From 466178aee4e5f8b63929eadddaaa915375a9b00f Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 07:16:01 +0100
Subject: [PATCH 01/19] Bumped build version for PJRT 0.98.X

---
 BUILDER_VERSION.txt | 2 +-
 XLA_COMMIT_HASH.txt | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/BUILDER_VERSION.txt b/BUILDER_VERSION.txt
index b8626c4..573541a 100644
--- a/BUILDER_VERSION.txt
+++ b/BUILDER_VERSION.txt
@@ -1 +1 @@
-4
+0
diff --git a/XLA_COMMIT_HASH.txt b/XLA_COMMIT_HASH.txt
index 30c9e7c..f62e5f4 100644
--- a/XLA_COMMIT_HASH.txt
+++ b/XLA_COMMIT_HASH.txt
@@ -1 +1 @@
-8ef2c8582a08a2b6aa5c74773421501e45e6be1c
+bb731838f24ec318ad410a4e5f73b6dcdd095d4c

From 2e97a5260e0347b224feb1ee1d2af22d555886c9 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 07:18:23 +0100
Subject: [PATCH 02/19] Updated AmazonLinux build instructions.

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index c50f188..7779fd3 100644
--- a/README.md
+++ b/README.md
@@ -36,5 +36,5 @@ It is also faster to run locally (if you have a good desktop) than in the GitHub
 To do that use:
 
 ```bash
-docker build -f Dockerfile --target export --output type=local,dest=. .
+docker build -f internal/al2023_amd64/Dockerfile --target export --output type=local,dest=. .
 ```

From 2fe10a78b1ebb769e09fb499c362e025d83f034e Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 08:03:22 +0100
Subject: [PATCH 03/19] Re-enabled required SIMD support.

---
 .github/workflows/windows_amd64_build.yaml | 23 ----------------------
 1 file changed, 23 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 94e3ec8..0fdeae5 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -75,29 +75,6 @@ jobs:
             --repo_env=BAZEL_COMPILER=C:/PROGRA~1/LLVM/bin/clang.exe \
             --repo_env=USE_CLANG_CL=1 \
             --copt=-march=native \
-            --define xnn_enable_arm_fp16_scalar=false \
-            --define xnn_enable_arm_fp16_vector=false \
-            --define xnn_enable_arm_bf16=false \
-            --define xnn_enable_arm_dotprod=false \
-            --define xnn_enable_arm_i8mm=false \
-            --define xnn_enable_arm_sme=false \
-            --define xnn_enable_arm_sme2=false \
-            --define xnn_enable_riscv_vector=false \
-            --define xnn_enable_riscv_fp16_vector=false \
-            --define xnn_enable_avxvnni=false \
-            --define xnn_enable_avxvnniint8=false \
-            --define xnn_enable_avx256skx=false \
-            --define xnn_enable_avx256vnni=false \
-            --define xnn_enable_avx256vnnigfni=false \
-            --define xnn_enable_avx512f=false \
-            --define xnn_enable_avx512skx=false \
-            --define xnn_enable_avx512vbmi=false \
-            --define xnn_enable_avx512vnni=false \
-            --define xnn_enable_avx512vnnigfni=false \
-            --define xnn_enable_avx512amx=false \
-            --define xnn_enable_avx512fp16=false \
-            --define xnn_enable_avx512bf16=false \
-            --define xnn_enable_hvx=false \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
 
       - name: "Package the binary (zip)"

From 629933c0bc47eb3ff0ad514cf506478a27b10cac Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 08:03:34 +0100
Subject: [PATCH 04/19] Updated .gitignore.

---
 .gitignore | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/.gitignore b/.gitignore
index de026a5..61799ba 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1 +1,3 @@
-pjrt_cpu_amazonlinux_amd64.tar.gz
+# Build artefacts.
+/pjrt_cpu_amazonlinux_amd64.tar.gz
+/pjrt_cpu_linux_amd64.tar.gz

From faac2f9504adca62b66455462eb23ae22472b29f Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 08:26:58 +0100
Subject: [PATCH 05/19] Disable XNNPACK altogether in Windows, since it fails
 to compile.

