)
+     // Use threads=cores.
+     auto result = cpuinfo_get_processors_count();
+-- 
+2.43.0
+

From b5a6b933d3bcdc2039dc0682dbe3071c6a4ed65c Mon Sep 17 00:00:00 2001
From: Ludovic Henry <git@ludovic.dev>
Date: Sat, 24 Oct 2026 11:58:16 +0800
Subject: [PATCH 4/5] executorch: Disable XNNPACK's RISC-V fp16 vector kernels
 (binutils 2.41 lacks zvfh)

---
 .github/workflows/build-executorch.yml | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/.github/workflows/build-executorch.yml b/.github/workflows/build-executorch.yml
index 2741ac794046..b880ff4a377a 100644
--- a/.github/workflows/build-executorch.yml
+++ b/.github/workflows/build-executorch.yml
@@ -185,9 +185,14 @@ jobs:
             pip install --only-binary=:all: "torch==2.13.0"
           # Without BUILD_VERSION, setup.py appends +<gitsha> to version.txt
           # (mirrors build-torchaudio.yml / build-torchcodec.yml).
+          # XNNPACK's rvvfp16arith microkernels are assembled with -march=..._zvfh, which
+          # the image's binutils 2.41 rejects (zvfh landed in 2.42; gotcha 420), so the
+          # fp16 vector family is switched off; plain RVV stays enabled. The XNNPACK
+          # pinned by 1.5.x has these kernels, 1.4.1's does not and ignores the option.
           CIBW_ENVIRONMENT: >-
             BUILD_VERSION=${{ env.EXECUTORCH_VERSION }}
             CMAKE_BUILD_PARALLEL_LEVEL=8
+            CMAKE_ARGS="-DXNNPACK_ENABLE_RISCV_FP16_VECTOR=OFF"
             PIP_EXTRA_INDEX_URL=https://pypi.riseproject.dev/simple/
           # The torch wheel already carries these and loads them RTLD_GLOBAL (gotcha 17).
           CIBW_REPAIR_WHEEL_COMMAND: >-

From f364d8bddfb95a0fc9a8278f4d3ff7d693a36757 Mon Sep 17 00:00:00 2001
From: Ludovic Henry <git@ludovic.dev>
Date: Sun, 25 Oct 2026 19:02:30 +0100
Subject: [PATCH 5/5] executorch: install pandas wheel-only in the test stage

---
 .github/workflows/build-executorch.yml | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/.github/workflows/build-executorch.yml b/.github/workflows/build-executorch.yml
index b880ff4a377a..aafc032c9430 100644
--- a/.github/workflows/build-executorch.yml
+++ b/.github/workflows/build-executorch.yml
@@ -205,6 +205,13 @@ jobs:
             --exclude libgomp.so.1
             --exclude libgfortran.so.5
             --exclude libopenblas.so.0
+          # The test install of the executorch wheel pulls pandas as a dependency;
+          # PyPI's newest pandas has no riscv64 wheel and its sdist fails to build
+          # (meson/xsimd), so install pandas wheel-only, with the registry's
+          # riscv64 wheels visible, and let pip back off to a release that has one.
+          CIBW_TEST_ENVIRONMENT: >-
+            PIP_EXTRA_INDEX_URL=https://pypi.riseproject.dev/simple/
+            PIP_ONLY_BINARY=pandas
           # Match the build-time pin (gotcha 615): the wheel's own metadata still
           # says "torch>=2.13.0a0", so without this the test install would resolve
           # a newer torch than the extension was actually compiled against.