---
 .github/workflows/windows_amd64_build.yaml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 0fdeae5..9b50a64 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -75,6 +75,7 @@ jobs:
             --repo_env=BAZEL_COMPILER=C:/PROGRA~1/LLVM/bin/clang.exe \
             --repo_env=USE_CLANG_CL=1 \
             --copt=-march=native \
+            --define=tflite_with_xnnpack=false \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
 
       - name: "Package the binary (zip)"

From 1a1cb54a212a645d9dc2c93cee1c5a3ba0ff07f8 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 10:10:12 +0100
Subject: [PATCH 06/19] Attempt to fix Windows build.

---
 .github/workflows/windows_amd64_build.yaml | 32 +++++++++++++---------
 1 file changed, 19 insertions(+), 13 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 9b50a64..8e452b9 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -45,37 +45,43 @@ jobs:
           echo "Building PJRT CPU plugin for version ${MAJOR}.${MINOR} (to be used in go-xla as v${MAJOR}.${MINOR}.${BUILDER_VERSION})"
           echo "RELEASE_VERSION=v${MAJOR}.${MINOR}.${BUILDER_VERSION}" >> $GITHUB_ENV
 
-      - run: |
+      - name: Configure XLA
+        run: |
           cd xla
           python configure.py --backend=cpu --os=windows --host_compiler=clang --clang_path=C:/PROGRA~1/LLVM/bin/clang.exe
           cat xla_configure.bazelrc
 
-      - name: Fix Windows paths in bazelrc
+      - name: Fix Windows paths and environment
         run: |
           cd xla
           # Fix Windows paths for Bazel compatibility
-          # Replace "Program Files" with short name (handles both / and \ separators)
           sed -i 's|Program Files|PROGRA~1|g' xla_configure.bazelrc
-          # Convert all backslashes to forward slashes
           sed -i 's|\\|/|g' xla_configure.bazelrc
-          cat xla_configure.bazelrc
+        
+          # Force Bazel to use the manually installed LLVM for everything
+          echo "build --action_env=CC=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
+          echo "build --action_env=CXX=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
+          echo "build --linkopt=/FORCE:MULTIPLE" >> xla_configure.bazelrc
 
           echo "Applying dllexport patch ---------------"
-          # --ignore-whitespace is required because of the CRLF vs LF difference in windows.
           git apply --ignore-whitespace --3way ../internal/patches/windows/export.patch
 
-      - run: |
+      - name: Build PJRT C API CPU plugin
+        run: |
           cd xla
           printf "\nStarting XLA bazel build:\n"
+          # 1. We use --config=clang_local to use the LLVM toolchain.
+          # 2. We add -mavx512f to copt to allow inlining of AVX512 kernels.
+          # 3. We disable the specific XNNPACK flags that were causing conflicts.
           bazel build \
             --config=clang_local \
-            --action_env=CC=C:/PROGRA~1/LLVM/bin/clang.exe \
-            --action_env=CXX=C:/PROGRA~1/LLVM/bin/clang++.exe \
-            --linkopt="--ld-path=C:/PROGRA~1/LLVM/bin/ld.lld.exe" \
-            --repo_env=BAZEL_COMPILER=C:/PROGRA~1/LLVM/bin/clang.exe \
-            --repo_env=USE_CLANG_CL=1 \
             --copt=-march=native \
+            --copt=-mavx512f \
+            --copt=-mavx512bw \
+            --copt=-mavx512dq \
+            --copt=-mavx512vl \
             --define=tflite_with_xnnpack=false \
+            --build_tag_filters=-no_oss,-gpu,-requires-mem-64g \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
 
       - name: "Package the binary (zip)"
@@ -96,4 +102,4 @@ jobs:
           tag_name: ${{ env.RELEASE_VERSION }}
           files: ${{ env.ZIP_NAME }}
         env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
\ No newline at end of file

From 3eafd46acd05445ef91efb6eb1e1bf0e4bc2b6f3 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 10:22:32 +0100
Subject: [PATCH 07/19] Attempt #5 in getting windows to compile.

---
 .github/workflows/windows_amd64_build.yaml | 20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 8e452b9..c56905d 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -54,14 +54,18 @@ jobs:
       - name: Fix Windows paths and environment
         run: |
           cd xla
-          # Fix Windows paths for Bazel compatibility
+          # 1. Fix Windows paths in the generated bazelrc
           sed -i 's|Program Files|PROGRA~1|g' xla_configure.bazelrc
           sed -i 's|\\|/|g' xla_configure.bazelrc
-        
-          # Force Bazel to use the manually installed LLVM for everything
+          
+          # 2. Force Bazel to treat Clang-CL as the primary compiler for all dependencies
+          # Using --compiler=clang-cl is crucial for external libraries like highwayhash.
+          echo "build --compiler=clang-cl" >> xla_configure.bazelrc
+          echo "build --linkopt=/FORCE:MULTIPLE" >> xla_configure.bazelrc
+          
+          # 3. Explicitly set the path to the compiler in the rc file
           echo "build --action_env=CC=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
           echo "build --action_env=CXX=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
-          echo "build --linkopt=/FORCE:MULTIPLE" >> xla_configure.bazelrc
 
           echo "Applying dllexport patch ---------------"
           git apply --ignore-whitespace --3way ../internal/patches/windows/export.patch
@@ -70,17 +74,17 @@ jobs:
         run: |
           cd xla
           printf "\nStarting XLA bazel build:\n"
-          # 1. We use --config=clang_local to use the LLVM toolchain.
-          # 2. We add -mavx512f to copt to allow inlining of AVX512 kernels.
-          # 3. We disable the specific XNNPACK flags that were causing conflicts.
+          # BAZEL_LLVM_RESPECT_CLANG_ONLY=1 helps prevent falling back to cl.exe
+          export BAZEL_LLVM_RESPECT_CLANG_ONLY=1
+          
           bazel build \
             --config=clang_local \
+            --define=tflite_with_xnnpack=false \
             --copt=-march=native \
             --copt=-mavx512f \
             --copt=-mavx512bw \
             --copt=-mavx512dq \
             --copt=-mavx512vl \
-            --define=tflite_with_xnnpack=false \
             --build_tag_filters=-no_oss,-gpu,-requires-mem-64g \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
 

From 043bbb5d9bb92f666143f0ebcca1f72f86c46709 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 18:09:37 +0100
Subject: [PATCH 08/19] Disabling AVX512 for Windows: it doesn't gracefully
 degrade to AVX2 if AVX512 not available, it was simply crashing.

---
 .github/workflows/windows_amd64_build.yaml | 19 ++++++++-----------
 1 file changed, 8 insertions(+), 11 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index c56905d..8c5891b 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -45,13 +45,13 @@ jobs:
           echo "Building PJRT CPU plugin for version ${MAJOR}.${MINOR} (to be used in go-xla as v${MAJOR}.${MINOR}.${BUILDER_VERSION})"
           echo "RELEASE_VERSION=v${MAJOR}.${MINOR}.${BUILDER_VERSION}" >> $GITHUB_ENV
 
-      - name: Configure XLA
+      - name: "Configure XLA"
         run: |
           cd xla
           python configure.py --backend=cpu --os=windows --host_compiler=clang --clang_path=C:/PROGRA~1/LLVM/bin/clang.exe
           cat xla_configure.bazelrc
 
-      - name: Fix Windows paths and environment
+      - name: "Fix Windows paths and environment"
         run: |
           cd xla
           # 1. Fix Windows paths in the generated bazelrc
@@ -70,24 +70,21 @@ jobs:
           echo "Applying dllexport patch ---------------"
           git apply --ignore-whitespace --3way ../internal/patches/windows/export.patch
 
-      - name: Build PJRT C API CPU plugin
+      - name: "Build PJRT C API CPU plugin"
         run: |
           cd xla
           printf "\nStarting XLA bazel build:\n"
-          # BAZEL_LLVM_RESPECT_CLANG_ONLY=1 helps prevent falling back to cl.exe
           export BAZEL_LLVM_RESPECT_CLANG_ONLY=1
           
           bazel build \
             --config=clang_local \
             --define=tflite_with_xnnpack=false \
-            --copt=-march=native \
-            --copt=-mavx512f \
-            --copt=-mavx512bw \
-            --copt=-mavx512dq \
-            --copt=-mavx512vl \
+            --copt=-mavx2 \
+            --copt=-mfma \
+            --copt=/MT \
             --build_tag_filters=-no_oss,-gpu,-requires-mem-64g \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
-
+            
       - name: "Package the binary (zip)"
         run: |
           # Create the tarball.
@@ -100,7 +97,7 @@ jobs:
           7z a "$ZIP_NAME" "${TARGET_NAME}"
           echo "ZIP_NAME=${ZIP_NAME}" >> $GITHUB_ENV
 
-      - name: Upload release asset
+      - name: "Upload release asset"
         uses: softprops/action-gh-release@v2
         with:
           tag_name: ${{ env.RELEASE_VERSION }}

From e204e59df792e954854d2232443c7a7de97d6c85 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 18:12:12 +0100
Subject: [PATCH 09/19] Updated name of a step of windows build.

---
 .github/workflows/windows_amd64_build.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 8c5891b..329cf8f 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -36,8 +36,8 @@ jobs:
           echo "Checking out XLA at commit: $HASH"
           git checkout $HASH
 
-      - run: |
-          # Extract version numbers
+      - name: "Extract version numbers"
+        run: |
           BUILDER_VERSION="$(cat BUILDER_VERSION.txt)"
           PJRT_C_API_HEADER="./xla/xla/pjrt/c/pjrt_c_api.h"
           MAJOR=$(grep -E '#define PJRT_API_MAJOR' "${PJRT_C_API_HEADER}" | awk '{print $3}')

From 34ee186a53d3826f22f9877b4fc53c3a6cdcc868 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 18:27:13 +0100
Subject: [PATCH 10/19] Another attempt at building for windows.

---
 .github/workflows/windows_amd64_build.yaml | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 329cf8f..b5cd541 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -54,16 +54,21 @@ jobs:
       - name: "Fix Windows paths and environment"
         run: |
           cd xla
-          # 1. Fix Windows paths in the generated bazelrc
+          # 1. Fix Windows paths for Bazel compatibility
           sed -i 's|Program Files|PROGRA~1|g' xla_configure.bazelrc
           sed -i 's|\\|/|g' xla_configure.bazelrc
           
-          # 2. Force Bazel to treat Clang-CL as the primary compiler for all dependencies
-          # Using --compiler=clang-cl is crucial for external libraries like highwayhash.
+          # 2. Force Clang-CL and fix "multiple source files" error
           echo "build --compiler=clang-cl" >> xla_configure.bazelrc
+          
+          # 3. Use features that force one-file-at-a-time compilation for safety
+          echo "build --features=-compiler_param_file" >> xla_configure.bazelrc
+          echo "build --features=-archive_param_file" >> xla_configure.bazelrc
+          
+          # 4. Linker safety
           echo "build --linkopt=/FORCE:MULTIPLE" >> xla_configure.bazelrc
           
-          # 3. Explicitly set the path to the compiler in the rc file
+          # 5. Compiler Paths
           echo "build --action_env=CC=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
           echo "build --action_env=CXX=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
 
@@ -76,6 +81,8 @@ jobs:
           printf "\nStarting XLA bazel build:\n"
           export BAZEL_LLVM_RESPECT_CLANG_ONLY=1
           
+          # AVX-512 is disabled because is doesn't gracefully fallback to AVX2 in runtime,
+          # instead it was just crashing. So we build only with AVX2.
           bazel build \
             --config=clang_local \
             --define=tflite_with_xnnpack=false \

From a69694ba5d555590c3d6161fd54a705a284eff14 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 18:28:16 +0100
Subject: [PATCH 11/19] win build update

---
 .github/workflows/windows_amd64_build.yaml | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index b5cd541..8a3dced 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -77,21 +77,21 @@ jobs:
 
       - name: "Build PJRT C API CPU plugin"
         run: |
-          cd xla
-          printf "\nStarting XLA bazel build:\n"
-          export BAZEL_LLVM_RESPECT_CLANG_ONLY=1
-          
           # AVX-512 is disabled because is doesn't gracefully fallback to AVX2 in runtime,
           # instead it was just crashing. So we build only with AVX2.
+
+          cd xla
+          printf "\nStarting XLA bazel build:\n"
+          export BAZEL_LLVM_RESPECT_CLANG_ONLY=1          
           bazel build \
             --config=clang_local \
             --define=tflite_with_xnnpack=false \
             --copt=-mavx2 \
             --copt=-mfma \
             --copt=/MT \
-            --build_tag_filters=-no_oss,-gpu,-requires-mem-64g \
+            --copt=-D_CRT_SECURE_NO_WARNINGS \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
-            
+
       - name: "Package the binary (zip)"
         run: |
           # Create the tarball.

From 9013d6a18e9c4a38d03d29d6d950a327a06dfb32 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 18:48:04 +0100
Subject: [PATCH 12/19] win build.

---
 .github/workflows/windows_amd64_build.yaml | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 8a3dced..9d94a70 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -61,14 +61,17 @@ jobs:
           # 2. Force Clang-CL and fix "multiple source files" error
           echo "build --compiler=clang-cl" >> xla_configure.bazelrc
           
-          # 3. Use features that force one-file-at-a-time compilation for safety
+          # 3. Prevent Git-Bash from expanding /MT into a path by using the -MT syntax
+          echo "build --copt=-MT" >> xla_configure.bazelrc
+          
+          # 4. Force one-file-at-a-time compilation to fix clang-cl /Fo error
           echo "build --features=-compiler_param_file" >> xla_configure.bazelrc
           echo "build --features=-archive_param_file" >> xla_configure.bazelrc
           
-          # 4. Linker safety
+          # 5. Linker safety
           echo "build --linkopt=/FORCE:MULTIPLE" >> xla_configure.bazelrc
           
-          # 5. Compiler Paths
+          # 6. Compiler Paths
           echo "build --action_env=CC=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
           echo "build --action_env=CXX=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
 
@@ -79,16 +82,17 @@ jobs:
         run: |
           # AVX-512 is disabled because is doesn't gracefully fallback to AVX2 in runtime,
           # instead it was just crashing. So we build only with AVX2.
-
           cd xla
           printf "\nStarting XLA bazel build:\n"
-          export BAZEL_LLVM_RESPECT_CLANG_ONLY=1          
+          export BAZEL_LLVM_RESPECT_CLANG_ONLY=1
+          
+          # Use -MT instead of /MT to avoid Git-Bash path translation
           bazel build \
             --config=clang_local \
             --define=tflite_with_xnnpack=false \
             --copt=-mavx2 \
             --copt=-mfma \
-            --copt=/MT \
+            --copt=-MT \
             --copt=-D_CRT_SECURE_NO_WARNINGS \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
 

From d008769d126e2b699cccbde4f00e4b023c022b3d Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Tue, 10 Mar 2026 19:08:51 +0100
Subject: [PATCH 13/19] win build

---
 .github/workflows/windows_amd64_build.yaml | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 9d94a70..e646eba 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -84,18 +84,18 @@ jobs:
           # instead it was just crashing. So we build only with AVX2.
           cd xla
           printf "\nStarting XLA bazel build:\n"
-          export BAZEL_LLVM_RESPECT_CLANG_ONLY=1
-          
-          # Use -MT instead of /MT to avoid Git-Bash path translation
+          export BAZEL_LLVM_RESPECT_CLANG_ONLY=1          
           bazel build \
             --config=clang_local \
             --define=tflite_with_xnnpack=false \
             --copt=-mavx2 \
             --copt=-mfma \
+            --copt=-mbmi \
+            --copt=-mbmi2 \
             --copt=-MT \
             --copt=-D_CRT_SECURE_NO_WARNINGS \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
-
+            
       - name: "Package the binary (zip)"
         run: |
           # Create the tarball.

From a2b7d9760fdff844a3b0ff9a1f4cf7769b117a74 Mon Sep 17 00:00:00 2001
From: Jan <janpfeifer@users.noreply.github.com>
Date: Tue, 10 Mar 2026 20:08:18 +0100
Subject: [PATCH 14/19] Add -mlzcnt option from Windows build workflow

---
 .github/workflows/windows_amd64_build.yaml | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index e646eba..e43d912 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -92,6 +92,7 @@ jobs:
             --copt=-mfma \
             --copt=-mbmi \
             --copt=-mbmi2 \
+            --copt=-mlzcnt \
             --copt=-MT \
             --copt=-D_CRT_SECURE_NO_WARNINGS \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
@@ -114,4 +115,4 @@ jobs:
           tag_name: ${{ env.RELEASE_VERSION }}
           files: ${{ env.ZIP_NAME }}
         env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
\ No newline at end of file
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

From e4faf2be18d07e1e701c106abbe73aa271eb60b7 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Wed, 11 Mar 2026 05:56:52 +0100
Subject: [PATCH 15/19] win build

---
 .github/workflows/windows_amd64_build.yaml | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index e646eba..fdd6469 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -55,7 +55,9 @@ jobs:
         run: |
           cd xla
           # 1. Fix Windows paths for Bazel compatibility
+          #    Replace "Program Files" with short name (handles both / and \ separators)
           sed -i 's|Program Files|PROGRA~1|g' xla_configure.bazelrc
+          #    Convert all backslashes to forward slashes
           sed -i 's|\\|/|g' xla_configure.bazelrc
           
           # 2. Force Clang-CL and fix "multiple source files" error
@@ -87,9 +89,14 @@ jobs:
           export BAZEL_LLVM_RESPECT_CLANG_ONLY=1          
           bazel build \
             --config=clang_local \
+            --linkopt="--ld-path=C:/PROGRA~1/LLVM/bin/ld.lld.exe" \
+            --repo_env=BAZEL_COMPILER=C:/PROGRA~1/LLVM/bin/clang-cl.exe \
+            --repo_env=USE_CLANG_CL=1 \
+            --copt=-march=native \            
             --define=tflite_with_xnnpack=false \
             --copt=-mavx2 \
             --copt=-mfma \
+            --copt=-mf16c \
             --copt=-mbmi \
             --copt=-mbmi2 \
             --copt=-MT \
@@ -114,4 +121,4 @@ jobs:
           tag_name: ${{ env.RELEASE_VERSION }}
           files: ${{ env.ZIP_NAME }}
         env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
\ No newline at end of file
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

From 708f0cf1945dd41c4647cbafb32053afc61eb649 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Wed, 11 Mar 2026 06:01:23 +0100
Subject: [PATCH 16/19] win bulid

---
 .github/workflows/windows_amd64_build.yaml | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index fdd6469..1075e50 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -87,8 +87,10 @@ jobs:
           cd xla
           printf "\nStarting XLA bazel build:\n"
           export BAZEL_LLVM_RESPECT_CLANG_ONLY=1          
-          bazel build \
+          bazel --output_user_root="C:/b"
+            build \
             --config=clang_local \
+            --features=archive_param_file --features=linker_param_file --features=compiler_param_file \
             --linkopt="--ld-path=C:/PROGRA~1/LLVM/bin/ld.lld.exe" \
             --repo_env=BAZEL_COMPILER=C:/PROGRA~1/LLVM/bin/clang-cl.exe \
             --repo_env=USE_CLANG_CL=1 \

From ab34df339181967e9c2edf715936086b9effc0fc Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Wed, 11 Mar 2026 07:23:17 +0100
Subject: [PATCH 17/19] win build

---
 .github/workflows/windows_amd64_build.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index fbe1d34..9dac376 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -87,7 +87,7 @@ jobs:
           cd xla
           printf "\nStarting XLA bazel build:\n"
           export BAZEL_LLVM_RESPECT_CLANG_ONLY=1          
-          bazel --output_user_root="C:/b"
+          bazel --output_user_root="C:/b" \
             build \
             --config=clang_local \
             --features=archive_param_file --features=linker_param_file --features=compiler_param_file \

From 8d65c5ca46f711d33b43ecf30890246c0f9aaa66 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Wed, 11 Mar 2026 08:39:44 +0100
Subject: [PATCH 18/19] style: Remove trailing whitespace from the Windows
 AMD64 build workflow file.

---
 .github/workflows/windows_amd64_build.yaml | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 9dac376..720f21e 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -59,20 +59,20 @@ jobs:
           sed -i 's|Program Files|PROGRA~1|g' xla_configure.bazelrc
           #    Convert all backslashes to forward slashes
           sed -i 's|\\|/|g' xla_configure.bazelrc
-          
+
           # 2. Force Clang-CL and fix "multiple source files" error
           echo "build --compiler=clang-cl" >> xla_configure.bazelrc
-          
+
           # 3. Prevent Git-Bash from expanding /MT into a path by using the -MT syntax
           echo "build --copt=-MT" >> xla_configure.bazelrc
-          
+
           # 4. Force one-file-at-a-time compilation to fix clang-cl /Fo error
           echo "build --features=-compiler_param_file" >> xla_configure.bazelrc
           echo "build --features=-archive_param_file" >> xla_configure.bazelrc
-          
+
           # 5. Linker safety
           echo "build --linkopt=/FORCE:MULTIPLE" >> xla_configure.bazelrc
-          
+
           # 6. Compiler Paths
           echo "build --action_env=CC=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
           echo "build --action_env=CXX=C:/PROGRA~1/LLVM/bin/clang-cl.exe" >> xla_configure.bazelrc
@@ -86,7 +86,7 @@ jobs:
           # instead it was just crashing. So we build only with AVX2.
           cd xla
           printf "\nStarting XLA bazel build:\n"
-          export BAZEL_LLVM_RESPECT_CLANG_ONLY=1          
+          export BAZEL_LLVM_RESPECT_CLANG_ONLY=1
           bazel --output_user_root="C:/b" \
             build \
             --config=clang_local \
@@ -94,7 +94,7 @@ jobs:
             --linkopt="--ld-path=C:/PROGRA~1/LLVM/bin/ld.lld.exe" \
             --repo_env=BAZEL_COMPILER=C:/PROGRA~1/LLVM/bin/clang-cl.exe \
             --repo_env=USE_CLANG_CL=1 \
-            --copt=-march=native \            
+            --copt=-march=native \
             --define=tflite_with_xnnpack=false \
             --copt=-mavx2 \
             --copt=-mfma \
@@ -105,7 +105,7 @@ jobs:
             --copt=-MT \
             --copt=-D_CRT_SECURE_NO_WARNINGS \
             //xla/pjrt/c:pjrt_c_api_cpu_plugin.so
-            
+
       - name: "Package the binary (zip)"
         run: |
           # Create the tarball.

From 54a17ea2c41a41e1ad0e2d0a9a4f8711b49b24f3 Mon Sep 17 00:00:00 2001
From: Jan Pfeifer <pfeifer@gmail.com>
Date: Wed, 11 Mar 2026 11:02:04 +0100
Subject: [PATCH 19/19] Adjust Bazel feature flags in
 `windows_amd64_build.yaml` by enabling `archive_param_file` and
 `linker_param_file` via `xla_configure.bazelrc` and removing explicit feature
 flags from the build command.

---
 .github/workflows/windows_amd64_build.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/windows_amd64_build.yaml b/.github/workflows/windows_amd64_build.yaml
index 720f21e..cc2f92b 100644
--- a/.github/workflows/windows_amd64_build.yaml
+++ b/.github/workflows/windows_amd64_build.yaml
@@ -68,7 +68,8 @@ jobs:
 
           # 4. Force one-file-at-a-time compilation to fix clang-cl /Fo error
           echo "build --features=-compiler_param_file" >> xla_configure.bazelrc
-          echo "build --features=-archive_param_file" >> xla_configure.bazelrc
+          echo "build --features=archive_param_file" >> xla_configure.bazelrc
+          echo "build --features=linker_param_file" >> xla_configure.bazelrc
 
           # 5. Linker safety
           echo "build --linkopt=/FORCE:MULTIPLE" >> xla_configure.bazelrc
@@ -90,7 +91,6 @@ jobs:
           bazel --output_user_root="C:/b" \
             build \
             --config=clang_local \
-            --features=archive_param_file --features=linker_param_file --features=compiler_param_file \
             --linkopt="--ld-path=C:/PROGRA~1/LLVM/bin/ld.lld.exe" \
             --repo_env=BAZEL_COMPILER=C:/PROGRA~1/LLVM/bin/clang-cl.exe \
             --repo_env=USE_CLANG_CL=1 \